A machine learning approach for EVCS integration in distribution network based on optimal investment actions

• Framework integrating supervised learning for optimal grid investment predictions. • Comparative analysis of four ML models based on simulation time, errors, and scalability. • Case study on CINELDI reference distribution network with different EVCS integration. • Performance analysis of ML models for two planning scenarios (Min. TOTEX and Max. Power). Achieving the European climate targets for 2050 requires the large-scale integration of electric vehicles and public Electric Vehicle Charging Stations (EVCS). This energy transition requires the reinforcement planning of the electrical infrastructure to ensure the quality of supply for the next years. Distribution operators need efficient planning tools able to evaluate reinforcement actions for each new connection of EVCS in MV distribution networks. This paper presents a Machine Learning (ML) approach for optimal EVCS allocation based on the Mixed Integer Second Order Cone Programming (MISOCP) formulation. First, a formulation for distribution network operational planning based on an MISOCP model is presented. This formulation aims to minimize total investments while considering operational constraints, the branch flow model, passive and active decision variables. Secondly, a large set of expansion scenarios are solved based on a sensitivity analysis and then used as training data for supervised learning models, allowing them to learn the mapping between scenario inputs and optimal investment outcomes. The strength of the trained model is the fast and accurate predictions of optimal reinforcement actions and investments from unseen new EVCS expansion scenarios, opening new opportunities for efficient large-scale EVCS assessments. Results on a 124-bus MV Network show that decision tree-based models can predict reinforcement actions and investments with a deviation of less than 1%.