Impact of electric vehicle charging – An agent-based approach

Several European countries have launched programs to increase the market penetration of battery electric vehicles (BEV). Similarly, politicians in Switzerland have targeted a 15% BEV share of new car registrations by 2022. As each electric car increases the power demand, new challenges are posed to the operation of existing distribution grid infrastructure. Here, a new bottom-up physical approach is presented that couples agent-based trafﬁc simulations through an unsteady vehicle power consumption model with distribution grid power ﬂow simulations. The impacts on hourly powerline loads from charging a car ﬂeet with an 8.5% BEV share are quantiﬁed in the real distribution grid for the canton of Zurich. The grid is composed of 12,000 buses and 9,800 powerlines, providing power to 398,000 individual customers. Results indicate that the risk of overloaded powerlines is highest in low-level distribution grids. In our most critical future scenario, with simultaneous 8.5% BEV charging at 8 pm with 11 kW, peak line loads reach up to 132% of rated capacity. Hence, in a potential energy transition towards a decarbonised future, each individual distribution grid could face critical loads at speciﬁc temporal and spatial bottlenecks. Thus, grids should be assessed individually to limit uncertainties and risks of critical power


INTRODUCTION
In 2017, the transport sector accounted for 27% of total EU-28 greenhouse gas emissions [1], of which 32% emanated from combustion-engine passenger cars. Therefore, increasing the share of battery electric vehicles (BEVs) is widely considered to be an effective measure to decrease CO 2 emissions in a decarbonised future. To this end, the Swiss government has established a political target that by 2022, 15% of newly registered cars are BEVs [2]. This increased BEV share will decrease gasoline consumption and thereby emissions in the transportation sector. In that each electric car increases the power demand, new challenges are posed to the infrastructure of the existing distribution grid. Therefore, it is necessary to assess the implications of a decarbonised transportation sector on the operational limits of the power system. Power systems can be divided into high-voltage transmission and low-voltage distribution networks. Consumers connect to the distribution grid, which consists of many short underground powerlines. Multi-megawatt generators are connected to substations that feed a few long overhead transmission grid power-This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2021 The Authors. IET Generation, Transmission & Distribution published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology lines. Regional and local distribution grids are typically operated below 36 kV by distribution system operators (DSO) and deliver electricity to household power sockets.
The tasks of DSOs include the provision of services for the secure, efficient and sustainable operation of distribution systems. Furthermore, DSOs face a legal obligation to provide high-quality and secure planning, as well as the operation and maintenance of the grid. Typical challenges for DSOs include integrating intermittent renewables, new consumers and grids with unknown behaviour. With the increasing number of BEVs, DSOs are faced with more requests for BEV chargers with unknown effects on the system. BEV chargers are normally connected to local distribution grids with private charger voltages of 230 or 400 V. These additional charging loads increase uncertainty and risk for the future operation of the existing grid infrastructure. Uncertainties in this infrastructure could be temporal or spatial, such as specific overloaded powerlines or transformers. To increase social welfare, expensive and timeconsuming underground distribution system extensions should be avoided. Therefore, there is a critical need for accurate and fast scenario simulations that quantify the impact of BEV charging on existing infrastructure. A detailed simulation and impact quantification of BEV charging is crucial for system operation in upcoming years. Shifting loads might be necessary to provide requested electricity without increasing peak demand because loads on the existing infrastructure can already be high. Moreover, load-shifting might offer a way to maintain customers' acceptance and comfort levels.
The power flow method is one of the most important tools for DSOs to maintain a secure distribution system. Different studies have applied power flow simulations of polar or rectangular formulations to determine system states of branches and buses. For example, Fonseca et al. [3] proposed a power flow methodology for reconfiguration analysis in distribution systems based on the extended fast decoupled Newton-Raphson method [4]. Their methodology was applied to a 38-bus distribution test system and a 1,531-bus distribution network. Thurner et al. [5] presented a power flow-based approach to determine the profitability of removing middle-voltage backup transformers in a medium-voltage network group. Adel Mohamed et al. [6] proposed a new rectangular branch-based loadflow algorithm applied to different test systems.
