Model Predictive Control for Building Energy Systems : A Comparative Study of Battery Integration and Machine Learning Forecasting in Virtual Power Plants

Sammanfattning

As buildings become increasingly active participants in modern power systems, leveraging distributed energy resources (DERs) to support grid stability and sustainability has become a key challenge. In this context, this study investigates the role of battery storage in enhancing building energy system performance under Model Predictive Control (MPC) within a virtual power plant (VPP) framework. Because MPC relies on future load information to optimize control decisions over a receding horizon, this study evaluates the impact of three forecasting models on control outcomes: Multiple Linear Regression (MLR), Long Short-Term Memory (LSTM) networks, and the Light Gradient Boosting Machine (LightGBM). These models are integrated into an MPC framework implemented in the CityLearn simulation environment, which manages energy dispatch for a community-scale building cluster equipped with rooftop solar panels, HVAC systems, and optional battery storage. Experimental results show that battery integration, when guided by MPC using machine learning forecasts, significantly improves system performance. Compared to a baseline without battery storage, the LightGBM-MPC strategy reduces electricity costs by 35%, carbon emissions by 17%, and improves grid efficiency by 13%. These gains are achieved by strategically scheduling battery operations in response to dynamic electricity prices and carbon intensity signals, effectively flattening the net load profile and enhancing renewable energy utilization. To understand the influence of forecast accuracy on control effectiveness, the study compares three machine learning models: MLR, LSTM, and LightGBM for 24-hour building load prediction. LightGBM consistently delivers the most accurate forecasts and enables the best control outcomes across all key performance indicators. These findings highlight the importance of high-quality predictions in predictive control and demonstrate the potential of integrating machine learning with MPC for intelligent, adaptive building energy management. Overall, the results demonstrate a practical pathway toward data-driven, grid-interactive control strategies for sustainable urban energy systems.