Spatial-temporal modeling for urban climate: Enhancing energy prediction with global temporal convolutional attention networks

Urban climate challenges and the growing significance of spatial-temporal big data have led to extensive exploration of integrating smart grids and renewable energy sources, such as solar electricity, in urban environments. Ensuring the stability and resilience of these systems is paramount, and accurately forecasting electrical energy demand and supply is a crucial element in achieving this objective. Temporal Convolutional Neural Networks (TCNN) have recently emerged as a promising method for spatial-temporal modeling. To further improve the accuracy and temporal modeling effectiveness of TCNN, the Temporal Convolutional Attention Neural Network (TCAN) was developed, enhancing the receptive field coverage of TCNNs. However, TCAN has a limitation in that it cannot effectively utilize earlier input time steps for later predictions in multi-step autoregressive forecasting, which constrains its application in electricity series forecasting. In this work, novel deep learning models, GTCAN-P and GTCAN-M, based on spatial-temporal big data are introduced to further improve TCAN. The models allow the network to access all past time steps by employing padding and masking mechanisms, respectively. The effectiveness of the proposed methods was evaluated on real-world electricity datasets, including two building energy consumption datasets and three solar power generation datasets, and compared with statistical models and state-of-the-art deep learning models. The results demonstrate the potential of GTCAN-P and GTCAN-M to address the limitations of TCAN and provide more accurate forecasting of electricity demand and supply in urban energy management scenarios.