Deep Learning-Based Spatial-Temporal Modeling for Energy Consumption Prediction

초록

Accurate short-term energy demand forecasting is a critical component of sustainable smart grid operation and building energy management. In this study, we propose a hybrid deep learning architecture that integrates two-dimensional convolutional neural networks (2D-CNN) with long short-term memory (LSTM) layers for high-resolution energy consumption prediction in educational buildings. Unlike conventional one-dimensional or purely sequential models, the proposed 2D-CNN + LSTM framework captures both local spatial correlations across weather and building metadata features and long-range temporal dependencies within energy consumption sequences. Using the ASHRAE public dataset, the model was trained and evaluated exclusively on the Education building type, incorporating features from building metadata, on-site weather conditions, and temporal indicators. Experimental results show that the proposed model achieves superior forecasting accuracy compared to benchmark models such as LSTM, GRU, Transformer, and CNN variants. Moreover, comprehensive analyses of computational cost, training time, and inference latency demonstrate that the proposed model achieves a balanced trade-off between accuracy and efficiency. This work provides a reproducible baseline for energy forecasting research using hybrid spatial-temporal architecture and offers valuable insights into model scalability for smart campus and district-level applications.