Power Load Forecasting Based On Echo State Network

Power load forecasting is critical for power department to make power grid planning, the historical load value is affected by temperature, season and many other factors, so its can be seen as the complex nonlinear time series with strongly nonlinear and non-stationary characteristics. At present, the main power load forecasting tool are neural network and support vector machine and other single computational intelligence methods. As a new dynamic recurrent neural network, echo state network(ESN) has aroused extensive attention by many researchers, and which has been used in power load forecasting. Compared with the conventional recurrent neural network, ESN only need to calculate the network output weights at the network training stage, has a strong ability of dynamic approximation.Therefore, the advantages of ESN is that its calculation is simple and effective, and has fast convergence speed.As a kind of adaptive signal decomposition method, empirical mode decomposition(EMD) can realize decomposition for the complex non-stationary load sequence, Contribute to grasp the inherent variation of the load sequence, so, which can improve the prediction accuracy effectively. Therefore, as an adaptive signal processing method, EMD has been successfully used in the load forecasting. On the basis of ESN and EMD, a kind of combined short-term power load forecasting method that based on complementary ensemble empirical mode decomposition(CEEMD)-fuzzy entropy(FE) combined echo state network(ESN) with leaky integrator(Li) neurons and a kind of combined mid-term peak load forecasting method that based on the complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)-permutation entropy(PE) combined LiESN are proposed in this thesis, both of the two method obtained a good prediction effect. The main contents of this thesis are as follows:(1) The ESN network and corresponding learning algorithm have been studied, and on this basis of ESN, focuses on the research of the LiESN learning algorithm, at the same time,the ridge regression learning algorithm is applied to calculate the output weights.(2) Empirical mode decomposition(EMD) and ensemble empirical mode decomposition(EEMD) learning algorithms have been studied, and on the basis of EMD and EEMD, in order to increase the integrity of the signal decomposition, further investigate the CEEMD and CEEMDAN method. In order to decrease the computing scale of the combined forecasting, using permutation entropy(PE) and entropy fuzzy(FE) to evaluate the complexity of the intrinsic model function(IMF) sequence.(3) The combination method of CEEMD-FE and LiESN has been given, and which is applied to the instance of short-term power load forecasting at New England; The comb-ination method of CEEMDAN-PE and LiESN has been given, which is applied to the instance of mid-term power load forecasting at New England and Smart grid in Europe respectively. The experimental results verify the effectiveness of the proposed combination forecast method through compared with the existing methods.