Dynamic thermal rating (DTR) of overhead lines can release a substantial share of
unused transmission capacity, yet it is most sensitive to the quantities that numerical weather prediction estimates least reliably at the level of an individual span:
wind speed and direction at the conductor. This thesis builds a computational chain
from a weather forecast to an empirically evaluated lower ampacity bound. Forecasts of the regional ICON-EU model are complemented with multi-resolution terrain
descriptions. A multi-branch convolutional model is developed and evaluated on a
four-year dataset of ten-minute observations at 266 German stations. It is trained
on data from 2021–2022, and its generalization is assessed on separate stations in
2023 and 2024. On the development set it reduces the wind speed forecast error by
roughly a quarter, from 1.373 to 1.037 m/s, and outperforms all reference models
considered in our implementation. On an independent test set comprising unseen
stations and an unseen year, it reduces the error from 1.271 to 1.056 m/s. A probabilistic extension fixes the location of the predictive distribution at the point forecast,
models speed with quantile regression and direction with circular quantiles, and
empirically calibrates the speed distribution with per-quantile corrections with the
median correction set to zero, followed by a width adjustment about the median.
Samples are drawn from the predictive distributions of wind speed and direction,
with the speed distribution calibrated. Propagating these samples through the standard CIGRE thermal model yields the ampacity distribution, from which a lower
bound at a prescribed risk level is taken. In the simulated test on unseen stations
the chosen classical static rating, assuming 0.6 m/s of perpendicular wind, exceeds
the fifth percentile of the reference ampacity computed from observed wind. At a
preset risk level of 0.05 the model attains an aggregate 4.26% violation rate and a
mean ampacity 14.5% higher than a retrospective static reference that is unavailable prospectively in operation, while at an exactly matched five percent empirical
violation rate the gain is 16.2%.
|