Several studies have assessed interactions between BEV charging, power demand and the power system. For instance, Andrade et al. [7] focused on a scalable methodology integrating a high-resolution traffic simulation to place fast charging stations with the objective of a least-cost solution. Fast chargers are connected to the medium-voltage grid with dedicated transformers. The power flow method was applied using OpenDSS [8], but individual line loads were not shown spatially or temporally. However, the methodology of Andrade et al. is substantially different to the work described in this paper. That is, the traffic and BEV power consumption in [7] were not physically coupled, incorporating velocity, acceleration, the drag coefficient, vehicle mass and recuperation. Furthermore, the traffic simulation was scaled down to 0.6%, with an 8,000-vehicle sample.
Veldman et al. [9] focused on network impacts and related financial impacts of various BEV charging strategies on medium-voltage distribution networks. BEV charging profiles were added to regular residential power demands to construct load profiles, which were then used in a loadflow analysis. The authors showed that if BEVs are charged based on electricity prices, high peaks in network loads result, leading to high costs of network reinforcement compared to the lower costs of a controlled charging strategy. However, instead of using actual locations and types of electric vehicles, each electric vehicle was represented with a battery capacity of 24 kWh and charged with 3 kW, reflecting 16 A at 230 V.
The project ARTEMIS [10] focused on the city of Zurich and used MATSim [11] for traffic simulations. However, the study used statistical driving cycles, and the power consumption of the electric vehicles was specified before the traffic simulation.
Multiple studies [12][13][14][15] have investigated the impacts of BEV charging using small generic IEEE 37-bus or 38-bus test systems. For example, Islam et al. [12] focused on the active and reactive power demands of BEV charging. Pirouzi et al. [13] evaluated the voltage security of distribution networks with BEV charging. The authors showed that employing a concurrent active and reactive control of BEVs offers more flexibility to the distribution network operation. Qian et al. [14] quantified the impacts of multiple controlled and uncontrolled charging scenarios. Stochastic BEV charging loads were applied, and results demonstrated an approximately 36% increase in peak load with a 20% penetration level in uncontrolled charging. In Mehta et al. [15], the authors evaluated smart grid-to-vehicle and vehicle-to-grid charging strategies for the optimal integration of electric vehicles. Two different strategies were considered for workplace charging scenarios to minimise the total daily cost and peak-to-average ratio. Results indicated that a 'minimisation of total daily cost' strategy provides significant economic benefits, while a strategy to minimise the peak-to-average ratio provides significant technical benefits, such as a reduction in peak loads.
Several studies have assessed the impacts of BEV charging on power demands without considering the grid infrastructure. For example, Muratori [16] quantified the impact of uncoordinated BEV charging on residential electricity demand using driving profiles generated from a Markovian stochastic tool. The author demonstrated that the shape of the aggregated residential power demand was impacted by BEV charging, yet the impact was limited. Zheng et al. [17] focused on a real-time scheduling strategy for integrating large-scale electric vehicles into a smart grid. Jian et al. [18] developed a valley-filling strategy for centralised, coordinated charging of large-scale electric vehicles. Their results demonstrated that coordinated charging could alleviate the negative impacts arising from BEV charging loads on power grids.
The literature review demonstrates that distribution grid power flow simulations for real-world, expansive grids have not been physically coupled with agent-based traffic simulations. Hence, this work aims to fill this gap by coupling an unsteady electric vehicle energy consumption model with agent-based traffic simulations.
This study integrates a bottom-up approach coupling agentbased synthetic population, daily-activity and traffic simulations through an unsteady power-based electric vehicle energy consumption model with distribution grid power flow simulations. The assessments conducted with this simulation framework can lower uncertainty and risk and could potentially reduce the necessity of time-consuming and expensive real-time grid measurements at each bus. Furthermore, in regions with many different private, commercial and industrial consumer behaviours, this framework can anticipate challenges at specific locations and times to circumvent critical infrastructure overloads.
The contributions of this work can thus be summarised as follows: Coupling: • Agent-based population, activity and traffic models are coupled with a power-based electric vehicle energy consumption model and distribution grid power flow simulations. • The exact positions of all BEV owners and car types in the canton of Zurich are used, and physical car properties such as Scale: • An agent-based traffic simulation is conducted at a national scale for Switzerland rather than using small samples or statistically average driving cycles. The simulation includes approximately 3.5 million car-driving agents with a temporal resolution of one second. • This integrated framework is applied to a genuine regional and local distribution grid that includes more than 12,000 buses and 9,800 powerlines. Each powerline and bus is simulated with the power flow method. Physics: • Instead of a statistical driving cycle, each individual BEV power consumption is physically calculated from the bottom up, including: • Agent-based traffic simulation results, such as acceleration, velocity, elevation and recuperation. • Individual car properties, including the mass and drag coefficient.

