VOLTAGE SECURITY ASSESSMENT USING ARTIFICIAL INTELLIGENCE TECHNIQUES

Abstract

Modern power systems are operating under much stressed conditions, i.e. closure to their security limits, because of economical and environmental pressure. The contingencies like line or generator outage, increase in load demand, change in system conditions or may be a combination of them, cause a progressive and uncontrolled decline of bus voltages leading to voltage instability or voltage collapse. Several events of voltage collapse have been occurred in the pastdecade; hence the study of voltage security has become an important area of research. The present work attempts to address the static voltage security assessment problem usingnovel methodologies. Voltage security function forms an integral part of the modern Energy Management System. The traditional methods of contingency selection based on approximate or full AC load flow are either inaccurate or time consuming. To overcome these difficulties, there is a pressing need to develop fast, accurate and transparent security assessment tools so that in real-time applications, necessary corrective actions can be initiated to avoid potentially dangerous situations of voltage instability within the given timeframe of interest. Machine learning isa broad area of artificial intelligence, which is concerned with design and development of algorithms and techniques that allow computers to learn. In a very broad sense machine learning isbased ontwo types of learners, firstly eager learners e.g. some of the artificial neural networks (ANN) and decision trees (DT), secondly lazy learners e.g. case-based reasoning (CBR). Inspite of good generalization ability, the ANNs have some shortcomings like, their opacity, slow training and inability to adapt a concept drift. On the other hand decision trees used for classification are fully transparent and instantaneous in training, hence they can be trained frequently to adapt the modifications in topology of the power system. In the last decade the classification type DT has been used for two-class and four-class classifications of power system operating states. For multi-class classification DTs are less accurate and their sizes become very large as compared to DT for two-class classification problem. The DTs are assumed to respond in terms of pre-defined classes in output, which Abstract are defined by the range of the output parameter values. This limitation of DT can be overcome by using model tree (MT), which gives continuous value in the output. In this work, the hybrid decision tree (HDT) model, which is a combination of filter and ranking modules, has been developed for fast voltage contingency screening and ranking. With HDT, in order to reduce the size and improve the accuracy of DT, the Kclass problem is converted into a set of K two-class problems, and separate DT modules are trained for each of the two-class problems. The critical contingencies screened out by filter module are presented to the ranking modules (DTs connected in parallel) for their further ranking. Real and reactive loads at each load bus and real load at PV buses along with a contingency number are selected as input features for the HDT. To measure the severity of contingencies, bus voltage violation based scalar performance index (VPI) is used. Full AC load flow is performed to generate the training and testing patterns for proposed HDT under each contingency. The continuous values of VPI are classified into five classes according to their severity and HDT is trained for this multi-class classification problem. Due to its self-feature selection ability, DT selects only relevant features during training. Once trained, a HDT method provides fast and accurate screening and ranking of contingencies for unknown load patterns. In this research work MT based approach has been developed for fast voltage contingency ranking of most severe contingencies. MT is developed to estimate the real value of VPI and rank the single line outage contingencies according to their severity. MTs under each of the contingencies are configured in parallel. During training of MTs only relevant features are selected, which are further used for linear regression models in the leaves of MT. Once trained, MT method gives fast and accurate estimation of VPI and hence ranking of contingencies for unknown load patterns. The proposed algorithms have been successfully applied to IEEE 30-bus and IEEE-118 bus systems. Fast estimation of post-contingent maximum loadability margin (MLM) under different contingencies is essential for evaluating on-line voltage security of a power system. To maintain security against voltage collapse, it is necessary to estimate the effect of contingencies on the voltage stability margin so that corrective measures can be taken to increase the margin to avoid blackout. 11 Abstract In this work, post-contingent MLM for voltage collapse has been used to determine the severity of a contingency and a MT method is employed to compute the postcontingent MLM. Database for training and testing of MT is generated by varying the real and reactive loads at each bus randomly and MLM is computed by continuation power flow under each line outage case. The MT method itself selects those features, which affects the MLM most. These selected features are used for linear regression models coupled to the leaves of DT. Once trained, the MT method is able to estimate the postcontingent MLM for unknown patterns instantaneously and with desired accuracy. Along with estimation of real values of MLM, MT also classifies the MLM into five classes (class-I to class-V) which define the different severity levels. Class-I belongs to insecure state defined by pre-defined critical value of MLM, below which the corresponding operating point can be assumed as an insecure operating