AI-Based Predictive Controller for Residential Heat Pumps: From Simulation to a Season-Long Field Trial
Heat pumps are essential for building decarbonisation, yet the fixed heating-curve controllers supplied with most units perform poorly under variable operating conditions. This study presents a vendor-independent neural-network controller that can be retrofitted to existing HPs through standard 0–10 V or Modbus interface, replacing the static control curve with a predictive, optimisation-in-the-loop strategy. A lightweight Transformer model forecasts indoor air temperature several hours ahead using existing heat-pump sensor data, including indoor, outdoor, return, and supply temperatures. A real-time capable optimiser selects the supply-water temperature that minimises a weighted sum of thermal discomfort and compressor energy, thus realising model-predictive-control benefits without requiring a physical plant model or cloud infrastructure. Simulations on nine German buildings from the TABULA database indicate electricity savings of up to 15 % and a seasonal COP increase of up to 0.3, while the duration of comfort violations is halved relative to a manufacturer standard heating curve. A field trial during the 2024/25 heating season in a renovated singlefamily house (radiators, 11 kW HP) confirmed a seasonal COP rise of around 0.2, an 8 – 11 % reduction in daily electricity demand, and a decrease in mean absolute comfort deviation. Over a 180-day season these figures correspond to approximately 387 kWh and 23 kg CO2 saved per building.