Data-Driven Multi-Objective Optimization of Low-Carbon Heating Systems in Cold Regions of China: A Case Study of Harbin

The transition to clean heating in cold regions faces challenges from extreme weather and high emissions. This study develops a dynamic modeling and surrogate-based multi-objective optimization framework for a solar-ammonia hybrid heating system. A lumped-parameter model simulates energy flows, and year-long Harbin meteorological data produce a physically consistent dataset via lumped-parameter simulation. A feedforward neural network with a residual autoregressive structure predicts temperature increments, preserving thermal inertia; it achieves RMSEs of 0.16 °C (indoor) and 0.61 °C (tank) over 168 h, including 48-h blind forecasting. This fast surrogate is embedded in NSGA-II to minimize carbon, cost, and discomfort via a climate-compensated curve and hysteresis dead-bands. Pareto analysis yields trade-offs: the carbon–comfort optimum attains 80.62 kg CO2/week and 104.85 °C·h discomfort; the cost–comfort optimum costs ~60 CNY/week. Three-dimensional optimization confirms aligned carbon and cost. TOPSIS selects a final strategy that, versus constant setpoints, reduces carbon by 30.3% and cost by 45.3% over 168 h by leveraging thermal mass and reducing boiler starts. Annual ammonia co-firing lowers CO2 from ~560 to 510 kg. The proposed integration of physics-informed simulation, residual autoregressive neural surrogates, and multi-objective evolutionary optimization provides an effective methodology for intelligent control of renewable hybrid heating systems in cold climates, contributing to deep decarbonization and energy savings.