The effective control of the steam generator water level is critical for the safe and stable operation of nuclear power plants. To address the poor performance of conventional cascade control, this paper investigates the classic Irving water level model and enhances the traditional cascade control by incorporating compensation for the water level controller's output control signal using an intelligent agent trained via deep reinforcement learning. Furthermore, a scheduling function is designed to integrate the intelligent agents applicable to typical power levels, aiming to alleviate the adverse effects of strong nonlinearities. Simulation results demonstrate that, compared to conventional cascade control, the proposed compensatory controller exhibits better adaptability. By collaborating with the cascade controller, it enables the water level to recover to its normal value more rapidly. The resulting control strategy shows strong robustness against disturbances, with significant improvements in performance indicators.