Methodology for real impact assessment of the best location of distributed electric energy storage systems

Tese de doutoramento em Sistemas Sustentaveis de Energia, apresentada a Faculdade de Ciencias e Tecnologia da Universidade de Coimbra


Introduction
The need to combine distributed generation, continuous load growth, increasing power quality requirements of electronic loads and higher regional power transfers in a (nowadays) largely interconnected network, presents new challenges to power systems.
These new challenges can lead to a complex and less secure power system operation, especially because power plants may not be able to follow new demands, as a result of technical, economic, environmental, and/or governmental regulation constraints (Ribeiro et al., 2001),(R.M. . In the context of smart grids and micro grids, distributed electric energy storage systems (DEESS) are presented as an option to enable the optimal use of resources, by providing the capability of effectively balancing supply and demand (EPRI, 2010). However, a methodology is needed to evaluate the impact of storage, as well as the best locations for DEESS units, in order to provide specific energy services to the network.

Motivation and research objectives
This thesis presents a methodology to assist decision makers on the assessment of feasible solutions regarding the integration of a DEESS in an urban environment, as a tool to provide power and energy services to the electric network. The developed methodology uses profile prototypes based on real data, obtained through clustering techniques. These profiles, of electricity demand, electricity prices and renewable electricity production, are used to optimize the placement of electric energy storage units by an improved genetic algorithm based on the Non-dominated Sorting Genetic Algorithm-II (NSGAII). The thesis considers expected attitudes of the main stakeholders towards DEESS implementation and discusses the possible regulatory framework options to define the DEESS business model.
The novelty of the proposed methodology is to combine the determination of the best location for the storage units with the definition of the best schedule of operation when simultaneously minimizing investment costs, energy losses and voltage deviations, as well as maximizing the economic benefit resulting from the energy exchange with the grid.
The proposed methodology aims to provide the decision-maker (DM) with insight on the impacts associated with possible energy storage systems (ESS) management schemes, as well as to propose a pricing scheme within the current legal framework for the ESS exploitation. Considering plausible management schemes based on possible private stakeholders objectives, the assessment of its impacts, both technical and economic, increases their perception of the characteristics of available DEESS options.
The specific goals of the research may be formulated as the following research question:  In which circumstances can urban DEESS be a feasible solution?, given: • Technological costs / Power Generation Costs; • Availability / Energy Services provided; • Location and sparsity of the storage units.

Scope and limitations
The developed methodology could be used in any type of network as long as a correct characterization is done. Considering the assessment of technical impacts in its current state, the proposed methodology is suited for energy services assessment.
The operation of ESS is constrained by the individual capacities of each storage unit and additional specifications, namely: the required voltage for the power inverter, the power limit considering the bus constraints in the network to be assessed, and the physical limit of each individual storage solution for a maximum volume of 1m 3 , specified in order to be easily installed within a typical distribution transformer station (EDP Distribuiçao, 2004).

Thesis structure
This doctoral thesis is organized in five chapters, each one describing specific work carried out during the development of the thesis. This chapter presented an introduction, discussing the motivation, objectives and research question to be answered, and the scope and limitations of the developed work.
The remaining chapters are organized as follows:  Chapter 2 presents a literature review on energy storage operation, possible applications, barriers and some principles for assessing the economic cost Continuous load growth, increasing power quality requirements of electronic loads and higher regional power transfers in a largely interconnected network, can lead to complex and less secure power system operation, especially because new economic, environmental, technical, and governmental regulation constraints make it difficult to follow the continuous growth of demand (Ribeiro et al., 2001), (Ipakchi & Albuyeh, 2009).
In the recent past, energy storage for daily periods seemed economically unfeasible (Barton & Infield, 2004). However, the rising cost of oil and the need to rethink the management of the electricity grid (to meet the demands of current loads) led to consider storage, either centralized or decentralized, as an opportunity to help solving problems of capacity and reliability of the network (Divya & Østergaard, 2009;Ribeiro et al., 2001). The SmartGrids European Technology Platform (SmartGrids, 2012) presents storage as a key solution, within five of the six research areas presented in Figure 2-1, in order to achieve the EU 20-20-20 goals, as well as to obtain the envisioned CO2 reduction of 80% by 2050, despite the high acquisition and maintenance costs. The integration of storage technologies, addressing different needs (e.g. quick response, high grid security, long-term energy storage, lowering peak demand, among others) will help tackling these challenges.

Electric energy storage operation
By the year 2035 it is expected that a large fraction of the generation capacity in Europe will have irregular and/or highly variable availability. Due to technological improvements, it is expected that ESS in Europe and USA will grow by 10 to 15% (and by larger figures in Japan), in a near future, motivating a major change in the electricity industry (H. Chen et al., 2009). The development of ESS is essential to the achievement of a more flexible and controllable electric network (e.g. smart grids), enabling higher levels of demand side management, increased proximity of generation to the loads, integration of non-dispatchable generation from renewable sources, and other forms of decentralized production (Ipakchi & Albuyeh, 2009), (Beer et al., 2012).
Although ESS are being envisaged as an important part of a smarter grid, their usefulness may apply to the existing grid technologies as long as there is a communication system that links them to the distribution system operator (DSO).
Nevertheless, the use of smart grids is crucial to make full use of the capabilities of smaller and larger generation and loads.
Some authors propose the introduction of an aggregator in order to fully realize the benefits of ESS (Battistelli et al., 2012), (Kristoffersen et al., 2011). However the need to develop the power system beyond the closed infrastructure of an aggregator is also defended, to address specific needs, in such a way that all generation and consumption units can participate automatically and individually as ancillary services, without the need for the control by the DSO, transmission system operator (TSO), utility or even an aggregator (SmartGrids, 2012).
Within a smart grid paradigm of network management, some authors defend that the increase of storage capacity, besides solving network problems, may facilitate the adoption of plug-in hybrid vehicles or electric vehicles (EV), mitigating some requests from the network (Sovacool & Hirsh, 2009), (Ipakchi & Albuyeh, 2009).
The use of intelligent charging/discharging (C/D) management schemes could release network capacity, minimizing or deferring supply-side capacity expansion, decreasing network losses and overall electric systems running costs (both in production and delivery costs), avoiding short interruptions and some voltage quality problems, shaving power peaks and smoothing load curves (Lassila et al., 2012), , , (Martins, 1999), (Sutanto & Lachs, 1998).
Different market operation frameworks within the smart grid concept are also presented as important tools to avoid other new types of problems such as potential shortage of generation capacity or the increased difficulty in optimizing the management of the grid to accommodate new dispersed loads and generators that could significantly influence the mix of generation and consequently, electricity prices and emissions (Hadley & Tsvetkova, 2009), (Rocha Almeida et al., 2010).
Within this paradigm the confidence on the impact assessment is influenced by the used grid control architecture as well as by other characteristics such as the network connections and used storage technology (Rocha Almeida et al., 2010).The key characteristics of available technologies of accumulators are their energy densities (up to 150 and 2000Wh/kg for lithium) and technological maturity. Generally the Ragone's diagram is used (Ibrahim et al., 2008) to represent the performance in terms of ratio of mass to energy and power. Although this diagram is useful when the size of the storage systems is important, it is not enough. Even considering the present technology limitations, instantaneous, short-term and mid-term applications seem to be commercially feasible. Although the storage of large amounts of energy would be desirable, even batteries of modest power and able to deliver only during seconds or minutes, seem to be cost-effective when designed for storage periods of less than 5 h (Divya & Østergaard, 2009).

