Abstract: In order to cope with the influence of power load fluctuation on the stability of power system when meteoro logical factors change, a CNN-LSTM combination model with attention mechanism was explored to predict the fluctuation of short-term power load. The power load was affected by the complex coupling of multi-dimensional climate factors, in order to effectively characterize these nonlinear and time-varying climate-load correlation characteristics, a multi-feature input model integrating temperature, precipitation, humidity and wind speed was constructed.This paper used the convolutional neural network(CNN) to capture the local climate patterns in the data, and the time-conditioning characteristics of key meteorological events were extracted through the sliding window mechanism. Then, the feature vectors were fed into the long short-term memory (LSTM) network, and its gating mechanism can effectively model the delay effect of climate factors and load response.Finally, the attention mechanism (AM) was introduced to dynamically quantify the temporal importance of each climate element. Through the comparison of simulation experiments, the results show that the CNN-LSTM-AM model has better prediction accuracy than the traditional LSTM and CNN-LSTM models.

    Key words: convolutional neural network; long short-term memory; attention mechanism; power load forecasting