Intelligent charging infrastructure

An increasing electric vehicle share calls for suitable, affordable and reliable support infrastructures, such as EV recharging. A major challenge in the integration of this support infrastructure into the existing grid is the increasing load demand, especially peak load demands are problematic. If many EVs are charged simultaneously, the utility grid has to serve great peak power demands. These can cause overload conditions for the grid’s infrastructural components, such as transmission lines and transformer substations. Grid peak loads occur, e.g. in the evenings as people plug in their EVs when returning home from work, or after plugging in commercial EV fleets, which have been used over the day.

Charging vehicles at times when there is a surplus of renewables in the energy mix increases utilization of renewable energy resources. In order to reduce the amount of energy that is transported, local energy resources should be preferred. This will prevent grid reinforcements.

What makes it intelligent ?

Simultaneously charging EVs can lead to undesirable and unmanageable peak loads. Peak loads can be avoided by sophisticated load management algorithms, based on mathematical optimization. A necessary prerequisite, though, is the availability of arrival and departure time information. Fortunately, our preliminary EV fleet load data analysis indicates that the peak load problem can be tackled by machine learning (ML) With the knowledge of recurring time patterns, it is possible to predict future (time versus load) states, as well as typical EV arrival and departure times. Note that, applying ML yields other beneficial byproducts: in principle, the EV types, their optimal demand profiles and the available renewable energy percentage time pattern can be learned. Machine learning algorithms have two phases in common: first the collected data is fed to the ML algorithm in the training phase. After that, the system enters a so-called test or prediction phase, where, based on the previously “learned” training data, decisions or predictions can automatically be made (see fig 20). The test data validates the prediction performance.

Try out some simulations on your own with our online tool: Smart EV fleet charging.

Following graph top: Preliminary data analysis: “Typical" mean daily energy consumption (orange line) based on hourly measurements over the year 2018 (“violin” plots showing the statistical energy consumption distribution for every hour of the day – typical peak load times recur at 12 noon, 1pm and at 11pm). Bottom: Tentative results showing that recurring energy consumption time pattern prediction can be predicted via Kalman filter. The data is based on EV charging data measured at Samsø.

Apart from an adequate choice of ML algorithms to serve the task, the reliability of such predictions depends critically on the available training data. To foster a precise arrival and departure prediction, we plan on collecting a vast amount of densely sampled data from our charging stations connected to the grid. Different recurrent neural networks can be trained to take seasonal as well as daily time patterns into account.