Dynamic security enhancement of power systems via population based optimization methods integrated with artificial neural networks

Abstract

Increasing consumption of electricity, penetration of renewable energy sources into electric power systems, uncertainties caused by the nature of these sources, enforcements of economics and energy markets, bring various and unplanned power flow patterns, which represents different and more stressed conditions than planned operating conditions of power systems. These new power flow patterns and their increasing diversity should be redesigned, to satisfy an acceptable security level for power systems. When a power system is operated under stressed conditions that are close to security limits, the system can be vulnerable from disturbances (contingencies), which can cause loss of synchronism and cascading blackouts. Therefore, online monitoring of dynamic security of the system becomes an important task for detecting an insecure current operating condition and applying proper control actions to regain the ability of the system to withstand contingencies such as faults, or sudden loss of any system component. Electric power is extremely important for economy and daily life. Safe and sustainable operation of power systems requires fast and accurate dynamic security assessment methods. Accurate methods with mathematical background exist for determining angular instability in power systems. Time-domain-simulation is the most straightforward method that directly shows the behavior of the system dynamics in time domain. However, dynamic model of power system involves linear and nonlinear equations that include large number of continuous and discrete state variables. When time domain simulations are applied for the stability analysis, the calculations required for recursively solving nonlinear differential equations of the system model over thousands of time steps, require significant time for the occurrence of one critical contingency. This drawback reduces the practicability of using time domain simulations. Although, direct methods based on energy function assess the stability of the system without solving differential equations, they involve modeling limitations for large-scale power systems, which make them impractical. Probabilistic methods are considered as a suitable method for system planning due to their detailed considerations and computation requirements. Measurable features of power systems can be associated with the stability of the system through machine learning methods. This study suggests using data collected from different buses of the system and processing these measurements by designed fast and accurate artificial neural networks to assess the dynamic security of the current operating point of the system. Several, recent studies suggest using machine learning methods for dynamic security assessment. Nevertheless, it is still a challenging task for large sized power systems. Large sized systems involve large number of critical contingency, which increase the computational complexity. In addition, large set of various system topologies and wide range of loading conditions should be taken into account by the designed security assessment tool. After the detection of insecure current operating point, system operator should apply control actions to move the system to a secure operating point, where the system can withstand contingencies. Objective of security enhancement is preventing the system from undesired situations, and avoiding large blackouts. Generally, security enhancement is categorized into two sub-categories such as preventive control and corrective control. The objective of preventive control is successfully preventing the power system from losing its stability against uncertain disturbances (contingencies). In this thesis, applications of generation rescheduling and load curtailment preventive control actions for enhancing the dynamic security of the power system are studied. Generation rescheduling is a useful preventive control action to restore and enhance the system security by shifting the generation among controllable generators. System operator can monitor the generation and demand of electric power. When an insecure operating point is detected, the system operator can request from available loads to curtail their electric power demand. This load curtailment action can collaborate with generation rescheduling and a proper adjustment of these preventive control actions can change the power flow patterns within the network. Then, power system can move to a secure operating point so that the stability of the system is preserved despite of the occurrence of a critical contingency. If preventive control actions cannot move the system to a secure point for any critical contingency, or the cost of preventive control is considered as high, power system can be armed by corrective control actions, which are ready to be applied in case of occurrence of any contingency. In this thesis, load shedding method is studied as a corrective control action to regain the stability of the system if preventive control action cannot be applied and protect the system against critical contingencies. Although both load shedding and load curtailment methods decrease the electrical power consumption in the related load buses, they are different methods by means of their application schedules, magnitudes and costs. As different from load curtailment method, load shedding is a fast switching operation that applied after the occurrence of contingency, and the amount of load to be shed is relatively larger than curtailment, since the maximum amount of load to be curtailed is based on the amount of available loads in the selected load buses. Dynamic security assessment and enhancement is under the responsibility of system operator. The complexity of the large sized power systems does not let the system operators to determine a proper control action, immediately. In addition, deregulation of power systems, integrates many participants into the market, and control actions designed for security enhancement, are generally contrary to the market-based strategies of participants. In other words, enhancing the system security brings a cost. Although, system operator may find a proper control action to enhance the system security, generally, the operator may not determine a cost effective control action, immediately. To overcome that problem, dynamic security enhancement can be considered as a constrained optimization problem to minimize the costs of required control actions. Due to the size of the real power systems, mentioned optimization problem involves multi-dimensional search space with many continuous and discrete control variables, many local optimum solutions and large number of equality and inequality constraints. In addition, constraints related with dynamic security of the system are highly nonlinear and too complex to be defined mathematically. Therefore, conventional optimization methods may not converge to a satisfactory solution for dynamic security enhancement. This study suggests using evolutionary algorithms to solve mentioned optimization problem. In addition, previously suggested artificial neural networks based dynamic security assessment methodology is integrated into the optimization process for estimating the violations of security based system constraints for candidate solutions. This integration can enable optimization methods to find a proper and cost effective solution for control actions within an acceptable time. Since the mentioned optimization problem involves too many constraints due to the size of the power systems, optimization methods should use constraint-handling methods. This study suggests applying an adaptive penalization technique to infeasible candidate solutions during the optimization. Adaptive penalization reduces the parameter adjustment process of static penalty function, which requires considerable effort and significantly affects the result of optimization. The main objective of this study is to research the applications of heuristic, population based optimization methods and their collaboration with artificial neural networks to develop a fast and powerful methodology to assess and enhance dynamic security of power systems. Different artificial neural networks are designed for dynamic security assessment, which are considered as both regression and classification problems. In regression approach, designed multi-layer perceptron neural networks estimates defined security indexes such as critical clearing time and minimum oscillatory damping value for each operating point. In classification approach, designed probabilistic neural network classify operating points as secure or insecure. Proposed artificial neural networks are designed to capture various topological changes in the system according to N-1 criterion and different loading conditions. For monitoring of a power system where all system parameters continuously change, ability of capturing topological and loading level changes is mandatory. In addition, artificial neural networks with that ability can directly be used during the optimization to enhance the security of insecure initial operating point with any topology and loading condition. Designed artificial neural networks process the data collected from different buses of the system. To increase the accuracy of the artificial neural networks, feature selection process is applied to all measurements. In addition, the feature set is enriched with an additional index that involves relevant information about the stability of the system such as inertia constants of synchronous generators, which cannot be obtained from measurements. The proposed methodology for dynamic security enhancement involves both preventive and corrective control strategies with mixture of various control actions such as generation rescheduling, load curtailment and load shedding. For the designed control strategies, various optimization methods such as genetic algorithms, differential evolution, particle swarm optimization, artificial bee colony, big bang-big crunch and mean variance mapping optimization are used. This study proposes artificial neural networks based dynamic security assessment methodology to estimate the security based constraint violations of candidate solution during the optimization of control actions. This will speed up the optimization, reduce the required time to find a cost-effective control action for enhancing the security of the system to an acceptable security level and increase the online applicability of proposed methodology. During the latter iterations of optimization, search methods tend to converge to an operating point near security boundary. To prevent the misdetection of an insecure operating point, artificial neural networks based dynamic security assessment tools are trained by using the training data that is enriched around the security boundary. Proposed dynamic security assessment and enhancement methods are demonstrated on the 16-generator, 68-bus system, 17-generator, 163-bus Iowa system and on the IEEE 50-generator, 145-bus system. The results of the studies are represented in Sections 4,5 and 6.