The rapid expansion of renewable energy generation in Germany has increased the frequency and severity of congestion events in medium-voltage (MV) distribution networks. Accurate short-term forecasts of line and transformer loading, together with reliable estimates of forecast uncertainty, are essential for grid operators to intervene proactively and minimize renewable curtailment. This work presents an advanced congestion-forecasting framework that combines artificial neural networks (ANNs) with uncertainty quantification (UQ) techniques.

A set of ANNs is trained on historical measurements and weather data to predict the active power flow at strategically selected measurement points (MPs) within an MV grid. To match this goal, variability in volatile electricity production by renewable sources as well as variations of the demands will be represented by suitable probability density approximations. In combination with the ANNs, which include uncertainty models either in the form of Bayesian neural networks or ensembles of different network models, these representations allow for directly linking the predicted uncertainty to physically interpretable probability distributions of input quantities.

The approach is validated on a camouflaged section of a real MV distribution network in Northwest Germany. The training data spans nearly one year (2025) in 1-minute resolution. Forecast performance is benchmarked against these data using AC load flow calculations performed in DIgSILENT PowerFactory.

The results show that early warning of impending overloads ahead of the actual violation is extending the decision horizon for redispatch actions.

By delivering both accurate point forecasts and well-calibrated uncertainty bounds, the proposed method directly supports the operator’s risk-aware decision making, fostering more efficient use of existing infrastructure and facilitating higher renewable penetration. The probabilistic insight allows operators to balance the trade-off between preventive redispatch and acceptable overload risk, thereby reducing unnecessary interventions and associated operational costs.

The integration of neural networks with uncertainty quantification provides a robust and scalable solution for short-term congestion forecasting in power grids. Adoptions of this approach promise to enhance redispatch management, minimize renewable curtailment, and contribute to the reliable operation of Germany’s evolving power system