JOVET: Jointly optimized multi-variational mode embedded Transformer framework for wind-to-hydrogen LCOH forecasting

This paper proposes a jointly optimized multi-variational mode embedded Transformer (JOVET) framework that improves levelized cost of hydrogen (LCOH) forecasting reliability for wind-to-hydrogen systems under uncertainty. In contrast to disjointed hybrid pipelines, JOVET is a unified embedded framework incorporating joint particle swarm optimization, jointly tuning multi-variational mode decomposition and Transformer hyperparameters. Embedding multi-variational modes into a multi-head-attention Transformer filters stochastic noise from supervisory control and data acquisition data, producing more stable LCOH forecasts. Benchmarked against time-series and decomposition-based models, JOVET achieves a mean absolute error of 77.34 kW, root mean square error of 124.38 kW, mean absolute percentage error of 7.36%, and coefficient of determination of 0.9914 for wind power forecasting, outperforming the best non-embedded baselines. Propagated to annual levelized cost metrics for alkaline and proton exchange membrane electrolyzers, JOVET yields an annual relative error of 0.007%, the smallest deviation among baselines, which propagates into the tightest LCOH distribution.