This master's thesis addresses the forecasting of the dynamic thermal rating (DTR) of overhead transmission lines using spatio-temporal deep learning models. DTR represents the maximum allowable current through a conductor under given weather conditions and, compared with traditional static limits, enables more efficient utilization of the transmission network. Since DTR strongly depends on weather conditions, we developed models that combine measurements from weather stations with forecasts from a numerical weather prediction model. Temporal dynamics were modeled using a bidirectional long short-term memory (LSTM) network, while spatial dependencies between measurement locations were modeled using three graph-based approaches: a graph convolutional network, a graph attention network, and a model with a learned adjacency matrix. The models were evaluated on a five-year dataset from 31 weather stations in the Slovenian transmission network. The results show that spatial modeling reduces the forecasting error on average, but its contribution is relatively small and varies across nodes and forecasting horizons. A substantially larger effect is achieved by including weather forecasts, which reduce the mean absolute error by approximately 14 percent across all considered architectures.
|
