A novel model for improving the prediction accuracy of the new heating station

Abstract A prediction model, which can be established with a small number of operating data, is the key to realize the on-demand heating as soon as possible for new heating stations. The existing physics-based methods with high accuracy are based on detailed building thermal parameters, which are difficult to obtain and have poor applicability. Data-driven methods have high prediction accuracy, but they require a large amount of historical data for model training. In this paper, a novel prediction model based on building thermal inertia was proposed, called κ -model. Analysis and comparison were conducted between κ -model, support vector machine (SVM), extreme gradient boosting (XGBoost) and Multiple linear regression (MLR) with different conditions for secondary average temperature prediction. Results showed that κ -model can explore the heating characteristics from 1/2 days of operational data, and predict secondary average temperature with high precision by one-step or multi-step not only hourly but also daily. While XGBoost and MLR required the operational data of 15/60 days for the hourly/daily prediction model training, the accuracy of multi-step prediction was higher than that of one-step prediction for XGBoost, and MLR was not suitable for daily multi-step prediction due to error accumulation. SVM required the training data of 15/75 days for the hourly/daily prediction, and one-step prediction had higher accuracy than multi-step prediction. The above mean that in practical engineering, only κ -model can be applied for hourly prediction from November 16 to November 29, and for daily prediction from November 17 to January 16 at the beginning of the heating season.