Wind cools down power lines and can double the capacity of the grid. Yet widely used weather models are not accurate enough for grid operators, since wind is very dependent on the detailed landscape surrounding the line. Further, wind must be determined accurately on every section of the line, since even a short segment might otherwise overheat (see here).
GridRaven applies machine learning for predicting wind speed and direction (see here). Previously we have reported an improvement of accuracy by 39% (here). We have continued to improve these models and have now achieved a 50% improvement in accuracy.
The ground truth are wind speed and direction measurements by weather stations that our models have not seen during training. We calculate the mean absolute error of our wind speed prediction versus the measurement, which equals 0.6 m/s. The most accurate numerical weather prediction in our region is MEPS (link). MEPS has a mean absolute error of 1.2 m/s in the same hours in the same stations. Our improvement in accuracy is therefore 50%.
We further validated our predictions against weather stations that are installed on high-voltage towers as shown on the image below.
GridRaven's models make predictions with confidence intervals. The graph below shows hourly wind speed predictions and measurements over one week. The best available public weather forecast significantly over-estimated wind speeds in this period (orange), while our model (green) was quite close to the measured values (black). The confidence intervals were maintained in 95% of the hours, as expected.
Overall this improvement in accuracy together with the confidence intervals makes fully sensorless Dynamic Line Ratings possible also on sections of the grid that cross forests. This is a significant advancement towards reducing grid congestion, improving transmission capacity and reducing energy prices.
