Observation-driven correction of numerical weather prediction for marine winds

TL;DR

Open-ocean wind forecasts suffer from sparse and heterogeneous observations, limiting NWP accuracy. The authors propose an observation-informed correction framework that feeds recent in-situ observations into a transformer to adjust GFS outputs, rather than predicting winds from scratch. The model uses masking to handle irregular data, cross-attention to condition on target points, and position/time embeddings for arbitrary coordinates, achieving 45% RMSE reduction at 1 h and 13% at 48 h over the Atlantic, with strongest gains along coastlines and shipping lanes. This approach offers a practical, low-latency post-processing tool that complements NWP for maritime safety, routing, and offshore operations, and scales to grid-wide and site-specific forecasts in a single pass.

Abstract

Accurate marine wind forecasts are essential for safe navigation, ship routing, and energy operations, yet they remain challenging because observations over the ocean are sparse, heterogeneous, and temporally variable. We reformulate wind forecasting as observation-informed correction of a global numerical weather prediction (NWP) model. Rather than forecasting winds directly, we learn local correction patterns by assimilating the latest in-situ observations to adjust the Global Forecast System (GFS) output. We propose a transformer-based deep learning architecture that (i) handles irregular and time-varying observation sets through masking and set-based attention mechanisms, (ii) conditions predictions on recent observation-forecast pairs via cross-attention, and (iii) employs cyclical time embeddings and coordinate-aware location representations to enable single-pass inference at arbitrary spatial coordinates. We evaluate our model over the Atlantic Ocean using observations from the International Comprehensive Ocean-Atmosphere Data Set (ICOADS) as reference. The model reduces GFS 10-meter wind RMSE at all lead times up to 48 hours, achieving 45% improvement at 1-hour lead time and 13% improvement at 48-hour lead time. Spatial analyses reveal the most persistent improvements along coastlines and shipping routes, where observations are most abundant. The tokenized architecture naturally accommodates heterogeneous observing platforms (ships, buoys, tide gauges, and coastal stations) and produces both site-specific predictions and basin-scale gridded products in a single forward pass. These results demonstrate a practical, low-latency post-processing approach that complements NWP by learning to correct systematic forecast errors.

Figures (11)

• Figure 1: Comparison of spatial and statistical wind patterns from GFS forecasts, in-situ observations, and model predictions.
• Figure 2: Spatial density of ICOADS observations by platform type (2015–2024). Higher densities occur near coastlines, while ship observations dominate the open ocean.
• Figure 3: Schematic representation of the spatio–temporal learning framework. Each sample consists of in-situ observations at time t (blue circles), corresponding GFS forecasts at t (red squares) and t + $\Delta$t (magenta squares), and target observations at t + $\Delta$t (green crosses). The lower panels illustrate how each input is encoded through cyclical time, geographic positional, and sequential embeddings. Past observations and past GFS forecasts are processed through self-attention to capture temporal and spatial dependencies and learn the correction pattern, while cross-attention integrates information from GFS forecasts to predict target wind components at t + $\Delta$t.
• Figure 4: Comparison of model performance across forecast lead times and observation platforms. (a) Overall accuracy relative to GFS and ERA5. (b) Error reduction by platform type.
• Figure 5: Spatial distribution of wind-speed prediction errors for the machine-learning model and the GFS baseline at lead times of 1 h, 8h, 24 h, and 48 h. Panels (a), (d), (g), and (j) show absolute errors for the model, while panels (b), (e), (h), and (k) show corresponding errors for GFS. Panels (c), (f), (i), and (l) display the difference between the two (model – GFS), where blue indicates lower errors for the model, and red indicates higher errors. Each panel corresponds to the corrected field at the specified lead time.