DDPG-Based Continuous Action Control for Hybrid Renewable Energy System Optimization in Multi-Energy Integrated Networks

Abstract

This paper addresses the research gaps in the resource allocation and collaborative scheduling of rural hybrid renewable energy systems, which are faced with high uncertainty and complex optimization dimensions. It proposes an artificial intelligence-enhanced cyber-physical system based on the Deep Deterministic Policy Gradient (DDPG) algorithm. This method utilizes measured data (with a sampling frequency of 15 minutes) from 50 households in a certain region of China from 2022 to 2023 to construct a test environment for a 33-node integrated energy system that couples electricity, gas, and heat. The model employs an actor-critic neural network architecture (with 2 hidden layers of 256 neurons), sets an experience replay buffer of 100,000, a batch size of 32, and uses the convergence criteria of a critic network loss change rate below 10⁻⁴ and cumulative reward fluctuation less than ±5%. Experiments show that the proposed method converges in only 480 iterations compared to baseline models such as rule-based, BCC, and DQN, and increases the renewable energy grid connection rate to 50.56%, reducing carbon emissions by 7.52 tons. The results indicate that this data-driven framework can effectively achieve real-time optimal scheduling for multi-energy complementary systems, providing a reliable solution for high-proportion renewable energy consumption.