A MPRF Residual Fusion Model for Short-Term Electric Load Forecasting

Abstract

Short-term electric load forecasting is crucial for the daily scheduling and market trading of power systems, ensuring their stable operation. However, the randomness and uncertainty of load series make accurate forecasting exceptionally challenging. To enhance the forecasting accuracy, a novel MPRF residual fusion model based on multilayer perceptron (MLP) and a periodic difference gated recurrent unit (PDGRU) is proposed in this paper. This model focuses on the important, detailed features of the power load series to better capture its volatility. Firstly, the main features are extracted by utilizing the nonlinear mapping capability of MLP. By separating these main features from the original data, more detailed residual features are obtained. Then, in order to make better use of the periodicity of the detailed component, a PDGRU model is proposed that uses periodic difference to optimize the GRU network. Finally, the extracted features are fused to produce the final forecasting result. Results validated by the actual data sets for Shaanxi province, China, show higher accuracy of the proposed model in comparison with other forecasting methods.