Optimising deep learning for smart building energy prediction: A particle swarm approach with real-world validation

Accurate short-term load forecasting (STLF) in smart buildings for the operation and control of HVAC and building systems is particularly challenging due to the complexity of system dynamics. While hybrid deep learning (DL) models, such as CNN- and LSTM-based architectures, offer high predictive performance, they are resource-intensive. Simpler models like multilayer perceptrons (MLP) are computationally efficient but typically lack accuracy for STLF prediction. This study proposes a novel approach, PSO-MLP, which leverages particle swarm optimisation (PSO) to enhance MLP performance by tuning architectural and learning hyperparameters. A novel incremental technique is employed to iteratively build deep MLP from previous optimal solutions, an area not commonly explored. The method is tested on a real-world smart building—‘Nanterre 3’ (N3) at CESI Paris—and benchmarked against advanced, hybrid DL models (MHA-CNN-LSTM and LSTM-DWT-CRT) using public datasets (UCI IHEPC and AMPds), demonstrating its effectiveness and generalisability across diverse datasets. The PSO-MLP model achieved ∼ 88% accuracy in predicting energy consumption 24-hour ahead in the N3 building, resulting in a 25.4% reduction in MAPE compared to the original MLP, with a runtime of 130 min. On public datasets, PSO-MLP outperformed both the original MLP (by 45.3%) and the MHA-CNN-LSTM (by 29.6%). It also consistently outperformed the LSTM-DWT-CRT model across all error metrics. Incorporating feature selection further improved accuracy by 5.49% on AMPds, making the proposed model a competitive, lightweight alternative for smart building energy forecasting with direct applications in building management systems and improving energy efficiency, and enabling edge computing and smart grid control. • Proposed particle swarm–based neural model for short-term load forecasting. • Developed deep neural network incrementally from optimised shallow models. • Validated on real-world smart building data and two benchmark public datasets. • Achieved 88% accuracy and 25.4% reduction in forecasting error. • Outperformed recent advanced deep learning methods across all datasets.