Energy storage peak-shaving demand prediction based on an improved neural network with hidden layers

Abstract

Abstract The result shows that the BP neural network has a very low convergent speed for load prediction, which leads to local optimization. This paper presents a modified GA for the optimization of the threshold value, the initial weight, and the addition of the smoothing coefficient. In addition, it improves the BP neural network’s adaptive learning speed and further improves the efficiency of the search. The proposed model is characterized by high convergent speed and global space retrieval capability. To validate the validity of the modified GA-BP model, the prediction of power load in some regions is made. Compared with the BP neural network forecast algorithm, this method has better performance in terms of prediction accuracy and takes less time in terms of prediction time than the BP neural network. In parallel, the realization of multiple sources of electricity from wind, solar, and heat storage is an important way to deal with large-scale new grid connections. Starting from an improvement in the efficiency of wind and PV generation, it is also possible to build a joint optimization scheduling model for wind, solar, and heat storage. The purpose of taking the minimum net load fluctuation is for optimization. Taking the typical output day of user scenarios as the optimization scenario, the scheduling strategies of wind power, PV, thermal units, and energy storage systems are introduced and simulated on the platform. It is proved that the proposed model can decrease the reduction ratio and the peak and valley load.