Time Series Optimal Control Of HVAC Systems

In China,building energy consumption has accounted for more than 25% of the total social energy consumption and it is on the rise.The energy consumption of heating,ventilation and air conditioning(HVAC)system accounts for about 50% of the total building energy consumption.By optimizing the control strategy of HVAC systems,the energy efficiency level can be improved and the development goal of carbon peak and carbon neutralization in China can be achieved.Based on machine learning,dynamic programming and reinforcement learning,this research proposes methods for utilizing historical data and instantaneous monitoring data,respectively,and then applies the time series data obtained by the building automation system to the optimal control of HVAC systems.The research results can provide guidance for the operation of HVAC systems,which is important for energy saving in the whole life cycle of buildings.The main contents include the following four parts.(1)Taking a near-zero energy residential building as the research object,the simulation system was built based on TRNSYS for the development and debugging of relevant control algorithms.Based on the TRNSYS simulation system,the simulation of the building and air conditioning equipment was carried out to study the operating parameters and operating energy efficiency of the HVAC system under different working conditions,and the control strategy that makes the HVAC system have the highest energy efficiency was selected as the baseline strategy.(2)In this research,a data-driven generic model identification method is developed.The method performs anomalous data identification and processing with the K-Means algorithm,dimensionality reduction with Pearson correlation coefficient and random forest algorithm in the data partitioning stage.The method uses the coefficient of determination and mean square error as evaluation metrics and conducts competitive learning with four machine learning techniques: multivariate linear regression,support vector regression,artificial neural network and XGboost in the modeling stage.The developed model identification method was applied to establish a load prediction model for a near-zero energy office building and a solar irradiance prediction model in Chaoyang District,Beijing,and the reliability of the method was verified by comparing the model prediction data with the actual measurement data.(3)The number of sensors and devices in the HVAC system is large,and it is difficult for the traditional optimization methods to obtain the optimal control strategy in time.This research proposed a dynamic programming-based HVAC system control method to address this problem.The method uses a model to predict the future state of the system and uses dynamic programming to obtain the HVAC system settings that result in the lowest overall cost.This method allows for the combined control of heat pumps,fresh air units and shading devices.To address the problem that traditional optimization methods are highly dependent on building models and are not able to self-adapt to the change of controlled environment,a reinforcement learning-based HVAC system optimization method is proposed,which does not require models and uses building real-time monitoring data to continuously iterate and update to obtain control strategies.(4)After the control strategies obtained from pre-training in the simulated environment are deployed to the real building,the performance needs to be verified.The actual validation results of a nearly zero energy residential building in Beijing show that compared to the baseline strategy,reinforcement learning and dynamic programming result in a 4.5 % and15.3 % reduction of the overall cost in typical winter conditions,respectively.Dynamic programming calculates the optimal control strategy based on the model,and the optimization effect in the simulated environment and the actual building is similar.Reinforcement learning requires continuous training using the building real-time monitoring data to get a new control strategy,which increases the time cost in the actual implementation process,so the optimization effect of reinforcement learning in the actual building differs greatly from that in the simulated environment.