acceptodds
Under review as a conference paper at ICLR 2027

FLOATSense: A Time-Series Dataset and Benchmark for Fatigue Load Reconstruction on Floating Offshore Wind Turbine Towers

Abstract

Fatigue limits the lifetime of an offshore wind turbine tower and is tracked from the bending moment along it, yet strain gauges are rare and cover one section, while a tower-top accelerometer is standard. Reconstructing the moment at every height from it is a time-series problem, and no public benchmark scores it by fatigue damage. We introduce FLOATSense, a dataset and benchmark for fatigue load reconstruction on 22 MW floating offshore wind turbines: 717,948 time series at 10 Hz from the 19,404 OpenFAST simulations of FLOATBench on three tower geometries, with accelerometer and SCADA inputs and the fore-aft moment at eleven heights as target. A reconstruction is scored by the fatigue damage it implies, not by waveform error, and the six realizations of an operating point are kept together, so a model that reads only the condition gets no credit on the within-condition metric. Twenty learned models and an adapted field-validated physics method are evaluated within tower, across towers and under sensor ablation. Sensor-level leaderboards mislead: base log-damage saturates among the leading models and its leaders collapse at the top, where at the declared budget only three learned models beat a one-gain floor and a condition-only control given the true operating condition exceeds them all in aggregate , so realization tracking must be read within condition; longer training recovers the top for the two TCNs, not for the tested attention and state-space models; five labeled simulations of a new tower recover its base by fine-tuning and its top by one scale per height, which fine-tuning often misses; and a second accelerometer axis improves the top at every budget. Limitations: the campaign is simulated, wind and waves are aligned, and only the fore-aft moment is scored. Resources: data and code at https://anonymous.4open.science/w/FLOATSense-56B2/review/index.html.

Primary area:
datasets and benchmarks
Keywords:
benchmark dataset, time-series regression, load reconstruction, inverse problems, virtual sensing, scientific machine learning, floating offshore wind turbine, structural fatigue

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_AKF9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces FLOATSense, a dataset and benchmark for reconstructing the bending moment at eleven heights of floating offshore wind turbine towers from a tower-top accelerometer and SCADA signals, scored by the implied fatigue damage rather than by waveform error. The data come from 19,404 OpenFAST simulations of FLOATBench on three 22 MW towers, keeping the six turbulence realizations of each operating point in one split.

Twenty learned models, a physics method and a one-gain floor are evaluated within and across towers and under sensor ablation. Base R² saturates among the leading models while the top separates them, and the two rankings disagree. A condition-only control beats every entry in aggregate top R² but has zero within-condition correlation, so aggregate R² measures the damage level rather than realization tracking.

Critical strengths
  1. Field towers rarely have strain gauges beyond one section, whereas a tower-top accelerometer is standard, and no public benchmark scored this task by fatigue damage.
  2. The condition-only control reaches a top R² of 0.95 with ρ_wc = 0, and adding colored noise to the TCN output raises its top R² from 0.83/0.64/0.71 to 0.95/0.88/0.88 with ρ_wc essentially unchanged.
Critical weaknesses
  1. Nearly all models use 50 epochs at 10⁻³ with no schedule. The authors note that "a top score measures convergence as much as architecture," yet the abstract states that longer training recovers the top "for the two TCNs, not for the tested attention and state-space models."
Questions for authors
  1. Were the Hybrid results produced with the public implementation? Was base_correction constrained in any other way? What is the actual range of |1+c| on the test set?
3 comments
Reviewer_AKF9OpenReview ↗

update

I have corrected the “Flag For Ethics Review” field, which I had accidentally marked as “Yes.”

AuthorsOpenReview ↗

Expanded HPO, Hybrid Verification, and Corrections

We thank the reviewer for the careful assessment and for examining our code. We particularly appreciate their recognition of the practical relevance of reconstructing tower loads from available sensors, the novelty of providing the first public benchmark for evaluating this task through fatigue damage, and how the condition-only and colored-noise experiments demonstrate that aggregate accuracy can hide poor tracking of individual realizations.

We address these concerns by revising the abstract and expanding the evaluation to 2,340 training runs across all 20 models, distinguishing training effects from architectural conclusions. We also clarified the Hybrid's unconstrained and bounded components and verified that the public code reproduces all reported results.

