A Short-term Residential Load Forecast Model Based on BiLSTM-MDN

Abstract

With the development of economy, residential power users account for a higher and higher proportion in the power system. The modern power system focusing on residential load needs to realize the stability of load demand changes by combining forecasting information with long and short term dispatching. However, residential micro grid load usually has high fluctuation, so it is a challenging problem to achieve accurate prediction. Based on the characteristics of residential power load, this paper studies the short-term forecasting task of residential power load. BILSTM-MDN hybrid prediction models were constructed by BiLSTM's ability to learn long-term dependence and underlying correlation logic. Finally, 50 apartment load data sets are used to verify the great potential of the model based on BiLSTM-MDN in residential short-term power load prediction with high fluctuation. The accuracy of prediction reached MAPE 18.25% and RMSE 30.53%.