Heating Load Prediction For District Heating Station Based On Transfer Learning

District heating is the main means to maintain the indoor thermal comfort in northern China in winter,but the current common real-time online regulation can not achieve the goal of on-demand heating,and the system operation efficiency is low,resulting in considerable energy waste.Accurate prediction of heating load can change the real-time online regulation mode into feed-forward regulation mode,provide support for the heating system on-demand regulation,improve the supply and demand ratio of heating energy,and reduce the heating energy consumption,which is the key to realize the refined heating management and on-demand heating.However,existing heating load prediction models have low generalization ability,and the prediction model can not be established or the accuracy of the model is poor when the training data quantity is small.Therefore,this paper adopted timeseries data processing,transfer learning,through extracting the effective knowledge contained in the data of the district heating station,established the heating load prediction model,which provided the theoretical basis for the model optimization of the cross heating season,the heating load prediction of the new district heating station,and the depth optimization of the heating load prediction model.The main research contents and achievements are as follows:(1)In addition to the traditional data preprocessing methods to clean,integrate and transform the original data,the original data was processed by timeseries processing method.Through the refinement of data timestamps to construct time features,using sliding windows and building thermal inertness coefficient based on professional knowledge as new input features to characterize the lag of building heat transfer.By timeseries processing of the original data before data input to the prediction model,the performance of the prediction model was improved.(2)In view of the problem that the prediction accuracy of the original prediction model decreases when it is used in the new heating season due to the change of the heating scenario after entering the new heating season,the layer transfer model and the fusion transfer model proposed in this paper can effectively optimize the original model by transfer learning based on a small amount of data in the new heating season,and can improve the prediction accuracy of the model after transfer learning in the new heating season to that of the original model in the historical heating season.The performance of the layer transfer model in the early stage of the new heating season is better than that of the fusion transfer model,which is more consistent with the prediction scenario of the cross heating season.(3)When there is no historical data for the new district heating station or the data quantity of the target district heating station is insufficient,the layer transfer model and fusion transfer model proposed in this paper can effectively learn and transfer the knowledge of other district heating stations.When the performance of the heating load prediction model based on a small amount of existing data is poor,the prediction accuracy of the model can be greatly improved by transfer learning;when the performance of the model is good,the model can still be further optimized in depth to improve the prediction accuracy.In addition,the fusion transfer model had good robustness in the case of insufficient data,and its prediction accuracy changes little when transferring knowledge from different transfer sources,which reduces the difficulty of selecting suitable transfer sources before transfer learning,and is more suitable for application in cross site transfer scenario.(4)In order to solve the problem of depth optimization of heating load prediction model when the data of multiple district heating stations are sufficient,this paper proposed a multi-task learning model.Results showed that simple combining data of other district heating stations in the original model can improve the prediction performance,but it is not stable and easy to produce negative transfer learning,which leads to the decline of the accuracy of the original model.The multi-task learning model can achieve the goal of depth optimization of the original model by mining the rich association information between different district heating stations.The CV-RMSE value of the original model of the three heating stations can be reduced by 2.50%,7.06%and 3.52%,respectively,which account for 20.74%,41.89% and 56.10%.Moreover,the multi-task learning model had good robustness,which can improve the prediction accuracy of the original model of each district heating station participating in the-multi task learning at the same time,and provide the basis for the depth optimization of the prediction model,the construction of the high-precision prediction model and the better realization of the precise adjustment and on-demand control of the district heating station.