Energy demand forecasting for electric vehicle charging infrastructures: a data-driven approach for an Italian use case

In recent years, rising temperatures and extreme weather events provide the most undeniable evidence of global warming due to greenhouse gases. This condition has led to the establishment of a number of treaties and countermeasures around the world to reverse this trend. Thanks to the Paris Agreement (2015), 196 world powers agreed to keep global temperatures rise below 2°C and limit the temperature increase to 1.55°C. EU countries’ have established ambitious policies to accelerate the green transition, specially for the road transport which is the cause of 12% of total emissions at European level. To deliver the European Green Deal (2020), European Commission set the target of 55% reduction of emissions from cars by 2030 and 0 emissions from new vehicles by 2035. To support the transition, the European plans should ensure at least the coverage with charging stations over all highway network and metropolitan territories. In the era of Big Data many companies are relying on data-driven decisions to contribute to the achievement of European targets. Data science methodologies assume a crucial role to create valuable information for the development of an accurate charging infrastructure. This work focuses on forecast analysis for the prediction of monthly energy demand for a single charge point over the Italian network. This evaluation supports growing demand trends with a consequent increase in customer satisfaction and profit for companies involved. By anticipating potential break-downs due to infrastructure overloads, maintenance costs are reduced. The methodology leverages charging sessions data together with external data (weather and population information and energy/fuel prices) collected from different sources. It consists of a building-blocks based approach inspired by the KDD process, a procedure for the discovery of knowledge from data. Different data science models (Linear Regression, Random Forest, XGBoost and Time Series), features and hyperparameters selection techniques (Correlation Analysis, Recursive Feature Elimination, domain expert support and Cross Validation) have been developed in order to reach the best combination model-purpose. Each model uses a moving window to analyze data of two sequential months to predict future values over a time period of one month, through a Sliding Window based approach. Data collection and data ingestion were the most time-consuming activities due to the several number of sources and different level of aggregation in input. The most promising model is the Random Forest Classifier with features in input selected with the support of domain experts. It has good performances in classifying charge points with high energy demand and is more advantageous to support the company’s decision-making process. Thanks to its level of interpretability, the model allows the visualization of the results on an interactive dashboard. Moreover, the presence of external data makes possible to highlight the influence that exogenous factors have on energy demand. Different environments were used to develop the project. First of all, the Amazon Web Services platform (AWS) was used to storage the collection of internal data then the Spyder environment with Python language for the data analysis part. Many possible developments are discussed to improve the prediction.