Multivariate Outlier Detection for District Heating Systems Improves Efficiency

Identifying irregular operations in District Heating Systems (DHS) is crucial for optimizing energy consumption and reducing CO2 emissions. This paper explores different methods for multivariate outlier detection within DHS data, specifically focusing on transmitted heat energy and outside ambient temperature in a selected substation. The study evaluated several techniques: univariate Z-score (as a baseline), Mahalanobis distances, Principal Component Analysis (PCA), Isolation Forest, and Hotelling's T-squared test. The researchers considered specific domain characteristics, such as the irrelevance of zero transmitted energy timepoints, which do not indicate an off-grid plant. Discussions with domain experts confirmed that PCA, Isolation Forest, and the Hotelling method yielded relevant results. Ultimately, an ensemble approach was adopted, where outliers were identified based on the agreement of all three effective methods. This combined strategy successfully uncovered irregular plant operations, pointing towards opportunities for reducing gas consumption in central heating plants and lowering CO2 emissions.