Research On Short-term Load Forecasting Of Power System Based On Feature Selection And Neural Network

Power system is one of the important infrastructures of modern society,which consists of power generation,transmission,distribution and consumption.Short-term load forecasting can provide guarantee for the stable operation of power system.In today’s vigorous development of new energy,the diversified forms of power generation and consumption make the short-term load forecasting more and more difficult,and the traditional forecasting models can no longer solve the problem of high accuracy requirements.In order to obtain accurate and effective forecasting models,scholars at home and abroad have conducted a lot of research,but the main way of current research is to increase the amount of data and data features,which makes the dimensionality of model training increasing.To address the above problems,by deeply analyzing the working principles of feature selection and neural networks,this paper proposes a prediction model based on feature selection and neural networks.The main contributions of this paper are:(1)The load is divided into four seasons of spring,summer,autumn and winter load considering seasonal factors.On this basis,in order to solve other irrelevant features that affect the prediction accuracy,Pearson correlation analysis and Spearman correlation analysis are used to synthesize the climate and tariff characteristic factors of the four seasons and give the features with large influence factors for each season.(2)A combination of Convolutional Neural Networks(CNN)and Long short-term Memory(LSTM)network is used.The load prediction model based on CNN-LSTM with feature selection is compared with the full-featured LSTM and feature-selected LSTM models.It is verified based on experiments that the models with feature selection and with CNN have a large improvement.(3)To enhance the training ability of the model,Bi-directional Long Short-Term Memory(BiLSTM)network is discussed,and a CNN-BiLSTM-Attention prediction model is proposed,which can fully learn the data in the sequence and can focus more on the important parts of the data and reduce the complexity of the model.Then,the data after feature selection is fed into the model.The experimental results show that the combined CNN-BiLSTM-Attention network has further improved the prediction accuracy in different seasons compared with the BiLSTM-Attention network.