Guardrail-Enforced Tabular Q-Learning for Energy-Efficient Aeration Control of a Pilot-Scale Biofilm A/O System Treating Piggery Wastewater

Aeration dominates energy use in biological wastewater treatment, but aggressive energy-saving control risks violating dissolved oxygen (DO) safety margins. We present a guardrail-enforced tabular Q-learning controller for a pilot biofilm anoxic/oxic (A/O) system (2 × 4.86 m3) treating piggery wastewater, compared online against always-on aeration and a hysteresis rule. A first-order-plus-deadtime DO model (R2 = 0.973) identified from pseudo-random excitation was used to pre-train a 108-state policy balancing energy, DO-band keeping, and switching wear, deployed as a 5 min decision loop nested inside a 1 min hard-guardrail monitor with command readback verification. In 24 h online campaigns, the RL controller cut metered aeration energy by 44.3% versus always-on (26.05 vs. 46.8 kWh·d−1; 95% CI 43.2–45.4%) with zero guardrail interventions and day-to-day reproducibility within 1%; the cheaper hysteresis rule (−55.2%) required 9–12 hard-guardrail interventions per day. Effluent COD and NH4+-N remained far below the enterprise discharge requirement under all controllers throughout the 24 h campaign horizons, with no detectable short-term effluent penalty, and a manual actuator-override incident was contained autonomously by the guardrail layer. Because the hydraulic retention time (≈17 d) far exceeds the campaign length, longer-horizon effluent effects remain open and are stated as a limitation. A deliberately simple, auditable tabular policy thus captures most of the achievable aeration savings on small decentralised systems while its safety case rests on hard guardrails that bound the worst case independently of what the learned policy outputs.