Cooling Load Prediction Model for Underactuated Zones Using a Multi-Layer Perceptron Artificial Neural Network

Yaddarabullah (2025) Cooling Load Prediction Model for Underactuated Zones Using a Multi-Layer Perceptron Artificial Neural Network. Doctoral thesis, Asia e University, Malaysia.

	Text Thesis Yaddarabullah.pdf Restricted to Repository staff only Download (5MB)
	Text Thesis Yaddarabullah-1-24.pdf Download (347kB)

Abstract

This thesis addresses the complex challenge of predicting cooling load in under-actuated zones, where variability in occupant behavior and environmental conditions limits the effectiveness of traditional models. To enhance predictive performance, a Multi-Layer Perceptron Artificial Neural Network (MLP-ANN) was developed, integrating a Leaky ReLU activation function with a trainable bias term and scaled Glorot Uniform weight initialization. The model was trained and validated on a time-series dataset collected from a controlled environment. Optimal time intervals were identified using Polynomial regression, with mixed intervals proving most effective in capturing dynamic occupant patterns and their impact on cooling load. This configuration yielded the highest Kaiser-Meyer-Olkin (KMO) score of 0.6237, outperforming fixed intervals by 7.3% in representing occupant-related variability. A comprehensive feature engineering strategy was employed, incorporating sine-cosine transformations, lag features, interaction terms, and temporal attributes to enhance data representation. These features enabled the model to capture cyclic patterns, historical trends, and complex inter-feature relationships more effectively. The developed model achieved strong predictive performance, with an RMSE of 255.751, MAE of 131.845, and R2 of 0.9962. Compared to the baseline, this reflects substantial error reduction and improved accuracy. Its stability was supported by a low R2 standard deviation (0.0017), indicating strong performance across varying conditions. The novel contributions of this study include the integration of a trainable bias into the Leaky ReLU function with customized weight scaling, the use of empirically derived mixed time intervals, and the development of a context-aware feature engineering framework tailored for under-actuated zones. These innovations enhance the model’s adaptability and generalizability, offering a reliable and scalable solution for occupant-centric cooling load prediction in modern HVAC systems.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	The under-actuated zone, occupant centric control, time interval analysis, polynomial regression, artificial neural network
Subjects:	T Technology > TH Building construction
Divisions:	School of Graduate Studies
Depositing User:	Muhamad Aizat Nazmi Mohd Nor Hamin
Date Deposited:	09 Apr 2026 08:28
Last Modified:	09 Apr 2026 08:28
URI:	http://ur.aeu.edu.my/id/eprint/1462

Actions (login required)

View Item