Improving Heat-Load Forecasting with Deep Learning

This thesis investigates short-term heat-load forecasting in the context of an internship at Optit, a company that develops decision-support systems across diverse sectors. The aim is to evaluate forecasting models alternative to those currently used in the proprietary OptiEPTM platform, namely XGBoost and Support Vector Regression, and to assess their potential to improve forecasting accuracy and the consistency of performance across horizons and series of varying difficulty. The analysis is conducted on five synthetic hourly heat-demand series, representative of different application contexts, and considers two forecasting horizons of operational interest: day-ahead and week-ahead. A per-site modelling strategy is adopted, with each model optimised separately for each series and horizon. The models examined include MSTL-ETS, SARIMAX, LSTM, the Temporal Fusion Transformer, and the foundation model Chronos-2 in a zero-shot setting. The evaluation is based on rolling cross-validation, seasonal analysis, residual diagnostics, and bootstrap significance tests that account for temporal dependence. The results show that there is no universally dominant model. On the more regular and structured series, LSTM emerges as the most reliable solution across both forecasting horizons. On the noisier and more challenging series, Chronos-2 proves particularly competitive, especially in the day-ahead setting. XGBoost nevertheless remains a strong benchmark thanks to its high computational efficiency, while SARIMAX offers greater interpretability despite weaker performance during transitional periods. Overall, the study shows that improvements in forecasting performance depend on a trade-off between accuracy, computational cost, and interpretability, and provides a methodological and empirical basis for the further development of the forecasting module within OptiEPTM.