PaperlayerView exact revision on arXiv

Source-linked AI summary

On the Limits of Univariate Deep Learning for Significant Wave Height Forecasting

Yilin Zhai, Hongyuan Shi, Zaijin You

arXiv:2609.30688v1physics.ao-phcs.LGphysics.comp-ph

TL;DR

Prior univariate SWH comparisons left open whether modern architectures gain skill at long contexts or across more buoys. This study systematically compares five architectures and context lengths, then evaluates configurations across a multi-buoy corpus; architecture differences are small relative to cross-dataset variation.

  • Problem

    Prior comparisons left open whether complex architectures outperform simple univariate SWH models at long contexts, beyond two tested lengths and a small buoy set.

    Source: §1 · p. 3

  • Method

    The study compares five forecasting paradigms across 1–168 h contexts and re-evaluates selected configurations on 47 NDBC buoys spanning 37 years.

    Source: §1 · p. 4

  • Results

    Between-family SD was 0.0014 m2 (0.8% of the grand mean), versus a 4.83× cross-dataset MSE shift; all 45 multi-buoy trials beat persistence (mean skill +0.062).

    Source: §5 · p. 26

  • Takeaways & Limitations

    For univariate Hs forecasting, architecture engineering has reached diminishing returns, with persistence capturing the dominant linear-inertial signal and deep models adding a smaller correction.

    Source: §5 · p. 27

    Single-buoy hyperparameter selection is tuned to a specific wave climate, and its model ranking does not transfer to heterogeneous buoy climates.

    Source: §4.3 · p. 22

Abstract

from arXiv · show

This study conducts a systematic hyperparameter search across five deep learning architectures, DLinear, LSTM, PatchTST, ResAttLstm, and Mamba2, and nine context lengths (1-168 h) for single-station significant wave height (Hs) forecasting on NDBC buoy 41009, followed by re-evaluation of the best configurations on a 47-buoy, 37-year corpus. The five families converge to a common performance level on the multi-buoy evaluation (between-family SD = 0.0014 m^2, 0.8% of the grand mean), a spread dwarfed by the 4.83x cross-dataset MSE shift between buoy corpora. All multi-buoy trials beat persistence (mean skill +0.062), but no architecture consistently outperforms the others. On the single-buoy experiment, skill peaks at 12-24 h where five trials fall below persistence, per-family Q4/Q3 test MSE ratios range from 2.4 to 2.6, and deep models underperform persistence for the most extreme 1% of waves. These findings are consistent with the interpretation that persistence already captures the dominant linear-inertial signal in univariate Hs, and that architecture engineering under this univariate input setting has reached diminishing returns: cross-buoy variance, not model class, dominates forecast error. Future work should prioritise atmospheric covariates, zero-shot cross-buoy transfer, and decomposition of Hs into swell and wind-sea components. By establishing a rigorous reference baseline for what univariate Hs models can and cannot achieve, this study provides a benchmark against which future multivariate and physics-informed approaches can be calibrated, and offers practical guidance for lightweight buoy-level forecasting in mid-latitude storm-dominated and swell-mixed environments.

1. Introduction

Significant wave height (Hs) is central to maritime and coastal operations, while atmospheric-forced numerical forecasts require resources that limit lightweight buoy-level deployment. This study tests whether architecture and longer input contexts improve univariate Hs forecasts beyond persistence, then evaluates whether single-buoy findings extend across a broader buoy corpus.

