Abstract
This study explores the feasibility of model-based predictive control (MPC) with a machine learning (ML) approach in commercial buildings. Grey-box and ML-based models were developed using experimental data from a test facility. Three different models were considered for ML-based model development: artificial neural network, gauss process regression, and support vector regression. In MPC simulations, the optimal solution for the grey-box model was achieved by applying linear programming, assuming a linear time-invariant model. On the other hand, the proposed ML-based method utilized predefined setpoint trajectories to achieve cost savings by load shifting from on-to off-peak. The estimated trajectory yielding the minimum cost was identified as the optimal trajectory, which was then input to the grey-box model to ensure a fair comparison of the performance of both MPCs against that of optimal feedback control. An average saving performance of 27.5 % and 23.7 % was achieved using the MPC with grey-box and ML approach over optimal feedback control. Near-optimal performance was achieved with ML approach without running the optimization. The comparable performance of the proposed method implies that the engineering cost in a typical MPC using a grey-box model can be significantly reduced by using the ML method with minimal engineering, which is easy to implement and scalable to other buildings.
| Original language | English |
|---|---|
| Article number | 110831 |
| Journal | Engineering Applications of Artificial Intelligence |
| Volume | 151 |
| DOIs | |
| State | Published - Jul 1 2025 |
Funding
This work was supported by the National Research Foundation of Korea (NRF) through a grant funded by the Korea Government (MSIT) (RS-2023-00253333). This manuscript has been co-authored by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the US Department of Energy (DOE). The US government retains and the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for US government purposes. DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan ( http://energy.gov/downloads/doe-public-access-plan ). Numerous AI studies on building performance focusing on modeling (Choi and Lee, 2023; Liu et al., 2025; Lu et al., 2023; Schreiber et al., 2021; Zhang et al., 2023) and control (Deng et al., 2023; Gao et al., 2023; Yang et al., 2023; Zsembinszki et al., 2023) have been presented. In up-to-date studies of MPC, Chen et al. (2023) presented the MPC framework to optimize air conditioning system performance. Energy consumption prediction models were developed by various data-driven techniques. Further, the optimization performance of each model was obtained by using different non-linear optimization algorithms. The optimal match (support vector regression (SVR) and particle swarm optimization (PSO)) achieved a notable energy saving of 7.1 % for the air conditioning system. A predictive model utilizing long short-term memory was developed to forecast the energy consumption, and an MPC simulation was derived using PSO to minimize the grid electricity usage for the following day (Jeon and Kim, 2021). The results showed a reduction of 30 % in comparison to the reference model. The MPC framework was implemented on nearly zero-energy building to reduce the heating energy cost (Aruta et al., 2023). Results show a substantial reduction of 26 % in energy consumption when compared to the reference strategy, which maintains a setpoint temperature of 21 \u00B0C. An artificial neural network (ANN) model was combined with a genetic algorithm to minimize overall energy usage (Reynolds et al., 2018). Hu et al. (2025) employs Flower Pollination Algorithm to solve the optimal chiller loading and parameter optimization problems to achieve energy saving in the HVAC system. Zheng et al. (2024) compared reduced-order model-based MPC with ANN-based MPC for utility cost reduction in residential buildings. The solution with reduced order MPC is based on linear programming and for ANN the Projected Stochastic Gradient Descent algorithm is employed. Energy savings between 17 % and 34 % were observed, with discomfort levels ranging from 30 % to 95 %. Sha et al. (2025), proposed an MPC strategy integrating encoder-decoder long short-term memory with multi-objective PSO to reduce the energy consumption of HVAC system. The total energy consumption was reduced by 10.2 % compared to the baseline control strategy. As described above, ML models predict the thermal load for the short or long-horizon leverage on a non-linear solver for control methods. This implies that a globally optimal solution cannot be guaranteed, and solving the optimization problem becomes computationally challenging due to nonlinearities, especially when the complexity of the model is high. Afram et al. (2017), employ a neural network as a representation of a plant model to design the MPC based on closed-loop optimization resulting in the reduction of operating cost by 6 %\u201373 %. Unfortunately, neural networks are non-linear, and MPC relying on such models also demonstrates non-linear behavior. Data-driven predictive control using random forest (Smarra et al., 2018) and regression tree (Jain et al., 2018) were employed for energy saving, demand response, and peak power reductions, considering linear models or switching linear models to setup quadratic and linear MPC depending on the choice of optimization. A novel approach to construct linear MPC with ML models was proposed, incorporating instantaneous linearization to handle non-linear models (Yang and Wan, 2022). However, high computational demand for solving optimization problems is involved, which hinders its implementation in actual building due to computing optimizer (e.g., PC). The optimizer needs to be computationally efficient because MPC iterates the optimization at every step. Determining the optimal control sequence is a major challenge in employing ML models for predictive control. One prospective approach relies on generating predefined set-point trajectories and evaluating the resultant energy/cost from the trajectories (Cotrufo et al., 2020; Zhu et al., 2025).This work was supported by the National Research Foundation of Korea (NRF) through a grant funded by the Korea Government (MSIT) (RS-2023-00253333). This manuscript has been co-authored by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the US Department of Energy (DOE). The US government retains and the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for US government purposes. DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-access-plan).
Keywords
- Artificial neural network
- Building automation and control system
- Cyber physical system
- Gauss process regression
- Grey-box model
- Model-based predictive control
- Multistep ahead prediction
- Support vector regression