TCAMS-Trans: Efficient temporal-channel attention multi-scale transformer for net load forecasting

Accurate net load forecasting contributes to increasing the integration of renewable energy sources and reducing the operating cost of the power grid. In recent years, deep learning models have been applied in net load forecasting. However, there is still a large scope for them to improve the extraction of complex multi-scale temporal features in the net load. In addition, most researches do not establish explicit mechanisms to consider the impact of relevant factors on net load forecasting, such as renewable energy outputs and load demand. Therefore, in this paper, we propose a multivariate Transformer-based model, named TCAMS-Trans, to achieve direct net load forecasting. First, TCAMS-Trans exploits multi-scale patch-based variable-independent embedding to model potential multi-scale patterns in the load data. Subsequently, TCAMS-Trans designs a new hierarchical encoder-decoder pair configured with temporal-channel attention block (TCA block) to further extract multi-scale features and capture the connections between the net load and the relevant factors. The temporal attention in TCA block is used to mine complex temporal dependency along the time dimension, and the channel attention is used to capture the inter-variable dependency along the variable dimension. Experiments conducted on three publicly available datasets show that TCAMS-Trans has higher prediction accuracy than the current state-of-the-art models at different prediction lengths. The model proposed in this paper provides a new powerful solution for net load forecasting.