Short-Term Load Forecasting Based On Phase Space Reconstruction And Support Vector Machine

Short-term load forecasting (STLF) in power system is an important work for power department, and its accuracy directly influences the safety and economic operation of power systems, and may also influence the power quality. It was drawn from a lot of researches that the short-term load time series has chaotic characteristics. It is difficult to forecast the load using traditional methods. In this thesis, based on the phase space reconstruction, add-weighted one-rank local-region multi-steps model (AOLMM) and support vector machine (SVM) model are used for the STLF. Theoretical analysis and testing results prove them to be valid and feasible. The main contents discussed in this thesis are described as follows:Phase space reconstruction theory is the foundation of short-term load chaotic time series forecasting. The choice of reasonable parameters, which is useful for the phase space reconstruction, makes it possible to get the information contained in the time series and to carry out the accurate load forecasting. Based on the phase space reconstruction, many load time series forecasting methods were established. The common used add-weighted one-rank local-region method (AOLM) is a single-step model, and it will result in large computational quantity and cumulative error for multi-steps forecasting. Due to the disadvantage, an AOLMM forecasting model is employed. When using the two methods for the STLF of the New South Wales, in Australia, 2006, satisfying results are reached. The chaotic characteristics of the load time series in this case are analyzed, and the phase space reconstruction parameters are deduced. Then two methods are used in the forecasting experiment for the load in working day and rest day. The simulation results show the superiority of the AOLMM model in STLF.The SVM method is used in STLF in the thesis. Aimed at solving the problems of choosing parameters of SVM and its kernel function, a novel optimization algorithm called stochastic focusing search (SFS) is proposed. The new algorithm is sorted to the swarm intelligence algorithm, which is based on simulating the behaviours of human randomized searching. It is simple with few parameters, and has the lower computational complexity than particle swarm optimization (PSO) and other modified PSO algorithms. The performance of this algorithm is investigated in optimizing" the typical benchmark functions and solving the reactive power optimization problem. After the comparison with differential evolution (DE) and three modified PSO algorithms, it is proved that SFS has a broader prospect and much more practical value.In order to choose the input parameters of the SVM in STLF, a SVM method combined with phase space reconstruction is proposed in the thesis. The historical load time series are used for phase space reconstruction, and the meteorological factors are ignored. The vectors in the load time series reconstructed phase space are treated to be the input of SVM, and then the nonlinear problems are transformed to linear ones in eigenspace mapped by kernel function. Simulation of daily peak load forecasting results show the validity of the model.Generally, imperceptible features are very important in processing load time series signal with chaotic characteristics. Some kinds of translation-invariant wavelet kernel SVM are established through combining wavelet analysis method with SVM kernel function method in this paper. Gaussian series wavelet kernel, complex Morlet wavelet kernel and complex Gaussian wavelet kernel are proposed, and some testification were done to prove that it satisfies translation-invariant kernel theorem. It is used in approximation of a single-variable function, two-variable function and forecasting chen's chaotic time series. Simulation results show the feasibility and validity of the three translation-invariant wavelet kernel functions, and they all perform better than the common used Gaussian kernel and Morlet wavelet kernel. Based on phase space reconstruction, the translation-invariant wavelet kernel support vector regression machine models are used in the forecasting experiment for the load in working day and rest day. The simulation results show that the wavelet kernel SVM model combined with phase space reconstruction is efficient in achieving the accurate STLF.