Transferability Investigation Of Building Energy Prediction Modeling Data And Control Strategy Of HVAC System

With the continuous advancement of sustainable urbanization and urban renewal process,the demand for building energy conservation and emission reduction is increasing.Reducing the energy consumption of building and HVAC systems,and improving energy utilization efficiency has become a hot research topic in the building energy field.Building energy consumption prediction modeling and HVAC system optimal control are the key foundations for optimizing the operation of building energy systems.The current building energy consumption prediction model and control strategy of HVAC systems have poor expansibility,which requires repeated modeling and training,greatly increasing the modeling cost and training time.This paper constructs an overall application framework of building energy consumption prediction model and HVAC system control strategy based on deep transfer learning methods,and systematically conduct research on the transferability of building energy consumption prediction model and HVAC system control strategy,and provides corresponding solutions for the application of deep transfer learning methods in the building energy field.The main research contents are as follows:A statistic informed neural network model(SINN),which combines statistic prior knowledge with deep neural network,is proposed for building energy consumption prediction modeling.Three different experiment scenarios are conducted to systematically investigate the comprehensive performance of SINN model from several dimensions,including model prediction accuracy,computation cost,model robustness and optimal critical training data volume.The results show that the proposed SINN model can effectively improve the prediction accuracy and model robustness,and reduce the computation cost and training data volume.Compared with the traditional neural network model,the MAPE performance of ARA-SINN model and EWA-SINN model increases by 11.90%～55.23% and 22.04%～55.58%,respectively.The optimal critical training data volume of the five traditional neural network models are 4-week,12-week,8-week,12-week and 4-week,respectively.Except for the EWA-SIFCNN model,the optimal critical training data volume of all other models for ARA-SINN and EWA-SINN is 2-week.That is the ARA-SINN model and EWA-SINN model can obtain higher model prediction accuracy with smaller training data volume.A deep transfer learning model based on LSTM-DANN is proposed for short-term cross-building energy consumption prediction.By conducting a series of experiment scenarios including similar buildings,different types of buildings,different prediction time intervals,and different network structures,the influences of the above factors on the prediction performance of the transfer learning model are systematically analyzed.The results show that LSTM-DANN deep transfer learning model can significantly improve the prediction performance of building energy consumption compared with the LSTM model which only uses target domain building data,source domain building data,target and source domain integrated building data.Compared with the other three models,MAE,MSE,MAPE and CV-RMSE all improve by more than 15%.The transfer learning method can overcome the domain shift of different data distribution.The performance of three neural network feature extractors,FC,LSTM and CNN,is also compared and analyzed.The LSTM-DANN model has short prediction time and the best prediction performance for building energy consumption prediction.A Sim2 Real deep transfer learning model is proposed to improve the performance of building energy consumption prediction in the target domain by using simulation building data.Three different experiment scenarios are conducted to analyze the influence of different transfer learning algorithms,building types,climate zones and training data volume on the prediction performance of Sim2 Real transfer learning models.Results show that R-DANN based Sim2 Real transfer learning model has the best performance.The Sim2 Real transfer learning models based on Finetune,R-DAN and R-Deep CORAL have a higher risk of negative transfer when the building types and climate zones of source and target building domains are different.Compared with the traditional LSTM model,when Sim2 Real transfer learning model is transferred between different building types,MAPE increasing percentage of R-DANN and RDAN models is 12.73%-29.05% and 2.34%-24.88%,respectively.When building data volume in source domains of different climate zones are used to assist building prediction modeling in target domains,the MAPE of Sim2 Real transfer learning model based on R-DANN increase by 1.06%～18.50%.When the amount of available data in the target domain is large enough,the traditional LSTM model can obtain better prediction performance without using transfer learning method.A general multi-source ensemble transfer learning(Multi-LSTM-DANN)model for building energy consumption prediction is proposed.By conducting experiment scenarios with different building combinations and number of buildings in different source domains to analyze their effects on the performance of the multi-source ensemble transfer learning model.Results show that compared with the model trained only with the target domain data,both the single-source and multi-source ensemble transfer learning models can improve the prediction performance of building energy consumption in the target domain.The MAE,MSE,MAPE and CV-RMSE performance metrics of all transfer learning models increase by more than 15%,and compared with the corresponding single-source transfer learning model,the performance of the multisource transfer learning model improves more obviously.The building similarity of the Multi-LSTM-DANN model also has a great influence on the prediction performance.When using the multi-source transfer learning model to select the buildings in the source domain,while ensuring the similarity between the source domain and target domain,it is necessary to maintain the differences of the buildings in different source domains,which can further improve the prediction performance of the multi-source ensemble transfer learning model.Different combinations of buildings and the number of buildings in source domains have a great influence on the performance improvement of building energy consumption for the target domains.There is an optimal combination of source domains and the number of source domains for the performance improvement of multi-source transfer learning model.The model prediction error decreases when the number of buildings in the source domain increases from 1 to 3.The model prediction error increases when the number of buildings in the source domain increases to 4.A multi-objective optimization control strategy based on DQN is designed for dynamic optimal control of temperature setpoint to balance the energy consumption of HVAC system and indoor air temperature deviation.Taking a variable air volume system of a multi-zone office building as an example for analysis,the optimal control performance of supply air temperature and chilled water temperature setpoint based on DQN strategy is analyzed and studied.An Energy Plus-Python co-simulation experiment platform is developed to train and test the performance of the DQN control strategy.The results show that the multi-objective optimal control strategy based on DQN can effectively determine the appropriate setpoint of supply air temperature and chilled water supply temperature,and achieve the balance between energy consumption and indoor air temperature.The DQN control strategy is compared with 14 control strategies with fixed temperature setpoint.The DQN control strategy is more energy saving than 11 control schemes with fixed temperature setpoint.At the same time,the indoor temperature can be maintained near the setpoint,and the maximum energy saving rate of the system can reach about 8%.The DQN control strategy can determine the optimal temperature setpoint sequence after 10 training epoches,and the temperature setpoint can always be relatively stable.A integrated transfer framework based on transfer learning and deep reinforcement learning(TL-DRL)is proposed to achieve the transfer of DRL control strategies in the building HVAC system level.By conducting simulation experiments of three different scenarios,the transfer performance of DQN control strategy of building HVAC system is analyzed.The results show that the TL-DRL model can effectively improve the training efficiency of DRL in the building HVAC system level.The performance of the TL-DRL model is sensitive to the number of DQN network transfer layers,and the transfer of different DQN network layers has a great impact on the system performance.The stability,DQN training efficiency and control performance of the transfer of different layers are comprehensively analyzed.The TL-2 model has the best comprehensive transfer performance.DQN training time decreases with the increase of network transfer layers.The training efficiency of TL-DRL model is about 3%～29% higher than that of DRL benchmark model trained from scratch.In three different transfer learning scenarios,the TL-2 model can improve the training efficiency of the control strategy by 13.28%,while maintaining the system energy consumption and indoor air temperature in the optimal range.The building climate zone has a great influence on the transfer performance of the control strategy of the HVAC system.The stability and temperature deviation of the transfer performance of the control strategy in the same climate zone are significantly better than the transfer performance in different climate zones.