Electric energy storage possible applications and services
The adoption of distributed energy storage allows an optimization of resources by providing the capability of effectively balancing supply and demand (Kintner-Meyer. et al., 2007), (EPRI, 2010), (Ibrahim et al., 2008), . The ideal ESS solution in urban environments should have means of rapidly damp oscillations, respond to sudden changes in load, supply load during transmission or distribution interruptions, provide spinning reserve, correct load voltage profiles with rapid reactive power control, and still allow the generators to balance the system load at their normal speed (Sutanto & Lachs, 1998), (Ribeiro et al., 2001), ,   (Divya & Østergaard, 2009).
According to Barton and Infield (2004), possible benefits of increased energy storage capacity (spanning a few hours) are both the relief of stress situations of the distribution electric grid and an improved use of reactive power. The first would allow the network to respond to higher demand levels without causing excessive load flow 9 in the transmission network, and the latter responds to voltage disturbances caused by embedded renewable generation.   We can observe that the capacity range for each application is not well defined in   Distribuição, 2011), the majority of distribution grid problems were related to short duration problems. Hereupon and given the following Table 2-3 information, we may conclude that electric ESS seems suited for applications of short and very short duration, such as transmission and distribution stabilization or renewable matching. Considering the possible exchange of ancillary services between the TSO and DSO, the existence of a storage system in the distribution grid might be a solution to perform ancillary services in a more cost-effective way (SmartGrids, 2012). Different technologies may be used for different applications, according to the discharge time: technologies with a long discharge time will be used to secure the capacity of renewable energy sources (RES), while those with a short discharge time will be adopted to perform frequency regulation or voltage control.
As small distributed generation systems are expected to spread in urban environments with their almost random, uncontrolled nature, the assessment of the value of small scale storage against supply interruptions will be very important (SmartGrids, 2012). In order to understand the technical and economic viability of each type of system, it is important to assess all the associated costs and benefits, and verify whether there is overall economic justification (SmartGrids, 2012), (Hittinger et al., 2012). This will certainly lead to further research development related to the distribution network management.
The energy service to be provided determines the technology to be used, which has to be assessed in order to ensure cost-effectiveness (Hittinger et al., 2012).
Considering the storage properties of li-ion batteries and Supercapacitors, the sensitivity analysis performed by (Hittinger et al., 2012) showed that capital cost, length of capital investment and efficiency have the largest potential in cost-ofservice reduction. Still according to the same authors, peak shaving and frequency regulation are the most capital cost dependent energy services.
Providing energy services such as frequency regulation to the grid is becoming more necessary as the use of renewable energy increases. In 2010, the frequency regulation market was worth US$495 million. This is a new market for flywheels, which are fast-acting regulators, needing just a few seconds to perform frequency regulation, which makes them a more efficient regulator than major conventional power plants (Fairley, 2011). Moreover, battery-based frequency regulators based on fixed utilities, such as AES 1 in Johnson city-N.Y. or in electric vehicles infrastructures, which have a lower unitary cost, are one of the present biggest competitors of the flywheel (Fairley, 2011).
The various battery technologies are different not only on their physical layout, chemical composition and energy density, but also on their voltage and current output characteristics. As such, the power conversion interface may differ for each technology, and each will have unique time-varying output that must be matched.
Other operational features, such as efficiency and size, also vary among technologies (Table 2-4). Despite being the weakest link of the energy domain, storage systems could play a very important role in maintaining system reliability, dynamic stability, enhanced power quality, transmission capacity and area protection (Ibrahim et al., 2008;F. Rahman et al., 2012;Ribeiro et al., 2001), presenting a positive cost and environmental impact by reducing fuel consumption and emissions through reduced line losses, and reducing the need of idle generation for frequency stabilization (H. Chen et al., 2009).

Barriers to electric energy storage
The operation of battery based ESS is constrained in capacity and storage balances, for grid connections and batteries. According to Kristoffersen et al. (2011) and Kempton and Tomić (2005) the constraints include:  nominal capacity of batteries;  battery wear costs, related to the life cycle of each technology;  capacity limits for the ESS connection to the grid, related to existing physical electric infrastructure interfaces, such as wires and other circuitry connection and/or the rated maximum power of the interface's power electronics.
One possible storage solution often referred in the literature is the EV storage infrastructure. However, it presents several limitations, namely: the unpredictability of vehicle (batteries) connection to the network, the uncertainty of the willingness of their users to present themselves as potential power providers, or the increase of harmonic and voltage problems, and of line losses, leading to a power quality degradation (Ipakchi & Albuyeh, 2009), (Gerkensmeyer et al., 2010), . Therefore, the use of stationary systems can be seen as more predictable, enabling the grid manager to choose the most effective location of ESS in the network. Ibrahim, Ilinca, and Perron (2008) pointed out the low durability for large-amplitude cycling as the main inconvenient, and Sovacool and Hirsh (2009) defends that a major issue to the adoption of a distributed storage infrastructure is also the existence of some resistance by network operators to integrate resources (e.g. batteries) which they do not manage.
Divya & Østergaard (2009) outline some of the reasons for a small deployment of storage systems such as the existence of a large number of conventional generators that can be adjusted to match the load demand (LD), and the interconnection of areas which can help to balance the LD. However, this scenario is changing due to the current high rate of deployment of renewable energy sources and the existing limits in the interconnections. Additionally, the lack of practical experience and availability of tools to perform optimization of operational costs, and to assess the benefits of storage technology (considering the market models) during planning, justify the motivation to develop such a methodology.

Economic ESS costs
The immediate success of battery ESS in rapid and short term applications is due to their capability of reacting instantly to system disturbances, which a conventional generator set cannot do (Divya & Østergaard, 2009). The mid-term battery ESS applications (less than 5 hours), based on dispersed ownership as in V2G applications, could become commercially viable if the utilities have a degree of control of the set of battery units in a way that make them useful to manage and control the power system, reducing operational costs. However, to make this idea feasible, an adequate framework needs to be developed (Divya & Østergaard, 2009), , (Rocha Almeida et al., 2010).
Aiming at the economic viability of an ESS to provide a specific energy service, some authors argue that ESS should perform all the possible services in order to obtain the highest possible profit (Hittinger et al., 2012). The research and technological innovation is proposed as a key element for decreasing the capital costs, being the perception of technologies with greater potential for reductions in capital costs seen as essential to the future of ESS (Hittinger et al., 2012).

= +
(1) The power conversion unit costs are proportional to the power rating of the system as shown in equation (2).
The energy storage unit costs are proportional to the amount of energy stored ( ) considering the round-trip efficiency 2 ( ), as presented in equation (3).
The annualized capital costs ( ) are obtained by multiplying the total capital cost with a capital recovery factor (CRF), as presented in equation (4).
= × The CRF can be obtained using the equation (5) formula: where, • is the discount rate, dimensionless; • is the expected lifetime of the equipment in years; The CRF converts the present value into a stream of equal annual payments over a specified time, at a specified discount rate, i.e., it is the ratio of a constant annuity to the present value of receiving that annuity during certain time.
Depending on the performed service, different parameters are used to assess the net benefit (subtracting the cost to the assessed benefits). Although with small differences, other authors aim to assess costs and revenues in a very similar way, as shown in Table 2-5 example. Table 2-5 -Two perspectives for peak power service revenue and costs Energy Service Peak power (Lassila et al., 2012) Peak power (Kempton & Tomić, 2005) In equations shown in Table 2 Kempton and Tomić (2005) present other expressions to quantify the costs and benefits for different services, based on the same general approach.
The inclusion of societal, environmental and technical contributions, when using ESS, could also be considered. Equation (7) shows one possible way to assess the benefit for peak power services, considering them at least equal to avoided grid reinforcements, using the EV available battery capacity (Lassila et al., 2012).
In equation (7), ∆ is the power capacity decreased on the feeder [kW] and is the AMC on the feeder [€/kW].
In order to augment the ESS aggregated benefit, several authors propose to combine the associated impacts with more or less overlapped benefits as shown in   (Eyer et al., 2004) In many cases more than one advantage is required from storage for benefits to exceed costs. However, careful consideration of operational, technical, and market details is required before considering all the possible benefits. Operational objectives involve conflicting uses for storage, e.g. between power output and duration of delivery (Eyer et al., 2004).
In some cases ESS are physically unable to serve more than one need. As an example, ESS are unable to simultaneously tolerate frequent deep discharges and/or significant cycling. Storage systems well suited to the transmission and distribution upgrade deferral applications, are probably not suitable for energy price arbitrage 3 (Eyer et al., 2004).
Another example is when storage cannot respond very rapidly to changing line conditions. Such systems may be suitable for energy arbitrage or to reduce demand charges but may not be able to provide transmission support or end-user power quality benefits. Eyer et al. (2004) provides a summary of key standard assumption values used for evaluating benefits, market potential, and total economic benefits from storage used for specific applications and/or for specific benefits, as shown in Table 2-7. 3 Advantage from buying and selling energy in different market price periods *Hours unless other units are specified. **Over ten years, based on lifecycle benefits times maximum market potential (market estimates will be lower). ***Placeholder values. The actual benefit was not estimated. #Does not include incidental energy-related benefit. ; ##Wind generation.
Whenever possible, it is very important to assess the AMC of a specific ESS, since the procedures of current energy markets are used to define which units are the first to provide services to the network. In order to provide data on the cost of selected storage technologies some guidance is presented in Table 2-8. The exact information depends on the manufacturer, the application and usage. However, it is a good basis for establishing some assumptions in further studies. It should be noted that the DEESS evaluation impact is not indifferent to the technology considered. Moreover, S. M. Schoenung & Hassenzahl (2003) also argue that the real benefit should be assessed considering the distribution of ESS in the network, given a specific technology and considering market parameters, such as electricity costs and interest rates. Finally according to Lassila et al. (2012), the ESS management can be more effective if the network operator owns the storage system.

