This paper based on the characteristics of plume diffusion, employs the Realizable \(\mathrm {k-\varepsilon }\) model for implicit unsteady-state calculations to simulate the diffusion process of pollutants. Additionally, it applies the SAC (Soft Actor-Critic) deep reinforcement learning algorithm to study its effectiveness in tracing pollution sources in reservoirs. The research results show that, compared to the Comb-shaped Search Path strategy, the SAC-based source tracing strategy achieves approximately 12 times improvement in time efficiency and converges faster. The capability of the SAC algorithm in handling continuous state and action spaces, combined with its entropy-maximizing objective and adaptive temperature coefficient ( \(\alpha \) ), provides strong exploration capabilities and robustness. Simulation results confirm the algorithm’s effectiveness: after an initial exploratory phase with Gaussian noise (first 50 episodes), the agent’s cumulative reward significantly increased, converging stably to around 1000 by episode 75. This indicates the discovery of a near-optimal policy for locating the pollution source from various starting points. The results validate SAC as a viable and efficient approach for this complex environmental task.