Accurate estimation of wind speed, wind direction, and ambient temperature at the span level is essential for DLR, especially in complex terrain where local microclimates vary substantially along overhead line (OHL) routes. We have developed a high-resolution machine learning-based forecasting approach that generates probabilistic, hyper-local weather predictions. This methodology allows us to produce dynamic ampacity estimates for every span without relying on physical sensors. The approach starts from the hypothesis that local terrain and land cover play a critical role in shaping near-surface meteorological conditions, particularly wind behavior. Traditional Numerical Weather Prediction (NWP) models-such as those from NOAA offer resolutions of 2.5 to 32 km (depending on the application), which are inadequate to capture weather variability at the 50-500 meter span scale. Dynamic and statistical downscaling methods are common workarounds but are limited by cost or location specificity. To overcome these limitations, a deep learning model was developed to learn a downscaling function from coarse NWP inputs to high-resolution predictions by incorporating topographic and environmental data. The model is a multi-branch neural network designed to process four types of input data:
- Numerical Weather Prediction (NWP): Core meteorological variables including air temperature, pressure, geopotential height, and u/v wind components at various atmospheric levels.
- Low-resolution remote sensing data: 1-arcsecond Digital Surface Models (DSM) and global land cover masks from satellite sources such as AW3D30.
- High-resolution topography: 1-meter resolution LiDAR-based DSM and Digital Terrain Models (DTM), along with derived features such as slope, aspect, and vegetation height.
- Metadata: Temporal and spatial context including forecast lead time, hour of day, month, region, and large scale weather model identifier
Training and evaluation
The model is trained using millions of historical samples from weather stations, typically covering several years of data.The training data includes:
- Core meteorological variables from archived NWP forecasts: In the US, we are using HRRR (High Resolution Rapid Refresh) and NAM (North America Mesoscale),
- terrain data,
- observed wind component values from land stations.
To promote location-independent generalization, critical for deployment across unseen
spans, the data was partitioned so that test stations were excluded from training (see
Figure 1 below).
Figure 1. Train-test-validation (eval) split of the whole dataset used in the machine learning model.
The training objective is to minimize the negative log-likelihood of observed wind speed and prediction values under the predicted target distribution. This allows the model to not only predict the expected wind vector, but also to quantify the associated uncertainty. Empirical results show that this approach reduces mean absolute error (MAE) and root mean square error (RMSE) in wind speed by over 35% compared to raw NWP values, while maintaining similar directional accuracy. To integrate these predictions into the DLR computation, we transform the predicted target distributions of wind speed and wind direction to predicted values and confidence intervals at 75%, 90%, 95%, 98% and 99% by using Monte Carlo sampling (N = 1000).
As illustrated in Figure 2 (below), the wind speed generated by our method is closely aligned with actual measured values at an example weather station, both in trend and magnitude. The model successfully captures diurnal fluctuations and localized wind dynamics, offering reliable confidence intervals. This is especially evident in forested or sheltered segments where conventional models often underpredict wind sheltering effects or misrepresent wind direction entirely.
Figure 2. 24-hour hyper-local wind forecast example with confidence intervals.
Predictions (i.e. inference) are done using the topography and new forecasts gathered each hour around the location of interest (Figure 3).
Figure 3. Schematic of the machine learning model predictions (inferences)