The hybrid active power filter (HAPF) emerges as a cost-effective remedy for power quality challenges in medium voltage power systems. The success of HAPF, crucially, hinges on the efficacy of its current control mechanism. This paper introduces a novel approach by proposing a deterministic policy gradient based reinforcement learning (DPG-RL) as the current control strategy for HAPF. In stark contrast to conventional model-based control methods, the DPG-RL leverages artificial intelligence (AI) technology, rendering it model-free and capable of dynamically seeking the optimal control policy to enhance HAPF performance. A notable advantage lies in its significantly lower computational burden during each sampling period, distinguishing it from other contemporary AI-aided control methods. The paper outlines a systematic design process, encompassing feature selection and reward function design, to formulate the RL problem. The comprehensive design procedure of DPG-RL is then detailed. Simulation results are subsequently presented, validating the effectiveness and reliability of the proposed DPG-RL across diverse operational scenarios.