Executive summary of an IEA DHC report (2023) on applying AI to forecasting heat demand/production and failure detection in district heating networks (DHN). The project developed and evaluated ANN-based forecasting (CNN, LSTM, autoregressive, single-shot) using real measurement data from the Stiftung Liebenau DHN (Germany) and explored ML/deep-learning approaches for fault detection using primarily synthetic data. Key findings: CNN (single-shot) performed best overall; total-network 12 h heat-demand prediction achieved ~5.84% relative MAE; prediction horizons of 4–6 hours are useful operationally and 12 h forecasts were demonstrated; conservative estimated energy savings from integrating forecasting into Liebenau DHN control are 3–4% (previous ANN work for single-family systems showed up to 12% savings); synthetic dataset and code were published to support further research. The report discusses cost/benefit considerations for sensor deployment, leak localization (thermography and drones), pre-training on synthetic data, and comparative ML performance across several fault types.
Questions this document answers
Kramer W; Vallée M; Lamaison N; Gölzhäuser S; Wissocq T; Gaoua Y; Bitterling M; Frison L: AI for Failure Detection and Forecasting of Heat Production and Demand in District Heating Networks – Executive Summary, IEA DHC Report, 2023.
Relative Mean Absolute Error (MAE) of 5.84% for next 12 h network heat demand.
Prediction horizon of 4–6 hours is recommended; experiments include forecasts up to 12 hours.
Conservative estimated savings of 3–4% of whole energy production; prior ANN work for single-family systems showed up to 12% savings.
Dataset: https://www.kaggle.com/datasets/mathieuvallee/ai-dhc ; Code: https://github.com/mathieu-vallee/ai-dhc