Research On Forecasting Method And Uncertainty Analysis Of Building Energy Consumption Based On Time Series

The power demand of the end-users is the main component of the building energy systems,whose accurate prediction is essential to the intelligent and efficient operation of the building energy system.Previous studies have mainly focused on deterministic power load forecasting.However,with the increasing complexity and intelligence of equipment systems,various uncertainties of building use and working conditions as well as weather conditions often lead to significant fluctuations of electricity load in the demand side,which brings great challenges to electric load forecasting.Therefore,it is necessary to explore a building energy consumption interval prediction method considering the uncertainty of input parameters to improve the forecasting performance of the model.Deterministic electric load forecasting is the basis for interval prediction of building energy consumption.In view of the urgent need to improve the accuracy and interpretability of electric load forecasting,this study has developed an ensemble algorithm combining a seasonal autoregressive integrated moving average（SARIMA）model with long-term and short-term memory network（LSTM）model,whose weight factors are dynamically updated using the least square method.A four-year energy consumption data of an office building in Hunan Province is used as a case study to train the model,and the short-term electric load forecasting results of SARIMA model,LSTM model,random forest（RF）model,support vector regression（SVR）model and integrated convolutional neural network and long short-term memory network（CNN-LSTM）model are analyzed.Compared with the optimal prediction model（CNN-LSTM）,the RMSE and MAPE values of the integrated SARIMA-LSTM model are reduced by 16.1%and 13.1%respectively,which demonstrates the integrated model developed has higher accuracy in this study.Furthermore,this study has adopted the improved Boruta algorithm with a mixing ratio of 0.5 for feature extraction,which not only reduces the complexity of shadow feature samples,but also extracts feature sets related to building energy consumption more efficiently.Meteorological prediction parameters are the main uncertain variables among the input characteristics of deterministic electric load forecasting model.In order to further explore the influence of the volatility of meteorological parameters and the uncertainty of prediction errors on the prediction accuracy,this study has explored a novel method for short-term building energy consumption interval forecasting,which converted deterministic load forecasting into interval prediction.The nonparametric kernel density estimation method is adopted to establish the probability distribution function of prediction error for temperature and relative humidity firstly.After that,the Monte-Carlo simulation method is introduced to obtain the prediction interval under different confidence levels.Furthermore,the predicted interval coverage probability（PICP）and average bandwidth(₈))are two of the criterions used to evaluate the uncertainty of the proposed model.It is indicated that with the increase of the confidence level,the range of the building energy consumption prediction interval gradually increases.Under different conditions,the actual power value of the building energy consumption is within the prediction interval with a probability greater than the confidence level.At the 90%confidence level,only 9.67%of the measured hourly electric loads lie outside the confidence interval.Finally,A short-term electric load forecasting interval prediction method considering the uncertainty of meteorological parameters,and another interval estimation method based on the distribution characteristics of real-time power prediction error have been compared to evaluate the prediction performance according to the uncertainty evaluation index.The comparison shows that at the same confidence level,the interval coverage ratios（PICP）of the two interval forecasting methods are roughly the same,but the interval average bandwidth(₈))of the former prediction interval estimation method is reduced by 35.4%,35.9%and 68.3%in summer,winter and transition season respectively,which indicates that the energy consumption prediction interval model proposed considering the uncertainty of meteorological parameters in this study has superior predictive performance.The deterministic point prediction is extended to the interval prediction to describe the distribution characteristics of power load at the next moment through the prediction interval estimation method,which can improve the accuracy and credibility of the short-term electric load prediction.The optimal scheduling of building energy systems based on the prediction interval can reduce the decision-making risk result from the large prediction error of the deterministic prediction model,thereby improving the reliability and accuracy of scheduling decisions.