Graph Neural Network-Based Multi-Objective Optimization for Renewable Energy Spatial Layout Design

Abstract

With the rapid development of Renewable Energy (RE), traditional layout methods struggle to integrate multi-source geographic data and avoid ecological conflicts. To address these issues, this paper proposes an intelligent solution combining a Graph Neural Network (GNN) and a multi-objective dynamic optimization framework to enhance planning accuracy and layout efficiency. Based on 1 km×1 km grid data of wind farms in northwest China (15,600 nodes/156×100 km²) and photovoltaic parks in eastern China (12,800 nodes/128×100 km²), a geographic spatial graph is constructed, allowing GNNs to identify spatial elements precisely. The framework integrates Genetic Algorithms (GA) with a Double Deep Q-Network (DDQN) to optimize RE layouts, balancing power generation efficiency, ecological impact, and construction costs through multi-objective dynamic optimization. Case studies of wind farms in northwest China and photovoltaic parks in eastern China are used to validate robustness and adaptability under complex terrains. Results show the proposed method achieves a spatial recognition accuracy of 0.93, 14.8% higher than traditional Convolutional Neural Networks (CNNs), ecological conflicts reduced by 29.6%, and cost savings reached 17.2%. This identification-optimization closed-loop framework effectively processes non-Euclidean spatial data and performs multi-objective collaborative optimization.