Вы находитесь на странице: 1из 10

762

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 3, JULY 2003

Minimum Losses Reconfiguration of MV Distribution Networks Through Local Control of Tie-Switches


Antonino Augugliaro, Luigi Dusonchet, Mariano Giuseppe Ippolito, and Eleonora Riva Sanseverino
AbstractThis paper deals with the problem of optimal reconfiguration of radial distribution networks for minimum loss operation. The proposed control strategy of the open-closed status of the tie-switches is distributed, since every MV/LV node is provided with local controllers having some measured entities as input. It also does not prevent the system from the future implementation of centralized control; instead it may represent the first step toward a complete automation of the distribution system. The proposed strategy is indeed organized in hierarchic levels, the highest of which may be in the future a central control. After introducing the general problem of networks reconfiguration, a review of the state of art on the subject is reported, even though the solution methodologies are usually related to the centralized formulation of the reconfiguration problem. Then, the proposed local control strategy is outlined and a detailed description of its different parts is reported, with special attention to all of the measures for a better performance of the system. Finally, results of a number of simulation runs are reported in order to test the behavior of the proposed local control system in different possible operating conditions. Index TermsReconfiguration, optimal control, distributed control, power distribution control.

I. INTRODUCTION HE reconfiguration of radial distribution networks with the aim of achieving minimum losses operation is one of the functions that can be addressed in distribution automation (DA). Networks reconfiguration is generally treated as a centralized control problem, having as input quantities the loads and the current tie-switches status and as output the set of status of the network tie-switches related to the minimum losses radial configuration. Therefore, the centralized control requires a data transmission system connecting the loads and tie-switches to the control center. Besides reconfiguration, other functions comprised in the general DA term should be carried out. Their implementation indeed produces economical advantages for utilities and customers, respectively, in terms of variable costs reduction and of better service quality. Designing automated distribution networks is certainly much easier and cost-effective than automating certain functions afterwards. Moreover, in the latter case, some difficulties related to technical problems can arise. In general, it is always possible to think about automation as a progressive process, in order to improve the system performances without hindering the future complete automation. In the case

Manuscript received June 22, 2001; revised June 14, 2002. The authors are with the Department of Electrical Engineering at the University of Palermo, Palermo 90128, Italy (e-mail: augugliaro@diepa.unipa.it; dusonchet@diepa.unipa.it; ippolito@diepa.unipa.it; eriva@diepa.unipa.it; riva@cere.pa.cnr.it). Digital Object Identifier 10.1109/TPWRD.2003.813801

of partial automation for network reconfiguration, the problem can be that of approximating the centralized control by means of local controllers that, on the basis of measures at the relevant nodes, decide about the status of the closest tie-switches. Each local controller acts on the basis of the local effects of the current status of the entire system, on which the input values of the local controllers depend on. Since the relations connecting the local quantities with the minimum losses setting of all the networks tie-switches are quite complex, a control strategy based on threshold values would not work. The approach followed here is based on the use of neural networks and a deterministic algorithm within a control strategy which would prevent from the risk of creating meshes and/or islands. During the neural networks training phase, it is necessary to know for each loading condition the relevant minimum losses configuration. The latter can be obtained for small size networks by means of exhaustive search algorithms and, for larger systems, using any of the heuristic methods for centralized network reconfiguration proposed in literature is used. The developed control strategy has been specifically designed for existing networks as the first step toward a centralized control system; the main advantage of the proposed system is its low cost and the guarantee, in normal operating conditions, of radial operation while supplying all of the loads. Of course, the local control is less accurate than the centralized control in the identification of the optimal solution. On the other hand, when using a centralized control system in a large network, it is not possible to perform an exhaustive search and the use of heuristic approaches for the minimum losses configuration search only allows to attain suboptimal solutions [1]. In the validation of the proposed strategy (Section IV), the possibility to get wrong configurations is put in evidence. The occurrence of these failures is due to local random variations of the loads; these produce the input values to the ANNs, whose outputs create a network configuration that is different from the one that would be outputted by a centralized system. The problem mostly dealt with in the literature on the subject is that of the identification of the radial configuration meeting one or more conditions such as minimum losses, voltage profile flattening, load balancing, etc. A mathematical programming technique is applied and a solution algorithm is developed by Aoki, Ichimori, and Kanezashi in [2]. The variables identifying the locations of the switchgears in the network are considered continuous; once the problem is solved by this approximation, the effective positions of the switchgears to be opened are determined. Civanlar, Grainger, Yin, and Lee [3] develop an expression for the variation of the total power losses in a radial network when, due to the closing of one switchgear and the opening of

