Safe reinforcement learning versus classical controllers for voltage regulation and power quality in the IEEE 33-bus distribution system

1 view

Abstract

Modern distribution networks face increasing challenges due to the integration of inverter-based resources and stochastic load variations. This study aims to evaluate whether advanced controllers, including a reinforcement learning based safe controller, can provide superior voltage regulation, stability, and power quality compared with conventional proportional–integral, droop, and predictive optimal strategies. The IEEE 33-bus feeder was simulated using realistic 24-hour demand data and sensitivity matrices derived from feeder impedances. Four controllers were implemented under the linearized DistFlow formulation with explicit constraints on voltage deviation and reactive power limits. The analysis focused on both control quality indices, such as settling time, overshoot, steady-state error, and integral squared error, and power system indices, including voltage deviation, regulation, total harmonic distortion, power factor, and network losses. Results show that the safe reinforcement learning controller achieved the fastest settling time of 16.8 hours, the lowest overshoot of 3.2 percent, and the smallest steady-state error of 0.0065 per unit, while also reducing power losses to 0.145 megawatts and maintaining voltage stability above 0.95 per unit. By contrast, the proportional–integral controller exhibited overshoot of 12 percent, harmonic distortion of 8.18 percent, and voltage drops below 0.90 per unit. These findings indicate that embedding safety guarantees into reinforcement learning yields controllers that not only surpass classical strategies but also meet utility-relevant criteria, providing a promising pathway for future deployment in smart distribution networks.