MoRAM for Multi-Step and Multivariate Short-Term Wind-Power Forecasting: A Reproducibility Audit and Corrected Attention Ablation

Accurate short-term wind-power forecasting supports renewable-energy integration. This study presents a reproducibility and implementation audit of MoRAM, an integrated architecture that combines feature-token attention, a residual Mamba layer, dense two-expert blending, and a one-dimensional convolution (Conv1D)–linear residual output path to forecast 12 h of normalized turbine power from 48 h multivariate histories in Dataset A (200 turbines; 8760 hourly records). Source reconstruction found that the historical PyTorch attention layer used its default sequence-major interface on batch-major data, thereby mixing samples; the dense-window record is therefore retained only as an implementation audit. We corrected the layer to batch-first feature-token semantics and ran an attention-only ablation using all turbines, a 12 h window-start stride, and five paired seeds. Corrected MoRAM achieved mean absolute error (MAE) 0.2428 ± 0.0029 per unit (p.u.) and root-mean-square error (RMSE) 0.3035 ± 0.0059 p.u.; bypassing only attention achieved MAE 0.2298 ± 0.0027 p.u. and RMSE 0.2839 ± 0.0028 p.u. (mean ± sample standard deviation). The attention-free condition had lower overall MAE and RMSE in every seed and, after averaging across seeds, at every forecast horizon. The principal validated contribution is therefore an auditable reconstruction—complete data protocol, source/log mapping, corrected tensor semantics, and a reproducible five-seed negative ablation—rather than evidence that every constituent module or the integrated architecture is superior.