1. Introduction

  • Atmospheric-forced numerical wave models achieve 0.95–0.96 correlation at Day 1 but require global atmospheric forcing and substantial resources, motivating forecasts from a buoy’s own Hs record.Hs is identified as a primary variable for ship routing, offshore operations, coastal flood warning, and harbour safety.Source: §1 · p. 2
  • Prior univariate SWH work spans recurrent, transformer, multi-scale convolutional, and selective state-space models, building on statistical methods and early LSTM studies.Transformer applications tested the hypothesis that self-attention can extract non-linear wave-field evolution over extended contexts.Source: §1 · pp. 2–3
  • Earlier evidence left open whether complex architectures outperform simpler models at long contexts, since Jiang et al. (2024) tested only 6 and 24 h on a small buoy set.Prior work also lacked unified controlled comparisons and multi-architecture evaluations across heterogeneous buoy climates.Source: §1 · p. 3
  • The study compares five architectures across contexts from 1 to 168 h and re-evaluates configurations on 47 NDBC buoys spanning 37 years.Its questions ask whether models beat persistence, whether longer contexts expose additional non-linear structure, and whether single-buoy findings replicate across buoys.Source: §1 · pp. 3–4

2. Methodology

The study first searches model families and context lengths on one buoy, then tests the selected configurations on a larger multi-buoy corpus. Its methodology defines the datasets, univariate six-step forecasting task, preprocessing, training, and evaluation protocol.

2.1. Experimental design

  • The two-phase design searches five model families across nine context lengths on station 41009, then re-evaluates the 45 validation-selected configurations on 47 buoys.The single-station phase uses 2,700 Optuna trials; the multi-buoy phase uses chronological splits and the same preprocessing, normalization, loss, and metrics.Source: §2.1 · p. 4

2.2. Datasets

  • The single-buoy dataset covers NDBC station 41009 from 1988–2025, while the multi-buoy corpus spans 47 Atlantic, Pacific, and Gulf of Alaska buoys.The corpus contains approximately 1.2 × 10^7 hourly samples, with record lengths from 12 to 37 years.Source: §2.2 · p. 5

2.3. Task and data preparation

  • The task predicts six successive hourly Hs values from Hs history alone, using context lengths from 1 to 168 hours and chronological 70/15/15 splits.Sliding windows are constructed after partitioning so none crosses a train-validation or validation-test boundary; log transformation and median/IQR scaling are used.Source: §2.3 · pp. 6–7

2.4. Models, training, and evaluation

  • Five architectures—DLinear, LSTM, PatchTST, ResAttLstm, and Mamba2—are trained with Adam, cosine annealing, and validation-MSE-based Optuna TPE search.The per-cell search uses early stopping; search spaces are held identical across context lengths.Source: §2.4 · pp. 8–9

2.5. Evaluation metrics and statistical analysis

  • Persistence is the primary baseline, and persistence skill is the primary evaluation metric, with MSE, R2, and Pearson correlation also reported.Metrics are reported after inverse transformation in square metres; p99 persistence skill evaluates the largest 1% of test waves.Source: §2.3 · p. 8, §2.4 · p. 9, §2.5 · p. 10
  • Effect sizes are used instead of formal significance tests, with standard errors for means and a bootstrap 95% interval for the cross-dataset MSE ratio.The bootstrap uses 10,000 resamples; the authors state that very large test sets could make trivially small differences statistically significant.Source: §2.5 · p. 10

3. Results

Across the single- and multi-buoy evaluations, the five architectures show little separation, while error changes substantially across buoy corpora. Persistence skill, seasonal performance, tail errors, and missing-data robustness vary with context and evaluation setting.

3.1. Hyperparameter Search

  • 0.0004 m2 was the single-buoy between-family test-MSE SD, while the 0.0330–0.0354 m2 context-length range exceeded the 0.0333–0.0343 m2 family range.The per-family results are means across nine context lengths.Source: §3.1 · p. 11, Table 3 · p. 11, Figure 4 · p. 12
  • Persistence skill peaked at CL = 12–24 h (+0.036 to +0.057); five trials had negative skill, only 6/45 had positive p99 skill, and per-family Q4/Q3 test MSE ratios were 2.4–2.6.At CL = 1, models added negligible value beyond the last observation.Source: §3.1 · p. 11, Table 3 · p. 11, Figure 5 · p. 12

