**1. Introduction**

With an increasing number of electric vehicles (EVs), an improved charging infrastructure is required in order to deliver the necessary energy for mobility while at the same time reducing the costs of operating said infrastructure. For the development of such charging systems, developers have to optimize a number of system parameters like the number, location, and type of charging stations (CSs). A key factor in this analysis is an accurate representation of customers and their charging requirements. In the majority of prior literature (for example [1,2], see also [3,4]), customers were modeled passively via their charging process with an arrival and departure time of the car plus an energy requirement. These values are typically described as a Gaussian probability distribution with a defined mean and variance. An example customer could be described as arriving at the charging station at 9:00 h ± 30 min, leaving at 17:30 h ± 10 min, and requesting 8 ± 0.5 kWh energy. For example, the work in [5] studied the psychological aspects concerning the charging behavior of electric vehicle drivers. During a six-month study with 79 EV users, charging data were collected. Typically, users drove almost 40 km per day and recharged three times weekly, with a high remaining battery energy level at the time of charging.

This approach has a number of benefits. It requires limited information about customers, can easily be scaled, has a low computational expense (data tables can be computed upfront) and is easy to investigate. The drawback is that with increasing numbers of customers, normal distributed variations level out, thus only covering average conditions, potentially overlooking rare, but serious extreme conditions. In addition, as we will show in this work, there are individual effects that cannot be

accurately modeled with the conventional approach. In order to show this, we extend the standard approach for modeling EV customers in two separate steps. In a first step, we introduce an EV sensitivity to the current weather conditions. We assume that under very cold or hot conditions, EVs have an increased energy demand, requiring more energy than usual (e.g., due to air conditioning, heating, or battery properties). The quality of the charging process is evaluated using a customer satisfaction indicator (CSI) that compares the result of the charging process (the reached SoC) with the customer requirements. We hypothesize that depending on the increase in energy demand, the mean CSI degrades on certain days. We call this the Temp scenario. The second studied extension (CSI sensitive) is that we assume that customers have a memory of past charging sessions, in the form of CSI values. If CSI values degrade below a threshold, customers will stop using the charging station, and the system will be used by fewer EVs.

The study system used in this work is a charging infrastructure of a medium-sized company. We assume that some of the employees are using EVs to commute and use the low-cost electricity at the company for most of their charging demands. In order to minimize CO2 emissions, the company uses a large-scale PV system to generate inexpensive, CO2 -free energy for the charging station, supported by a low level of grid power. The drawback of the approach is that during days without sunshine, only a small level of extra charging power from the grid is available. This approach minimizes CO2 emissions and energy costs (especially peak load costs), but is sometimes not able to satisfy all customer (EV user) demands.

The practical implementation of customer models that respond to external conditions and historical data is done via a multi-agent system (MAS) approach, where each customer, electric vehicle (EV), and charging station (CS) is represented as a computer program that can respond to current and past conditions. Using this approach, it is possible to replicate the conventional approach with normal randomly-distributed requests, but also to investigate other more realistic scenarios. In our study, we employ the NetLOGO framework.

Using the baseline and the two extended scenarios, we show how the characterized modeling framework can lead to new perspectives and insights on the more complex interactions between charging infrastructure systems and their customers.
