Lightweight approximate model predictive control for grid-responsive building air conditioning using quantized neural networks

We propose a lightweight approximate model predictive control (MPC) framework for grid-responsive building air conditioning systems, replacing computationally expensive high-fidelity solvers with quantized neural networks (QNNs) to enable real-time decision-making under grid constraints. The system integrates an offline-trained QNN that approximates optimal heating ventilation and air conditioning (HVAC) control actions from building states and grid signals, achieving a 4–8× reduction in computational overhead through 8-bit quantization while maintaining near-optimal performance. A transformer-based architecture captures temporal dependencies in grid signals and thermal dynamics, with QNN outputs discretized into setpoint adjustments compatible with legacy HVAC controllers. Additionally, a low-order RC network validates QNN predictions at runtime, ensuring stability by reverting to a conservative MPC policy when deviations exceed predefined thresholds. The method uniquely combines quantization-aware training with hybrid data-driven and physics-based validation, addressing the trade-off between computational efficiency and control accuracy. Implemented on edge devices with sub-10 ms inference latency, the system supports dynamic grid interactions like demand response and real-time pricing without compromising thermal comfort. Experimental results show up to 15% energy cost reduction compared to rule-based controllers, with robustness against grid disturbances. This work advances practical MPC deployment in building energy management by bridging the gap between theoretical optimality and real-world computational constraints.