A short-term wind power forecasting model based on neural ordinary differential equations

Accurate short-to-medium-term wind power forecasting is essential for grid dispatching and safe operation. To address limitations in the modeling of endogenous variables, cross-channel semantic propagation, and time continuity modeling, this study proposes ODE-EMAformer, a hybrid framework integrating continuous-time latent state evolution with conditional attention. It combines an ordinary differential equations (ODE) feature extractor with kernel-based aggregation and observation-triggered updates, a conditional attention module with channel-phase encoding, and an ordinary differential equations-based prediction head for multi-step forecasting via explicit Euler integration. Experiments on the France and Turkey datasets show lower mean absolute error (MAE) and root mean square error (RMSE) than baseline methods, with 43.50 and 75.17 on France, and 97.63 and 179.43 on Turkey. Relative to the 2DXformer baseline, ODE-EMAformer reduces MAE/RMSE by 21.2%/17.8% on France and 19.2%/18.8% on Turkey. These results indicate that combining time continuity with data-driven adaptability is effective for multivariate wind power forecasting.