Assessment methodologies
Previous research assessing the impact of energy storage systems (ESS) on the power system operation and economics has been focused on economic/optimal sizing. As such, ESS have been modelled from the point of view of cost (economic models) or with a focus on the assessment of operational benefits, modelling the ESS response to power system disturbances at appropriate time scales or operational models, as shown in Table 2-9 (Divya & Østergaard, 2009).  Demand side applications.
 BESSpower system reliability analysis;


BESSpower system stability analysis.
The majority of studies performing economic analyses of ESS considered vertically integrated utilities, not incorporating market models to assess the benefits of ESS for the current deregulated market. From the operational side, most studies on the impact on power systems generally do not refer to any particular battery technology, or to any limitation of the performance of ESS (Divya & Østergaard, 2009).
Focusing on the economic side, some authors presented methodologies for evaluating the associated costs and benefits of energy storage by analyzing single specific services (S. Schoenung, 2011), (Kempton & Tomić, 2005). However, none of these studies consider the relation of an ESS management scheme with its associated impacts on the electric grid performance.
Aiming to increase wind farms power dispatchability (power firming) Le and Nguyen (2008) developed an analytical approach to find a rating for the most-profitable ESS.
Some of the considered benefits were the revenue from higher renewable integration into the network (higher capacity), the revenue from generation capacity firming, the revenue from improved grid voltage stability, the revenue from improved grid reliability and the revenue related to environmental benefits such as reduced greenhouse gas(GHG) emissions. Those authors also concluded that the quality of results depends on the measurement and verification methods, and on the accuracy of the calculation methods, which should be established within the research community and the power industry.
Concerning operational benefits, the advantages of ESS led to the development of the V2G concept, a long-held idea of an option to provide specific energy services to the grid based on connected EVs capacity. Different methodologies were then applied to analyze their contribution to the grid (Lassila et al., 2012), , (Kristoffersen et al., 2011), (Balcells & Garcia, 2010). However, the uncertainty associated with the location of EV led some authors to focus on this constraint (Battistelli et al., 2012). To assess the possible contribution in terms of energy exchange to the electricity network, Lassila et al (2012) tested the effect of different C/D management schemes to provide specific energy services.
Considering the technical characterization requirements to perform power services studies (e.g. regulation services), some authors have used island case studies or simplified grid models to ease the application of those methodologies (Sigrist et al., 2013), , (Lim et al., 2011).
As mentioned previously, most studies in the literature focused on only one type of assessment (either technical or economical), and usually did not provide more than one option for a DM. However, effective decisions over the subject will need to simultaneously consider both kinds of objectives (and even different objectives of each kind), and should imply a multiobjective optimization procedure.
Le et al (2012) presented a multiobjective methodology for calculating a reference output profile, using a C/D scheme for guiding the ESS operation, together with an economic assessment for compensating wind power variation through storage, and an optimization-based method for determining the ESS optimal rating. Regarding the technical side, and envisaging to fill the gap between the real and the desired profiles, these authors used an optimum power flow analysis (OPF) program to define a suitable storage power output for a given network. Regarding the economical side, they determined the ratings of the ESS by an optimization-based method, considering a cost-based objective function that maximizes the net benefit and takes into account five possible benefit factors, namely the revenues earned from: saving and integrating an amount of wind power into the grid, wind farm capacity firming, improved voltage stability, improved reliability, and environmental considerations.  Analysis of impact method on the distribution system User defined periods based on technical data.  Framework for determining the optimal charging/discharging of BES Using predicted production profiles valued with day-ahead and realtime prices (Ippolito et al., 2014) Definition of some stratagies for design and the optimized management of ESS Multiobjective optimization using different design strategies Table 2-10 summarizes the approaches used in the literature in order to establish the operating scheme of storage units. Modified profiles are commonly used to assess the impact of C/D batteries (e.g. EVs) in the power grid, or evaluating impacts on load diagrams (T. Chen et al., 2013;Koyanagi & Uriu, 1998;Lassila et al., 2012;Le & Nguyen, 2008;Le et al., 2012;Leou, 2012;Nguyen et al., 2012;Qian et al., 2011;S. Rahman & Shrestha, 1993). In order to avoid new peaks, S. Rahman and Shrestha (1993) and Koyanagi and Uriu (1998) proposed the need of proper incentives and market regulation. Both the creation of limits to use the excess energy available, according to grid requirements and the time shift control, regulating the charging schedules in the network, were presented as adequate mechanisms to avoid new network power peaks.
As already stated, other ESS advantages are the ability to postpone grid reinforcements, to avoid short interruptions and voltage quality problems, to shave power peaks and to smooth load curves (Lassila et al., 2012). However, the economic attractiveness of ESS is not guaranteed, especially due to the high capital cost (when compared to alternative technologies). Therefore, the proposal of a feedin tariff scheme is justified to increase the interest of investors (Dufo-López & Bernal-Agustín, 2015).
The main obstacle to ESS deployment seems to be the corresponding investment costs and the reluctance of network operators to integrate resources that they do not manage (Sovacool & Hirsh, 2009). To overcome these barriers, it is crucial to improve the assessment of application and benefits of these systems, namely by developing operational tools to select the quantity and location of these resources, optimize the operational cost and assess the benefits of storage during the planning phase (EPRI, 2010), (Divya & Østergaard, 2009).
Drawing on the existing literature, several potential simulation tools for network analysis may be applied, considering distribution networks, EVs and energy storage applications. Investigates the impacts of increasing EV loads in a typical Danish primary distribution network (S. Rahman & Shrestha, 1993) Generic Mathematical modeling Investigate the impact of EV load on the electric utility distribution system. Table 2-11 summarizes a selection of studies applying potential software tools. In most studies, load power flow analysis was combined with power demand profiles of the distribution grid to quantify the technical impact of each solution. In order to assess the best location, avoiding pre-defined scenarios, the use of techniques that can combine grid characteristics, load diagrams, supply diagrams and C/D diagrams of ESS, will be important to obtain the maximum benefit from storage. Leou (2012), and T. Chen et al. (2013) used genetic algorithms (GA) whose results only differ in the fitness function factors. The use of GA enables a multiobjective approach especially useful to problems that do not have a unique optimal result but a set of non-dominated or efficient solutions (the choice depending on the relative importance of each objective to the DM).
As a result of using a multiobjective approach the possible use of interactivity should be explored, helping the DM to obtain the desired results without the need to map his/her preferences to hard values assigned to weights of each objective function (Gonçalves et al., 2013).  In the multiobjective approaches presented in Table 2-12, all authors presented fitness functions where the objectives were equally weighted. The main difference between studies regarded the used objectives to perform the impact evaluation.

Evaluation objectives
More concern with economic aspects, (Lassila et al., 2012;Leou, 2012) focused their assessment in the associated benefits from charged and discharged energy, between ESS and the network, in different time periods. Leou (2012) additionally included the economic benefits from reducing the transmission access costs and from deferring facility investments. Moreover, both authors considered the investment costs and the maintenance costs of the ESS and for the network. 4 A broader analysis could be found in (Lassila et al., 2012) The use of network data not easily available, and in some cases restricted, such as the investment and maintenance cost of the network or the benefits from deferring facility investments turn those objectives difficult to use.
T. Chen et al. (2013) developed a multiobjective methodology that enables the use of public available or easy to obtain data, to assess the ESS impact from the point Finally, and focused on network reconfiguration Sahoo and Prasad (2006) developed a fuzzy genetic approach using only the voltage stability index to assess the technical impact on the network.
Drawing on the literature review, one can conclude that investment and operation costs, together with energy tariffs, are important factors for the economic analysis and voltage stability, as well as active power losses are important factors for technical analysis.

Multiobjective optimization and genetic algorithms
A diversity of methodologies have been used for solving optimization problems, leading the classical mathematical operations towards a knowledge based on human behavior intelligence and Nature. Evolutionary algorithms (EA) are one of the solutions developed to combine search techniques and optimization. Inspired on natural evolution mechanisms and genetics, the EA works with several solutions simultaneously, providing a set of individuals generally designated by population.
By applying the algorithm, a set of individuals with a starting low rate of survival or environmental fitness, can evolve over several generations. Based on nature evolution process, the most fit individuals are retained by natural selection passing their traits to the next generation, combining features in so-called offspring, or leading to new specimens through a mutation process that change their characteristics for better environment adaptation.
Most of the real problems usually involve multiple objective optimization and decision making, and engineering is one of the areas where the existence of conflictive objectives is most likely to appear, causing limitations on one or more objectives and associated difficulties in the optimization process. The ability to work with a population of potential conflictive solutions in every generation, enables EA to effectively solve many real problems, as shown its wide application in the literature.
An operational optimization tool that comprises a search technique for exact or approximate solutions, based on the evolution of organisms, is the Genetic algorithms (GA), which belong to the larger class of EA.

