Deep Learning Framework Using Transformer Networks for Multi Building Energy Consumption Prediction in Smart Cities
The increasing complexity of urban building energy systems necessitates advanced prediction methods for efficient energy management. Urban buildings account for approximately 40% of global energy consumption, making accurate prediction crucial for sustainability goals. This research develops a novel...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-03-01
|
| Series: | Energies |
| Subjects: | |
| Online Access: | https://www.mdpi.com/1996-1073/18/6/1468 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The increasing complexity of urban building energy systems necessitates advanced prediction methods for efficient energy management. Urban buildings account for approximately 40% of global energy consumption, making accurate prediction crucial for sustainability goals. This research develops a novel transformer-based deep learning framework for multi-building energy consumption forecasting. Despite recent advances in energy prediction techniques, existing models struggle with multi-building scenarios due to limited ability to capture cross-building correlations, inadequate integration of diverse data streams, and poor scalability when deployed at urban scale—gaps this research specifically addresses. The study implemented a modified transformer architecture with hierarchical attention mechanisms, processing data from 100 commercial buildings across three climate zones over three years (2020–2023). The framework incorporated weather parameters, occupancy patterns, and historical energy consumption data through multi-head attention layers, employing a 4000-step warm-up period and adaptive regularization techniques. The evaluation included a comparison with the baseline models (ARIMA, LSTM, GRU) and robustness testing. The framework achieved a 23.7% improvement in prediction accuracy compared to traditional methods, with a mean absolute percentage error of 3.2%. Performance remained stable across building types, with office complexes showing the highest accuracy (MAPE = 2.8%) and healthcare facilities showing acceptable variance (MAPE = 3.5%). The model-maintained prediction stability under severe data perturbations while demonstrating near-linear computational scaling. The transformer-based approach significantly enhances building energy prediction capabilities, enabling more effective demand-side management strategies, though future research should address long-term adaptability. |
|---|---|
| ISSN: | 1996-1073 |