Q1. Were the Hybrid results produced with the public code? Yes. We reran all nine Hybrid models with a fresh copy of the public code on the full test set, and every result matches the paper within 3×10⁻⁶. The models are included in the released data (see our comment "Complete Dataset and Evaluation Materials"), so these results can be checked directly.

Q2. Was base_correction constrained in any other way? No. Our original description was incomplete: it mentioned only the bounded term, which implied that the entire correction was bounded by ±0.5. In fact, base_correction is a free per-frequency term, while only the condition-dependent component is bounded through 0.5 × tanh. We have clarified this in all four relevant locations: Section 4, Table 10, Appendix F, and Appendix G. The implementation and reported results remain unchanged.

Q3. What is the actual range of |1+c| on the test set? Across the nine Hybrid checkpoints, we computed |1+c| over the full test set and the scored frequency band (0–3 Hz). The observed values range from 0.00 to 3.39. Although 65% of frequency bins have gains outside [0.5, 1.5], these bins account for only 8% of the signal energy. In other words, the largest deviations from the physics gain occur mainly at frequencies with little signal energy, while the energy-weighted median gain is 0.97. This behavior is observed, not imposed by the model (Appendix G.1).

W1. The 50-epoch recipe and the abstract claim. We addressed the reviewer's concern in two ways: revising the abstract and expanding the evaluation to all 20 models. The abstract now limits its longer-training claim to the two convolutional models under a fixed training budget, without drawing conclusions about attention or state-space architectures.

We conducted validation-based hyperparameter optimization, varying learning rates, weight decay, schedules, and architecture-specific parameters. Each model received 30 trials per tower (1,800 trials). The two best configurations were confirmed over three seeds at 300 epochs, and the selected configuration was retrained for 300 epochs with three seeds. This required 2,340 training runs and at most 1,478 GPU-hours (Appendix K).

Top R² for the five best tuned models, median of three seeds, averaged over the three towers (all 20 in Table 34):

Model50 epochsTuned
Prob-TCN0.860.98
TCN0.760.95
U-Net0.210.95
LSTM0.420.91
FNO0.050.75

Tuning substantially improves performance at the tower top: 18 of 20 models improve, with a median R² gain of 0.23, compared with 0.03 at the base. Prob-TCN remains first at the top, while U-Net and LSTM become competitive. Attention and state-space models also improve: S4D rises from 0.44 to 0.70, Mamba from −0.29 to 0.47, and iTransformer from −0.14 to 0.38.

These comparisons still use a common training loss. With a damage-aware loss, S4D reaches 0.88–0.92 at the top (seed 0; Appendix J.2), so we avoid broader claims about architectural superiority.

The validation-tuned track complements the fixed-budget analyses and confirms their main within-tower findings (Appendix K); rerunning the cross-tower, few-shot and sensor analyses with the tuned configurations is our next step. The central findings remain: performance varies substantially with tower height, and strong aggregate accuracy can still hide poor realization tracking.

Summary: We have addressed all three questions and corrected the Hybrid description without changing any result. We also tuned all 20 models (2,340 training runs), and the abstract no longer draws a conclusion about architectures from a single recipe. The central findings remain robust. We hope these improvements warrant reconsideration of your overall assessment.

Thank you for the detailed response and additional experiments. I maintain my weak-accept recommendation.

Reviewer_KxmGOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FLOATSense, a time-series dataset and benchmark for reconstructing fatigue-relevant bending-moment histories along floating offshore wind turbine towers from observable sensor signals. The dataset is derived from 19,404 OpenFAST simulations of three 22 MW tower geometries and contains 717,948 individual channel time series, with fore-aft bending moments at eleven tower heights as reconstruction targets. The benchmark evaluates the physical consequences of reconstructed load histories through fatigue damage rather than relying exclusively on conventional waveform reconstruction metrics. Its evaluation protocol includes operating-regime generalization, cross-tower transfer, few-shot adaptation, and sensor ablations. Importantly, it distinguishes aggregate damage prediction from within-condition tracking of stochastic realizations.

The authors evaluate twenty learned models from multiple architecture families, together with an adapted physics-based baseline and diagnostic controls. The experiments reveal that strong aggregate performance at the tower base does not necessarily translate into reliable fatigue estimation near the tower top. They also show that checkpoint convergence, training budgets, and additional accelerometer measurements can substantially influence performance.

