Neural network-based model predictive control for HVAC and ice storage in a high-fidelity virtual testbed

Model predictive control (MPC) has been widely recognized for its potential to improve the energy efficiency and flexibility of building heating, ventilation, and air-conditioning (HVAC) systems with thermal energy storage (TES). However, a practical deployment remains hindered by the complexity of high-fidelity modeling, computational burden, and the lack of reproducible testbeds that integrate TES and advanced controls. This paper presents a high-fidelity commercial building HVAC–TES virtual testbed. This testbed’s functionalities were demonstrated through an MPC case study based on surrogate artificial neural network (ANN) models. The virtual testbed was developed in Modelica and represented a medium-size office building served by a chiller and ice storage tank, with the TES subsystem validated against experimental data. ANN models were trained in Tensorflow/Keras to predict system power consumption, zonal air temperatures’ dynamics, and TES state-of-charge, and were implemented in CasADi, a symbolic framework for optimal control. The resulting nonlinear, integer optimization problem was solved using an evolutionary algorithm to handle the non-convexity of ANN surrogates. Simulation studies across multiple TES capacities and prediction horizons demonstrated that MPC effectively leveraged TES to shift loads away from high-price periods. For the most cost-effective configuration in this case study, MPC with a 20-h prediction horizon reduced energy costs by 25.2%, peak demand by 41.9%, and improved the energy flexibility factor by 154% relative to a storage-priority rule-based baseline over two testing days in the cooling season. The open-source framework is presented as a research prototype to evaluate ANN-based MPC for one office building. With additional work on controller interfaces and documentation, it is anticipated that this open-source virtual testbed would form the basis for future reproducible and extensible benchmarking studies of HVAC–TES control strategies.