point. Based on percentage of insecure operating points for a fixed set of testing patterns, each of the contingencies are ranked. The contingency having highest percentage of insecure states will be on top of the list. MTs are fully transparent and can easily be understood by operators. The MT method has been successfully applied to IEEE-30 bus and IEEE-118 bus systems. Database of a power system has valuable knowledge in it. In past decades, because of limitations in computer technology, engineers were unable to use these databases in real time security analysis and control. Using this knowledge to shorten the cycle of problem solving is challenging and has led to many different approaches: one of these approaches is the case-based reasoning (CBR). Using solved problems in the past for solving current problem is fundamental to the CBR. In this work, CBR, a novel approach, for real-time voltage security assessment of the power system is developed. Because of faster training, probabilistic neural network (PNN) is used for similarity matching and determining solution of new cases, using old cases from the case-base of the CBR system. To reduce the input dimensionality, an information gain based feature selection technique has been employed to determine relevant input features for PNN. By using relevant features, the accuracy enhances and execution time reduces as compared to the case when all the features are used as inputs to solve the new case.A casebase for CBR is generated by computer simulation and MLM of the power system is used as a voltage security index. Case-base is organized into different groups of cases; each in Abstract group of case-base belongs to a particular contingency, which makes retrieval process faster. After solving the new case the CBR system compares the new case with the most similar case from the case-base for its inclusion (learning) into case-base for solving the similar cases that may occur in future. Theaccuracy of the technique depends on number of solved cases in the case-base. As the number of solved cases is added to the case-base, accuracy enhances greatly because the probability of finding similar case to the new case is increased. The proposed Approach has been successfully applied to IEEE 30-bus and IEEE-118 bus systems. The proposed CBR system is very fast and accurate enough to determine potentially dangerous operating states of the power system. Analytical methods to compute minimum load shedding require large computational time and therefore not suitable for real-time applications especially during emergencies. Detection of dangerous operating states of the power system is not enough; hence it is essential that effective and practically justified solutions should also be developed to mitigate voltage collapse during emergencies in real-time. Load shedding is the most practical and effective solution to mitigate the voltage collapse. In this work a CBR system is developed for on-line voltage security assessment and optimal load shedding to mitigate the voltage collapse under different line outages. DTs are used for case indexing as well as for determination of optimal load shedding plan for voltage insecure cases. CBR system also ranks the line outages in the order of their severity. The line outage having highest percentage of insecure states will be kept on top of the list. The proposed approach has been successfully applied at IEEE 30-bus and IEEE- 118 bus systems. The proposed CBR system is very fast and accurate in assessing security and giving optimal load shedding plan for insecure cases and found to be most suitable for real-time application in modern energy management systems. In heavily stressed systems, most of voltage collapse incidents are caused by insufficient reactive power supply. Therefore cost of reactive power support to alleviate the voltage insecurity is very important for the transmission operator (TO) in today's competitive electricity market. In this work, reactive power generation cost in addition to real power generation cost is simultaneously minimized, taking into account voltage security margin and other operational constraints. Quadratic opportunity cost functions is derived for reactive power IV Abstract generation in terms of real power generation cost coefficient and day-ahead electricity market data. The optimization problem is a nonlinear mix-integer programming problem and hence hard to solve by conventional methods. To overcome this, an algorithm based on particle swarm optimization (PSO) and sequential quadratic programming (SQP) is developed to solve the formulated optimization problem, further it alleviates the use of integer variables in the optimization. The proposed algorithm has been successfully applied to IEEE 14-bus and IEEE-30 bus systems. The various contributions, which have been made in this thesis work, are summarized as follows: > The HDT and MT methods have been developed for fast and accurate voltage contingency screening and ranking. > MT model is developed for fast estimation of post-contingent MLM. Contingencies are ranked in the order of their severity. > For fast and accurate detection of potentially dangerous situations of power system operating points, a CBR method is developed which in turn ranks the severe contingencies according to their severity. > The CBR system is developed for voltage security assessment in addition to give optimal load shedding plan to mitigate the voltage collapse conditions. DTs are used for case indexing as well as for determination of optimal load shedding plan in the proposed CBR system. > An optimization problem is formulated to minimize the cost of reactive power generation considering voltage security margin. Quadratic opportunity cost function is derived for reactive power generation cost. To solve the optimization problem a PSO-SQP algorithm is developed.