Simultaneous Feature Selection and Network Optimization in HVAC Systems Using a Hybrid Evolutionary Mating Algorithm-Artificial Neural Network

Accurate prediction of supply water temperature is essential for optimizing energy consumption and maintaining thermal comfort in heating, ventilation, and air conditioning (HVAC) systems. However, selecting relevant input features and determining an appropriate neural network architecture remain challenging tasks that significantly influence model performance. This paper proposes a hybrid framework that integrates the Evolutionary Mating Algorithm (EMA) with an Artificial Neural Network (ANN) for simultaneous input feature selection and network hyperparameter optimization. Each candidate solution is encoded as a chromosome representing the input feature subset, number of hidden neurons, and activation function, which are jointly optimized through the EMA to minimize a regularized fitness function balancing prediction accuracy and model complexity. The proposed EMA-ANN framework was evaluated on a real HVAC dataset and benchmarked against two established metaheuristic approaches, namely Particle Swarm Optimization ANN (PSO-ANN) and Genetic Algorithm ANN (GA-ANN), under identical experimental conditions across 10 independent runs. The results demonstrate that EMA-ANN achieved the best predictive performance, attaining an RMSE of 0.9740, an MAE of 0.6898, and an R² of 0.9919 on the unseen test set, outperforming both PSO-ANN and GA-ANN in terms of best, average, and worst-case metrics. Furthermore, the EMA-ANN identified a more compact and physically interpretable feature subset of 8 variables compared to 11 and 10 selected by PSO and GA, respectively. These findings confirm that the proposed framework offers a reliable and efficient solution for data-driven thermal modelling in real-world building energy management applications.