Particle-swarm-optimized neural networks for baseline energy modelling in institutional buildings

Accurate baseline energy models are essential for quantifying savings and guiding efficiency measures in institutional buildings under IPMVP-aligned M&V. This study evaluates multiple linear regression (MLR), a gradient-trained single-hidden-layer artificial neural network (ANN, tanh), and a particle-swarm-optimized ANN (PSO-ANN) for hourly electricity prediction in a Malaysian academic building using 1,584 hourly observations of energy use, outdoor temperature, and occupancy. Data are randomly partitioned 70/15/15 (train/validation/test). Each model is trained with 15 random seeds; for ANN and PSO-ANN, the hidden-layer size is swept from 5 to 30 neurons for each seed, retaining the configuration with the lowest validation-set MSE. A swarm size of 200 and a maximum of 1000 iterations are used to maintain search diversity and support convergence when optimizing the full weight–bias vector. On the test set, MLR, ANN, and PSO-ANN achieve RMSE of ≈23.5 kWh, 15.4 kWh, and 14.5 kWh, respectively (MSE ≈ 553.8 kWh², 238.6 kWh², and 210.0 kWh²). Relative to MLR, ANN reduces test MSE by 57 % and PSO-ANN by 62 %; the additional improvement over ANN is modest. Across seeds, PSO-ANN shows slightly lower central error; however, the improvement over ANN is not statistically significant (p > 0.05). Overall, PSO-ANN provides an accurate and reproducible baseline for institutional buildings, enabling benchmarking, retrofit planning, and IPMVP-oriented M&V within smart-campus operations.