Electricity consumption prediction using an advanced spatial-temporal deep learning framework

Reliable forecasting of electricity usage is vital to support economic growth and economic performance. The inherent non-linearity and complex temporal dependencies in energy demand data pose significant challenges for traditional forecasting methods. This report details an extensive investigation into the application of advanced deep learning (DL) architectures for this task. We propose spatial temporal model, namely Spatial-Temporal GRU (ST-GRU), for predicting the electricity consumption. In this work, the term spatial refers to inter-feature relationships within the multivariate input space rather than geographic spatial information. Utilizing a multi-source 15-minute interval dataset, we employ cyclical feature engineering and a sliding window approach to structure the sequential inputs. We conduct a comparative analysis of a diverse suite of sequence models, including baseline GRU and LSTM networks, alongside more sophisticated architectures such as Bidirectional RNNs, Attention-based models, Temporal Convolutional Networks (TCN), Transformers, and Spatial-Temporal models. Our empirical evaluation on a reserved test set reveals that specialized architectures, particularly the Spatial-Temporal Gated Recurrent Unit (ST-GRU), deliver superior predictive accuracy. The proposed ST-GRU reduced RMSE by 6.4% compared to standard GRU and by approximately 25% compared to LSTM on Dataset 1, while maintaining consistent generalization performance across cross-dataset evaluation. The ST-GRU model demonstrates a notable ability to effectively model the inter-feature relationships (spatial dynamics) before capturing their evolution over time (temporal dynamics), resulting in the lowest prediction error. This outcome highlights the substantial potential of architecturally specialized DL frameworks to advance the precision of energy.