Improved NSGAII
GA have been widely used in several fields, such as economic dispatch, reconfiguration of power networks, optimal placement of capacitors or inductors, and optimal power flow analysis (T. Chen et al., 2013). Conventional GA, however, may get stuck at a local optimum, or have slow convergence speed. In order to overcome these shortcomings, many researchers have aimed to improve the performance of GA (R. M. , (Sahoo & Prasad, 2006), (J. A. Vasconcelos et al., 2001).
A GA multiobjective approach allows the integration of different non correlated objectives, but the existence of multiple conflicting objectives prevents the achievement of an optimal. In such a case, the DM has to analyze a set of nondominated solutions, or solutions which are no worse than all others at least in one aspect (objective) in order to establish tradeoffs and to choose the option that better represents his/her preferences.
The Non-dominated Sorting Genetic Algorithm-II (NSGA-II) is a particular case of a multiobjective GA implementation which has proved to be efficient, especially in power distribution operation and planning problems. Detailed information regarding the working principle of this tool can be found in (Deb et al., 2002).
GA use genetic operators, such as crossover, cloning and mutation, to perform genetic manipulation in order to generate new individuals with characteristics of both parents.
An improved NSGA-II (iNSGA-II) was proposed in (   Usually the crossover operator is predominant (thus with a higher probability of occurrence) and the mutation operator has a low probability. However, the mutation operator is highly significant, because it is the operator which is responsible for introducing new genetic characteristics in the population. This operator randomly changes characteristics of one specific individual, avoiding a premature convergence and guarantying a non-null probability of reaching any point of the search space. Therefore, each time the genetic characteristics tend to become homogeneous, an attempt to increase diversity is made by increasing the mutation probability.
In order to improve the computational efficiency and accuracy of the search tool, a black list of infeasible solutions was used to record all the individuals whose characteristics could imply negative impacts beyond the admissible limits.
Methodology for real impact assessment of the best location of DEESS 30 The choice of crossover and mutation probabilities (pc and pm, respectively) affects the behavior and performance of the GA (Romeu M . The heuristic updating principles of the pc and pm probabilities, as a function of the genetic diversity (Gdiv) are as follows: 1. Use large pc and small pm when in the current generation is large; 2. Use small pc and large pm when in the current generation is small.
pc and pm are adjusted according to the following conditions, presented in (8) and (9), and considering bounds and : If the genetic diversity ( ) is within the considered bounds, pc and pm are calculated through a linear interpolation as shown in equations (10) e (11).

Genetic diversity in the population
The genetic diversity of a population, i.e., the genetic variability of individuals, is responsible for the dispersion of solutions in the feasible space. To measure the resemblance of individuals, they must be regarded as multidimensional vectors and a distance calculated from one vector with the differences may be used. If the Where g is the chromosome gene of individual "i" and g the chromosome gene of individual "j".
To measure the genetic diversity (Gdiv), the following equation is used: it is a variable in the range [0,100] meaning that when the value is zero all individuals are similar and when it is 100 all individuals in the population are dissimilar.

Selection mechanism
The selection mechanism on iNSGA-II is performed using a "Binary Tournament Selection" based on rank and crowding distance. An individual is selected if its rank is lower than the rank of the other or if its crowding distance is greater than the crowding distance of the other, in case they share the same rank. Parents are selected using non-domination, with assigned crowding distance, expressed through the crowded-comparison-operator (≺ ) that provides a fast and easy way to estimate the spread in obtained solutions (Deb et al., 2002).
The crowded-comparison-operator provides the estimation of the perimeter of the cuboid formed using the nearest neighbors as the vertices. This operator will provide the needed information to choose regions with lesser density. The comparison is carried out in two steps for every individual in the population: 1. Assignment of non-domination rank ( ); 2. Assignment of crowding distance ( ) for individuals on the same front: The crowding distance represents the sum of absolute normalized differences in the function values of two adjacent solutions. For boundary solutions, namely solutions with smallest and largest function values, an infinite distance value is assigned.
Considering the non-domination rank ( ) and the crowding distance ( ) the partial order for the crowded-comparison-operator (≺ ) is: Since all individuals are included in population "Rt" by using the binary tournament selection with crowding-comparison-operator, elitism is ensured. solutions (Klamroth & Miettinen, 2008).

Multiobjective optimization and decision making support
The interactive methods help to overcome the above-mentioned constraints enabling the DM to learn about the problem, its possibilities and limitations, as well as the interdependencies among the objective functions (Klamroth & Miettinen, 2008). These methods can be applied to decision making or optimization processes in which the preferences are interactively refined at interleaved steps (Kacprzyk; & Pedrycz, 2015). However, it may be worth to first get an overview of what the Pareto set looks like, in order to acknowledge the feasibility of solutions (Klamroth & Miettinen, 2008).
The available methods to introduce the DM preferences in the optimization process are not conflictive and each has advantages and limitations. Moreover, the potential of combining these tools and approaches enable to develop hybrid approaches, with better performances when solving multiobjective optimization problems (Klamroth & Miettinen, 2008), (J. Clímaco & Antunes, 1994).
In the second phase the final selection of non-dominated solutions from the Pareto Set might be a difficult task, mainly due to the fact that the set of available solutions, and the corresponding representative Pareto Frontier in the objective function space, may be crowded (Iris & Asan, 2012).

Externalities: CO2 impact evaluation
Assessing externalities can be a very difficult task. In fact, all the issues regarding environmental impacts are critical due to the difficulty to monetize them, as the environment and its resources are generally considered abundant, renewable, without boundaries and free of charge.
Many authors aimed to minimize environmental impacts by assessing not only their emissions but also the linkages between acquisition and consumption, giving a price to the associated greenhouse gas (GHG) emissions and welfare loss (Paltsev et al., 2008), (Milne et al., 2013), (Grant et al., 2006), (Commission, 2005). A well-known example is the ExternE methodology (Commission, 2005) , which comprises a framework to transform different impacts into a common unit, using specific conversion factors (Neves, 2004).
The difficulty to implement and quantify impacts resulted in many methodologies being unused, such as the ExternE methodology, replaced to some extent by the CO2 equivalent valorization within the European emissions trading system (EU ETS). The CO2 equivalent is an internationally recognized term, used by the Kyoto protocol (Ki-moon, 2008) to describe different GHG in a common unit, enabling their comparison in terms of their global warming impact. As a tool to help accomplishing the Kyoto protocol targets, the EU ETS is the first (and still the largest) international system for trading GHG emission allowances, covering 28 EU countries.
The EU ETS sets a cap (or limit) on the total amount of GHG emissions that can be emitted by factories and reduces the cap over time, so that emissions fall. Despite its limitations, this approach is under consideration among several non-European countries (and countries which did not ratify the Kyoto protocol for political or economic reasons) (Paltsev et al., 2008), (Commission, 2015d).
Launched on the 1 st January 2005, the initial trading period ran for three free trading preparation years in a "learn by doing" process. Since initial emission projections revealed to be wrong, the first phase enabled to assess the correct allocation of allowances among member states and sectors. After this period, and using the assessed data under the EU ETS, the second phase began on the 1 st January 2008 (until the end of 2012), matching the first commitment period of the Kyoto Protocol.
In order to guarantee the EU as a whole, as well as each individual member state Kyoto commitments, the EU ETS emissions cap had been settled 6.5% below 2005 levels. The second phase showed that harmonization (regarding the cap on overall allowances settlement) within the EU ETS is imperative to guarantee the EU emission reduction objectives at least cost and with minimum competitive distortions (Commission, 2015d).
With a timespan of eight years, the third stage started on January 2013, and introduced a more efficient, harmonized and fair system. The cap was settled in less 21% of emissions compared to 2005 and the amount of auctioning is expected to be more than half. A single EU-wide cap was introduced and allowances are now allocated on the basis of harmonized rules, avoiding the need of national allocation plans (used in phases 1 and 2). The cap was settled considering the average total quantity of allowances to be issued by member states in phase 2, the broadened scope of the system that was introduced in phase 3, and the overall reduction of GHG emissions, compared to 1990. Considering the target EU emissions reduction, the EU defined a wider cap, as follows (Commission, 2015d): Due to its effectiveness on promoting the reduction of GHG emissions, as well as searching a cost reduction of cutting emissions (increasing the market liquidity and making the carbon price more stable), the EU ETS is currently being integrated with another cap-and-trade systems. This will enable participants to comply with their obligations in a more cost effective way, promoting a global cooperation on climate change, and leveling the international playing field (Commission, 2015d).
Even with adjustments in the market rules (as a result of stage 1 and 2 learning process), there is a surplus of allowances since 2009, affecting the orderly functioning of the carbon market (e.g. excessive reduction in the carbon price). As a result, the commission postponed the auctioning of 900 million allowances until 2019-2020. The impact analysis regarding this option showed the ability to rebalance supply and demand in the short term, as well as to reduce price volatility without compromising competitiveness (Commission, 2015b). Moreover, the commission proposed to make a structural change in the ETS in the long term, establishing a market stability reserve. This reserve would address the present surplus of allowances, improving the system resilience (Commission, 2015b).
In spite of its pros and cons, the EU ETS is part of a continuous progress towards a low-carbon society which objectives are to cut emissions to 80% by 2050 compared to 1990 levels (with intermediate milestones of 40% by 2030 and 60% by 2040) (Commission, 2015b).  (Commission, 2015a). For transportation and building sectors, the cuts are expected to achieve almost 60% and 90%, respectively, by increasing efficiency in current technology, replacing current vehicles with hybrid or electric technologies, and introducing biofuel in aviation and haulage. In buildings, energy performance is expected to improve with the use of passive architecture, retrofitting of existing buildings and replacement of domestic fossil fuels use by electricity and renewable sources (Commission, 2015b), (Commission, 2015c).
Guided by the ETS, the EU cap-and-trade system is an important instrument currently available and feasible for GHG impact assessment, and therefore the DEESS associated externalities will use this system. Using the EU ETS will contribute to the development of a methodology that uses existing and available data within the energy sector.

