A comparison of prediction and forecasting artificial intelligence models to estimate the future energy demand in a district heating system. (15th April 2023)

• Record Type: Journal Article
• Authors: Runge, Jason; Saloux, Etienne
• Is Part Of: Energy. Volume 269(2023)
• Publication Date: 2023-04-15
• Subjects: District heating -- Prediction -- Forecasting -- Machine learning -- Deep learning -- Heating load; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2023.126661
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: Forecasting the short-term future energy demand in buildings and districts is a vital component towards the optimization of energy use and consequently the reduction in greenhouse gas emissions. This paper explores artificial intelligence approaches applied to estimate the future heating load in a district heating system. A distinction is made within thisd work between a prediction and forecasting based approach; a comparison is then accomplished by applying each method with prominent Machine Learning and Deep Learning based algorithms to estimate the future heating demand over 6 h and 24 h ahead. This analysis used available data from a Canadian district heating system in Quebec and actual weather forecasts obtained from Canadian meteorological services. All models within this work applied a grid search in order to calibrate their respective hyperparameters. Results of this work indicated that the prediction-based approach (with forecasted inputs) obtained a higher accuracy than the forecasting approach. All the machine learning models obtained good accuracy with errors not exceeding 16% CV(RMSE) and closer to 10% CV(RMSE) for the top performing models. Furthermore, the LSTM and XGBoost were consistently among the top performing algorithms and provided good performance over a variety of hyperparameters. The biggest difference between the two algorithms was the computational times; it was observed that the XGBoost was significantly faster to train . Highlights: Artificial Intelligence models are investigated to forecast district heating demand. Prediction and forecasting based approaches are compared. Model performance is estimated using accuracy, training time and stability. The prediction approach using forecasted inputs shows slightly better results. LSTM and XGBoost models outperform other techniques with an error of 11%.