Smart grids represent an evolution of traditional electrical systems, integrating new technologies to enhance the monitoring and control of electrical system operations. In this context, machine learning techniques, particularly deep learning algorithms or ensemble methods, offer valuable tools for extracting information from the vast amount of data collected by smart grid components.

Additionally, the use of ontologies capable of representing the topology of energy networks enables the application of processing techniques and artificial intelligence algorithms specific to graph structures. In this study, we introduce a novel machine learning-based approach for estimating node voltages in Medium Voltage distribution networks, typically derived from power flow computations.

Our solution employs a hybrid strategy characterized by stacking a Graph Convolutional Network model and a Gradient-Boosted Decision Tree model. This method achieves an estimation error that adheres to constraints, thus mitigating the need for penalties imposed by the power supplier, while also substantially reducing computational time compared to conventional solutions.

Our approach could be more relevant for obtaining fast responses in scenarios not easily solvable through numerical techniques, such as in the case of extended and multi-energy networks.