0885-8977/03$17.00 2003 IEEE

AUGUGLIARO et al.: MINIMUM LOSSES RECONFIGURATION OF MV DISTRIBUTION NETWORKS THROUGH LOCAL CONTROL OF TIE-SWITCHES

763

another, some loads are transferred to a contiguous feeder. The expression above is applied to establish a criterion for the identification of the pair of switches (one closed the other opened) giving rise to the maximum reduction of losses. Liu, Lee, and Vu [4] propose two algorithms for the minimization of losses in distribution networks, both valid in the case of loads modeled as constant current sinks, uniformly distributed as well as concentrated. Baran and Wu [5] deal with the problem of reconfiguration, both for the reduction of losses and for the load balancing at HV/MV substations. The minimum loss radial network proposed by Shirmohammadi and Hong [6] is obtained by means of an iterative process, the initial stage of which consists in the closing of all switchgears and then solving the resulting meshed network. The method proposed by Civanlar et al. is adopted and modified by Huddleston, Broadwater, and Chandrasekaran in [7], in order to operate multiple exchanges rather than single exchanges. Glamocanin [8] formulates the problem of the optimum reconfiguration of distribution networks as a transhipment problem with quadratic costs. A new formulation of the reconfiguration problem for network loss reduction and load balancing at the substations is proposed by Chiang and Jumeau in [9]. A two-step methodology based on a simulated annealing technique is developed and presented. The solution algorithm and application of the procedure are reported in [10]. A modification of the Baran and Wu [5] method is proposed by Jasmon, Callistus, and Lee [11] in order to simplify and speed up the search for the minimum loss configuration. A mathematical approach is proposed by the authors in [12]; simulating the switchgears as fictitious resistances and starting from the meshed network obtained closing all switchgears, appropriate constraints lead the solution process toward the identification of a radial structure. The methodology proposed in [6] is applied and modified by Goswami and Basu in [13]; in the radial network, switchgears are closed one at a time and the optimal flow pattern of the network is assessed; the choice of the branch to be opened in order to make the network radial is made using the same criterion developed in [6]. Chen and Cho [14] describe an hourly optimal switching plan by means of a branch and bound technique and binary integer programming. A heuristic method based on the branch exchange approach is presented by Cherkaoui, Bart, and Germond in [15]; a number of research strategies are described which include the principles of a new general method of combinatorial optimization. In [16], a two-stage heuristic procedure based on a minimum flow rule is proposed. A natural algorithm is developed in [17] for the solution of the minimum power loss reconfiguration of a MV radial distribution network problem. An application of neural networks in the reconfiguration for minimum losses is reported in [18] and [19]. The first paper considers the minimum losses reconfiguration as a pattern recognition problem. The system is divided in areas having homogeneous loads and, for each of these areas, a neural network is trained in order to find the relevant load value on the basis of the measures made at some sections. Then, on the basis of the obtained results, another network identifies the radial topology satisfying the optimality condition. In [19], the described procedure is modified with the introduction of a neural network which, on the basis of the input data related to the real and reactive power injections at the nodes, finds out the status of network tie-switches.

The minimum losses reconfiguration solution strategy is here divided into two different parts, one part, dealing with the closing phase, is given in charge to neural networks, the other one, controlling the opening phase, is based on a deterministic algorithm, which locally checks the existence of a mesh in the current network configuration. The paper is organized as follows. After the introduction of the general problem of reconfiguration in automated distribution networks, the control strategy here proposed is deeply described as well as the different parts composing it, putting into evidence some additional operators that ensure the correct working condition of the entire control system. The neural network architecture used is then presented and some features related to the training phase are commented. Finally, test results of the behavior of the entire control system in different working conditions are presented and discussed. II. PROBLEM OF CONFIGURATIONS CONTROL DISTRIBUTION SYSTEMS
IN

In electrical distribution systems, power losses depend on the real and reactive power flows, which are related to the real and reactive loads. Minimum power losses reconfiguration by means of controllable tie-switches installed on the network branches is an intrinsically distributed problem, since both the loads and the tie-switches status are physically distributed. In the general formulation of the reconfiguration problem, the load values are the independent variables, whereas the tie-switches statuses are the optimization variables. The objective can be fulfilled performing a centralized control strategy in which the status of each tie-switch depends on the current loading condition of the network. In this way, the network can be optimally operated and it is not necessary to know the load in advance, and therefore, not introducing any restricting hypothesis about the nature of the supplied load. The hypotheses introduced in the system and loads modeling only influence the optimal solution research phase carried out in a centralized way. For the real implementation of a such centralized control system, the following elements are necessary: a measurement system for real and reactive loads; a transmission system for the load data connecting each node to the control center; a host computer able to identify the minimum losses configuration on the basis of the current loading condition; a transmission system for sending the input signals to the tie-switches located at the MV/LV substations. The centralized control cannot start if the above described components and systems are not properly installed and in correct working condition. The transmission systems for data and signals must be highly reliable, since their working condition directly affects the centralized control system. A distributed control system can instead be implemented at each tie-switch location on the basis of the information collected in the same place. The greatest advantage of this configuration is the initial requirement of low investment and the possibility afterwards to be integrated into a centralized control system. The local controller must be able to take the same decisions that could be taken in a centralized control system. This condition requires

764

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 3, JULY 2003

that each controller must be trained to recognize, for all of the values of the monitored quantities, the optimal status of the relevant tie-switch. If the problem is formulated as above outlined, a neural network should be able, if well trained, to control the status of the relevant switch with a given precision. However, the systems working condition should meet some strict constraints about radial topology for correct operation of the protection system and continuity of load supply in normal working conditions. Local control must also ensure that even due to possible errors in any taken decision about the tie-switches status, the network keeps the radial topology and all of the loads remain energized; any misoperation should not lead to the creation of meshes or islands in the system (an island is a set of nodes and branches not connected to any supply point).
Fig. 1. Simplified scheme of MV bus of a MV/LV substation. All of the afferent branches (1; 2; . . . ; n) can be sectionalized by means of the tie-switches SWT.

