GNN-PPO-Based Multi-Stage Security-Constrained Transmission Expansion Planning for Renewable Energy Integration

Abstract

The increasing penetration of renewable energy introduces additional complexity into long-term transmission expansion planning because investment decisions must simultaneously account for network topology, nonlinear operating constraints, renewable-generation uncertainty, and contingency security. This study proposes a Graph Neural Network-assisted Proximal Policy Optimization (GNN-PPO) framework for multi-stage security-constrained transmission expansion planning (SC-TEP). A two-layer edge-aware GNN with a 16-dimensional embedding is employed to encode the transmission-network topology, while PPO sequentially selects reinforcement projects over the planning horizon. The framework incorporates AC power-flow constraints, three renewable-availability scenarios, candidate-specific investment costs, and N−1 security requirements. To reduce computational burden, the five most critical credible contingencies are screened during training, whereas all credible non-islanding N−1 contingencies are evaluated during post-training validation. A stage-aware security-repair procedure followed by backward project pruning is introduced to eliminate residual security violations and unnecessary reinforcements. The framework is applied to the Southern Sulawesi 150 kV transmission system over the 2025–2034 planning horizon. The raw GNN-PPO portfolio requires 64.17 million USD (MUSD), compared with 97.10 MUSD for the Multilayer Perceptron-based PPO (MLP-PPO) benchmark, but full-horizon validation identifies one residual contingency violation in 2025. After security repair and pruning, the final solution consists of eight transmission reinforcement projects with a total planning investment of 86.72 MUSD. In 2034, the final system achieves a minimum voltage of 1.0038 pu, maximum branch loading of 100.00%, transmission losses of 41.31 MW, and renewable curtailment of 495.42 MW, with zero failures among 62 credible non-islanding N−1 contingencies. The results demonstrate that topology-aware reinforcement learning can provide an effective candidate-search mechanism for multi-stage SC-TEP, while explicit post-training security verification remains essential for obtaining an engineering-feasible expansion plan.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint