Short-term Power Load Forecasting Based on TCN-BiLSTM-Attention and Multi-feature Fusion

Abstract

Accurate power load forecasting provides reliable decision support for power system planning and operation, however, only using the load data for prediction is not enough, since it is influenced by electricity demand, electricity behavior, electricity prices, etc. Inspired by this, this paper proposes a hybrid model to promote the short-term power load forecasting performance by integrating such external factors and power load as multivariate time series. The proposed model, TCN-BiLSTM-Attention, combines two temporal convolutional network (TCN), two bidirectional long short-term memory (BiLSTM), and attention mechanism. Wherein, TCN uses parallel convolution kernels to extract temporal features from preprocessed each subsequence, and then BiLSTM further captures the long and short-term dependencies of these features. Further, the flatten and fully connection layer with Attention discovers the correlations between multivariate time series and improves the predictive performance by giving higher weights on the important information. The extensive experiment results show that TCN-BiLSTM-Attention is superior to the state-off-the- art, and the addition of multiple factors enables it to learn more useful information, and thus improving the prediction performance. All suggest that there is a strong correlation between the power load and external factors, and the proposed model can effectively obtain the long and short-term dependencies of single sequence and the correlations between multivariate time series, and this advantages makes it have excellent predictive performance and strong robustness in short-term load forecasting.