Local Search Strategy And Multiobjective Approach In Evolutionary Computation For Reactive Power Optimization

In recent years, evolutionary computation (EC) for reactive power optimization (RPO) has drawn extensive attention due to its features of high flexibility, high robustness, and strong adaptability. Such reactive power optimization algorithm is a hot topic of current research because it does not need any gradient-based information, does not depend on the functional state of the objectives, and does not has any trouble to handle both continuous and discrete reactive power and voltage control equipments. As a class of stochastic optimization methods, however, EC still has some disadvantage when solving complex problems such as RPO, i.e., premature, slow convergence, low local search ability. Although many modification strategies have been proposed, there is still large margin for improvement, and many issues need further discussion, i.e., how to obtain overall upgrade of convergence when applying local search strategies (LSSs), how to group the control variables in coevolutionary computaion, how to balance between the objectives in multiobjective optimization, etc. To solve these problems, the strategies of improvement for EC in RPO is chosen to be the research topic, including four aspects, say, EC with multiple LSSs for RPO, multiobjective reactive power optimization and its two-stage optimization strategy, robust reactive power optimization model and optimization strategy, and grouping methods of control variables in coevolutionary computation for RPO. The outputs of the thesis are listed as follows.Memetic algorithm based on multiple LSSs for RPO is proposed. Based on the existing LSSs, correcting memes, directed memes and stochastic memes are proposed to form the meme pool for RPO, by which the multiple LSSs can be combined together. The algorithm is applied in IEEE 30 bus system and numerical simulations demonstrate that the algorithm combines all the advantages of LSSs and shows the best performance of convergence. In addition, simulation results with different percentage of population participating in LSSs process are compared and relationship between percentage of population and efficiency of the algorithm is analyzed.A contour-line analysis approach for objective-converting methods (OCMs) in RPO is proposed. With the contour line concept in geography area, contour analysis can obtain the search directions of various OCMs in the objective space, and uncover geometric meaning of their optimal solutions. According the the observations, the intrinsic link between OCMs are studied, and OCMs are compared and classified.Five state-of-the-art multiobjective evolutionary algorithms (MOEAs) are compared for RPO. Pareto front, the outer solutions, and C measure are used to analyze the performance of the five MOEAs in RPO, which are classified into five performance levels. Strength Pareto Evolutionary algorithm 2 (SPEA2) and Nondominated Sorting Genetic Algorithm-II (NSGAII) are demonstrated to be the best algorithms that can obtain Pareto front with higher quality.A two-stage optimization strategy is proposed for MOEAs to solve multiobjective RPO. The strategy divides the search process into two phases, global search and focused search. The first stage uses NSGAII to search for a rough Pareto set, from which the decision maker (DM) obtain the information of the approximate distribution range of the objective functions, their approximated relationships, and then draw the requirements and preferences of the optimal solution in a simple and visualized way. On this basis, the second phase set preference parameters, and then adopts weighted method, goal attainment method,ε-constraint method, to focus on the key areas and directions in the optimization. Simulation analysis showed that the new strategay can combine the advantages of NSGAII and OCMs and it is a flexible and practical approach with high optimization efficiency.Robust RPO based on Monte Carlo integration is proposed. It adopts a Monte Carlo integral form of the objective functions, calculating the approximate expectations of system loss and voltage deviation in the presense of load fluctuations and searching for robust solution that is immune with load perturbations. To reduce the error of Monte Carlo integration, a samplingetechnique is proposed for RPO.It expresses a sample solution as two parameters, the percentage of the system load increase and the direction of power growth, and then select the appropriate sample solution in accordance with load changes. Genetic algorithms and NSGAII is used to solve robust RPO on IEEE-118 test system and the results show that during the fluctuations of the system load, the robust solutions can better maintain their quality with higher performance.A genetic-algorithm-based approach is proposed for the grouping of control variables in the coevolutionary computation for RPO.In this approach, the control variable grouping problem is converted to a reduced network partition problem, and a mathematic model is formulated. The new model not only adopted the typical cluster validity index, and taken the uniformity of group size of control variables into account. On this basis, a genetic-algorithm-based partition method is introduced to solve the new model. This method employs a seed encoding scheme that can adaptively choose the partition number with high flexibility. The simulation result on IEEE 118-bus system demonstrates that the new partition approach can automatically determine the partition number, grouping the control variables fast in a reasonably way. It can further improve the parallel efficiency of coevolutionary computation for RPO.