Overall, the paper contributes a domain-specific scientific machine learning benchmark with a well-motivated physical objective and several informative findings regarding the limitations of existing time-series models.

Critical strengths

S1. Unlike conventional reconstruction benchmarks that primarily evaluate pointwise waveform accuracy, FLOATSense evaluates reconstructed bending-moment histories through their implied fatigue damage at multiple tower heights. The distinction between aggregate damage prediction and within-condition realization tracking is particularly important. The finding that a condition-only predictor can achieve high aggregate damage accuracy while failing to capture stochastic realization differences demonstrates a genuine limitation of commonly used evaluation metrics.

S2. The benchmark combines operating-regime shifts, cross-structure transfer, few-shot adaptation, sensor ablations, and comparisons against a physics-based baseline. The grouping of stochastic realizations by operating condition is an appropriate experimental design choice. The extensive analysis of tower-height-dependent performance, training-budget sensitivity, and additional sensor measurements provides actionable insights for future scientific machine learning research.

Critical weaknesses

W1. The paper's principal contribution is a reusable dataset and benchmark, but during review only 47 simulations per tower are available as synchronized time series; the full release is deferred. The complete score distribution, split manifest, and headline model results therefore cannot be independently reproduced from the review materials. This substantially limits assessment of the central resource claim.

W2. All target moments come from one simulated 22 MW turbine family with a common platform and controller, aligned wind and waves, and supervised labels at eleven heights. The default model ranking is also a fixed 50-epoch track, with longer runs changing the top performance of selected TCNs. These choices define a useful simulation benchmark but do not establish field transfer or a stable architecture ranking.

Questions for authors

Q1. What evidence supports the physical representativeness of FLOATSense beyond the OpenFAST simulation environment? Although the physics-based baseline is adapted from a field-validated method, the benchmark itself contains no measured load histories. Can the authors provide additional validation against independent physical measurements, higher-fidelity simulations, or established engineering reference quantities?

Q2. Can the complete dataset and evaluation pipeline be made available during the review process? The manuscript indicates that only a fixed subset of 47 simulations per tower is accessible during review, with full release planned upon publication. Since the dataset is a central contribution, access to the complete benchmark would be important for verifying the data partitions, evaluation metrics, and reported leaderboard results.

1 comment
AuthorsOpenReview ↗

Physical Validation, Dataset Release, and Training Robustness

We thank the reviewer for the thoughtful and positive assessment. We particularly appreciate their recognition of the value of fatigue-based evaluation, how the condition-only baseline reveals limitations of aggregate metrics, and the insights provided by our generalization experiments, sensor ablations, and physics-based comparisons.

Following the reviewer's feedback, we have made the complete dataset and evaluation materials available, provided further evidence of the physical representativeness of our simulations, and clarified their limitations. We have also substantially strengthened the experimental evaluation, conducting 2,340 training runs across all 20 models to address the concerns about the fixed training budget.

Q1. Physical representativeness beyond OpenFAST. The wind and wave conditions used in our simulations were derived from 22 years of ERA5 reanalysis data for a real offshore site west of the Isle of Barra. At the 22 MW scale, equivalent field measurements of tower loads are unavailable. Simulations are therefore essential for designing and assessing these turbines before construction, following established engineering practice (IEC 61400-3-2).

OpenFAST has been compared with experimental wave-tank measurements in the OC5 and OC6 projects, showing generally good agreement, with some documented discrepancies. We also compared our simulations with the published IEA 22 MW reference, finding agreement within 0.4% for power and rotor speed above rated wind (Appendix E.4). Together, these comparisons support the physical consistency of our simulations, including the operating conditions that drive tower loads, although direct load validation at 22 MW remains unavailable. Higher-fidelity FEA and CFD comparisons remain future work. Section 6 now clarifies this evidence and its limitations.

Q2/W1. The complete dataset during review. Initially, we provided 47 simulations per tower due to the size of the complete time-series dataset (24 GB). Following the reviewer's feedback, we have now made all 19,404 simulations available in an anonymous Hugging Face repository (see our comment "Complete Dataset and Evaluation Materials").

The release also includes the data splits, reference fatigue damage, model predictions, trained checkpoints, and evaluation code. Reviewers can therefore independently reproduce every leaderboard result without retraining. We thank the reviewer for highlighting the importance of full access during review.

