Communication Dans Un Congrès Année : 2025

A generic substation heating power forecasting approach using machine learning

Résumé

To improve the energy performance and decarbonize the district heating (DH) systems, optimal control of systems is essential. A better energetic control of DH networks requires the forecast of the heating demand in substations. Among Machine Learning (ML) methods, regression techniques have found especially useful in short-term heat demand forecasting for DHNs, identifying patterns that assist in optimizing energy distribution. This study aims to provide a robust and generic approach to build statistical predictive models applicable to diverse building types. First, the selection of features is analyzed, and then various machine learning models are explored such as Artificial Neural Networks (ANNs), Linear Regression (LR), Ridge Regression (RR), Support Vector Regression (SVR), and Extremely Randomized Trees Regression (ETR). The approach is based on the analysis of 10 substations (SSTs) in 3 diPerent DHNs. Initially 13 features are used to forecast the heating power. The diPerent models are compared using Mean Absolute Percentage Error (MAPE) and R2 values. In addition, this study investigated the influence of threshold values on model accuracy, finding minimal impact, and examined how reducing data availability aPected predictions. The performance metrics underscore the ANN and ETR’s abilities to capture complex, non-linear heating power demand patterns, particularly in standard building categories like residential and tertiary buildings with predictable heating demands. While large datasets over several years are typically required for accurate forecasting, more recent advancements in ML and deep learning have shown promise in achieving similar accuracy with less data. On the opposite, the irregular patterns of certain SSTs highlighted the need for further features and additional data to enhance predictive accuracy. Feature selection and data availability played critical roles in model performance specifically in these SSTs. While tertiary substations demonstrated resilience to reduced datasets, the swimming pool substation's performance notably deteriorated, emphasizing the importance of comprehensive datasets and robust feature engineering

Fichier non déposé

Dates et versions

hal-05451178 , version 1 (09-01-2026)

Identifiants

  • HAL Id : hal-05451178 , version 1

Citer

Allan Iraqi, Pascal Stabat, Cong-Toan Tran. A generic substation heating power forecasting approach using machine learning. 11th International Conference on Smart Energy Systems, Aalborg University, Sep 2025, Copenhagen, Denmark. ⟨hal-05451178⟩
33 Consultations
0 Téléchargements

Partager

  • More