Short- Term Load Forecasting Model Based on Multi-Feature Fusion of Long Short-Term Memory Network

Abstract

With the continuous development of China's economy and the popularity of Internet of Things and intelligence, the use of power load forecasting in production an life has deepened in proportion. In this paper, multiple influencing factors of electric load are considered, and in terms of model selection, a multivariate Long Short-Term Memory (LSTM) model considering external multi-feature factors is adopted based on the LSTM network model in deep learning in order to avoid phenomena such as long-term dependence on the network and gradient disappearance. By adding a fully connected layer, it is made to analyze all external feature factors in a weighted manner and fuse the whole features through the connected layer before output to achieve the optimal output. To verify the accuracy and improvement advantages of the model, historical electricity load data of a region in 2012 are used for training forecasts. A comparison test is set up to predict the load of the last month in the sample by the improved model and the LSTM model respectively, and the results show that the former has better prediction accuracy with a MAPE value of 4.94%.