Uncertainty-aware day-ahead thermal-load forecasting in expanding district heating networks with attention-enhanced encoder-decoder LSTM

Abstract Reliable day-ahead thermal-load forecasts with quantified uncertainty can support risk-aware control decisions and reduce unnecessary safety margins. This paper applies an Attention-Encoder-Decoder Long Short-Term Memory (AEDL) architecture with direct multi-horizon decoding to rolling 24-hour-ahead thermal-load forecasting and compares three variants: deterministic AEDL, Gaussian-AEDL, and Quantile-AEDL. Gaussian-AEDL predicts a parametric Gaussian distribution, while Quantile-AEDL predicts non-parametric quantile levels. Two AEDL versions handle the growing load: version A scales the load of each input window, and version B adds a linear shortcut from the load of the last 24 h. The methods are evaluated on data from an expanding district heating network in Germany, using two rolling test periods that cover six months. Using the weather forecasts available at each forecast origin as inputs over the 24 h forecast horizon, the mean absolute percentage error (MAPE) ranges from 6.95 to 7.10%. Both versions outperform seasonal-naive, LSTM, and Temporal Fusion Transformer (TFT) baselines. Compared with TFT, they reduce the point-forecast error by 4 to 6% at about a quarter of the computational cost. Gaussian-AEDL achieves the lowest calibrated continuous ranked probability score (CRPS) of 161 kW, compared with 166 kW for Quantile-AEDL. Validation-based calibration raises central 80% interval coverage to 77 to 78%. Without weather forecasts, the forecast errors increase by 16 to 27%. Overall, the results show that probabilistic AEDL models retain point-forecast accuracy while providing useful uncertainty information, but their prediction intervals require calibration before operational use.