Research on Impulse Power Load Forecasting Based on Improved Recurrent Neural Networks

Abstract

Deep learning is good at extracting the required feature quantity from the massive input information through multiple hidden layers and completing the learning through training to achieve the task of load forecasting. The impulse power load data contain a lot of noise, burrs, and strong randomness. As an improved recurrent neural networks, the output of long short-term memory (LSTM) network is not only related to the current input, but also closely related to the historical information, which can effectively predict the impact power load. An impulse power load forecasting model based on improved recurrent neural networks is proposed. To solve the training difficulties caused by deep networks, database is divided into training data set and test data set. To accelerate running speed and improve accuracy and reliability, parameter setting in deep learning neural network is analyzed. The proposed load forecasting model is verified by simulation and compared with the existing methods. Taking the average relative error as the standard, the effectiveness of the proposed model for the forecasting of impulse power load connected to the bus is verified.