III. PROPOSED LOCAL CONTROL STRATEGY The control strategy below outlined is referred to a network in which the tie-switches are installed at the MV/LV substations, Fig. 1; in this way, each line can be sectionalized at both ends, where the substations are located. As it can be noted, network reconfiguration can be easily performed by means of the opening/closing operation of only one of the switches for each branch. Nevertheless, in this formulation, both ends of each feeder are sectionalizable for the following reasons. i) This network setting is quite common. ii) The opening of the switch at one end of the branch allows the voltage measurement at the other end (in the control strategy here set up, a voltage signal different from zero activates the closing control and this closing operation is effective only if the branch is open only at one end). iii) The operations of the two tie-switches in each branch can be easily and reliably performed. The operating conditions of the tie-switches shown in Fig. 1 are the following: at least one closed tie-switch and all the others in open position; the load at the node is attributed to the end of the branch whose tie-switch is closed; the power flowing toward the node and through the latter branch equals the power required by the load; all of the tie-switches are in closed position; in only one of the branches the power flow direction enters the node, in the others it comes out of it, due to the radial topology of the network; some of the tie-switches are closed and some others in open position; the conditions that are related to the power flows of the branches are the same of the case reported before. The operating condition with all of the tie-switches in open position is not allowed in normal conditions, since it corresponds to the disconnection of the load at the node. In order to prevent the formation of islands, the proposed control strategy entrusts the closing phase to a neural network trained on purpose, and the opening phase to a deterministic algorithm based on the search of a current inversion point. Indeed, radial networks do not have current inversion points; only the closing of one tie-switch can cause the formation of a mesh, and therefore,

of a current inversion point. Opening the tie-switch in the mesh, that is closest to the current inversion point, makes the network topology radial with reduced losses. In this way, it is guaranteed that the opening maneuver is performed after the closing one and, then, that the sequence of networks operations, in general, is the following: 1) closing of one of the tie-switches in the network; 2) creation of a mesh; 3) opening of one tie-switch in the mesh; 4) radiality restoration. It can be noted that the network reconfiguration can be started on the basis of the neural networks response, that may be wrong, but always attains deterministically a radial topology, so as to restore the correct network working condition (all of the loads are energized and the network topology is radial). Since the switches working states are binary, and the transition from one state to the other is subjected to two logical mechanisms, one excluding the other, it is necessary to design a higher local control level that, on the basis of the tie-switches status, can coordinate their activation. In the proposed strategy, the operation of the tie-switches status is then operated according to two hierarchically organized levels, see Fig. 2, as a node supervisor, NS, which selects the type of control to be activated on the basis of the status of each tie-switch; it also controls some additional functions for the correction of the errors; a switch controller, SC, of hybrid type acting on the basis of two different algorithms and whose activation is controlled by NS; one of these algorithms is a neural network, the other is a deterministic algorithm based on the current inversion point. The control strategy outlined before, even if strictly meeting the network radiality and the islands prevention constraints, may give rise to working conditions that are different from the desired ones at minimum loss level, for two main reasons. differences between the actual load distribution at the nodes and the one taken as a reference for the neural network training;

AUGUGLIARO et al.: MINIMUM LOSSES RECONFIGURATION OF MV DISTRIBUTION NETWORKS THROUGH LOCAL CONTROL OF TIE-SWITCHES

765

Fig. 2. Hierarchical structure of the control system for the tie-switches control.

when multiple contemporary operations (more than one pair of switches changes its status) there is no possibility to coordinate, along time, the tie-switches manoeuvres; this characteristic, related to the intrinsic nature of the local control, may lead to wrong configurations. When multiple contemporary switching operations produce more than one mesh in the network, opening the tie-switches according to the current inversion point criterion does not guarantee that the resulting radial network is operated at minimum loss level. In the same way, if multiple switching operations are required to go from one radial configuration to another, the intermediate radial configurations may not be included into the training set and therefore the neural networks may give erroneous outputs. In the proposed strategy, the node supervisor implements some functions devoted to the correction of these errors. With reference to a generic load node connected to lines, all of them sectionalizable, Fig. 3 shows the set of electrical (voltage, current, and status quantities and signals of each tie-switch) sent to the local control system. In Fig. 4, the flow of the same quantities and signals within the control scheme of the tie-switches status is shown. A. Control Strategy for Tie-Switch Closing The control of one tie-switch, during the transition from the open status to the closed one (box CL in Fig. 4), has been entrusted to a neural network with the permissive signal of the node supervisor. The neural networks setting up first requires their architectural design, then the electrical quantities to be measured at each load node (inputs of the neural networks) must be adequately chosen as well as the training strategy. 1) Neural Networks Architecture: An artificial neural network (ANN) is a parallel distributed architecture processor having a specific attitude for the storage of empirical knowledge, making it usable when needed. A neural network is a set of elementary units, neurons, having local calculation ability, connected by means of communication channels (connections). Knowledge can be acquired through a learning procedure and the elements connecting two neurons (known as synaptic weights) are used to store empirical knowledge [20][22]. In the power systems field, such a processor can be used to solve different problems, often quite complex, and for this

Fig. 3.

