Explainable and transferable heating load forecasting

The increasing electrification of the heating sector and growing renewable energy integration require accurate heating load forecasts to integrate buildings as flexible consumers into energy management systems. The expanding use of heat pumps and the need for load shifting necessitate accurate predictions for grid stability and cost-effective operations. Although machine learning models have achieved significant improvements in forecast accuracy, they face two key limitations: First, limited model explainability, which reduces transparency for building operators. Second, restricted data availability in many buildings, particularly in new or renovated buildings. This dissertation addresses these challenges through systematic integration of Explainable Artificial Intelligence (XAI) with Transfer Learning (TL), developing heating load forecasts that combine high accuracy, explainability, and transferability. The work pursues this approach through three interconnected studies. The first study establishes the methodological foundation by demonstrating for the first time a systematic relationship between building characteristics and feature importance using 15 552 simulated room variants. Outside temperature dominates with an average of 55 % of global SHapley Additive exPlanations (SHAP) feature importance, with feature importance distributions varying significantly according to parameters such as air exchange rate, U-values, and g-value. Building on this foundation, the second study transferred this approach to multi-step forecasts using Long Short-Term Memory (LSTM)-based Encoder–Decoder (ED) models and two real-world building datasets. Approximately 75 % of SHAP feature importance concentrated within the last eleven to thirteen input hours before the forecast hour, highlighting the critical importance of current system states. Attention maps and DeepSHAP provided consistent yet complementary explanations for the temporal weighting of input variables. Targeted, SHAP-based reduction of irrelevant features improved forecast accuracy by 8.1 % and reduced training effort simultaneously. The third study addresses scalability through a large-scale transfer learning benchmark using 116 real-world building datasets and multiple model architectures. Fine-tuning strategies reduced median forecast error by 22 %, while zero-shot approaches achieved additional improvements. Transformer-based models such as PatchTST and TiDE proved particularly robust under data scarcity and with minimal computational requirements. The integration of SHAP-based features did not yield consistent improvements and proved highly model-dependent. This work establishes the first comprehensive framework for robust, explainable, and transferable heating load forecasts. The framework both advances the scientific knowledge base on XAI in building technology and opens direct application opportunities in energy management systems and Demand Response (DR) programmes.