| 29 Jun 2026
Brief communication: Enhanced wind turbine fatigue load estimation using digital shadows with data-driven bias correction
Abstract. This study proposes a digital shadow framework for wind turbine load estimation that integrates a linearized industrial-grade aeroelastic model with a deep learning–based bias correction (BC) method. To address model mismatches and limited inflow representation, a learning-based bias correction strategy is introduced, where static bias terms are first calibrated via wind-speed-dependent fitting, followed by perturbed correction profiles and parametric simulations to construct a digital shadow dataset. A neural network (NN) is then trained to map operating conditions and bias parameters to load estimation errors, enabling adaptive correction under unseen conditions.
The proposed method is validated using field data spanning diverse inflow conditions, achieving a reduction in blade bending moment DEL prediction errors at the 25 % span location from 15–25 % to below 5 %. This demonstrates strong robustness and improved capture of inflow–structure interactions. Overall, the framework provides a scalable pathway to data-driven digital shadows and a foundation for future digital twin applications in real-time load estimation and operational optimization.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Wind Energy Science.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.
The manuscript proposes a learning-based bias correction for a wind-turbine digital shadow and reports reductions in damage-equivalent-load errors. However, the validation does not presently support the stated claims of robustness, generalization, or broad applicability.
The main concern is that the correction is trained using blade-root bending-moment errors, while blade-root bending moments are also included among the measured outputs assimilated by the Kalman filter. The method therefore depends on instrumented blade-load measurements rather than ordinary SCADA data. The manuscript should clearly distinguish this from a SCADA-based load estimator and explain what practical benefit remains when blade strain-gauge measurements are already required. The estimated 25\% span load is not used directly in the filter, but it is strongly related to the measured root load used to construct the correction. The reported accuracy cannot therefore be interpreted as independent reconstruction of unmeasured loading from routine turbine signals.
The experimental validation is also too limited for the conclusions drawn. All results come from one turbine and approximately two weeks of measurements. The nominal validation sets contain only 6 h of simple inflow and 3 h of complex inflow, while the two additional tests contain 4.5 h and 3 h. These periods are drawn from the same turbine, site, campaign, and season as the training data. No independent turbine, campaign, or seasonal dataset is used. Consequently, the assertion that the method is robust under unseen operating conditions is not demonstrated.
The construction of the training and validation data requires more transparency. The baseline correction is obtained from a “short representative dataset,” but the data included in this calibration are not given. It is therefore impossible to determine whether the reported validation periods are fully independent of bias calibration, architecture selection, feature selection, and hyperparameter tuning. A locked test set, selected before model development, is required.
The damage-equivalent-load calculation is insufficiently documented. The manuscript does not state the Wöhler exponent, rainflow-counting procedure, reference cycle count or duration, treatment of means, filtering, or uncertainty of the reported DEL errors. DEL values computed from intervals of only 3–6 h may be highly sensitive to the particular load sequence. Reporting a percentage difference for one short period is not sufficient to establish fatigue-load accuracy.
The comparison also lacks appropriate baselines. The learning-based correction is compared mainly with the uncorrected digital shadow. There is no comparison with a wind-speed-dependent lookup-table correction, direct regression, conventional Kalman bias estimation, or another simple data-driven model. Since the initial correction is already fitted as a function of wind speed, an ablation study is needed to show that the neural network provides a material improvement beyond the underlying fitted correction and the use of measured blade-root loads.
Several statements and results are internally inconsistent. The abstract and conclusions state that DEL errors are reduced from 15–25\% to below 5\%, but Table 3 reports a baseline error of 14\% for simple inflow, and Table 5 reports only 6\% for the near-sunrise case. For the post-sunset case, the text states that the network trained on complex inflow gives a DEL error of 5\%, whereas Table 4 reports 3\%. Figure 6 also labels two different panels as panel (b). These discrepancies affect the headline quantitative claims.
The measurements are described as being used without calibration or post-processing, apart from removal of invalid operating periods. Strain-gauge bending moments cannot automatically be treated as an exact reference without documenting calibration, offset and drift correction, synchronization, temperature compensation, and measurement uncertainty. The reported differences of only a few percent are comparable to plausible measurement and calibration errors, yet no uncertainty bounds are provided.
Finally, the proprietary encrypted aeroelastic model, normalized correction factors, and unavailable underlying field dataset prevent independent reproduction of the principal results. Providing plotted or processed data does not resolve the inability to reconstruct the model, training process, or reference loads.
These issues concern the central validation strategy rather than presentation. Establishing the claimed performance would require an independently defined test set, complete documentation of the DEL calculation and measurement calibration, comparisons against simpler correction methods, and validation on additional turbines or campaigns. The present evidence is insufficient to support the main conclusions, and a decline-to-publish decision is warranted.