District heating systems in Finland are undergoing structural transition as combined heat and power (CHP) plants are decommissioned, creating non-stationary forecasting conditions that challenge conventional models. This paper presents a probabilistic forecasting pipeline for national Finnish CHP-based district heating using a Temporal Fusion Transformer (TFT) trained on nine years of hourly Fingrid open data. A residual-from-persistence target reformulation is introduced to handle non-stationarity under structural regime shift, significantly improving forecast accuracy. The proposed TFT-resid model achieves a mean absolute error of 87.4 MWh/h and a sMAPE of 21.2%, outperforming XGBoost baselines and naive persistence benchmarks. This work forms the research foundation of the ET-Design Lab (Aurinkolab Community), where the forecasting methodology is applied in the construction of an AI-driven Uusimaa Heating Digital Twin — a real student-built tool at the intersection of cutting-edge energy research and hands-on engineering education for teenagers.