Electrical features and signals sent to the local control system.

Fig. 4. Logical scheme of the proposed control strategy.

reason it has recently found many applications. As it was pointed out before, literature reports the use of neural networks for minimum loss distribution networks reconfiguration [18], [19]. However, these applications refer to centralized control strategies using electrical measures executed all over the network and collected at one site. In the proposed strategy, each neural network must control the closing phase of the switching operation, for all of the tie-switches that are currently opened; the decision is then taken on the basis of local measures, directly executed on the sectionalizing node with no remote control. For this application, since the need for the definition

766

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 3, JULY 2003

network behavior must be executed in order to define the configuration of all the local controllers. In particular, this analysis allows the identification of a set of meaningful cases of possible working conditions, in open position, in which each tie-switch can be during the normal operating conditions. Moreover, for each of these conditions, it is necessary to identify the status of all the switches. For this reason, it is necessary to 1) implement a network model able to accurately simulate its working condition as the switches configuration varies; 2) fix the possible load levels (real and reactive) for each load; 3) assess, for each loading condition, the voltages and currents distribution for a nonoptimal network setting; 4) identify, for each of the above loading conditions, the relevant minimum loss configuration (by means of an algorithm used for the centralized control); 5) associate the input state vector of each tie-switch, detected in 3, with the relevant optimal state identified in 4. This procedure allows the attainment of a matrix DM ( rows and columns) for each controller, being the number plus one output (the desired of components of the vector state for the tie-switch) and the number of executed simulations. The content of the DMs is the knowledge base that must be transferred to the controllers during the learning phase. The weights of the connections can be identified using these data and the optimal number of neurons of the hidden layer. For each of the considered loading conditions, the minimum loss configurations have been also assessed. The supervised learning has been done by means of the conjugate gradient method. This method, as compared to the back propagation algorithm, shows a much faster convergence to the absolute minimum of the error function. During the training phase, the automatic update of the number of neurons in the hidden layer is possible, when after a given number of cycles, improvements of the error index become negligible. Moreover, the maximum number of these neurons has been fixed so as to prevent the so-called overfitting, namely the situation in which the network learns too much loosing the ability to generalize. In addition to the initial training phase, the knowledge base of each neural controller is automatically enriched during the networks operation, by means of the acquisition of particular working conditions. In this aim, as it will be clarified in what follows, are quite meaningful the working conditions in which the overall control logic can identify and correct the errors caused by the purely neural control. In the following Section IV, for a given network assumed as test case, the autonomous property of each ANN to close the relevant tie-switch when required, for loading conditions that are different from those included into the training set, has been verified. B. Control Strategy for Tie-Switch Opening The state transition of any tie-switch from the closed position (ON status) to the open position (OFF status) is controlled by a current comparator device (box CCD in Fig. 4),

Fig. 5. MLP network scheme with three layers adopted for the generic tie-switch controller.

and setting up agents having easy and cheap implementation is considered to be primary, the architecture chosen for the neural networks is a multilayer perceptron (MLP) with three hidden layers Fig. 5. In particular, the network has input neurons, being n the a first layer made up of number of the network branches afferent to the controlled MV/LV substation. The following column vector has been assumed as input of the generic tie-switch SW controller: (1) constituted of the RMS value of the voltages detected and , above and below the controlled tie-switch, and of the currents in each of the branches connected to the node, except that flowing in the branch where the tie-switch is located; an intermediate layer with a number of neurons depending on the controlled tie-switch; an output layer giving out the output of the network in the range [0, 1], obtained by the following function: if if (2)

If indicates the net input (weighed summation of the inputs, is the transfer function including the bias) of each neuron, of sigmoid type. Its expression is (3) having continuous values between 0 and 1, derivability and . asymptotic saturation for 2) ANN Training: According to the proposed control strategy, first an offline analytical software simulation of the

