Answers are grounded in this paper and cite their sources.
Data-driven models can achieve high accuracy in district heat load forecasting, yet no single model performs best across tasks and systems, motivating ensemble learning based on diverse base learners. Studies predominantly evaluate fixed ensemble configurations, leaving component contributions unquantified and informed training data partitioning uncontrolled against random partitioning. To quantify these contributions to day-ahead district heat load forecasting, we deconstruct ensemble models into base learner algorithm, ensemble structure, and combination rule, comparing four regression algorithms, homogeneous and heterogeneous structures, and three dynamic combination rules across five training data partitioning strategies, including a random control and a temperature-based seasonal rule. All configurations are benchmarked prequentially on a controlled synthetic and an operational real-world dataset. Homogeneous ensembles of non-linear algorithms under random partitioning with dynamic error-based weighting reduce MAE by 7.3 % on synthetic data and 15.7 % on real-world data over the respective single models, with improvements persisting across partition realizations. On both datasets, the heterogeneous ensemble trained on the complete dataset remains within the statistically best model group, indicating a favorable default configuration. Linear regression benefits most, reducing MAE by up to 52.4 % through informed partitioning and dynamic selection, narrowly outperforming the best single model on real-world data. Performance is governed not by the cohesion and separation of load profiles across partitions but by the balance between specialization and available training data, rendering the value of partitioning algorithm-specific. The findings yield practical recommendations for batch-based ensemble model initialization and combination rule selection ahead of data stream deployment.
No takes yet. Share an insight, caveat, or question.
Derzsi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: