This study proposes a hybrid neural network architecture that integrates error backpropagation (EBP) optimization with TRNSYS co-simulation, specifically targeting the performance enhancement of air-source heat pump (ASHP) systems in public buildings. The calculated optimal dynamic return water temperature setting was input into an established low-temperature air-source heat pump model for feasibility verification. Results showed: fan energy consumption rose by 13.0% vs. the traditional control method, while unit energy consumption dropped by 8.60% and per-unit-area system energy consumption decreased by 6.99%. The optimized method was applied to a typical air-source heat pump heating system in Changchun during the heating period on high-conditioning days. In low-conditioning mode, daily energy consumption fell from 287 kW•h (traditional) to 262 kW•h (optimized) - an 8.7% reduction. In high-conditioning mode, it decreased from 737.3 kW•h to 710.5 kW•h, a 3.6% drop. When used in three typical such systems across three cities in severely cold regions, the optimized method cut per-unit-area unit energy consumption by 5.15%-9.31% and per-unit-area system energy consumption by 6.15%-7.37% compared to the traditional method. By dynamically controlling the optimal return water temperature of the simulation system, energy consumption has been reduced, which has contributed value to achieving China's dual carbon goals.