| In order to meet the high requirements of high-efficiency,flexible and reliable operation of building energy system with high penetration of renewable energy,decentralized supply and multi energy coordination and complementarity for building load forecasting in the future,data-driven building load forecasting model has been widely valued and studied by virtue of its advantages of high precision,high computational efficiency and low cost.In engineering practice,however,in engineering practice,the data-driven model is difficult to perform in many scenarios with incomplete basic data,which usually shows weak fitting ability when the data of feature variables is incomplete and weak generalization ability when the data of training samples is incomplete.In the field of building energy,domain knowledge(such as physical analytical model)is rich,and the domain knowledge proved by theory or practice usually has strong universality and generalization.If domain knowledge can be integrated into the data-driven modeling,the robustness,accuracy and applicability of data model will be enhanced.From the perspective of the development trend of artificial intelligence technology,the previous artificial intelligence technology based on data,algorithm and computing power has shown the problem of insufficient processing capacity in many application scenarios.In the future,artificial intelligence(AI)technology will move towards a new generation of AI era with knowledge,data,algorithm and computing power.In this context,this paper takes the short-term building thermal load forecasting modeling as the research object,aiming to establish a dataknowledge fusion driven forecasting modeling method based on data-driven modeling and aided by domain knowledge.The main research work is as follows:(1)Data completeness analysis of feature variables in data-driven building thermal load forecasting model.The completeness of feature variables greatly affects the fitting error of data-driven model.It is the basis of developing accurate data-driven model to clearly understand the demand of data-driven model for the completeness of feature variables.This paper presents a method to determine the optimal set of feature variables by using the maximum information coefficient and the maximum correlation minimum redundancy multivariable selection strategy.The results of case study show that,compared with the common filtering feature variable selection method,this proposed method can effectively eliminate the redundancy of feature variable set on the premise of ensuring that the important variables are not missing,and ensure to obtain the sufficient and necessary feature variable set of the data-driven model.It provides a method basis for analyzing the complete feature variable set of data-driven model for building thermal load forecasting.The results of complete feature variable set analysis of six typical types of buildings by using this method can provide some data collection guidance for the development of building load forecasting model in engineering practice.(2)The inverse disaggregation of internal disturbance feature based on calibration of simulation model.The lack of internal disturbance feature variable is a common feature variable incomplete situation in engineering practice.Therefore,this paper proposes a non-invasive method to obtain the internal disturbance characteristic variable to enhance the completeness of the feature variable set.In this method,the internal disturbance features are extracted from the total load sequence by using the simulation model calibration.The results show that the artificial internal disturbance feature variables obtained by this method can effectively increase the information completeness of the feature variable set,and reduce the fitting error of ANN model from 21.71%(when the internal disturbance data is completely missing)to 10.25%.Compared with the data-driven model with simple feature variable set using calendar information,it has higher prediction accuracy(10.25% vs 16.76%).Although the accuracy of this method is not as good as that of the internal disturbance feature variables obtained by accurate monitoring,this method achieves the balance between the information richness and the difficulty of data monitoring by means of non-invasive data acquisition,and is an effective substitution method of internal disturbance feature variables.(3)Data completeness analysis of training samples in data-driven building thermal load forecasting model.The completeness of the training sample set used in the datadriven prediction model directly determines the generalization performance of the model.Based on the computational learning theory of data-driven predictive modeling,this paper summarizes the general requirements of data-driven model for the completeness of training samples,which are summarized as follows: Based on the samples in global sample space,based on the independent and identically distributed sampling method,and guaranteed by sufficient sample capacity.The research results on the specific problem of data-driven forecasting modeling of building thermal load show that in engineering practice,when a complete cooling / heating season data is accumulated with the system operation time,the training sample set can basically meet the requirements of the data-driven model for its completeness,and achieve the same forecasting effect as the model under the ideal condition of independent sampling in the global sample space.However,when the building load boundary conditions change in the future,the prediction effect of the data-driven model based on the complete training sample set will be affected,and the larger the load boundary changes,the worse the applicability of the data-driven model.It should be considered and handled as a incomplete training sample set.(4)Research on the method of data-driven model learning guided by domain knowledge.In engineering practice,training samples are often incomplete.Aiming at the problem that incomplete training samples lead to weak generalization ability of data-driven model,this paper proposes a technical framework of using domain knowledge to guide data-driven model learning,which is composed of two stages of domain knowledge optimization and knowledge transfer,so that the data-driven model can also achieve effective training and learning under the condition of limited training samples.This technical framework can not only take into account the robustness,reliability and strong generalization performance of domain knowledge,but also make use of training data to make the data-driven model retain the flexibility and difference of mining uncertain and unknown rules,which has certain theoretical advantages.The framework can effectively utilize two kinds of domain knowledge,one is similar building data,the other is physical analytical model.For the former,a proposed similarity measurement index(SMI)is used to select the optimal source task data.The actual case results show that using SMI can select the data which is most beneficial to the target building from 55 available source tasks for transfer learning.Compared with the transfer learning model which directly uses the limited sample data and without optimal selection of the source data,SMI achieves the best prediction effect,and the prediction error and model stability are higher than the other two.For the use of domain knowledge of physical analytical model,this paper improves the quality of virtual simulated sample data by using model calibration method,and then improves the prediction effect of transfer learning model.The actual case study results show that,in the comparison of four different sample data augmentation modeling strategies,the transfer learning strategy using calibrated simulation sample data achieves the best application effect.Therefore,the proposed framework of using domain knowledge to guide data-driven model learning in this paper can ensure that the data-driven model can still obtain excellent generalization ability under the condition of limited training samples. |
PDF Full Text Request