3.2. At-scale evaluation

  • 0.1631 m2 was the multi-buoy five-family grand mean test MSE, with a 0.0014 m2 between-family SD (0.8% of the mean); all 45 trials had positive persistence skill (mean +0.062), with no consistent architecture winner.The reported test MSE distribution and persistence-skill figure accompany these comparisons.Source: §3.2 · p. 13, Figure 6 · p. 14, Figure 7 · p. 14
  • The five-family grand mean p99 persistence skill was −0.040; deep models still underperformed persistence on average for the most extreme 1% of multi-buoy samples.Only 15 of 45 trials achieved positive p99 skill, up from 6/45 on the single-buoy set.Source: §3.2 · p. 13, Figure 7 · p. 14
  • Q4 test MSE exceeded Q3 by a factor of 2.8–3.2 across families, and Q4 skill was below Q3 skill in 42 of 45 cells.The five-family grand mean skill was Q3 = +0.084 and Q4 = +0.055; the pattern was not solely an artefact of higher SWH variance.Source: §3.2 · p. 13, Figure 8 · p. 15, Table 4 · p. 16
  • No configuration simultaneously minimised MSE and maximised persistence skill: short-CL ResAttLstm/PatchTST and long-CL PatchTST/DLinear primarily formed the Pareto front.Figure 9 plots test MSE against persistence skill for each family/context-length cell.Source: §3.2 · p. 13, Figure 9 · p. 15

3.3. Cross-dataset transfer

  • 4.83 was the mean ratio by which 47-buoy test MSE exceeded single-buoy test MSE; family/context rankings also reversed between datasets.The ratio spans 4.26–5.15 across 45 pairs; the cross-dataset ratio is shown in Figure 10.Source: §3.3 · p. 17, Figure 10 · p. 17, Figure 11 · p. 18

3.4. Robustness to missing observations

  • With 5% and 15% missing observations linearly interpolated, the five-family mean test MSE decreased by 0.6% and 2.7%, persistence skill shifted by −0.0002 and −0.0006, and family ranking was preserved.The configurations were re-evaluated without retraining.Source: §3.4 · p. 18, Table 5 · p. 18

4. Discussion

The discussion interprets architectural convergence as a ceiling imposed by univariate Hs information, while showing that context-length behavior differs between the single-buoy and multi-buoy settings. It bounds these findings to the study’s data, environments, and methodological choices.

4.1. Why the five families are of no practical consequence

  • The authors hypothesize that all five families learn the same linear-inertial Hs signal already captured by persistence, but the experiment cannot distinguish this from a swell/wind-sea explanation.Under the two-component hypothesis, wind-sea becomes unpredictable from Hs alone below approximately 24 h, while swell may provide a small, slowly saturating residual signal; explicit decomposition is left to future work.Source: §4.1 · pp. 19–20
  • Single-point Hs histories omit spatial processes such as harbour resonance, Bragg reflection, and gap resonance, imposing a separate ceiling that architecture alone cannot overcome.The forecaster lacks wave direction, basin geometry, and inter-body spacing available to phase-resolving models.Source: §4.1 · p. 20

4.2. Context-length scaling

  • On 47 buoys, mean skill rises from +0.023 at CL = 1 to +0.077 at CL = 96, then plateaus at +0.076 through CL = 168, unlike the single-buoy 12–24 h peak.The best single-trial skill is +0.1011 at CL = 96 h (PatchTST); the authors caution that context-length trends on one buoy do not transfer to the multi-buoy setting.Source: §4.2 · pp. 20–21