W2. One turbine family and a fixed training budget. We agree that the benchmark's scope should be explicit. FLOATSense focuses on generalization across operating conditions and three tower designs within one turbine family, sharing the same turbine, platform, and controller. Wind and waves are aligned, and training uses reference loads at eleven heights. Section 6 now clarifies these limitations and that transfer to other turbine families or field measurements remains untested. The evaluation protocol can be extended to additional designs and operating conditions.

On the 50-epoch budget, we conducted a new validation-tuned evaluation of all 20 models, with 1,800 hyperparameter optimization trials, followed by confirmation and final training at 300 epochs over three seeds. In total, this required 2,340 training runs and at most 1,478 GPU-hours (Appendix K).

ModelTop R² (50 ep.)Top R² (Tuned)
Prob-TCN0.860.98
TCN0.760.95
U-Net0.210.95
LSTM0.420.91
Mamba−0.290.47

Across all 20 models, 18 improve at the top, with a median R² gain of 0.23. Rank correlations are 0.71 at the top and 0.85 across the eleven heights (Table 34). Prob-TCN stays first at the top, while U-Net and LSTM become competitive.

The main findings remain after tuning: the tower base is substantially easier than the top, and strong aggregate accuracy can hide poor realization tracking. The validation-tuned track complements the fixed-budget analyses and confirms their main within-tower findings (Appendix K); rerunning the cross-tower, few-shot and sensor analyses with the tuned configurations is our next step.

Summary: We have addressed both questions by releasing the complete dataset and evaluation materials and clarifying the physical validation evidence. We have also made the benchmark's limitations explicit and substantially strengthened the experimental evaluation with 2,340 training runs across all 20 models. The central findings remain robust. We hope these improvements address your concerns and warrant reconsideration of your overall assessment.

Reviewer_LLPNOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents FLOATSense, a time-series benchmark dataset designed to reconstruct the history of bending moments along an offshore wind turbine tower from sensor measurements and other readily available operational data, with the ultimate goal of estimating fatigue damage. Essentially, inferring or reconstructing a latent physical state from indirect/imperfect observations. It offers a substantial simulation-derived dataset and evaluates various reconstruction approaches under different operating conditions, sensor configurations, and tower geometries.

Critical strengths

This is a useful dataset and a relevant evaluation problem.

This addresses a concrete problem, in a specific form: certain physical quantities of interest cannot be measured everywhere, or would require costly instrumentation. I particularly appreciate that the authors do not limit themselves to evaluating the reconstruction error of the time series, but also take into account the final quantity of interest, namely fatigue damage. The analysis of generalization across different operating conditions, sensor configurations, and tower geometries makes this benchmark much more informative than a simple comparison of predictive accuracy.

The analysis of reconstruction performance by location and operating condition is valuable. In particular, distinguishing the ability to predict typical damage levels from the ability to reconstruct variations associated with specific operating situations highlights limitations that aggregated metrics could mask.

Critical weaknesses

The general problem and the broader relevance to machine learning (ML) could be made more explicit. The paper would benefit from at least some effort toward framing its contribution as a general problem of reconstructing latent physical responses from indirect observations. This approach extends beyond wind turbines: many applications require inferring unobserved physical quantities from sensor measurements, with the reconstructed signals then used for downstream analyses or decision-making. Making this abstraction explicit would help readers distinguish what in FLOATSense is specific to wind turbine fatigue analysis from what makes this dataset and evaluation protocol useful to a broader ML community. In addition, a writing effort more oriented toward the human reader would be beneficial. Without having read the paper or being an expert in the wind energy domain, the second half of the abstract is difficult to grasp, which is likely to discourage some readers.

The nature of the 'ground truth' and the scope of the conclusions drawn from it should be made explicit. The target signals of this benchmark are generated by simulation, which is a reasonable and potentially useful approach for building a large dataset but it does determines what the benchmark can actually evaluate: the performance of reconstructing simulated responses, and not necessarily the fidelity of reconstructions applied to real turbines. A more explicit discussion of this distinction, and of course of the extent to which conclusions drawn from this dataset are transferable to real-world measurements, would strengthen the paper; yet still avoiding treating simulation-derived labels as inherently problematic - which is not the case.

Questions for authors

Can we say that the evaluation protocols and metrics are appropriate, reproducible, and informative enough to support meaningful future comparisons?

Can you clarify exactly how is the data split to test generalization?

