A Hybrid LSTM‐KAN Framework With Dynamic Residual and Adaptive Grid for Robust Wind Speed Forecasting
Wind Energy · 2026-09 · 29 (9)
DOI: 10.1002/we.70146 ↗Kurz gefasst
Fachartikel in der Zeitschrift Wind Energy, online erschienen am 11.08.2026. Verfasst von Guangjie Liu und 3 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Natural Science Foundation of Jilin Province.
Abstract (Originalsprache)
Accurate wind speed forecasting is indispensable for the efficient integration of wind power into the electrical grid, optimizing turbine scheduling, and ensuring grid stability. Existing models, such as Long Short‐Term Memory (LSTM) networks, often struggle to capture the highly nonlinear relationships and long‐range dependencies present in wind speed data. To address these limitations, this study proposes a hybrid model, LSTM‐FDA‐KAN, which integrates the sequential learning strength of LSTM with the nonlinear approximation capabilities of the Kolmogorov–Arnold Network (KAN). The model is further enhanced by three specialized modules. The Feature Cross (FC) module captures complex interactions between input variables, the Dynamic Residual (DR) module adaptively corrects prediction errors in real time, and the Adaptive Grid (AG) module enables flexible spatioemporal discretization, making the model more robust to uneven data distributions. We evaluate the model using the LandBench 1.0 dataset, which combines diverse meteorological and land‐surface variables across global regions. LSTM‐FDA‐KAN consistently outperforms baseline models, achieving Pearson correlation coefficients nearly 0.91 and reducing root mean square error (RMSE) by 2%–3% . Gains are most pronounced in coastal and high‐latitude regions, where prediction accuracy improves by up to 15% under highly variable weather conditions. Beyond overall accuracy, the model demonstrates stability in capturing long‐range temporal dependencies. It performs reliably during prolonged extreme events, such as consecutive heatwaves and extended rainfall periods, where conventional models typically degrade. Together, these results establish LSTM‐FDA‐KAN as a scalable and effective solution for geophysical time‐series forecasting. By combining sequential modeling, nonlinear function approximation, and adaptive modules, the framework advances both the accuracy and reliability of wind speed prediction.
Förderung
- Natural Science Foundation of Jilin Province
Zitieren
APA
Liu, G., Wang, X., Li, Q., & Xu, B. (2026). A Hybrid LSTM‐KAN Framework With Dynamic Residual and Adaptive Grid for Robust Wind Speed Forecasting. Wind Energy, 29(9). https://doi.org/10.1002/we.70146
Vancouver
Liu G, Wang X, Li Q, Xu B. A Hybrid LSTM‐KAN Framework With Dynamic Residual and Adaptive Grid for Robust Wind Speed Forecasting. Wind Energy. 2026;29(9). doi:10.1002/we.70146
BibTeX
@article{Liu2026,
author = {Liu, Guangjie and Wang, Xinyu and Li, Qingliang and Xu, Bo},
title = {A Hybrid LSTM‐KAN Framework With Dynamic Residual and Adaptive Grid for Robust Wind Speed Forecasting},
journal = {Wind Energy},
year = {2026},
volume = {29},
number = {9},
doi = {10.1002/we.70146}
}
EUVIMED