Dynamic KNX-based energy management system with machine learning for energy optimization

Efficient energy management in residential buildings is essential for reducing consumption and supporting sustainability initiatives. Conventional Energy Management Systems (EMS) often rely on static control strategies that do not reflect dynamic occupancy patterns, limiting their capacity to optimize performance. This study presents a dynamic KNX-based EMS that integrates real-time occupancy sensing with machine learning models to support forecast-assisted energy management of building energy systems. A residential building in Jeddah, Saudi Arabia, equipped with a 200-kW photovoltaic array and a 100-kW inverter, was simulated using HOMER Pro. High-resolution data from lighting, HVAC, and occupancy sensors were analyzed separately and compared after hourly aggregation to characterize building energy demand. The Solace dataset was then separated into 70% training and 30% testing subsets for ML model development and validation. Four predictive models: Support Vector Regression (SVR), Random Forest (RF), Gradient Boosting (GBM), and Multilayer Perceptron (MLP) were trained and tuned using Bayesian optimization and five-fold cross-validation. The integration of KNX occupancy controls resulted in an average hourly energy reduction of 15.33% over the 41-day monitoring period relative to the HOMER baseline, while the SVR-predicted annual energy consumption was 313,239.12 kWh, corresponding to a difference of 61,798.88 kWh (16.48%) relative to the HOMER baseline. Among the evaluated models, RF achieved the highest large-scale accuracy, while SVR offered the best balance between accuracy and computational cost at the moderate data scale. These findings suggest that integrating dynamic KNX automation with machine learning supports forecast-assisted, data-driven energy management of residential energy systems, providing a scalable strategy for enhancing energy efficiency in smart buildings.