Dual Attention Mechanism Based EMD-GRU for Electricity Load Forecasting

Abstract

In order to fully exploit the correlation features of the data on power load time series and input data, and to increase the non-stationary power load’s prediction accuracy, an empirical mode decomposition (EMD) – gated recurrent unit (Gru) short-term power load forecasting model based on double attention mechanism is proposed. Firstly, the raw load data are decomposed by EMD to obtain the Eigen-modal component (IMF). Secondly, the relevance between the historical load data and the IMF component is mined using a feature attention mechanism, the component is trained using a multi-layer GRU network, the temporal correlation features of the historical load data are mined using a temporal attention mechanism, and the forecasting model is generated. Finally, the model is tested and the output prognosis results are reconstructed to gain the ultimate load forecast values. In this paper, the actual load data from 2017 to 2021 in Gansu Province are used for experimental analysis, which verifies that high prediction accuracy is achieved by the method.