The performance of solar heating systems is significantly influenced by outdoor weather fluctuations and building heating loads, leading to dynamic variations that undermine the efficacy of rule-based control (RBC) strategies. Additionally, the hydraulic and thermal time-delay characteristics frequently lead to delays in control points for real-time optimization (RTO) control strategies. While Model Predictive Control (MPC) effectively addresses these dynamic and time-delay issues in solar heating systems, its substantial computational demands limit its real-world applications. To overcome these challenges, this study proposes a Model-Free Predictive Control (MFPC) approach utilizing Deep reinforcement learning (DRL). Through TRNSYS simulations, the study conducts a comparison of the performance and energy consumption of RBC and MFPC systems, focusing on a residential solar heating system in Lhasa, Xizang as a case study. The results demonstrate that the MFPC method reduces unmet heating demand by 31% compared to traditional RBC strategies, improves solar collection efficiency by nearly 12%, and decreases tank heat loss by 2.2%. When accounting for thermal storage effects, the optimized MFPC strategy achieves a reduction in net energy consumption of 25.6%.