Sewer overflow from urban sewerage systems (USS) represents a primary driver of coastal pollution worldwide. Urban expansion and climate change further exacerbate overflow contamination in coastal urban areas. Real-time control (RTC) of USSs has been widely validated as an effective strategy to exploit network spare capacity and curtail overflow events, thereby alleviating coastal ecological degradation. The development of reliable, stable control algorithms is therefore pivotal for robust pollution mitigation via RTC. Deep reinforcement learning (DRL), a state-of-the-art AI intelligent control technique, has emerged as a powerful tool for USS RTC, exhibiting remarkable efficacy in sewer overflow reduction. Nevertheless, training DRL agents from scratch for complex USSs suffers from prohibitive time costs and unstable convergence, limiting its practical deployment. Here, we propose a knowledge-data fusion DRL training framework that integrates multi-source urban drainage engineering expertise and mechanistic knowledge into DRL optimization. Specifically, we translate diverse forms of prior engineering knowledge into knowledge-embedded datasets via systematic simulation, and pre-train DRL agents using a newly designed offline supervised learning paradigm. The pre-trained model is further fine-tuned through standard reinforcement learning workflows. When applied to a benchmark USS, our framework achieves comparable performance in mitigating inundation and sewer overflow to fully trained conventional DRL models, while reducing training time by nearly 90%. This technology enables the field deployment of intelligent RTC algorithms in coastal cities and provides robust support for coastal urban pollution control.