An XGBoost-Based predictive control strategy for HVAC systems in providing day-ahead demand response

Abstract

Heating, ventilation and air-conditioning systems (HVAC), at demand side, are regarded as promising candidates to provide demand response to smart power grids by adjusting indoor temperature setpoint. However, the optimal control of indoor temperature setpoint with the constraint of indoor environment is a key issue. In this study, an XGBoost-based predictive control strategy is proposed for HVAC systems in providing day-ahead demand response. XGBoost (Extreme Gradient Boosting) models are developed to predict the variation of power use and indoor temperature. Several innovations are implemented to improve the performance of the control strategy, including dedicated generation of the initial population of genetic algorithm (GA) in rolling optimization, the use of a step-decreasing control horizon, and the separation of the sampling period and control action period. The strategy is validated on a TRNSYS-Python co-simulation test platform. Results show that the developed XGBoost models can achieve a high prediction performance of power use (CV-RMSE of 4.52%) and indoor temperature (CV-RMSE of 0.40%) in the coming 10 min with only using one-week historical data. The computation cost is reduced by 78% due to the dedicated generation of GA initial population while the control performance is not affected. Compared with the traditional operation, the proposed control strategy can reduce the total cost by as much as 27.30% and 33.6% while the weighted indoor temperature only increases by 0.14 K and 0.62 K, respectively. In test cases, the highest Predicted Mean Vote (PMV) is 0.72 and the corresponding Predicted Percentage of Dissatisfied (PPD) is 15.9%.