A light wind doubles grid capacity
Out of all the weather parameters wind can have the biggest effect on the ability of high-voltage power lines to transport energy. An increase of 2 m/s at right angles nearly doubles the capacity of a power line to transmit energy (source).
Accurately predicting the wind is therefore critical in order to unlock the maximum potential of transmission and distribution assets. Unfortunately, existing weather forecasts are not accurate enough, especially in complex terrain, to allow utilities to take advantage of this data.
Machine learning for wind prediction
Grid Raven is tackling the wind prediction challenge with the help of machine learning. Our model takes as input the widely used numerical weather predictions (see for example https://weather.us/model-charts) and downscales them with the help of detailed landscape data.
We set out to benchmark the ability of our model to improve wind forecasts. To do this we obtained data from 37 official weather stations by Estonia’s Estonian Environment Agency. We trained our model on data from 2020 and 2021 then validated it against data from 2022. We then benchmarked our results against Met.no’s forecast for the same locations.
Improving accuracy by 39%
The mean absolute error for wind speed in Grid Raven’s forecast was 0.79 m/s, while the error in Met.no was 1.30 m/s. This is an improvement in accuracy of 39%.
To illustrate the case we chose data from the weather station in Ristna. It is known to meteorologists in Estonia that Ristna is a challenging weather station for wind forecasts because it is sheltered to northerly winds.
These above graph shows hourly wind speeds over three days in January 2022 at Ristna. The numerical weather prediction predicted strong northerly winds for 17th of January 2022. However, the actual measured wind speeds remained at around 6 m/s, which was correctly predicted by our model.
This is a significant improvement in accuracy of predicting wind. Accurate wind prediction is central to safely increasing the amount of power transmitted over the high-voltage network.
The next step is to demonstrate the accuracy of the model for predicting wind in locations that it has not encountered during training.
