Cooling Load Prediction in the Under-Actuated Zone with Multilayer Perceptron Artificial Neural Network

This study focuses on addressing the challenge of predicting cooling loads in under-actuated zones, where the variability in occupant behavior, environmental conditions, and electronic usage creates complex dynamics. Traditional models like Support Vector Regression (SVR) and E-Lastic-Net (ELN) often struggle to capture these non-linear relationships, leading to inefficient Heating, Ventilation, and Air Conditioning (HVAC) management and increased energy consumption. To overcome this, the research proposes a hyperparameter-tuned Multi-Layer Perceptron-Artificial Neural Network (MLP-ANN) model, enhanced by integrating trainable bias and custom weight scaling. The results show that the proposed model significantly outperforms both baseline models and state-of-the-art techniques. The model using leaky ReLU with trainable bias and a weight scale of 2.0 achieved superior performance, with an RMSE of 128.26, MAE of 90.65, and an R² of 0.9992. In comparison, the baseline models demonstrated RMSE values between 1906 and 1919 and R² scores ranging from 0.8105-0.8141, showcasing the proposed model's effectiveness. Furthermore, activation function performance showed substantial improvement, particularly in reducing dead neurons and training loss. ReLU with trainable bias and a weight scale of 2.0 had a final training loss of 1,034,874.61 and 0.83% dead neurons, while PReLU and leaky ReLU with trainable bias had 0% dead neurons. These enhancements, along with improved smoothness scores (ranging from 0.84-1.24), contributed to more stable and accurate predictions, highlighting the benefits of trainable bias and custom weight scaling in improving model performance and generalization.