4.3. Limitations and generalisability

  • The best architecture ranking on station 41009 does not transfer to the 47-buoy corpus, and longer-context test windows are composition-dependent subsets of shorter-context windows.The authors recommend multi-buoy validation for model selection; CL = 168 requires 174 consecutive hours without gaps, compared with twelve hours for CL = 6.Source: §4.3 · pp. 22–23
  • Best-trial selection, log-MSE training, and six-lead averaging constrain interpretation, while extreme-wave performance remains uniformly poor.A full multi-buoy cross-validated search was not feasible; best-trial selection inflates apparent gains relative to median trials, and lead-time-specific differences may be obscured by averaging.Source: §4.3 · p. 23
  • Autumn and winter skill degrades across families, and the univariate study omits atmospheric covariates that may help address the information deficit.The authors note that winter wave evolution is not inferable from local Hs history alone and leave testing whether even one wind-speed channel lifts the 6-hour barrier to future work.Source: §4.3 · pp. 23–24

4.4. Geographic and dynamical scope

  • The findings are bounded to NDBC environments dominated by mid-latitude storms or mixed swell; tropical, strongly tidal or shallow-water, and polar ice-zone regimes require dedicated validation.The authors describe the univariate formulation as ill-suited to semi-enclosed bays and strong-tide estuaries, where bathymetry, tidal harmonics, and spatially resolved forcing are needed.Source: §4.4 · pp. 24–25

4.5. Practical and methodological scope

  • Conclusions are conditional on complete, regularly sampled univariate Hs inputs, the selected architectures, and the study’s evaluation and preprocessing choices.The study’s missing-data tests preserve architecture ranking, but irregular sampling and burst dropouts remain unmodelled; the five-family comparison is not exhaustive, and richer inputs or different objectives could change rankings.Source: §4.5 · pp. 25–26

5. Conclusion

Across 47 buoys, the five model families converge, with corpus differences dwarfing between-family variation; deep models beat persistence on average but do not consistently outperform one another. The results also expose tail, seasonal, and transfer limits, supporting a shift toward richer inputs and cross-buoy evaluation.

5. Conclusion

  • Between-family SD is 0.0014 m2 (0.8% of the grand mean), dwarfed by the 4.83× cross-dataset MSE shift; all 45 multi-buoy trials beat persistence, but no architecture consistently leads.Mean skill was +0.062; the cross-dataset shift was an absolute increase of 0.129 m2, approximately 95× the between-family SD.Source: §5 · p. 26
  • Deep models underperform persistence on the most extreme 1% of waves (grand mean p99 skill −0.040; only 15 of 45 trials positive), while Q4 test MSE is 2.8–3.2× Q3 across families.The ranking of (family, CL) pairs also reverses between the single-buoy and multi-buoy experiments, so single-buoy hyperparameter selection does not transfer to heterogeneous wave climates.Source: §5 · p. 27
  • The convergence across five inductive biases and nine context lengths indicates a task-level ceiling for univariate Hs regression on mid-latitude NDBC buoy data, rather than a model-class-specific ceiling.Only DLinear and PatchTST gain more than +0.02 skill from 6 h to 168 h; even PatchTST and Mamba2 do not substantially exceed the best-observed ceiling.Source: §5 · p. 27
  • The authors recommend atmospheric covariates, zero-shot cross-buoy transfer, Hs decomposition into swell and wind-sea components, and tail-aware objectives as future directions.These directions target missing forcing information, generalisation without per-buoy retraining, physical drivers hidden in univariate inputs, and underperformance on the largest waves.Source: §5 · pp. 27–28

7. Author Contributions

Yilin Zhai led the study’s conceptualization, analysis, experimental design, and drafting; Hongyuan Shi and Zaijin You contributed guidance, revisions, data discussions, and manuscript sections. All authors approved the final manuscript.

7. Author Contributions

  • Yilin Zhai was responsible for conceptualization, data analysis, experimental design, and drafting the manuscript.Source: §7 · p. 28
  • Hongyuan Shi participated in research discussions and provided guidance and revisions.Source: §7 · p. 28
  • Zaijin You assisted with data discussions and contributed to several manuscript sections; all authors read and approved the final version.Source: §7 · p. 28

Related papers

Loading 2609.30688v1…