Methodology proposal
The proposed methodology is outlined in six steps, as presented in Figure 3-1 (Gonçalves et al., 2015b). The objective is to combine the determination of the best location for the storage units with the definition of the best schedule of operation, giving the DM insights on his choice (and on the associated impacts of possible tradeoffs), as well as testing possible management schemes that may be proposed for a stationary DEESS within a smart grid environment.  The network characterization block defines the daily demand diagram of the distribution substation and the active and passive elements of the electricity grid.
This step characterizes the entire network, describing buses, cables and interconnections. The information is compiled in data matrixes used to perform a power flow analysis with the help of the Matpower toolbox (Zimmerman et al., 2011) and Matlab™. Although the methodology is here applied to study the best ESS location within a distribution network (considering different management schemes), it may be applied to any type of grid, as long as the correct characterization is provided. The methodology does not consider potential problems arising in the operation of protection devices due to reversal of power flows.
The Working definition block, presented in Figure 3- The present work also implements a possible pricing scheme to be used for promoting DEESS exploitation, because the existence of a regulatory framework may stimulate the existence of market players intending to invest on energy storage.
Since one of the main objectives is to balance the surplus/deficit periods of renewable energy (RE) availability, the energy recovered from energy storage is assumed to be rewarded on an equivalent basis to the energy that is displaced as explained in section 3.2.
The iNSGAII block uses the genetic algorithm presented in (Gonçalves et al., 2015a), to search for non-dominated solutions. The tool uses a "Binary Tournament Selection" based on the rank and crowding distance to choose the best individuals for the evolution process. An individual is selected if its rank is smaller than the rank of the other, or, if its crowding distance is greater than the crowding distance of the other in case they share the same rank. (Figure 2-9, section 2.3.3). For the considered case study, the iNSGAII population of possible solutions was assumed to be composed of 150 individuals for a maximum number of 100 generations (both values that were consistently above those needed for convergence to be attained).
The evaluation functions were integrated in the algorithm using powerflow analysis to assess the impact of each individual solution on the network performance.
The Externalities block represents the assessment of possible avoided greenhouse gas (GHG) emissions regarding the chosen solution (or set of chosen solutions), as well as a surplus tariff proposition that can improve the associated economic benefit.
The new surplus tariff proposal is based on the solution's annualized cost and the annual economic benefit needed to obtain profit. After doing this economic analysis the DM will increase the perception of possible viable choices.
All the possible non-dominated solutions are presented in the Pareto optimal front as result of the optimization process. The final preferred solution must be chosen by the DM considering its own criteria, namely the tradeoffs between objectives. The stakeholders that can benefit from the developed methodology are the DSO, a private investor or any authority acting according to the societal interest.

Data clustering
The need for evaluating the DEESS contribution within a smart grid environment, presents an opportunity for data classification, considering the large data sets for dynamic grid analysis, such as the annual monitoring of distribution LD, energy market prices and wind power generation.
The following sub-sections present a methodology 5 to group information by similarity, enabling the analyses of the associated impacts on the electricity grid of the deployment of a DEESS, by providing reliable daily simulation profile prototypes of market price, LD and wind power generation (Miguel et al., 2016).

Clustering techniques
Cluster analysis is a convenient method for identifying homogenous groups of objects called clusters, aiming to group objects by similarity, but keeping a significant difference between them (Mooi & Sarstedt, 2010).
Clustering techniques could be divided in two main methods, namely the Partitional clustering and the Hierarchical clustering, as presented in Figure 3-3. The Partitional algorithms use an interactive optimization paradigm. It starts with an initial partition and by applying an interactive control strategy, it tries to swap objects to improve the quality of the clustering. The quality of the clusters will strongly depend of the initially selected partition. In the k-means algorithms a cluster is represented by its center of gravity while the K-medoid algorithms use the closest object from the center as the center representation.
The Hierarchical clustering methods create sequenced partitions, each one containing the previous one in a hierarchic way. There are two types of hierarchical methods, the agglomerative clustering algorithms and the divisive clustering algorithms.
The agglomerative clustering method starts with an equal number of clusters and objects (each cluster has only one object) that will be merged in pairs until the number of clusters reduces to k partitions. The merging procedure always considers the nearest pair of available clusters and terminates when the hierarchy of clusters is built with just a single cluster containing all objects (Pujari et al., 2001).
The divisive clustering method is an inverse process to the agglomerative clustering method, starting with all the records in one cluster and iterating to split each cluster into smaller groups.
An alternative method introduced by Teuvo Kohonen in the early1980's, the selforganizing maps (SOM), is a type of artificial neural network (ANN) with an unsupervised learning process because the classes for the output vectors are not initially known (Kohonen, 1998). The Kohonen SOM could be compared to conventional clustering methods, considering the internal allocation rules and their performance (Sousa, 2006).
The cluster analysis used in the current methodology makes use of a hierarchical clustering method (HM) and an artificial neural networks method, more precisely the Kohonen SOM, as detailed in the following sub-sections. The final objective of the author is to develop cluster prototypes of daily price profiles, load demand profiles and wind generation profiles

Hierarchical Method (HM)
As previously mentioned, the hierarchical clustering method investigates possible groupings of data, by creating a structure similar to a hierarchical tree (Sousa, 2006). Such tree is comprised with a multiple level clusters hierarchy, where clusters at one level are grouped in clusters of a higher level. The adopted procedure to create the cluster profile prototypes is here only generally described, but a detailed description may be found in (Sousa, 2006).
First the input data are evaluated (in this case, market price diagrams, LD and wind power generation profiles), and the type of clustering results that are valuable to the analysis are identified. For the specific case, the assessment is based on either the magnitude of the difference between cluster members (defined as the difference between extreme values of a daily profile), or on the diagram shape.
The diagram shape comparison is performed by using the Euclidian distance between individual normalized vectors or/and intermediate vectors that represent the data clusters in the tree dendogram. Initially the assessment is performed between individual profiles but as a result of the data clustering, this comparison will be also performed between individual profiles and clusters or between clusters, until the hierarchical tree is composed of one final cluster.
The Euclidian distance between two profiles ( and ) is measured as shown in equation (15) The authors performed an analysis to the meaning of the obtained clusters by attempting to match with day types, namely weekdays, Saturdays, Sundays and holidays. After several tests, while for electricity profiles (LD and wind power generation) the most relevant was the shape of the profile, in the case of the price profiles, the magnitude of the difference assumed greater relevance to differentiate daily diagrams.
The sequence of partitions is assessed according to the distance between objects, namely their linkage distance that is characterized as single, complete, average, centroid or ward distance, the latter being used in the present case.
The following table presents the different ways to measure the linkage distance between objects, which is the main difference between existent clustering methods (Sousa, 2006).

Single linkage
Uses the minimum distance between objects under analysis. It tends to produce unbalanced and scattered clusters especially in large sets of data, without taking into account the structure of the formed clusters.
Complete linkage Uses the maximum distance between objects under analysis. It tends to produce compact clusters with equal diameters not considering the structure of the formed clusters.

Average linkage
Uses the average distance between objects under analysis. It tends to group clusters with small variances. It could be seen as an intermediate method between single method and complete method. It considers the structure of clusters and reveals to be a relatively robust method.

Centroid linkage
Uses the distance between the clusters centroids or means. This method is one more averaging technique, assuming that the points can be represented in a Euclidean space for geometric interpretation.

Ward linkage
The distance between two clusters is the sum of squared deviations from points to centroids. The goal is to minimize the within-cluster sum of squares. Tends to produce clusters with a spherical shape and similar numbers of observations (representations), being also sensitive to outliers. It assumes that the points can be represented in Euclidean space, providing a geometric interpretation.
After calculating distances between objects it is possible to group profiles considering their proximity, using the shortest distance. The distance between clusters is determined each time a new cluster is created, comprising new objects or group of objects, until the hierarchical tree is composed of one single cluster.
More detailed information regarding this procedure can be found in (Everitt et al., 2011).

