SCG and LM Improved BP Neural Network Load Forecasting and Programming Network Parameter Settings and Data Preprocessing

Data pre-processing in modeling of neural network (NN) is relatively more complicated and usually manual. Trial and error method is commonly used to determine the number of hidden layer neurons, which is easily affected by human factors and is opportunistic. Relevant training parameters using default value commonly result in lower model accuracy. In this paper, a NN load forecasting model with higher accuracy was established using the actual historical load, meteorological data in Yichang, by means of the Scaled Conjugate Gradient (SCG) and Levenberg-Marquardt (LM) improved BP algorithm which is more suitable for modeling of large or moderate size network with fast convergence. The procedures for data pre-processing and program determining the optimal number of hidden layer neurons to reduce man-made interference or contingency are presented. In order to improve generalization ability, an early termination method is used in network training. This paper proposes it is necessary to reset mu factor and the relevant learning parameters in LM training. Illustrations inform that the initial mu should be relatively larger and mu increase factor and mu decrease factor should be close to 1. The result shows that the NN intelligent forecasting model is valid and feasible.