Optimized Prediction of Airflow Volume in Under-Actuated Zones through Multilayer Perceptron Artificial Neural Network

This study addresses the challenge of predicting airflow volume in under-actuated zones, where occupant behavior and environmental factors complicate standard models.To improve prediction accuracy, we propose the Sigmoid Parametric Shifted ReLU (SPS-ReLU) with custom weight scaling as a novel activation function within a Multi-Layer Perceptron Artificial Neural Network (MLP-ANN) model.The model was trained and tested on a time-series dataset from a controlled environment, using optimal time intervals (5, 15, and 30 minutes) identified through polynomial regression analysis.These intervals best capture airflow patterns: the 5minute interval effectively handles rapid fluctuations in Zones 1 and 2, while the 15-minute interval is better suited for the gradual changes in Zone 3. Results show that SPS-ReLU, particularly with a weight scale of 1.5, significantly improves accuracy, achieving an RMSE of 2.3891 and R of 0.9974, outperforming both standard and advanced activation functions.Comparatively, DPReLU achieved an RMSE of 3.0469 and R of 0.9957, while ReLU's RMSE was 22.5458 with an R of 0.7741.This demonstrates SPS-ReLU's capability to balance smoothness and flexibility, enabling it to capture intricate airflow dynamics within dynamic environments.The findings highlight SPS-ReLU with custom scaling and optimal time intervals as an effective solution for enhanced airflow predictions in under-actuated zones.