Hybrid attention-based improved temporal convolutional BiGRU approach for short-term load forecasting

Abstract

Accurate load forecasting can ensure the safe and reliable operation of power systems, reduce generation costs, and improve economic efficiency. To improve the accuracy and performance of short-term load forecasting, this paper proposes a hybrid short-term load forecasting method composed of an improved temporal convolutional network (TCNPlus) with an attention mechanism and a bidirectional gated recurrent unit (BiGRU). Firstly, the collected pre-processed training data is reconstructed using a fixed-length sliding window. Secondly, using the self-attention mechanism (SA) in the improved TCN to further enhance the weight of key features, and introducing residual connections can allow the input to propagate forward faster and improve the representation ability and efficiency of error backpropagation of the network, to eliminate the impact of interference signals. Finally, BiGRU is used to learn the forward and backward dependencies of the load sequence in both directions and predict the true load value. Based on the real load data of a national power grid in South China, through experimental comparison of multiple models, the results show that this model still has higher short-term load forecasting accuracy with fewer input features.