Implementation and Evaluation of a Limited-Area Artificial Intelligence Wind Speed Forecasting Model

The increased penetration of variable wind power production poses challenges for transmission system operators (TSOs) and wind power producers (WPPs), creating the need to generate accurate weather forecasts to mitigate the effects of volatile wind speed and associated wind power changes. While Numerical Weather Prediction (NWP) has traditionally been the cornerstone of weather forecasting, its high computational demands limit accessibility, making it primarily available to specialized weather forecasting companies. Artificial Intelligence-based Weather Prediction (AIWP) has emerged as a more efficient alternative to NWP, providing forecasts of similar accuracy while drastically lowering computational costs. This advancement is driving the liberalization of the weather forecasting market, enabling a broader range of users, including smaller energy producers and grid operators, to access and generate reliable, cost-effective forecasts. In particular, limited-area AIWP models (LAMs) are accelerating this transformation by offering high-resolution regional forecasts without needing large-scale supercomputing. This study develops a custom LAM based on a graph neural network (GNN) architecture for the South African region, hypothesizing that LAMs can serve as a cost-effective alternative to both NWP and global AIWP models by producing comparable forecasts (within 10% of other models’ accuracies). The proposed LAM produces forecasts that outperform the baseline global AIWP model. However, the proposed LAM underperforms compared to NWP by 16% in accuracy, concluding that the LAM still requires development as an accurate alternative to NWP, and a comparison of the state-of-the-art GNN AIWP model proves that the LAM has the potential to outperform NWP with enough training resources. In a case study for a selected wind farm, the LAM’s forecasts have an 8.63% increase in percentage error compared to a leading NWP model while operating at a fraction of the computational time required by the NWP, making it a computationally efficient alternative for regional forecasting. The results conclude that by lowering financial and technological barriers, LAMs contribute to the liberalization of weather forecasting, enabling more decentralized and democratized access to accurate, timely predictions.