Answers are grounded in this paper and cite their sources.
The share of wind energy in global electricity generation is rising, driving demand for accurate and interpretable short-term generation forecast models that combine physical consistency with the ability to capture the nonlinear dynamics of meteorological time series. We propose a hybrid architecture: a physics-informed Kolmogorov–Arnold network (KAN) paired with an XGBoost-based residual corrector and a residual damping mechanism. The KAN core encodes wind turbine operational constraints—such as the Betz limit, power curve monotonicity, and cut-in/cut-out thresholds—via learnable B-spline activation functions on the network edges and structured loss function terms. XGBoost acts as a weak corrector: it learns from the core’s residual errors and relies on exogenous physical features (ERA5 reanalysis, the Jensen/Park wake model, air density, and inter-turbine wake effects), while a damping coefficient regulates its contribution. The evaluation was conducted on 58,496 hourly records from four wind farms. Two different problem formulations can be considered: one that accounts for lags of the output variable, and another where such lags are intentionally excluded from the set of input data. In our study, we compare the proposed model with baseline models XGBoost and LightGBM. The metrics on the test set for excluded lagged values of the target variable are MAE = 1385 kWh and R2 = 0.846 for XGBoost and MAE = 1460 kWh and R2 = 0.842 for the proposed interpretable hybrid model, where the kernel’s contribution was 91.1% and the MAE of the standalone KAN improved by 6–8%. For the alternative problem formulation that incorporates lagged values of the target variable, the proposed hybrid model yielded MAE = 790 kWh and R2 = 0.947, while XGBoost and LightGBM achieved an MAE of approximately 767–770 kWh and R2 = 0.948–0.949 on the same test set. The proposed model offers greater transparency and interpretability, allows for the inclusion of components based on physical principles, while remaining comparable in accuracy to XGBoost or LightGBM models in both problem formulations. Thus, the proposed architecture offers a controlled trade-off between accuracy and interpretability.
No takes yet. Share an insight, caveat, or question.
Kochueva et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: