Dynamic Line Rating in Germany: Integrating Machine Learning and Terrain Data for Improved Forecasts

In Germany, wind energy is reshaping the electricity landscape, notably by increasing energy input in the northern regions. This renewable energy must traverse the occasionally saturated electrical grid to the south, testing the system’s capacity limits. An effective solution to this problem of lacking ampacity is Dynamic Line Rating (DLR), which makes it possible to increase the load on the power lines under certain weather conditions.
Numerical weather predictions (NWP) are used to forecast the potential grid load for the weather-based DLR and predict key parameters such as wind speed and temperature for the downstream processes. However, the accuracy of the forecasts is sometimes impaired by the complex topography and dense vegetation along some overhead power lines. To enhance these predictions and tailor them to local nuances, a post-processing method has been developed.
This post-processing model works universally, i.e., it can be applied at any location. A two-stage method is used: First, ordinary kriging performs a downscaling from the model grid points to the prediction point. In the second step, a universal machine learning model (ML model) is trained that integrates not only weather model predictions but also static terrain information such as Terrain Position Index (TPI), slope, aspect, additionally land cover classes and statistical wind atlas data. The forecast horizon for the evaluation refers to the following day.
We investigated and tested two approaches to develop the optimal ML model: A simple artificial neural network (ANN) with two hidden layers served as a starting approach. To take terrain properties into account more effectively, an advanced models such as Long-Short-Term-Memory (LSTM) was tested and fine-tuned by hyperparameter optimization. The quality of these models is verified using independent measurement stations.