Short-term Power Load Forecasting Model Based on CNN-GRU and Dual Attention Mechanism Hybrid Neural Network Model

Abstract

This paper proposes a CNN-GRU and dual attention mechanism (DAM) hybrid neural network model (CNN-GRU-DAM) for short-time power load forecasting. First, historical data is transformed into a time-series form, and time-series data is used as the input of the model. The relevant time-series features that affect the power load forecasting are extracted through CNN. Then, the extracted time-series features are input into GRU for power load forecasting. The GRU adopts an encoder-decoder framework composed of DAM, where the feature attention mechanism (FAM) is used as an encoder to mine the dynamic correlation between historical data information and the input features. Additionally, the temporal attention mechanism (TAM) is used as a decoder to assign the weights of the hidden state of the GRU to automatically extract key historical temporal information. Finally, actual power load in a certain region is taken as a case to compare the power load forecasting accuracy of CNN-GRU-DAM, CNN-GRU, GRU, and DNN. It is verified that the forecasting mean absolute percentage error (MAPE) of CNN-GRU-DAM is 6.55%, which is better than those of CNN-GRU, GRU, and DNN.