A1673
Title: Data-driven forecasting of wind vectors over the Tohoku region by convolutional LSTM and copulas using NWP outputs
Authors: Keigo Sato - Tohoku University (Japan) [presenting]
Yusaku Katsuragi - Tohoku University (Japan)
Yasumasa Matsuda - Tohoku University (Japan)
Abstract: A data-driven probabilistic forecasting model for station wind vectors over the Tohoku region is developed, trained on AMeDAS observations. The model post-processes wind fields predicted by numerical weather prediction (NWP) and jointly outputs prediction intervals of wind speed relevant to wind-power operation. As inputs, lead-time sequences of wind fields predicted by NWP on the Tohoku grid are used. A convolutional LSTM is trained on AMeDAS station observations and maps these wind fields to a bivariate Gaussian predictive distribution for the wind vector (u,v) at each lead time and station. The (u,v) formulation yields a predictive distribution of wind speed $s = \sqrt{u^2 + v^2}$ and prediction intervals computed from the predictive distribution via sampling, without any iterations to obtain multi-step-ahead forecasts, while retaining the physical dynamics already embedded in NWP. Furthermore, the model is extended to multivariate probabilistic forecasting using copulas to account for spatial dependencies, enabling the modeling of joint probabilities and spatial uncertainty structures across the region. The accuracy of predicted wind-speed intervals is evaluated using empirical coverage and average interval width, together with reliability diagnostics. Results are compared against quantile regression baselines that directly predict wind-speed quantiles from station-mapped NWP predictors.
