Electricity price spike forecasting via multivariate adaptive visibility graph neural networks

• A novel MAVG-DiffPool with an oversampling framework is developed for electricity price spike forecasting. • MAVG captures multivariate interdependencies through adaptive visibility graph construction. • A balanced epoch sampler eliminates stochastic variance and ensures stable training convergence. • Balanced performance is achieved across multiple spike conditions in the Victorian region of Australia’s National Electricity Market. Accurate forecasting of electricity price spikes is critical for energy market risk management, as these rare yet disruptive events cause substantial financial losses and arise from complex interactions among multiple market variables. An enhanced Graph Neural Network (GNN) framework is proposed that captures interdependencies among these Multivariate variables via an Adaptive Visibility Graph (MAVG) network. A hierarchical Differentiable Pooling (DiffPool) GNN architecture is then applied to coarsen this graph and forecast the next-step price spike probability. To address the extreme class imbalance inherent in spike data, a novel oversampling method called the balanced epoch sampler is also introduced. The proposed MAVG-DiffPool framework with oversampling is validated using real electricity prices from Australia under multiple spike thresholds. The framework achieves an F1-score of 0.84 under the moderate threshold. Under the extreme threshold, the framework outperformed LSTM by achieving nearly double the F1-score, demonstrating its effectiveness and robustness in addressing the challenge of spike forecasting.