Data-driven stnet and stprophet models for secure edge-based indoor air temperature prediction in smart buildings

Energy efficiency is vital yet underutilized in buildings. Reducing energy consumption while maintaining human-level comfort within certain boundaries requires accurate indoor air temperature (IAT) modelling. IAT prediction models support HVAC optimization, setting operational limits, and detecting discrepancies between predicted and actual conditions for predictive model control. However, accurately predicting IAT in large-scale smart buildings is challenging due to numerous complex factors. To address this issue, this paper presents two data-driven hybrid models for accurate IAT prediction. The first model, STNet, integrates a CNN with a Bi-LSTM, while the second model, STProphet, combines a CNN with Transformers to capture spatial–temporal dependencies. Both models are deployed on an edge device to enhance data security and privacy. Experimental evaluation shows significant improvements over a baseline method. STNet reduces MAE, RMSE, and MAPE by 75.74%, 68.58%, and 76.92%, respectively. STProphet achieves reductions of 72.44%, 66.58%, and 73.76% for the same metrics. Inference efficiency also improves substantially: STNet reduces latency by 53.64% (to 51 ms) and STProphet by 68.18% (to 35 ms), compared with the baseline's 110 ms. The results confirm the effectiveness of the proposed models for real-time IAT prediction, supporting more reliable energy modelling and optimization in large-scale smart buildings.