Loading...
A Dynamic KNX-Based Energy Management System Utilizing Machine Learning and Probabilistic Models for Renewable Energy Optimization
Fatayerji, Hala
Fatayerji, Hala
Citations
Altmetric:
Type
Supervisor
Date
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This thesis presents a comprehensive analysis and optimization of an intelligent Energy Management System (EMS) specifically developed for residential buildings, integrating real-time occupancy-driven controls and advanced machine learning (ML) predictive modeling. Initially, a baseline energy scenario was established using HOMER Pro, simulating a grid-connected hybrid renewable energy system featuring a 200 kW photovoltaic (PV) array and a 100 kW inverter for a mid-rise residential building located in Jeddah, Saudi Arabia. Comparative analysis revealed significant hourly energy savings of approximately 15.33% through the implementation of occupancy-responsive automation based on KNX control systems.
To further refine energy predictions and optimization, predictive modeling using Support Vector Regression (SVR) and alternative ML techniques was conducted utilizing the high-resolution Solace dataset. Bayesian hyperparameter optimization was systematically applied, enhancing the predictive reliability of SVR. At a moderate dataset scale (10,000 sampling points), SVR achieved robust predictive with an RMSE of 0.0781 kWh and R² of 0.9195) and practical computational efficiency. However, at a larger scale (100,000 sampling points), SVR accuracy declined notably with an RMSE of 0.0904 kWh and R² of 0.8854, highlighting an inherent performance trade-off related to dataset size and computational demand. In contrast, Random Forest maintained superior accuracy with an RMSE of 0.0702 kWh and R² of 0.9310 at large scale, albeit with significantly increased computational overhead.
The SVR predictive optimization (at moderate scale) further enhanced annual energy savings by approximately 16.48% compared to the baseline scenario, totaling approximately 61,799 kWh/year. These findings validate the clear economic, operational, and environmental benefits of incorporating predictive machine learning into intelligent occupancy-driven EMS frameworks, emphasizing the importance of strategically balancing predictive accuracy with computational resources and dataset granularity.
Department
Publisher
Sponsor
Copyright
Attribution-ShareAlike 4.0 International