Observation-driven correction of numerical weather prediction for marine winds

Technical Summary: Observation-driven correction of numerical weather prediction for marine winds

Problem Statement
Accurate marine wind forecasts are critical for navigation, ship routing, and offshore energy operations. However, generating these forecasts is challenging due to the sparsity, heterogeneity, and temporal variability of observations over the ocean. While global Numerical Weather Prediction (NWP) systems like the Global Forecast System (GFS) provide coherent large-scale dynamics, they exhibit systematic biases at many locations. Existing bias correction methods, ranging from classical Model Output Statistics (MOS) to deep learning approaches, generally assume fixed observation locations or regular grids. They fail to handle the irregular, time-varying geometry of marine observations, where the set of available data changes hourly as ships move, buoys go offline, and reporting intervals vary. Furthermore, current frameworks often lack the ability to perform inference at arbitrary coordinates without retraining or interpolation, a requirement essential for operational marine wind correction.

Methodology: ORCA
The authors propose ORCA (Observation-informed Real-time Correction with Attention), a transformer-based deep learning architecture designed to correct NWP outputs by assimilating irregular, time-varying in-situ observations. Rather than forecasting winds directly, ORCA learns local correction patterns to adjust GFS output.

The core technical components include:

• Correction-Based Formulation: The model frames the task as learning an additive adjustment $h_\theta$ conditioned on past observations ($w$) and NWP guidance ($f$), such that the corrected forecast is $w(t+l\Delta t) = f(t+l\Delta t) + h_\theta(\dots)$.
• Set-Based Attention: To handle the dynamic number and spatial distribution of observations, the model treats each forecast time step as a variable-size set of heterogeneous reports (ships, buoys, tide gauges, coastal stations).
  • Self-Attention: The encoder applies self-attention over historical observation-forecast pairs to build a context representation, learning which past pairs are most relevant for correcting the forecast.
  • Cross-Attention: The decoder uses cross-attention to condition predictions at new target locations on this dynamic context, allowing the model to query the most informative past data for any specific target.
• Irregular Observation Handling: Binary masks are employed during attention layers and loss computation to ensure the model only processes valid tokens and accumulates errors only over available targets, eliminating the need for imputation or fixed grids.
• Embeddings: The architecture utilizes:
  • Cyclical Time Embeddings: Four features capturing daily and annual periodicities ($\sin/\cos$ of day and hour).
  • Geographic Positional Embeddings: A hybrid scheme combining spherical harmonic basis functions with a Sinusoidal Representation Network (SirenNet) to respect spherical geometry and avoid distortions at high latitudes or across the date line.
• Inference Capability: The model supports single-pass inference at arbitrary spatial coordinates, enabling both site-specific predictions and basin-scale gridded products without fixed neighborhoods.

Experimental Setup
The study focuses on the North Atlantic domain (2°–64°N, 98°W–11.5°E) from April 2015 to September 2024.

• Data: Observations are drawn from the International Comprehensive Ocean-Atmosphere Data Set (ICOADS), covering ships, moored/drifting buoys, C-MAN stations, coastal stations, and tide gauges.
• Baselines: The model is compared against raw GFS forecasts, ERA5 reanalysis (as a benchmark), and a MOS-style linear regression.
• Training: The model is trained to minimize the mean vector wind error between predicted and observed 10-m wind vectors. Data is split chronologically (80% training, 10% validation, 10% testing) to prevent leakage and ensure evaluation on unseen future periods.

Key Results
ORCA demonstrates consistent improvements over the GFS baseline across all lead times up to 48 hours:

• Error Reduction: ORCA reduces GFS 10-meter wind error by 45% at a 1-hour lead time (from 2.75 to 1.51 m s⁻¹) and by 13% at a 48-hour lead time (from 3.44 to 2.99 m s⁻¹).
• Comparison to Baselines: ORCA outperforms the linear regression baseline by an additional 7–17% across all lead times and surpasses ERA5 reanalysis up to 12 hours.
• Platform Performance: Improvements are observed across all platform types. The largest gains are seen for stationary platforms (coastal stations, tide gauges, C-MAN), followed by fixed platforms, moored buoys, and ships. Drifting buoys show the smallest improvement, likely due to the difficulty of generalizing to unseen ocean regions as they drift.
• Spatial Analysis: The most persistent improvements occur along coastlines and major shipping routes where observations are dense. While performance degrades slightly in data-sparse open-ocean regions, ORCA generally maintains gains over GFS. In high-wind regimes (top 5%), the model improves short-term forecasts but shows diminishing returns and eventual degradation relative to GFS at longer lead times (36–48 hours).
• Operational Efficiency: Inference on a 0.25° grid across the Atlantic takes under 5 minutes on a single NVIDIA Tesla V100 GPU, demonstrating feasibility for operational workflows.

Significance and Claims
The paper claims that ORCA addresses a critical gap in marine forecasting by jointly handling three requirements previously unmet by a single framework: (i) assimilating irregular, time-varying multi-platform observations; (ii) learning state-dependent NWP error patterns conditioned on the most recent data; and (iii) producing corrections at arbitrary ocean coordinates in a single forward pass.

The authors emphasize that this approach does not replace NWP but serves as a practical, low-latency post-processing layer that leverages the physical grounding of NWP while correcting systematic biases using local observations. The work demonstrates that integrating global weather models with heterogeneous observational data and transformer-based attention mechanisms can produce more reliable marine wind forecasts, particularly in regions with abundant in-situ data. The study also highlights the continued importance of robust marine observing networks, noting that performance is inherently tied to observational density.