Study On Parallel Reactive Power Optimization Based On New Evolutionary Algorithms And PC-Cluster

The development of automatic reactive power and voltage control in power systems can be roughly divided into four phases, namely the apparatus level local dispersed control, the power plant and substation level local coordinated control, the regional coordinated control and the global coordinated control. And the highest objective is the reactive power optimization (RPO) based global coordinated optimal control that addresses both the security and the economic issues. Its precondition is maturating with the improvement of the data accuracy of SCADA and the practicability of EMS, and the demand for it is becoming more and more urgent. Therefore, RPO is one of the hotspots in the power engineering research field. The contents of research mainly consist of two aspects: the detailed modelization considering more practical requirements and the fast as well as accurate solution. This thesis concentrates on studying the solution methods of RPO.Mathematically, RPO is a non-convex and multimodal complex optimization problem involving nonlinear objective function, nonlinear equality and inequality constraints and both continuous and discrete variables. It is quite difficult to be solved quickly and accurately. Currently, mainly two classes of methods are used to solve RPO, namely the gradient-based mathematical programming methods and the intelligent optimization methods. The new development of the former class is the interior point methods (IPM) and the typical representatives of the latter class are evolutionary algorithms (EAs). Each class of methods have its own advantages and disadvantages: the former is fast, but is theoretically easy to converge to local minima and has difficulties in handling discrete variables and infeasibility problems; the latter can find global solutions in high probability and is good at handling discrete variables and infeasibility problems, but is slow and suffers from the problem of premature convergence.In order to overcome the disadvantages of EA-based RPO, many researches have been conducted, from which three directions can be summarized: (1) constructing hybrid algorithms using the complementary features between EAs and other intelligent optimization methods or IPMs; (2) simplifying the model and reducing the problem size using RPO-related experiences and knowledge in the field of power system calculation and operation; (3) accelerating the computation using parallel computing technologies. The work of this thesis roughly follows these three directions. Chapter 2 starts from evolutionary programming (EP) that has good global search ability. The performances of four EP schemes on solving RPO problems are first compared. The effectiveness of the so called adaptive fast EP on solving RPO problems is then studied. Finally, according to the principles summarized from the comparison study, successful improvements on two of the four EP schemes are made. As a whole, the researches in this chapter show that EP is generally too slow for solving RPO problems.Chapter 3 applies differential evolution (DE) to solve RPO problems for the first time. The mechanism and parameter setting of DE are first analyzed. Performance of DE on RPO problems is then studied with comparison to other EAs and the particle swarm optimization algorithm. The study shows that generally, DE is an excellent new EA for solving RPO problems, it is worthy of further studies and applications. However, it is also found that DE requires relatively large population size to avoid premature convergence. When the target power system is large, this will make the computational time too long to be acceptable for online RPO.Chapter 4 manages to improve the speed of DE for solving RPO problems by using parallel computing technologies, and parallel computing is implemented on a PC-cluster. Case study shows that parallelization does significantly improve the speed of DE for solving RPO problems; it is possible to realize online RPO with clusters of moderate size. However, it is also found that the efficiency of parallelization saturates quickly with the increase of the cluster size. Therefore, it is necessary to improve the algorithm itself to reduce the required population size and hence to further accelerate the computation or enable the use of clusters of smaller sizes for economic consideration.Chapter 5 first analyzes the complementary feature of DE and EP. This feature is then utilized to design a hybrid algorithm named DEEP. DEEP maintains the main body of DE, while uses the EP-style random mutation to introduce new genetic information to mitigate the pressure of premature convergence. Case studies show that DEEP has three advantages. First, it can effectively overcome the disadvantage of DE that requires relatively large population size to avoid premature convergence, which can greatly save the computational time. When extended to master-slave parallel computing, the reproduction step of DEEP can also be carried out dispersedly by slave processes without significant deterioration of solution quality, which can further save computational time. Second, DEEP is a universal algorithm. Its performance is not parametrically sensitive. Its only parameter can just be set to a fixed value. Third, due to the adoption of a novel scheme of primary population plus auxiliary population, and fitness evaluation is not arranged for the auxiliary populations, the additional time consumption of DEEP is negligible. So it is very suitable to use DEEP to solve optimization problems like RPO that consume most of the computational time on fitness evaluation.Chapter 6 utilizes the architecture provided by cooperative co-evolution to introduce the decomposition and coordination technique into DE. The local property of the relationship between reactive power and voltage is also utilized to decompose a power system into several sub-systems that are as independent as possible to reduce the work of coordination. Based on these techniques, a method combining cooperative co-evolutionary DE and power system decomposition (CCDE-PSD) is proposed for solving RPO problems. According to the characteristic of the CCDE-PSD, a three-leveled master slave parallel computing topology is also designed to improve the computational speed. Case study shows that the design of both the CCDE-PSD and its parallelization scheme are effective. The CCDE-PSD is superior to basic DE with respect to both solution quality and computational speed. It can reach better solutions with smaller population size and fewer generations.Chapter 7 concludes the thesis and points out some directions for future research.