A novel CNN-LSTM-based forecasting model for household electricity load by merging mode decomposition, self-attention and autoencoder

Accurate electricity load forecasting can maintain the supply-demand balance, and ensure the safe and stable operation for the power grid system. However, household electricity data usually exhibits intricate temporal dependencies and multi-scale patterns, thereby hindering the comprehensive extraction and effective utilization of its underlying temporal dynamics. To address this issue, a novel electricity load forecasting model, named as CEEMDAN-CNN-LSTM-SA-AE, is developed in this study by integrating complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), convolutional neural network (CNN), long short-term memory (LSTM), self-attention (SA) mechanism, and autoencoder (AE). The proposed model first utilizes the CEEMDAN, a signal decomposition method, to separate the original load data. Then, a CNN model is established to capture local temporal features for the samples constructed from decomposed sub-signals. Next, the output features of CNN are passed to a designed LSTM-SA-AE model and thus the ultimate prediction results are obtained. In this stage, the embedding of SA between the long short-term memory encoder and decoder can automatically extract representative features. Finally, the forecasting performance of CEEMDAN-CNN-LSTM-SA-AE has been successively validated through numerical experiments on two household electricity load datasets. Experimental results show that the proposed model significantly outperforms existing baseline models, achieving a minimum 28 % improvement in R2 and a maximum 52.36 % reduction in MAPE on the first dataset.