Artificial Neural Networks (ANN) Method
The ANN method is based on the Kohonen SOM (Kohonen, 1998). This algorithm is able to capture eventual nonlinear statistical relationships between elements of data into a simple geometric relationship (neural network) on a low dimensional display, based on similarity and topology, a feature that is one of the most important advantages of the method.
Similar to the hierarchical method, this clustering method is also an example of unsupervised learning because the classes of the output vectors are not initially known. The used data in the training stage is crucial since the neural network is iteratively adapted in order to minimize a performance function, the result being an automatic mapping of the relationships between the inputs and outputs (Sousa, 2006). The neural networks contain two layers: an input layer of x-observations with a p dimension and an output layer representing k nodes for the k clusters, each one associated to a weight w with dimension equal to p. The clustering process occurs when an input vector is assigned to an output node characterized by a weight synaptic vector w with dimension equal to p. (Sousa, 2006). According to Everitt et al. (2011), the random weights initially assigned to nodes change with the learning (training) process, representing the allocation of input data to clusters. The stabilization of the iterative process eventually occurs when weights corresponding to cluster prototypes ensure that similar clusters are represented closely on the map (Sousa, 2006).
Methodology for real impact assessment of the best location of DEESS 47 The ANN application was performed as shown in Figure 3-5 using the available MATLAB toolboxes. The Clustering block is where data will be clustered based on similarity and topology, with a preference for assigning the same number of instances to each class. This step considered a total of 1000 training steps input space coverage, an initial neighborhood size of two, and a layer topology function that calculates neuron positions for layers whose neurons are arranged in an N dimensional grid (e.g.

Profile prototypes definition
The proposed clustering methodology, presented in Figure 3-8, was fed with one year of data on a quarter hour time step. The LD and wind power generation diagrams were obtained directly from the Portuguese electric energy sector databases. The electricity wholesale market prices for one year were converted from hourly data into quarter hour time step data.
In order to define the appropriate use of gathered data, an experimental setup was performed using two approaches for the data clustering techniques, one using the input data without any treatment and a second one where the input data were subjected to a normalization procedure such that the sum of all the elements of each vector equals 1, in order to cross-check their validity (Sousa, 2006).
The normalization process was performed using equation (16) where is the value of the i th element of the vector to be normalized and is the number of elements (m=96).

= ∑
This experiment showed that in the case of the electricity profiles, the use of the shape of the diagram resulted in data being differentiated in workdays and weekends. However, in the case of the market price diagrams, the magnitude similarity between diagrams resulted in more discrimination than when using the shape of the diagram. An important issue was the need to adjust the data to the two clock changes of the daylight savings scheme, namely on the last Sunday of March and October.
The hierarchical method helped to choose the appropriate number of clusters graphically, using the tree dendogram and the inflexion point of the curve that relates the distances between clusters with the number of clusters (cf section 4.1 related to the case study example) (Salvador & Chan, 2004). A comparison was made between the output of both methods regarding the assignment of day types (working day, Saturday, Sunday and holiday) to clusters.

Proposal of a pricing scheme for the energy storage service
Since the research aims to contribute to the integration of electric ESS within a radial distribution system (RDS), a definition of the regulatory framework for the management of the infrastructure is required.
Being the purpose of promoting ESS the benefits these systems could bring to maximize the generation of electricity from renewable sources, it seems plausible to consider electric ESS on an equivalent basis to a renewable generator, or what in Portugal is defined as a Special Regime Producer (SRP). In fact, a better match between supply availability and demand needs is expected to help maximizing the interest on investing in renewable generation. However, this feature is not yet accounted for in the current legislation.  The last resort retailer, when making purchase offers on the Iberian electricity market (MIBEL), takes into account the energy acquired in the special regime.
Hence, the special regime production does not appear explicitly in the energy market, but it has influence on the price since it influences the volume of the purchase offers.  The proposed pricing scheme should consider the electric ESS as a SRP that buys energy at the daily market price and sells it at SRP prices, in high demand periods.

Optimization method
The following sections detail the approach briefly described in section 2.3, namely the population definition, the crossover method, the coding and decoding technique, termination operator and the adopted objective functions to measure the DEESS network impacts of each solution.

Initial population generation
The initial population is randomly generated considering a predefined number of individuals (Nind) with a vector of length equal to the total available nodes predefined by the user. Moreover, during the random generation, it is assured that the individuals of a certain population are all different in order to guarantee a high genetic diversity and to avoid repeated computation. Therefore, it should be guaranteed that the total number of individuals for each population are less than the possible combination of available nodes (expression (17)).

≤ 2
With -The total number of individual in each generation; -The total number of available nodes for ESS installation. (17)

Multi-point crossover
Since the coding technique is based on a binary assignment of the entire buses of the network, the chromosome tends to become very long. For that reason, and ambitioning the best performance of the algorithm, the author uses a multi-point crossover.
On the chosen multi-point crossover, the cutting point divide the chromossome in 2 bit parts, leaving a last part of 3 bit for odd chromosome lengths. The proposed method allows a fast crossover over the wide length chromosomes, as well as maintaining the descendants inherit characteristics from both parents according to a defined probability (pc).

Coding and Decoding Technique
The definition of the available nodes for the installation of storage units is made by the DM, based on technical criteria (e.g. which buses have critical load factors). The DM can easily define the availability of a specific node for ESS installation using a binary number (1 or 0). This technique will create a matrix of chromosomes with total length equal to the predefined number of individuals (Nind), as presented in Table 3-3. Each chromosome has a number of bits equal to the number of available nodes (nnos) for ESS installation. The correspondence between the nodes in the network and the chromosome element is achieved through the use of a reference vector with the same size of a chromosome, as depicted in Table 3-4.
The use of a reference vector (as presented in Table 3-4) will enable the use of shorter length chromosomes because it will assure the generation of valid individuals. After the creation of each population, the network model must be updated with the availability of storage.

Termination
Genetic algorithms are stochastic search tools and for that reason it is difficult to specify convergence criteria. However, improved NSGA-II uses as a first termination criterion the maximum number of generations (Npop), and secondly a specified convergence threshold (Cth). The motivation for using the diversity of population lies in the fact that the individuals of generation "Rt+1" may have similar performance to generation "Rt" implying a progressive reduction of genetic diversity. If the diversity of population does not suffer any changes, we may assume that GA converged.

Network Energy Losses (NEL)
The NEL objective function to be minimized, described in equation (18), is the sum of the network power losses (PL) in all the n branches of the medium voltage (MV) distribution network during the whole day. The elementary time interval is a quarterhour (tj=0.25h) so the data set has 96 values (m=96).

Network Voltage Quadratic Mean Deviation (NVqmd)
The NVqmd objective function to be minimized, described in equation (19)

Network Storage Annualized Cost (NSAC)
The third objective function, to be minimized, is the network storage annualized cost (NSAC) for installing x units of DEES with an individual capital cost ( ), as presented in equation (20).
The is calculated considering the capital costs ( ) and the capital recovery factor ( ) as presented in equation (21). Where is the dimensionless discount rate and the expected life of the equipment, measured in years;

Network Energy rate benefit (NERB)
The fourth objective function, to be maximized, is the network energy rate benefit (NERB) considering the energy tariff ( ) and the required energy ( ) for charge ( ℎ) and discharge( ℎ) periods during the 96 elementary time intervals, as presented in equation (22).

Case study
In order to test it, the methodology was applied to a case study. Due to the lack of detailed data with enough quality, the network characterization of the physical infrastructure was based on a standard IEEE network. All data used in the optimization process was based on the Portuguese energy market, and treated using clustering methods according to the previously developed clustering methodology.

Data clustering process
The first step to define the appropriate number of clusters was the analysis of the dendogram or tree of clusters that resulted from the clustering process.

Clustering of wholesale market price profiles
In order to compare the data allocation, verify the similarity of results with both methods, and understand the meaning of the clusters, the obtained outputs were analyzed in terms of the total number of allocated days (hits), on weekdays, weekends and holidays. As an example, Table 4-1 and Table 4-2 provide the comparison of electricity price profiles. The resulting clusters for the electricity prices did not show a particular correspondence to the type of day, no cluster being formed more with a type than with other. Although the current market situation creates conditions for this to happen, it suggests also that the natural consumption variation between workdays and weekends is not influencing prices. Merge Distance The resulting clustering price prototypes can be visualized in Figure 4-3. Both methods present similar types of price prototypes and the approximate same number of hits. This resemblance ensures that any of the clustering methods may be used for generating reliable simulation prototypes.   If the stimulus to use storage consists in economic incentives based on market price diagrams, the viability of investments in storage might be dependent of technology development to increase total system efficiency. Since technology improvements are costly, the stimulus to investors may require a feed-in tariff scheme, such as the one proposed in section 3.2.

