As renewable energy penetration grows across Europe, transmission networks face increasingly frequent and unpredictable congestion, a problem costing an estimated 4.2 billion euros per year in Europe alone. Yet most data-driven forecasting tools either focus on generation or demand in isolation, or require access to proprietary grid models unavailable to market participants. This thesis addresses the question of whether interzonal power flows, and by extension congestion, can be forecast purely from publicly available market and weather data.

The proposed approach is a spatio-temporal graph neural network that operates directly on transmission edges rather than nodes. It combines an LSTM encoder for temporal dynamics, a Transformer-based graph message-passing module with updatable edge representations, and a future-aware decoder that ingests day-ahead prices and weather forecasts. The model is trained and evaluated on Italy's seven-zone electricity market over a 2025 test year.

Against baselines ranging from naive persistence to gradient-boosted trees and LSTM, the model achieves the best normalized absolute error, directional accuracy, and congestion detection F1 score, with advantages that persist across all six forecast horizons. Critically, the proposed model achieves the best AUROC, demonstrating its ability to rank truly congested hours above non-congested ones regardless of where the threshold defining congestion is placed.

Through comprehensive experiments and analyses, the model proves to be more accurate than standard industry methods and domain-specific models. Further experiments demonstrate the quality of the forecast per edge and horizon, and break down the contribution of the different choices regarding its design. Moreover, the impact of future features is assessed, showing a significant performance increase in congestion detection.

The central contribution to the renewable energy field is a reproducible, open-data framework that transforms observable market and weather signals into physically grounded flow forecasts, without access to network topology or TSO-proprietary models.