Machine Learning-Based Forecasting for Transmission Grid Stability under Heat Stress

Abstract

Transmission grid stability under heat stress presents a critical challenge as climate driven heatwaves intensify, simultaneously elevating electricity demand for cooling and degrading key infrastructure components through elevated conductor temperatures, increased line losses, reduced ampacity due to sag, and derated generation capacities. Traditional physics-based models for dynamic line rating (DLR), thermal limits, outage prediction, and stability assessment are computationally intensive and often rely on conservative static assumptions that fail to capture spatiotemporal variability in real-time or near-real-time operations. Machine learning (ML) approaches, including Long Short-Term Memory (LSTM) networks, Graph Neural Networks (GNNs), hybrid GAT-LSTM architectures, and reinforcement learning variants, offer powerful data-driven alternatives by learning complex nonlinear mappings from historical weather, SCADA, PMU, and outage datasets to forecast conductor temperatures, dynamic line ratings, fault probabilities, load under heat, and stability margins. This research paper develops a comprehensive ML-based forecasting framework that integrates hyper-local weather predictions with grid topology for probabilistic DLR forecasting, heat-induced fault and outage risk prediction, and stability indicators such as voltage collapse proximity or frequency response under derated conditions.