How representative/generalizable is the fixed training budget (of 50, as stated in the paper)?

2 comments
AuthorsOpenReview ↗

Evaluation Protocols, HPO, and Broader ML Relevance

We thank the reviewer for the constructive and encouraging assessment, particularly for recognizing FLOATSense as a useful dataset addressing a relevant problem, and its contributions to downstream fatigue evaluation, generalization, and revealing limitations hidden by aggregate metrics.

We address the three questions with clearer evaluation protocols and a substantially stronger experimental evaluation, including 2,340 training runs across all 20 models.

Q1. Are the protocols and metrics appropriate, reproducible and informative? Yes. We evaluate models on fatigue damage, the quantity engineers use to assess structural lifetime. We also report how well they track individual runs of the same weather condition (Table 24), since a good aggregate score can hide poor tracking. Confidence intervals resample whole weather conditions because their repetitions are not independent. The complete dataset, splits, scoring script, code and trained models are available (see our comment "Complete Dataset and Evaluation Materials"). Future models can therefore be compared on both damage accuracy and run tracking.

Q2. How exactly is the data split? We use the regime-aware split of FLOATBench, the same on the three towers. We split the data by weather condition, so all six realizations of a condition stay together. Each tower has 22 wind speeds and 49 sea states (7 wave heights × 7 wave periods), giving 1,078 conditions and 6,468 simulations.

SplitConditionsSimulationsSelection
Training2881,72818 wind speeds × inner 4 × 4 sea states
Validation (within training)5834820% of training conditions
Test7904,740All remaining conditions

Each test simulation is classified into nine wind–wave regimes (IT, IP, EX for each), using standardized distances to training samples and an alpha-shape boundary (Appendix E.2).

Q3. How representative is the fixed budget of 50 epochs? The 50-epoch results are broadly representative at the base, but less so at the top. To assess this, we conducted a new validation-tuned evaluation with hyperparameter optimization (HPO) for all 20 models (Appendix K). Each model received 30 trials per tower (1,800 trials). The two best configurations were confirmed over three seeds at 300 epochs, and the selected configuration was retrained for 300 epochs with three seeds. This required at most 1,478 GPU-hours.

Tuning improves performance, particularly near the tower top:

ModelTop R² (50 ep.)Top R² (HPO + 300 ep.)
Prob-TCN0.860.98
TCN0.760.95
U-Net0.210.95
LSTM0.420.91
Mamba−0.290.47

Across all 20 models, median R² gains are 0.03 at the base and 0.23 at the top, with 18 improving at the top. Rank correlations between the two tracks are 0.82 and 0.71, respectively. Prob-TCN stays first at the top, while U-Net and LSTM become competitive (Table 34).

The central findings remain: the base is easier than the top, performance varies along the tower, and aggregate accuracy can hide poor realization tracking. The validation-tuned track complements the fixed-budget analyses and confirms their main within-tower findings (Appendix K); rerunning the cross-tower, few-shot and sensor analyses with the tuned configurations is our next step.

W1. Framing for a general ML audience and readability. We revised the abstract and introduction to frame FLOATSense as a general inverse time-series problem: reconstructing unobserved physical responses from sensor measurements for downstream decisions. We clarify the broader ML challenges: downstream evaluation, realization tracking, and generalization across conditions, sensors, and structures, and distinguish them from the wind-specific application. We also rewrote the second half of the abstract in simpler language and improved readability in Sections 3–5.

W2. Simulated ground truth. We agree that the scope of simulation-derived conclusions should be clearer. Sections 3 and 6 now distinguish simulated reconstruction accuracy from real-world performance. Floating turbines at the 22 MW scale do not yet exist physically, and equivalent public field data are unavailable. Simulation is therefore essential for designing and assessing such turbines before construction. OpenFAST has also been compared with experimental measurements, with generally good agreement but known limitations. FLOATSense supports the development of virtual load sensors and sensor selection for future turbines.

Summary: We have addressed all three questions, substantially strengthened the experimental evaluation through hyperparameter optimization of all 20 models, and clarified FLOATSense's broader ML relevance and real-world potential. The central methodological findings remain robust. We hope these improvements warrant reconsideration of your overall assessment.

Thank you to the authors for providing the clarifications, answers to questions, and for enacting changes to the manuscript to address comments. I remain positive about this paper and its contributions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.