Clustering of LD profiles
Both methods presented rather segmented clusters in the electricity demand profiles (based on weekdays and weekends). The differences between methods were essentially related to the total number of allocated days. For different distributions, another measurement method may be used (e.g. complete linkage). The holidays were not represented in any particular cluster, most likely because these days normally present patterns similar to weekends. The results for both methods, regarding the 2008 high voltage (HV) LD are presented in Table 4-3 and Table 4-4.   In order to develop a prototype, the clustered data were used to obtain an average profile of each cluster. The developed prototypes for each method are presented in   In order to determine the profile variation range, the average profile for all data was compared to all five obtained prototypes. This process allowed to obtain the average variation range of the profiles, as presented in Figure

Clustering of wind generation profiles
As one of the major motivations for storage is to increase the potential of renewable energy by decoupling the periods of usage from the periods of generation, the clustering of historical data from renewable energy generation allows the analysis of different daily patterns and provides insight on their relevance. The obtained prototypes were dissimilar as expected, due to the unpredictability of this energy source, as shown in Figure 4-

Maximum variation (maximum) from the demand average profile in [%]
Maximum variation (minimum) from the demand average profile in [%] values in Figure 4-9 shows that a single average profile is of limited use. ANN cluster #1 (17.21%) and ANN cluster #5 (13.11%), representing 30.33% of the wind pattern profiles in one year time, show clearly the relevance of the integration of energy storage systems. On ANN cluster #1, the energy generated from 00:00 to 08:00 can eventually be stored to be used in periods of higher consumption, while in cluster #5, the interest in storage is not so obvious, as the maximum generation occur in periods when consumption is traditionally high.

Technology characterization
The data used to perform the simulation of storage systems was obtained from a manufacturer regarding a nanophosphate lithium ion battery with the respective power converter (Systems, 2013), (Vartanian, 2010), (Systems, 2012).
According to these data, a total energy of 61.13Wh/cell was assumed, corresponding to 95% of depth-of-discharge. This option was based on the manufacturer information indicating more than 7000 cycles even with 100% depthof-discharge (Systems, 2013).
Given the individual cell dimensions of 7.25x160x227mm (Systems, 2012), the proposed solution required roughly 0.1m 3 for the battery systems (a battery pack of two Rows of 180 cells, with a total capacity of 22 008 Wh) without power converter, a volume which could be easily integrated in any power transformer facility.
According to several manufacturers, a plausible 90% efficiency can be expected for the power converter (NEC, 2013) when performing a complete charge/discharge cycle, assuming a nominal discharging capacity of 20.85 kW and requiring a charging power of 23.17kW per power transformer.
The author considered three different ways to obtain the unitary DEESS system cost, considering different sources of data.
The first source of data (J. Vasconcelos et al., 2012), presented in Table 2-8  Using equation (4) and (5) presented in section 2.1.4, the author obtained the annualized capital costs presented in Table 4-6.

Network characterization
As an example of a distribution network, the case study network was based on the IEEE 69 bus, 12.66kV RDS (Sahoo & Prasad, 2006), a frequently used test network

ESS working profiles
The energy storage charging and discharging daily behavior can be very diverse depending on the DM preferences. Based on the literature review, plausible DM goals were considered, according to three possible management schemes: 1. Management scheme A -To maximize profit from daily energy spot market rates; 2. Management scheme B -To minimize daily energy distribution network losses; 3. Management scheme C -To maximize the societal value of wind generation.
Considering the three previous stimuli, the author proceeded to the definition of a C/D binary vector. The five prototypes obtained at the clustering stage for energy market rates, distribution network LD and wind generation were used for the charging/discharging profile definition.
In order to increase the C/D daily periods the author considered, as a general approach, for management schemes A and B, the use of the average value of each daily data set as a reference. The storage units should charge and discharge when the daily profile presents the extreme values below and above the reference value, respectively. The designation of "-1", "1" and "0" was adopted, to define if the storage system is charging, discharging or on standby mode, respectively.  The minimization of losses was performed using the load flow calculations for the network load profiles. For the case of defining the C/D profile when attempting to minimize losses, the LD can be used as a proxy due to the fact than network losses are usually strongly dependent of demand.  The renewable generation integration, in particular wind generation, required a different type of analysis. The main motivation for using energy storage for renewable generation resides in storing energy during high power generation periods and delivering it during low power generation periods, in order to stabilize the net distributed power injection profile. Since both generation and load diagrams were subject to a normalization procedure during the clustering process, a measure of the renewable generation availability was obtained from the difference between the average LD profile and renewable generation prototypes as represented in  Based on the variation, the following procedure to obtain the C/D binary vector is similar to previous scenarios as presented in Figure 4-15.   It is important to note that for the third case the C/D developed binary vectors resulted from combining the five wind generation prototypes with the annual average energy LD.
Considering the time definition for charging and discharging the storage units, the information was converted into a C/D matrix to be used in the iNSGA-II during the optimization process.  Difference between RE and load demand profile

Model constraints
The operation of ESS is constrained by the individual capacities of each storage unit which was specified considering the required voltage for the power inverter, a power limit which considers the minimum values of the bus capacities in the case study, and the physical limit of each individual storage solution for a maximum volume of 1m 3 .

Results and discussion
Due to the important influence of the data used in the study, the clustering process is detailed in section 3.1, showing to be an adequate method to obtain representative data of real and specific location behaviors. Nevertheless, in order to help assessing the results associated to each management scheme, a data analysis of the used data should be performed.
The combination of prototypes, e.g. those regarding LD and energy market rates, does not have a specific annual representativeness, although each considered prototype represents real behaviors. Therefore, the performed combinations should be considered as possible scenarios whose relevancies will be different and can be assessed. For the sake of demonstrating the application of the proposed methodology, and for the sake of simplicity, the weights assigned to the scenarios were considered equal.
Searching a similarity among obtained prototypes revealed that the market prices prototypes #4 and #3 were similar to the LD prototype #4, as shown in Figure   Moreover, the average annual LD profile and market price prototypes also showed some dissimilarity, as shown in Figure 4-18, mainly due to the nonrepresentativeness of the LD of the particular substation used regarding the global LD which influences the national energy market prices.
Since most players of the energy sector react to price stimuli, this dissimilarity may lead to non-optimal grid management on a local level. This phenomenon indicates that it may be more adequate to use a local price scheme, adapted to the cost of supply near consumption in order to avoid the use of DEESS with negative consequences to the electricity grid.   When using the aggregation of objectives, the DM has to establish tradeoffs or preferences by setting weights for the different objective function. An interactive search is the most suited way to perform a multiobjective analysis(J. N. Clímaco et al., 2003) as it enables the DM to assess the impact of each pair of objective function results without hiding information. It was the option selected for this study because it provides the DM with all possible options and does not require the DM to freezes his/her preferences as a set of numbers or as a set of mathematical expressions defining utility functions. This method is further detailed in the following subsection.

DM interactive search
As a first step, an analysis of the Pareto fronts is performed, providing the DM with 2D representations, combining different pairs of objective functions. This procedure helps the DM in his assessment of associated impacts, especially when more than three objective functions are present (where the amount of data can hinder the analysis).  Although different DMs were considered, i.e., possible stakeholders/agents in the energy market each with its own multiple specific preferences, the most plausible situation is that only one of such agents will decide to invest in a storage system in a particular network. As such, this analysis is intended to solve the problem considering only one perspective at a time, defining choices regarding the definition of the C/D schedule, as well as the interactive analysis of the resulting Pareto front.
To provide the DM with reference points in the Pareto front, two possible methods are described in the next subsection. ii) providing the solution which has the minimum Euclidian distance to the optimal Pareto front available point.

Crowding distance method
The crowding distance, presented in (Deb et al., 2002),

Euclidian Distance method
Searching for the minimum Euclidian distance to the optimal Pareto front available point might be easier to understand. However, its calculation may need the definition of different weights for each objective function, something that can only be defined by the specific DM.
For the current case study only three objective functions were considered, namely the NEL, the NSAC and the NERB, as the NVqmd has a low variation among the majority of scenarios. For the sake of simplicity equal weights were assigned for the different objectives.
Applying the Euclidian distance method to find a reference point, as presented in equation (23), requires a normalization of all the tree objective functions.
The Euclidian distance to the optimal point of the Pareto front defined by the NEL, NSAC and NERB objective functions (‖ ‖ 2 ) used the variation range of each objective function ∆ , such as the NEL and the NSAC, which is calculated using equation (24), and a weight factor ( ) to consider the annual representativeness of each scenario.
The variation range is calculated considering each objective function value divided by the same objective function range of the Pareto front results, namely the maximum minus the minimum obtained value. Table 4-8 presents the results of the described normalization process for the same scenario 28 example: the C/D profile obtained from combining the LD prototype#3 and the energy rate prototype#1 for the MSch B goal, is applied. Solution number 28 had presented the smallest distance to the "ideal optimal point" (meeting the considered weights), providing a "balanced" solution in the Pareto front but which acceptance by the DM will depend of the implied tradeoffs, as shown in Comparing the solutions obtained with the Euclidian distance and crowding distance method, both methods pointed to nearby areas. Both methods give an important reference to the DM, the crowding distance not requiring additional data manipulation or weight definition.

