Machine-learning-based multi-step heat demand forecasting in a district heating system

• Published in: Energy and buildings Vol. 233; p. 110673
• Main Authors: Potočnik, Primož, Škerl, Primož, Govekar, Edvard
• Format: Journal Article
• Language: English
• Published: Lausanne Elsevier B.V 15.02.2021; Elsevier BV
• Subjects: Comparative analysis; Demand; District heating; Economic forecasting; Feature extraction; Forecasting; Gaussian process; Gaussian process regression; Heat demand; Irradiation; Learning algorithms; Machine learning; Mathematical models; Model accuracy; Radiation; Regression analysis; Short-term forecasting; Weather forecasting; NN; ANN; LM; CNN; DH; SVMlin; UTC; LR; ELM; GPR; LM–BR; CPU; SVMquad; CHP; ALADIN; SVM; LRord; ARSO; ECMWF; ARMA; ML; INCA; LSTM; District heating; AI; MANE; GP; LRrob; MAPE; Short-term forecasting; Heat demand; RVFL; TE-TOL; CET; CV; Machine learning; Gaussian process regression; TOS; LRstep
• ISSN: 0378-7788; 1872-6178
• DOI: 10.1016/j.enbuild.2020.110673

[Display omitted] •Methodology for 48-hour multi-step heat demand forecasting in a DH system.•Gaussian process regression outperforms considered machine learning methods.•Accurate temperature forecasts are important, solar irradiation forecasts are not.•Forecasting errors for 48 h ahead below 3% of the max. heating power.•Proposed forecasting solution can be fitted to different DH systems. Short-term heat demand forecasting in district heating (DH) systems is essential for a sufficient heat supply and optimal operation of the DH. In this study, a machine learning based multi-step short-term heat demand forecasting approach using the data of the largest Slovenian DH system is considered. The proposed approach involved feature extraction and comparative analysis of different representative machine learning based forecasting models. Nonlinear models performed better than linear models, and the best forecasting results were obtained by Gaussian process regression (GPR), where the mean absolute normalized error was 2.94% of the maximum heating power of the DH system. The analysis confirmed the importance of accurate temperature forecasts but did not confirm the relevance of using future solar irradiation forecasts. The optimal length of training data is shown to be 3 years, and past data of up to 4 days can be used as input to increase the forecasting accuracy. The forecasting model (GPR) proposed in this study can be fitted to different DH systems. In the presented case study, it was selected to implement the online forecasting solution for the DH of Ljubljana and has been generating forecasts with a mean absolute normalized error of 2.70% since November 2019.