AUGUGLIARO et al.: MINIMUM LOSSES RECONFIGURATION OF MV DISTRIBUTION NETWORKS THROUGH LOCAL CONTROL OF TIE-SWITCHES

767

which detects the currents in the branches connected to the node , and compares their phase angles for possible inversion point identification. The node supervisor controls the CCD, which allows the phase angles check for those switches through which nonzero current flows. The CCD identifies which couple of switches has both current flows entering the node, and then it gives the opening command for the switch of the couple through which the smallest current flows. The presence of an inversion point produces a positive re, Fig. 4), otherwise, sponse of the CCD ( the tie-switch stays in closed position. The adopted control logic is based on the principle that, for a line supplied at both ends and constituted of branches connecting load nodes, the minimum loss radial setting is the one obtained sectionalizing the branch with the smallest current; this branch is always adjacent to the currents inversion point. One of the greatest advantages of such control logic, during the opening phase, is that to inhibit the formation of islands. When a radial configuration is reached; indeed, there is no possibility to have new current inversion points, and therefore, possible openings. This logic ruling the opening phase may lead to nonminimum power loss network configurations when the closing operations have produced more than one mesh in the network. In this case, sectionalizing many lines with the criterion of the minimum current entering the node, where currents inversion occurs, does not give any guarantee about reaching optimality. Moreover, depending on the particular chronological sequence of the openings in each mesh, final configurations that are nonoptimal can be reached. With reference to this aspect, many simulations with two or more meshes have been performed, also changing the order of execution of the opening operations. As expected, the most critical situations are those for which the difference between the modules of the currents entering the inversion node is small. On the other hand, if the opening operation for these nodes is executed at last, the situation can still be handled, since the minimum current switch between the two that are adjacent to the inversion node can change. In this case, indeed, the error in terms of power loss reduction is quite limited and can be further contained by means of some additional control functions performed by the local node supervisor. C. Node Supervisor Functions The primary tasks of the node supervisor are selection of the type of control to keep activated, deof each tie-switch of the node pending on the status (box SEL in Fig. 4); in particular, it activates the phase checking control for the closed switches and the neural control for the opened ones. When the tie-switch is open, only on one end of a branch, the voltage measure at the other end of the same branch is available at the tie-switch; in this way, this value can be used to decide whether closing it, or not, later (in other terms, the presence of nonzero voltage below the open tie-switch makes effective a possible closing operation). the supervisor indeed activates the possibility for opening only if the switch is closed and nonzero current flows

through it. In this case, for the closed switch belonging to the branch that is open at the other end, the opening stays inhibited, since zero current flows through it. preventing the node from working with all the switches in open position; this is taken into account using the voltage measurement at the load; if the voltage at the load, , reaches zero, the supervisor commands the sudden closing (box FC in Fig. 4) of all the switches and the simultaneous phase check by the CCD. Other than the above mentioned primary functions, the node supervisor must also operate some correction manoeuvres in order to limit possible control errors by means of the execution of some random closing operations on the open switches (box RG in Fig. 4); the probability of the execution of this operator is as much higher as more robust the control has proved to be during the test phase; in this way, the reliability of the neural networks can be verified and, if needed, the network configuration can be brought closer to optimality; II) the control of the opening times of the tie-switches (box OT in Fig. 4), so that the most critical ones, due to small differences between the currents modules, are operated with delay. With reference to the first action, it is important to point out that, since each neural processor must use the knowledge, acquired during the training phase, within new scenarios, in some cases, it can give a wrong output causing errors in the network configuration. These errors are of two different types: I) a) undesired closing of a switch; b) staying into the opened state, instead of changing to the closed state. Since the adopted control logic, the errors of a) type do not have a permanent effect on the network configuration. The undesired closing of a tie-switch induces indeed the activation of the opening controller; the latter, if recognizes the required condition for opening in that point, will suddenly command the opening. In these cases, there is a conflict between the two algorithms; therefore, if the supervisor does not detect any outage in both controllers, prefers the CCDs deterministic output to the neural network one. The b) type of mistakes instead is more serious; their automatic detection and correction are indeed not possible. In order to limit the effects of such errors, the node supervisor periodically activates the random closings of the open switches. The number of these maneuvers and the times at which they are executed depend on how robust the control system has proved to be during the long run test phase. In this way, it is possible to assess the number of wrong operations, and therefore, associate an error probability to each switch. This additional function can handle neural networks errors, improving the quality of the reached solution. The random closing function is devoted to the correction of possible opening errors addressing the evolution of the opening operations toward optimality. Every time the described measures allow the identification and correction of an error made by the neural controller, the node supervisor uses that particular working condition and the corrected information on the status of the tie-switch to enrich

768

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 3, JULY 2003

the knowledge base of the neural network with the acquisition of such particularly meaningful cases in order to refine the training phase. This technique allows an additional reduction of the errors due to the neural control along time. Measure II (control of the opening times) is performed to address the correct sequence of openings, taking into account the effects of other possible openings in the network. It can be observed that, during the control of the opening operations, first the currents inversion node is identified, then, among the branches connected to the node in which the current flows enter, the minimal current is found and, finally, the relevant switch is opened. When more than one mesh is present, in the proposed strategy the opening times inversely depend on the absolute value of the difference of the current modules flowing in the pairs of branches connected to the inversion nodes. In this way, first the most robust operations are executed taking time then to modify the other openings on the basis of the new flows distribution caused by the executed opening operations. Once the two switches adjacent to an inversion point and the one to be opened are identified, the supervisor delays the opening in a way that is inversely proportional to the absolute value of the currents modules difference. At the starting time of the in, the relations between the modules terval , that is, for and the phases of the two currents are valued and stored (box in Fig. 4); at the end of this delay time, the relations between the modules and the phases of the two currents are again in Fig. 4); if these relations are unchanged, the checked (box tie-switch is opened, otherwise, the phase check is reinitialized. This operation does not guarantee optimality, but reduces the error possibility. As the difference between the currents modules gets lower, the possibility of the variation of the minimum current branch due to other switching operations in the network increases; correct timing of the opening maneuver, and its possible interdiction, allows the most sensitive switches, in terms of currents, to be opened in the end. In a future scenario of complete development and implementation of distribution automation, each node supervisor can be connected with a control center, constituting the highest hierarchical level in the control system of the network. In this operating condition, the control of the configuration can still be at a local level, giving the central controller the following tasks: i) verifying the correct functioning of all the remote units; ii) starting new training phases for the neural controllers; iii) performing reconfiguration during service restoration.

