A novel trilinear deep residual network with self-adaptive Dropout method for short-term load forecasting

Abstract

Short-term load forecasting underlies the effective energy management. However, accurate short-term load forecasting is a big challenge owing to the significant uncertainty and volatility of load demand. Deep learning has successfully been applied for short-term load forecasting, but may encounter the problems such as vanishing gradient, exploding gradient, and overfitting of the neural networks, which affect the robustness of the forecasting results. Herein, a novel deep residual network with self-adaptive Dropout method is proposed for short-term load forecasting. First, a new trilinear deep residual network with three routes of stacked residual blocks is proposed to solve the problems of vanishing gradient and exploding gradient of neural networks. Second, a new self-adaptive Dropout method that considers the effects of both redundant features and interference terms of the neural networks is proposed for automatically setting the neuron drop ratio so as to improve the robustness of the model. Finally, an improved neural network ensemble method is applied to further enhance forecasting accuracy without additional training costs. The proposed model has been proven to achieve better forecasting accuracy and robustness than state-of-the-art baseline models on two public datasets. Therefore, the proposed model is beneficial in promoting the development of practical smart grids.