This project explores the application of the RLlib-IMPALA framework to address complex problems in the smart grid domain, particularly focusing on Volt-VAR optimization in electrical distribution networks. Our primary objective is to leverage this advanced reinforcement learning framework to reduce training times and computational demands, which are typically significant in smart grid applications.
We employ the RLlib-IMPALA framework, known for its efficiency in handling large-scale and computationally intensive training tasks.
The main application is the Volt-VAR problem in electrical distribution networks. However, the methodology is applicable to other challenging problems in the smart grid domain.
We integrate OpenDSS, a comprehensive electrical power system simulation tool, for realistic modeling and simulation of power system behavior.
Our models are trained using real-world data, including irradiance and load profiles from Santa Clara, California.
All simulations and training are conducted on the RAY platform, known for its scalability and efficiency in distributed computing environments.
We experimented with several DRL agents to identify the one that offers the optimal balance between training time and accuracy.
Our findings highlight the most effective DRL agents in terms of reduced training time and enhanced accuracy in solving the Volt-VAR problem.
The RLlib-IMPALA framework significantly decreased training times for the applied problems, demonstrating its effectiveness in the smart grid context.
This work showcases the potential of the RLlib-IMPALA framework in addressing complex, computationally intensive problems in the smart grid domain. By reducing training times and resource requirements, our approach makes it feasible to apply advanced machine learning techniques to optimize electrical distribution networks and potentially other areas in smart grid technology.