Research On The Building Heat And Cool Load Short-time Prediction Method

At present,the operation of building heating and air-conditioning system is generally inefficient,and the rough regulation cannot meet the needs of building energy saving.Short-term load forecasting of heating and air-conditioning system is the basis of optimal operation of the system,and is an important means to improve system operation efficiency and reduce energy consumption.Load forecasting algorithms are multitudinous,while keep evolving.The performance and characteristics of different algorithms differ gre atly,as well as the application difficulty.So,it is necessary to compare short-term load forecasting algorithms from multiple aspects.This research takes the building block C of a Beijing public building as the research object,establishing a model which is in line with the actual building.By using the De ST to simulate the building cold and heat load,it is possible to use these simulation data to compare the performance of four algorithms: Random Forest,Support Vector Machine,Artificial Neural Network and XGBoost.The purpose of this study is to give the characteristics and application scenarios of each algorithm,and give some suggestions for the application of each algorithm.Firstly,the data source,training and parameter adjusting steps of the algorithm are explained,and the methods for evaluating the algorithms are determined.Then,the requirements of each algorithm for input data are analyzed,and the combination of input features suitable for each algorithm are given.At the same time,it can be concluded that the neural network has a high demand for input data dimension,while SVM has the opposite effect.Under the condition of rich input dimension,the Coefficient of Determination of XGBoost can reach 0.96.By constructing a variety of defective data,the robustness of the four algorithms is compared.Random Forest has the strongest robustness among the four algorithms,and it has outstanding resistance when facing data quality defects.Compared with the four algorithms,neural network is very sensitive to data quality.In the comparison of data set scale,the experiment shows that the performance of support vector method is very good when the data set time span is small(one month).At the same time,the data set time span suitable for short-term load forecasting is at least one month,and the ideal time span is two months.For the neural network algorithm,it needs at least one-year-long data.In this paper,the parameters tuning of each algorithm is analyzed,and the search range and method of parameters recommended for load forecasting are given.Meanwhile,the tuning complexity of each algorithm is obtained: Random Forest and XGBoost tuning is simpler,support vector method is complicated,and neural network is extremely sophisticated.Combined with the requirement and adaptability of the algorithm for input data,the comprehensive performance evaluation and recommended application scenarios of the four algorithms are given.