A novel reinforcement learning-based framework for Data-enabled predictive control (DeePC) hyperparameter tuning, designated SARSA-DeePC, is proposed in this work to achieve model-free operation, real-time adaptability, and computational tractability. DeePC relies on system input-output (I/O) measurements to characterize dynamics, where control performance critically depends on the regularization parameter \(\lambda _g\) . Under noisy conditions, existing approaches often exhibit prohibitive computational burdens or excessive conservatism. In this framework, the hyperparameter tuning problem is formulated as a Markov Decision Process (MDP), establishing a quantitative mapping between I/O behavior and \(\lambda _g\) . An optimal tuning policy is acquired through offline training, while continuous Q-table updates enable robust online adaptation to non-stationary noise. Theoretical guarantees for convergence and stability under stochastic exploration are derived. Experimental validation over 10 independent trials demonstrates that SARSA-DeePC significantly outperforms baseline methods. Quantitatively, the proposed approach achieves Mean Absolute Error (MAE) reductions of 72.5% (95% Confidence Interval (CI): [65.7%, 79.3%]) under baseline noise and 50.5% (95% CI: [38.8%, 62.3%]) following noise amplification. Compared to standard and competing adaptive DeePC methods, the MAE is reduced by up to 37.5% and 21.4% under varying noise intensities, respectively. With an average per-cycle computational time of 0.0292 seconds, the framework maintains balanced multi-metric performance and statistical robustness, validating its efficacy for real-time industrial applications.