Fig. 6.

Test system.

IV. STRATEGY VALIDATION The test system, Fig. 6, also used [18], [19] for centralized minimum loss reconfiguration using neural networks, is a megavolt system with 14 load nodes and 17 branches. The network is supplied by three HV/MV substations and each branch has been considered with two tie-switches at the ends. For the present application, a small size network has been chosen in order to make faster the calculations for the neural networks training. It is indeed clear that the local control logic is not influenced by the size of the network where the controller is

installed; the size of the network influences the ANNs training time and the way to obtain the optimal solutions necessary to build the training set. In any case, the ANNs answer is of binary type, since it has to give a response about the status (ON/OFF) of the relevant switch. During the training phase, also the optimal configuration calculated on the basis of a centralized logic is given as input to the ANN; the identification of such configuration is attained by comparing one another different solutions on the basis of the losses associated to each of them. In this case, the order of the considered solutions does not depend on the modeling provided that all of the evaluations are executed using the same models. In paper [23], the influence of the use of different load models on distribution systems reconfiguration has been investigated; since this influence can be neglected, the one allowing the fastest calculations has been chosen. As far as the lines are concerned, the effects of shunt capacitors have been neglected, even if their influence can be easily included in the model by two lumped admittances at the ends of each branch. The electrical data are reported in Table I, where and are the real and reactive power required by the substations located at the ending bus of each branch, EB, in normal operating conditions and at rated load levels. The network model, used for the execution of the calculations required for the training set and the test set, is linear with lumped parameters. In the performed simulations, the loads have been grouped into three sets conformally varying: nodes 2, 5, 12, and 13; nodes 7, 6, 3, 10, and 8; nodes 4, 9, 11, and 14; for each set of loads, five different load factors have been considered variable between 0.1 and 0.9, assuming a constant power factor. The working condition of the network has been simulated with reference to random sequences of loading conditions among 125 possible conditions (125 is 5 to the power of 3). From the obtained results, different training sets for the neural networks have been built. Two general conditions are to be considered for training sets creation. 1) equilibrate representation of each class within the training set; generally, the training sets are constituted of different subsets, each representing a particular class; 2) adequate representation, within each class, of the statistical variations in presence of noise.

AUGUGLIARO et al.: MINIMUM LOSSES RECONFIGURATION OF MV DISTRIBUTION NETWORKS THROUGH LOCAL CONTROL OF TIE-SWITCHES

769

TABLE I ELECTRICAL DATA OF THE TEST SYSTEM IN FIG. 6

With reference to the first criterion, the different classes have been obtained with various starting configurations and with different load levels, so as to obtain an equilibrate data set with respect to the different operating conditions. Concerning the second criterion, since the considered problem is of discrete nature, any verification in presence of noise has been considered useless, the noise being in this case errors in the used model and in measures. During the training phase, the behavior of each neural net. work has been considered good for Square During the following test phase, with a new sequence of loading conditions, generated simulating random variations of the loads at all the nodes of the network (in the interval % as compared to the training levels), the output of the neural controllers has been verified as compared to the relevant optimal configurations. The resulting setting in each condition has been assumed as input for the following evaluation, for which a new loading condition was set. In Table II, Node ID is the node identifier, Total err is the total number of errors executed by the neural network out of the 125 possible cases, Serious err is the number of errors of b) type already classified as serious (see Section III.C); finally Switching oper is the number of operations required by the neural controller out a total number of 125 conditions. As the obtained results show, the number of errors made by the neural controllers is quite low, also considering that the performed test is quite hard. The existence of errors is due to the introduction of a strong random variation of the loads, with perturbations that reach the 30% of the rated values, randomly and independently generated at each load node. In the worst case among the few ones in which one of the neural controllers fails, the increase in power loss, as compared to the optimal configuration, is about 2%. In order to clarify the procedure followed in the validation phase of the control strategy, in Fig. 7, the starting radial layout, the intermediate meshed layouts created by the ANN intervention, and the final radial layout created by the CCD intervention are shown.