Results analysis
In order to have comparison values for each scenario, a first analysis was performed considering the system without energy storage and obtaining the reference values of losses (NEL) and voltage quadratic mean deviations (NVqmv), as presented in Table 4-9.  (Table 7-1, Table 7-2, Table   7-3, respectively). In order to simplify the numerical analysis, only the maximum and minimum values for each scenario are presented in the summarized tables. Table   4-10 presents the global analysis.
Results show that the MSch B returns higher NERB results when compared to MSch A. Although not strong, the bias between the substation LD (and consequently with network power losses) and energy market prices will contribute to reduce the reluctance of DSO in integrating resources out of his control (Sovacool & Hirsh, 2009), even if the stakeholder objective is only economic. MSch C had the smallest economic income and NEL reduction. These results can be explained considering the obtained prototypes during the clustering stage, in which the obtained wind generation prototypes showed no correlation with market energy prices and the substation LD. Therefore, the suited C/D periods may introduce some pressure into the network in periods of high demand.    On average, the DEESS produced an increase of the NVqmd (by 0.0050 pu), while the remaining scenarios showed a slight reduction of the NVqmd (by 0.0007 pu).
The increase was higher (about 0.0168 pu) for the MSch A scenarios 15, 20 and 25.
In the MSch C, the NVqmd increased (by 0.0011pu) in 36% of the studied scenarios.
All the mentioned C/D profiles have in common the existence of two charging and discharging periods, increasing the probability to charge and discharge in critical capacity periods.
In the MSch B scenarios, the NVqmd did not present a     The results for scenarios 21, 22 and 24, showed that, for solutions with similar NEL results of 7 205.0 kWh, only 9 out of 48 buses were identical (Table 4-12). The results showed that predefining the admissible buses for DEESS installation and capacity optimization, as performed in previous studies (Ippolito et al., 2014), (Fossati et al., 2015), might have limitations, because different buses may present different advantages for different energy services.   Figure 4-30 or as presented in Table 4-12, where the assignment of units is marked with "1".
For the specific DM preferences used in the example, repeated DEESS locations were obtained for different management schemes (namely in nodes 54, 55, 59, 61, 62, 64, and 69).This behavior pattern reveals the most critical nodes in the network in which storage obtained the best results, according to the stated preferences. The results are aligned with previous proposals (Eyer et al., 2004), (Kempton & Tomić, 2005) to combine benefits, in order to increase the final energy storage benefit (e.g., reducing the GHG emissions or providing regulation services).

Externalities: avoided GHG
The EEX auction platform is hereby used as the reference for emission allowances, in order to obtain real data of GHG costs commonly accepted by the majority of European countries. The EU ETS is a policy to combat climate change and reduce industrial GHG emissions. This is currently the largest international system for trading GHG emission allowances, covering more than 11 000 power stations and industrial plants (in 31 countries), and it uses auctioning as the basic principle to allocate allowances (Commission, 2015e), (EEX, 2014). In 2013, more than 40% of all allowances were auctioned, and these figures are expected to increase in the following years. The role of transitional common auction platform to deploy the use of allowances has been awarded by EEX (on behalf of 24 Member States and Poland) during a transitional period.
The environmental assessment is performed based on the following assumptions: • The reference price for emissions is an annual average price; • The avoided emissions are established a priori by the DM, considering available conventional power plants for energy regulation (such as coal and natural gas); and • Avoided emissions calculation should use the efficiency of the specific power plant; Since efficiency depends on specific power plant characteristics, the efficiency indicators for Portuguese power plants (ERSE, 2012) are considered, to obtain the associated emissions. Therefore, in order to calculate GHG emissions for the case study task, the following data and assumption were considered: • Allowances annual average price ( ) in 2013: 4.19 [€/t CO2]; • Avoided power plant technology: Coal; • Power plant emission factor ( ): 897 [g/kWh]; • The charged energy in the DEESS is obtained from a RE source.
Choosing a technology to be partially replaced by DEESS should account for the market offer. However, due to data limitations, the study assumes an avoided power plant technology.
The avoided GHG emissions are calculated using equation (25) where ℎ relates the DEESS discharged energy to the network.
[ 2] = × 10 (25) Hence, the daily avoided cost in GHG emissions ( ) is calculated using equation (26).   Even without economic attractiveness for simply buying and selling energy, results show how important DEESS may prove to be for a network, providing power services which can be paid for. An example may be the possibility of replacing at least part of the existing conventional generators which are being used simply as backup generators. However, due to the lack of regulation on this issue, an adequate framework should be developed to deploy energy storage applications.

Sensitivity analysis
Using the optimization process results as reference, several feed-in tariffs were simulated to verify the associated NERB and also to make a preliminary assessment of the values required for obtaining a positive net benefit of deploying DEESS.
Considering non-dominated solutions associated to each MSch, the solutions with the highest NERB were selected, i.e., scenarios 15, 36 and 66 with 23, 24 and 21 batteries to install in buses, comprising MSch 1, 2 and 3, respectively.   Table 22 shows a sensitivity analysis on the influence of the surplus feed-in tariff on the results. Benefits may have higher relative impacts than the relative variation of the tariff. As an example, for a 36% variation of the surplus feed-in tariff, the network energy rate benefit increased by 30%, 39% and 34%, depending on the MSch being 1, 2 or 3, respectively.
The results showed the importance of market prices to ease the DEESS interest. for MSch 1, 2 and 3, respectively).
The results show that DEESS can only become attractive if the economic benefit results from more than just energy purchasing and selling. Future viability should also include power demand services besides energy service, taking profit of its fast response capability to network requests of power demand.

Conclusions and future work
This thesis presents a methodology to provide the decision maker ( problem. The consequence is that, instead of being presented with an "optimal solution", the DM will have to choose the preferred solution from the highest quality set of non-dominated solutions presented by the search tool, assuming tradeoffs between the objectives.
The research question to be answered took into consideration the circumstances in which urban DEESS would be feasible. The answer to this question strongly depends on the associated technological costs, being this feature one of the most important issues for DEESS economic viability and attractiveness. Results showed that a higher technological efficiency of DEESS devices would be very important to achieve a positive economic benefit. Even with the possible future use of a suitable feed-in-tariff, DEESS feasibility may depend of combining energy, power and environmental services for increasing its economic attractiveness.
The economic CO2 impact assessed within the EU ETS market was not enough to obtain a positive economic benefit within the chosen case-study. The low market value of the considered emissions limited the contribution of this potential benefit to the viability of the solution. Therefore, the viability of DEESS is highly dependent of the consideration of all possible positive impacts.
The possibility of assessing power generation costs could have helped to increase the DEESS interest. However, this possibility proved to be unfeasible since the current energy market runs on daily energy auctions, the market players not being required to inform on their assumed costs when delivering the contracted energy.
The availability of DEESS depends on the envisaged management scheme since the ESS can only provide a service at a time. The energy services to be provided are restrained by the considered technology, with different performance impacts (more or less severe) and different costs (e.g. those dependent on the expected lifetime), being battery storage technologies able to provide most of the energy services to the network.
The location and sparsity of storage units strongly depends on the used management scheme. Different impact results and solutions are obtained for different management schemes, suggesting the need to incorporate also this level of decision on the multiple objective formulation. A definitive solution implies naturally the analysis of the possible tradeoffs according to the preferences of the DM, which could be any private investor, the DSO or an authority. Depending on the stakeholder, different objectives could be pursued, leading to a different final "best" choice among the non-dominated solutions.
The developed work enlightened the need to have a framework that stimulates existing or new market players to invest on ESS as a way to increase the share of renewable energy, by storing energy generated on surplus periods to be used when generation is limited. One possible approach to this objective could consist of using a feed-in-tariff scheme, rewarding energy recovered from energy storage on an equivalent basis to the energy delivered directly to the grid by renewable generators.
The MSch profile definition for renewable integration should take into consideration the real behavior of the wind, in order to obtain a plausible management scheme of the DEESS, as well as, to calculate the associated benefit. Although initially unforeseen, a data clustering process enabled the use of real data prototypes of prices, load diagrams and wind generation, allowing a more accurate definition of the impacts of the use of storage devices.
In order to improve the developed methodology, future work should include the following directions:  To incorporate a dynamic C/D optimization routine, searching for the best ESS management scheme, given the economical and technical contraints;  To consider the deployment of distributed renewable generation within distribution networks, a future research should include the impact assessment of DEESS, when facing the non dispatchable injection of small power generation. This feature will enhance the suppletive nature of DEESS by characterizing and additional element of the complex energy market environment.
 To increase the number of objective functions in order to assess the influence of DEESS capacity on the profile diagram, and to include the assessment of regulation services such as the secondary and tertiary frequency regulation control;  To include a second optimization procedure to define the optimal capacity of individual units per installation site. This feature may introduce some insights regarding the assessment of best location of DEESS as well as give insights on how storage capacity could influence the network global performance.    Electricity price difference to the average DEESS charge/discharge          Difference between RE and load demand profile DEESS charge/discharge