Adaptive model predictive control with ensembled multi-time scale deep-learning models for smart control of natural ventilation

Model predictive control (MPC) offers promising solutionsfor the smart control of natural ventilation. However, challenges arise in constructing precise models for such nonlinear systems, preserving accuracy over a long-prediction horizon while capturing short-term dynamics, and handling disturbances uncertainties. This study proposes a novel Ensembled Multi-time scale deep-learning-based Adaptive Model Predictive Control (EMA-MPC) system. It integrates an ensembled Long-Short-Term Memory (ensembled-LSTM) model, comprising two LSTM models for fast and slow time-scale dynamics, respectively, and continuously evolving to changing conditions through online model adaptation. A bound control module is incorporated as an additional safety mechanism ensuring the environmental control within the desired threshold during unforeseen scenarios. A multi-objective optimization problem is formulated to maintain indoor air temperature and CO2 concentration within the predefined comfort range while optimizing energy efficiency by controlling automated windows in a naturally ventilated room in winter. The EMA-MPC system demonstrates superior performance in balancing indoor air quality, temperature regulation, and energy efficiency. The ensembled-LSTM model significantly reduces the mean absolute error by 78.6% and 88.9% for CO2 and indoor air temperature predictions respectively, against a single LSTM model. The proposed EMA-MPC system achieves an 86% reduction in unmet CO2 hours compared to rule-based control and reduces occupied hours with temperature below 19 °C by 44% and 93% compared to enhanced MPC and basic MPC, respectively, while maintaining similar heating demand as other controllers. In conclusion, the proposed EMA-MPC system reduces modeling efforts and provides an effective approach towards reliable use of ML models in smart building control.