TABLE II TEST RESULTS OF THE NEURAL CONTROLLERS

In Fig. 7(a), the radial configuration correspondent to a given action of the local controllers for a fixed loading condition is represented. In this working condition, for the nodes 2, 4, and 12, having the relevant switches in open position (on the branches 25, 49, and 1213), the closing control due to the ANNs is active. The new loading condition requires a new minimum losses configuration having the switches 38, 67, and 1314 in open position. The new loading condition changes the local controller ANNs input variables; the ANNs outputs are the closing of the tie-switches located on the branches 25, 49, and 1213. In order to verify the behavior of the control system, when more than one mesh is present, the contemporaneous closure of the three switches has been hypothesized giving out the meshed network in Fig. 7(b). The currents flows, for the network with three meshes, indicate the presence of three inversion points at the nodes 3, 6, and 13. The current flows toward node 3 in the branches 13 (800 A) and 8-3 (36 A); the current flows toward node 6 in the branches 3-6 (591 A) and 7-6 (274 A); the current flows toward node 13 in the branches 12-13 (105 A) and 14-13 (31 A). The difference between the modules of the two currents entering the node is: 764 A for node 3; 317 A for node 6; and 74 A for node 13. Since the opening time is inversely proportional to this difference, the switch at node 3 opens first on the

770

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 3, JULY 2003

Fig. 7. Intermediate layouts during a change of configuration of the test system.

branch 3-8 since the current entering it is smaller. The network has therefore the configuration with two meshes as Fig. 7(c) shows. The currents flows in such network indicate the presence of two inversion points at nodes 6 and 13. The current flows toward node 6 in the branches 3-6 (582 A) and 7-6 (282 A); the current flows toward node 13 in the branches 12-13 (99 A) and 14-13 (33 A). The difference between the modules of the currents entering the node is: 300 A for node 6 and 66 A for node 13. Since the opening time is inversely proportional to this difference, first it opens the tie-switch at node 6 on the branch 6-7 since the current entering it is smaller. The network therefore has only one mesh as Fig. 7(d) shows. The currents flows for the network with one mesh indicate the presence of one inversion point at node 14. The current flows toward node 14 in the branches 13-14 (50 A) and 11-14 (155 A); the difference between the modules of the two currents entering the node is then: 105 A. With an opening time, which is inversely proportional to this difference, the controller opens the tie-switch at node 14 on the branch 14-13 since the current entering it is smaller. The network has the radial configuration of Fig. 7(e); this is the optimal configuration previously identified using a centralized approach and an exhaustive search. V. CONCLUSION Distribution networks reconfiguration can produce good savings in variable costs through the power loss reduction. The main obstacle in a large application of such DA function resides in the installation costs, when a centralized controller is required. The distributed control as here suggested allows network reconfiguration without any physical connection between sub-

stations; moreover, this system does not prevent the system from evolving into a centralized control system. The local control strategy here proposed guarantees that the systems behavior is quite close to that attainable using a centralized control strategy; the executed simulations show that the wrong configurations are not so frequent and that the damage they bring is negligible in power losses terms. The control logic here developed also prevents the formation of islands and meshes and contains some additional function able to correct possible errors due to loads variations and to the intrinsic nature of the distributed control. The simulation on a small size network has proved the feasibility of the proposed local control, the obtained results are quite good and they encourage the implementation of the strategy on a large size network and without any of the simplifying hypotheses here introduced. In particular, items of greater interest are: i) to realize the control in a network where other control and supervision systems not dedicated to reconfiguration functions are installed; ii) studying the influence of loads unbalances on the measures to be used for local control; iii) finding the most feasible type of ANN for systems with high loads variability; iv) evaluating costs and economical benefits of such a solution. REFERENCES
[1] A. Augugliaro, L. Dusonchet, and E. R. Sanseverino, Genetic, simulated annealing and tabu search algorithms: Three heuristic methods for optimal distribution networks reconfiguration and compensation, Europe. Trans. Elect. Power Eng., vol. 9, no. 1, pp. 3541, Jan./Feb. 1999. [2] K. Aoki, T. Ichimori, and M. Kanezashi, Normal state optimal load allocation in distribution systems, IEEE Trans. Power Delivery, vol. PWRD-2, pp. 147155, Jan. 1987. [3] S. Civanlar, J. J. Grainger, H. Yin, and S. S. H. Lee, Distribution feeder reconfiguration for loss reduction, IEEE Trans. Power Delivery, vol. 3, pp. 12171223, July 1988.

