Short-Term User Load Forecasting Method Considering Meteorological Factors Based on Conv1D-LSTM with Attention Mechanism

With the energy system transition to more intellectualized, more flexible and more interactive, power companies need more accurate load forecasting results to design effective dispatching and operation schemes to improve competitive in market. Load forecasting, especially short-term load forecasting for a single power user, plays an increasingly important role in power grid planning and stable operation in the future. A Conv1D-LSTM neural network model based on attention mechanism is established in this paper, considers a time of the day index, a day of the week index and seasonal factors, which is shown to increase the forecasting accuracy, so as to avoid the negative impact of environmental variables. The experimental results show that the Conv1D-LSTM neural network model based on attention mechanism can absorb the advantages of CNN network automatic feature extraction and LSTM network efficient processing of time series, and improve the prediction accuracy. Moreover, in order to evaluate the performance of the model in this paper, the performance of the other machines, namely a Recurrent Neural Network (RNN) and a Long Short-Term Memory (LSTM) and a Gated Recurrent Unit (GRU) is also shown. The results show that this model can reduce the generalization error of the model, it is verified that the method based on Conv1D-LSTM with attention mechanism can faster model training speed, higher prediction accuracy and lower prediction error.