Abstract
Demand-side flexibility from buildings is becoming increasingly important in Nordic local flexibility markets (LFMs), where credible participation depends on accurate estimation of both the counterfactual baseline load and the flexibility that can actually be delivered. This study investigates indoor thermal modulation as a flexibility resource in four large public heat-dominated buildings in Ålesund, Norway, representing three building types and three common electric heating systems in the Nordic context. First, the flexibility potential and economic value of one-hour heating setpoint relaxations are quantified through physics-based simulation and annual Monte Carlo market analysis. Second, six forecasting approaches are compared for predicting baseline and flexible electrical loads. To address the limitations of conventional single-mode forecasters, an end-to-end pipeline, Interpretable Machine Learning for Energy Management Systems (IML-EMS), is proposed. The framework combines a differentiable gradient-boosted model with physics-informed ARX regularization derived from an RC thermal representation. Results show that directly heated buildings provide the highest flexibility, reaching 11.73 Wh/m2, with average LFM benefits of 0.108 NOK/m2. Although several benchmark models achieve competitive load-forecasting accuracy, IML-EMS provides the most reliable joint prediction of baseline and flexible load and achieves the highest bidding precision (BP) across all pilots. In addition, the model identifies interpretable RC parameters and generalizes successfully to real smart meter data, where it significantly outperforms benchmark methods.