What is Hyper-local wind forecasting?
Hyper-local wind forecasting is a specialized machine learning-based downscaling process. It translates low-resolution weather parameters from global models into high-precision local conditions at the exact height and orientation of the power line.
Existing weather forecasts lack the resolution to ensure correct wind speeds in difficult terrain, such as valleys or forests. Gridraven’s approach is location-independent, meaning the model does not just memorize historical data for a specific spot; it learns the fundamental physical relationship between topography and wind flow. This allows for an accuracy improvement of over 35% in wind speed prediction compared to raw NWP values.
How it works?
Gridraven’s proprietary machine learning architecture, integrated into the Claw platform, synthesizes several high-volume data streams to produce a single, reliable forecast
- Numerical Weather Prediction (NWP): The model ingests core meteorological variables, including air temperature, pressure, and u/v wind components from models like ICON, GFS, or HRRR.
- Terrain Data: High-resolution LiDAR DSM/DTM (1-meter resolution) or Satellite DEMs are used to define the physical obstacles.
- Neural Processing: ResNet blocks learn local features (slopes, aspect, vegetation height), which are then concatenated and processed through MaxViT blocks to model the complex interactions between moving air and the landscape.
- Probabilistic Output: Instead of a single "best guess," the model outputs a Gaussian Mixture Model (GMM). This provides a probability distribution of possible wind values.
- Confidence Intervals: Using Monte Carlo simulations, these distributions are converted into actionable Confidence Intervals (e.g., 95%, 98%, or 99%). This allows operators to choose a risk-aware rating that accounts for atmospheric uncertainty.
This diagram illustrates the multi-branch architecture of our machine learning model. By processing Numerical Weather Predictions (NWP) and high-resolution terrain data (LiDAR or Satellite DEM) in parallel, the system effectively models the physical interactions between topography and wind flow to generate accurate downscaled forecasts.