Short-term power load forecasting based on DQN-LSTM

In order to improve the accuracy of short-term load forecasting and solve the problems of time sequence, nonlinearity and multidimensionality of short-term power load data, this paper proposes a multi-step short-term power load forecasting method based on reinforcement learning Deep Q Network (DQN) and long-short-term memory network (LSTM). Firstly, the normalized historical data are sampled by time sliding window method to construct feature map as input. Then, LSTM is used as the main body to build a prediction model to capture long and short time series features. DQN is combined with LSTM network to form a skip layer to capture super-long time series features, so that DQN-LSTM can capture features of different time scales. Finally, the full-connection layer is introduced to fuse the multi-scale temporal features extracted from each layer, and the autoregressive model (AR) is used as the linear component to obtain the final prediction results. Experimental results verify that the MAPE and RMSE of the proposed prediction method are lower than those of other algorithms, and it can effectively capture the multi-scale time series features of power load data, thus improving the prediction accuracy.