Research On Detection Of VRF Air Conditioning System Refrigerant Charge Using Stacking Ensemble Learning

The HVAC system is an important part of the building’s electrical system and the main energy-consuming module.Maintaining its stable and efficient operation plays an important role in building comfort and energy saving.Aiming at the common refrigerant charge failures in air conditioning systems,this thesis proposes a VRF refrigerant charge detection model using stacking ensemble learning method.At the same time,the semi-supervised model is used as the base learner of the stacking integrated model,so that the model can use a large amount of unlabeled data generated in the actual operation of the air conditioner.This thesis verifies the various models in the research through a VRF refrigerant charge failure experiment.The experimental data includes three indoor and outdoor environmental conditions under heating conditions.10 charge levels with refrigerant charge range from 63%to 140%.The input data of the model is preprocessed,including data cleaning,removing variables that cannot be used by the data-driven model,and normalizing the data,using recursive feature elimination method for feature selection.Use correlation analysis methods to reduce the coupling between model input variables.Finally,10 variables were selected as the input of various subsequent detection models,and the trend of model input variables with the level of refrigerant charge was analyzed,and the reasons were explained.Established 5 simple classification models of refrigerant charge: SVM,BPNN,RF,GBM and MLR,optimized the parameters of each model.The results show that among the5 simple supervised classification models of refrigerant charge,2 models based on ensemble learning methods(RF and GBM)have the best fault diagnosis performance,and GBM has the largest classification accuracy rate(97.62%)in the training set,RF has the largest classification accuracy rate(95.37%)in the testing set,followed by the SVM model,and the least effective models are MLR and BPNN.The Stacking integration method is used to combine these several models.After structural optimization,there is the largest classification accuracy rate in the training set and the testing set(99.45% and 98.38%).In order to realize the application of Stacking ensemble learning model in actual scenarios,use the original data set to simulate the actual situation,discuss the limitations of the supervised algorithm SVM in the actual scene,it is found that as the amount of labeled data in the training set decreases,although the classification ability of the supervised SVM in the training set hardly changes,the classification accuracy rate in the testing set drops sharply.And propose 4 semi-supervised refrigerant charge detection models based on selftraining and co-training methods: Self-knn,Self-ksvm,Tri-training and Co-bc,instead of the supervised learning methods as the base learner to establish the VRF refrigerant charge stacking integrated detection model in actual scenarios.The classification accuracy of the model in the training set and the testing set can be improved to 97.28% and 93.91%.In summary,stacking multi-model integration can effectively improve the classification performance of the VRF refrigerant charge detection model.Semi-supervised base learners can also be used to make the model use a small amount of labeled data measured in the laboratory and a large amount of unlabeled data generated in the actual operation of the air conditioning system in the actual scene.It shows the practical applicability of Stacking ensemble learning refrigerant charge detection model.