Application Of Energy Consumption Prediction Model In The Detection Of Abnormal Energy Consumption Of Water Chiller In The Air-conditioning System

Abnormal energy consumption detection of building air-conditioning systems has been a research hotspot in the field of building energy consumption analysis,while popular anomaly detection methods cannot trace the source of anomalies in the detection.This thesis,by using data-driven technology to complete the research of future value prediction and anomaly detection for energy consumption of chiller in the air conditioning system,builds prediction models based on input variables regression model and target variables autoregressive model,anomaly detection models based on statistical analysis and prediction models,aiming to solve the problems of energy consumption prediction and abnormal detection of building air-conditioning system.The thesis takes the water chiller of the air conditioning system of a hospital building as the research object,fully verifies the feasibility of the above energy consumption prediction and anomaly detection model,and finally,a complete and reliable abnormal energy consumption detection and anomaly traceability system is formed.The main research contents of this thesis are as follows:Aiming at research on energy consumption prediction of chillers in the air-conditioning system based on input variable regression,this thesis proposes a prediction model based on grid search optimization support vector regression(GSO-SVR).Firstly,support vector regression was used for fitting the hyperplane,and then grid search optimization algorithm under 10-fold cross-validation was used to optimize hyperplane parameters under support vector regression.Finally,the energy consumption prediction model of input variable regression is constructed by the optimized model.In order to verify the prediction performance of the model,two prediction models,negative feedback neural network(BPNN)and long and short memory neural network(LSTM)are introduced to compare and evaluate by the performance index.The results show that the prediction performance of the basic GSO-SVR is better than that of the unoptimized SVR,BPNN and LSTM.As for the research on energy consumption prediction of chillers in the air-conditioning system based on target variable autoregression,this thesis introduces the autoregressive moving average model(ARIMA),the multivariate autoregressive moving average model(ARIMAX)and three parameters exponential smoothing model for the prediction research.Firstly,the stationary test is carried out for the time-series data of energy consumption,then,the sliding window method is used to construct the single-step and multi-step prediction strategies of different models,finally,compares the performance evaluation indexes of the prediction results of models in different strategies.The results show that the prediction results of ARIMAX and three-parameter exponential smoothing model are better than the ARIMA model,and affected by the accumulation of forecast errors,the results of multi-step prediction of the three models are all worse than the single-step prediction results,while the three-parameter exponential smoothing model is the least affected.Aiming at the problems of poor accuracy of the current abnormal energy consumption detection methods for air-conditioning systems and the inability to trace the source of the abnormalities,this thesis proposes differentiated abnormal energy consumption detection strategies based on statistical analysis,abnormal energy consumption detection and traceability strategies based on predictive models respectively;For the former,energy consumption patterns are classified by centroid clustering,and then use support vector multiple classifiers to recognize the energy consumption patterns,finally,statistical anomaly analysis under different patterns is carried out to realize differential detection of abnormal energy consumption.Comparing the detection threshold obtained by the proposed strategy with the normally used detection method,the results show that the differentiated anomaly detection strategy has greater potential for anomaly detection.In addition,comparing the detection thresholds of different statistical analysis methods,the results show that although the box-plot analysis has greater potential to detect anomalies during periods of steady energy consumption than the 3-sigma,it is less reliable than 3-sigma during periods of drastic changes in energy consumption;As for the abnormal energy consumption detection and traceability strategies based on predictive models,this thesis completes the anomaly detection based on the GSO-SVR prediction model and the Holt-Winters three-parameter exponential smoothing model,anomalies detected by these two methods correspond to the abnormal operation of equipment and abnormal energy consumption behaviors respectively,finally combines these two detection results,the cause of abnormal energy consumption can be found out.This thesis introduces a variety of data mining algorithms to complete the research on energy consumption prediction and abnormal energy consumption detection of water chiller of the air conditioning system.It enriches the research of non-stationary time series analysis in building air conditioning energy consumption prediction,optimizes the application of statistical analysis method in abnormal detection of building air conditioning energy consumption,and realizes abnormal energy consumption detection and anomaly traceability at the same time.