Short-term load forecasting scheme based on improved deep residual network and LSTM

Abstract

Short-term load forecasting with higher accuracy can help the power system to design reasonable operational planning of generation units. However, the features of load data in deep neural networks are not fully extracted, which affects the prediction accuracy. In this study, the authors propose a new short-term load forecasting scheme based on a neural network hybrid model. The authors first choose a similar day from the historical hourly load data and forecast the hourly load of the day. Especially, in order to extract more features, the authors input the hourly load to the improved deep residual network model (ResNetPlus), which is improved on the deep residual network (ResNet). Finally, the authors combine the correlation factor data with the feature vectors from the ResNetPlus network and enter the data into long short-term memory (LSTM) network. The experimental results prove the effectiveness of the proposed model. Compared with the deep residual network, improved deep residual network, and LSTM network based on electricity load forecasting, the authors reduce the mean absolute percentage error with higher prediction accuracy.