AUGUGLIARO et al.: MINIMUM LOSSES RECONFIGURATION OF MV DISTRIBUTION NETWORKS THROUGH LOCAL CONTROL OF TIE-SWITCHES

771

[4] C. C. Liu, S. J. Lee, and K. Vu, Loss minimization of distribution feeders: Optimality and algorithms, IEEE Trans. Power Delivery, vol. 4, pp. 12811289, Apr. 1989. [5] M. E. Baran and F. F. Wu, Network reconfiguration in distribution systems for loss reduction and load balancing, IEEE Trans. Power Delivery, vol. 4, pp. 14011407, Apr. 1989. [6] D. Shirmohammadi and H. W. Hong, Reconfiguration of electric distribution networks for resistive line losses reduction, IEEE Trans. Power Delivery, vol. 4, pp. 14921498, Apr. 1989. [7] C. T. Huddleston, R. P. Broadwater, and A. Chandrasekaran, Reconfiguration algorithm for minimizing losses in radial electric distribution systems, Elect. Power Syst. Res., vol. 18, pp. 5766, 1990. [8] V. Glamocanin, Optimal loss reduction of distribution networks, IEEE Trans. Power Syst., vol. 5, pp. 774782, Aug. 1990. [9] H. D. Chiang and R. J. Jumeau, Optimal network reconfiguration in distribution systems: Part 1: A new formulation and a solution methodology, IEEE Trans. Power Delivery, vol. 5, pp. 19021909, Nov. 1990. , Optimal network reconfiguration in distribution systems: Part [10] 2: Solution algorithms and numerical results, IEEE Trans. Power Delivery, vol. 5, pp. 15681574, July 1990. [11] G. B. Jasmon, L. H. Callistus, and C. Lee, A modified technique for minimization of distribution system losses, Elect. Power Syst. Res., vol. 20, pp. 8188, 1991. [12] A. Augugliaro, L. Dusonchet, and S. Mangione, Optimal reconfiguration of distribution network for loss reduction using nonlinear programming, Europe. Trans. Elect. Power Eng., vol. 1, no. 6, pp. 317324, 1991. [13] S. K. Goswami and S. K. Basu, A new algorithm for the reconfiguration of distribution feeders for loss minimization, IEEE Trans. Power Delivery, vol. 7, pp. 14841491, July 1992. [14] S. C. Chen and M. Y. Cho, Energy loss reduction by critical switches, IEEE Trans. Power Delivery, vol. 8, pp. 12461253, July 1993. [15] R. Cherkaoui, A. Bart, and A. J. Germond, Optimal configuration of electrical distribution networks using heuristic methods, in Proc. 11th Power Syst. Comput. Conf., Avignon, France, 1993, pp. 147154. [16] A. Augugliaro, L. Dusonchet, and S. Mangione, An efficient greedy approach for minimum loss reconfiguration of distribution networks, Elect. Power Syst. Res., vol. 35, no. 3, pp. 167176, 1995. , MV radial network optimal reconfiguration: A new approach [17] based on genetic algorithm, in Proc. Arabian Union Producers, Transporters, Dist. Elect. - Int. Conf. Elect. Dist., Amman, Jordan, June 36, 1996. [18] H. Kim, Y. Ko, and K. Jung, Artificial neural-network based feeder reconfiguration for loss reduction in distribution systems, IEEE Trans. Power Delivery, vol. 8, pp. 13561366, July 1993. [19] M. A. Kashem, G. B. Jasmon, A. Mohamed, and M. Moghavvemi, Artificial neural network approach to network reconfiguration for loss minimization in distribution networks, Elect. Power Energy Syst., vol. 20, no. 4, pp. 247258, 1998. [20] T. Masters, Practical Neural Network Recipies in C++. New York: Academic Press, 1993. [21] S. Haykin, Neural Networks: A Comprehensive Foundation, New York: Macmillan, 1994.

[22] T. Masters, Advanced Algorithms for Neural Networks A C++ Sourcebook. New York: Wiley, 1995. [23] A. Augugliaro, L. Dusonchet, and S. Mangione, Load influence on minimum loss reconfiguration of automated distribution networks, in Proc. Stockholm Power Tech., Int. Symp. Elect. Power Eng., Stockholm, Sweden, June 1822, 1995.

Antonino Augugliaro was born in 1949. He received the Ph.D. degree in electrical engineering from the University of Palermo, Italy, in 1975. Currently, he is Full Professor of Electrical Power Generation Plants at the Faculty of Engineering at the University of Palermo, Italy. He was Associate Professor at the University of Palermo from 1978 to 1994. His main research interests include the simulation of electrical power system, transmission over long distances, mixed three-phase/six-phase power system analysis, optimization methods in electrical distribution systems design and operation, and distribution automation.

Luigi Dusonchet was born in 1948. He received the Ph.D. degree in electrical engineering from the University of Palermo, Italy, in 1975. Currently, he is Full Professor of Industrial Electrical Systems with the Faculty of Engineering at the University of Palermo, Italy. He was Associate Professor at the University of Palermo from 1978 to 1990 . His main research interests include simulation of electrical power system, transmission over long distances, mixed three-phase/six-phase power system analysis, optimization methods in electrical distribution systems design and operation, and distribution automation.

Mariano Giuseppe Ippolito was born in 1965. He received the Doctor degree in electrical engineering in 1990 from the University of Palermo and the Ph.D. degree from the University of Palermo, Italy, in 1994. Currently, he is Associate Professor of Power Systems with the Faculty of Engineering at the University of Palermo, Italy. He was a Researcher with the Department of Electrical Engineering at the University of Palermo from 1995 to 2001. He has also made some contributions to the advancement of power quality study. His main research interests include the areas of power systems analysis, optimal planning, design, and control of electrical distribution system.

Eleonora Riva Sanseverino was born in 1971. She received the Doctors degree in electrical engineering and the Ph.D. degree from the University of Palermo, Italy, in 1995 and 2000, respectively. Currently, she is an Associate Professor of Power Systems at the University of Palermo, Italy. During 2002, she was a Full Time Researcher of the National Council of Research. Her main research interest is in the field of optimization methods for electrical distribution systems design, operation, and planning.

Вам также может понравиться