Voltage control in distribution networks using a model free OLTC transformer based on reinforcement learning techniques

This paper proposes a model-free methodology that leverages a Double Deep Q-Learning algorithm to optimize operation of transformers equipped with on-load tap changers, relying solely on minimal, reliably available data. By continuously learning from system states, the algorithm effectively regulates the voltage profile, prevents overvoltage, and reduces energy losses. The approach has been tested in a rural European low voltage network under various levels of photovoltaic penetration, comparing results with conventional techniques. Results indicate a reduction in both overvoltage events and the number of tap changes, leading to lower system losses and an extended transformer lifespan. This solution offers a robust, data-driven alternative for next-generation low voltage distribution systems, where measurement scarcity and uncertain topology often limit traditional control methods.