Short-Term Load Forecasting Based On Double Hidden Layer Neural Network And Chaotic Time Series

With the development of the power system, high-quality short-term load forecasting more and more important and urgent. So short-term load forecasting (STLF) plays a key role in the power system operation is a prerequisite for safe and economic operation of power system. This paper set electricity load forecasting as the object of study, conducted in-depth research on how to improve the accuracy of the power load problem. Through the characteristics analysis of power load data, I proposed a new feature extraction method based on chaotic time series, DFNS. Secondly double hidden layer neural network combined with Softplus, used in power load forecasting in this field was proposed. Finally in the algorithm optimization question, I used the improved DE, which called BMDE with high speed and accuracy to solve it.Firstly, determine power load value is chaos, so the treatment can be used to deal with the chaotic time series data in line. Then consider the current mature approach chaotic time series, and from the perspective of both theoretical and experimental comparison of the pros and cons of various methods, and ultimately determine the use of the complex autocorrelation method and the G-P algorithm to strike a phase space reconstruction parameters. Finally, based on phase space reconstruction, a new feature extraction method was proposed, which is called dynamic feature selection method 0 In order to improve the accuracy of forecasting power load based on the input vector nonlinearity.Taking the actual operation of the computer defects in the process into account, it will make a single hidden layer neural network accuracy can not reach the level of theory. This article intends to improve the prediction accuracy by using double hidden layer neural network, and set Softplus function as activating to ensure prediction algorithm reliability. From the simulation results, double Softplus hidden layer neural network’s fitting accuracy compared to the single hidden layer Sigmoid type has improved greatly.For the gradient descent algorithm fall into local minimumin neural network’s optimization process. I propose to join DE ensure global optimization algorithm. And for DE"s shortcomings of slow convergence. I proposed the use of variable-bound search and copying the first generation of high-quality individuals to improve the way through the simulation results show the effectiveness of the improvement of DE. Then, feature extraction, neural network structure and DE combined to form the overall structure prediction algorithm of this paper. Finally use this forecasting method to predict the real power load data from the US PJM electricity market.