Deep Reinforcement Learning Applied to Electric Distribution Network Applications Using Real Time Simulation

The continued growth in size and complexity of electricity distribution systems poses significant challenges for utilities, who struggle to maintain optimal control of service levels in line with established power quality regulations. Recently, the incorporation of distributed generation systems into the grid has increased the complexity of electrical distribution systems, rendering conventional power control systems inefficient. The present work proposes a strategy for effectively controlling the 34-bus IEEE power grid using a combination of deep learning and reinforcement learning techniques. The aim is to maintain voltage levels within the regulatory limits set by the governing framework. The adopted approach utilizes the Actor-Critic algorithm, which combines policy learning (actor) with state valuation (critic) to guide control decisions. This methodology is implemented in a Real-Time (RT) simulation environment using OPAL-RT's Synchronized software, ensuring the necessary accuracy and adaptability for safe and efficient grid operation. The proposed architecture interfaces with the power grid through transformers, enabling precise adjustments to the transformer tap settings for optimal voltage regulation. The developed LabVIEW graphical interface provides an intuitive tool for operators, facilitating RT interaction and control. Experimental results showcase the effectiveness of the proposed approach in maintaining voltages within permissible ranges, even under dynamic conditions and network disturbances.