Research On Heat Load Prediction Of District Heating Systems Based On Data-driven Models

District heating system is an important urban energy facility.Short-term load forecasting is of guiding significance for the district heating system to adjust the heat supply in advance.Accordingly,the research object in this thesis is a district heating system in Tianjin,and the district heating load forecasting method is researched.The main contents of this thesis are as follows:(1)Temporal Convolutional Network(TCN)and Cat Boost are applied to the heat load prediction tasks,respectively.To explore the performance of the TCN and Cat Boost,additional multiple linear regression and decision tree models are built for comparison.The experimental results show that the prediction results of the TCN and Cat Boost are more accurate,and the modelling of Cat Boost is easier and more feasible.(2)The prediction performance of the TCN is heavily dependent on the setting of hyperparameters.For this reason,a hybrid model based on Sand Cat Swarm Optimization(SCSO)and TCN is proposed for heat load prediction.SCSO is used to optimize the hyperparameters of TCN.To verify the effectiveness of this hybrid model,TCN,PSO-TCN(optimizing the hyperparameters of TCN using Particle Swarm Optimization(PSO)),and SSA-TCN(optimizing the hyperparameters of TCN using Sparrow Search Algorithm(SSA))are built for comparison.The experimental results show that the proposed hybrid model outperforms the other models.(3)A hybrid model combining a similar hour selection method and TCN is proposed for energy savings prediction of district heating loads.A Random Forest(RF)weighted Euclidean Norm(EN)is used to assess the similarity between the forecast hour and historical hours.To achieve further energy conservation,the minimum heat load among the predicted and similar hours is used as the prediction target.But when the difference between the actual heat load and the minimum heat load is greater than0.6 MW,the actual heat load is taken as the forecast target to ensure adequate heat delivery.TCN is utilized to predict the optimized heat load.The values of the coefficient of variation of the root mean square error of this hybrid model on different test sets are0.0847,0.1103,and 0.1017,respectively,and the values of the energy-saving rate are6.73%,7.32%,and 7.25%,respectively.