METHODOLOGY
This work extends our in-house, agent-based, bottom-up and integrated simulation framework 'EnerPol', which is used to support decision-making through scenario-based assessments. Specifically, in this work, the novel extensions to the EnerPol framework are: (i) integration of a new vehicle power consumption model that accounts for the type of BEV, how the BEV is driven, and for the topography and weather in which the BEV is driven; (ii) integration of a new distribution grid power flow model that is used to quantify the impact of BEV charging on the local grid; and (iii) integration of a new forecast model of BEV owners that accounts for the effects of neighbourhood clustering.  Figure 1 displays a workflow showing how the data and models are connected. Although BEV locations are known precisely, they need to be coupled to a synthetic population to identify the corresponding agent. EnerPol's activity model [19] generates daily activity plans from the population model for the entire synthetic population. These individual agent-based travel demands are used as input data for the traffic model [24,25] to simulate the overall road traffic in Switzerland. All BEV agents are therefore identified, and simulation results of acceleration, velocity, street slope and time can be extracted for each agent. These physical parameters, coupled with individual car properties, are then fed into the car power-consumption model to calculate the battery state of charge (SoC) as soon as each agent returns home. This charging demand is then added to each customer's power demand to conduct distribution grid power flow simulations and quantify the impact of BEV charging on existing infrastructures.
The workflow in the present study is described below and summarised in Figure 1.
Agent-based population, activity and traffic simulations are coupled through Fiori et al.'s [22] unsteady power-based electric vehicle energy consumption model with distribution grid power flow simulations. As a test-case, the entire Elektrizitätswerke des Kanton Zürich (EKZ) regional distribution grid on 16 kV level is assessed together with the local distribution grid of one village at 400 V. Thus, a total of 12,000 buses and 9,800 powerlines providing power to 398,000 customers are simulated. The power flow is simulated with hourly resolution for a year-long period for each BEV total car share case from today's 0.5% to a future 8.5%.

Battery electric vehicle ownership
The adoption of BEVs is modelled based on observations from previous studies [23][24][25][26][27][28] that the effects of neighbour-hood clustering have a significant influence on the adoption of BEVs. Thus, for a given share of BEVs, the socioeconomic and behavioural characteristics of individual agents are assessed to determine potential BEV owners. The current locations of BEV owners and of car models, Figure 2, are obtained from the road traffic office of canton Zurich. The socioeconomic and behavioural characteristics of potential BEV owners are derived from the attributes of agents in the synthetic population that is generated in the EnerPol population model [29]. Specifically, the potential BEV owners are considered: (i) To live in geographic proximity to current owners of BEVs; (ii) to possess a driver's license; (iii) to commute by car to/from work; and (iv) to have sufficient income to buy a BEV. Furthermore, as observed in the case of Switzerland, a household may have no more than one BEV.

Activity model
The daily-activity model is based on the Swiss Transportation Census 2015 [30] where about 57,000 Swiss residents were asked in detail about their travel behaviour to determine the following: when, where and why is travel undertaken; what modes of transportation are used; and what is the duration of stay at each location. The daily-activity model is described in detail in [19], but for sake of completeness some salient features are presented here.
The inputs to the daily-activity model, Figure 3, are the population agents and households that are generated in the EnerPol population model, and data that is comprised of: the Microcensus; quality of public transit; typology of municipality; and locations of places of activity, such as shopping, education etc. Each agent of the synthetic population, generated in the Ener-Pol population model, is characterised in parameters including age, gender, employment status, salary and location of job.

FIGURE 3
Structure of the daily-activity model. Adapted from [19] Furthermore, the population agents are linked into households that live in geo-referenced dwellings.
The public transit quality of each dwelling is ranked on the basis of the diversity of transport modes, proximity of stop facilities and frequency of service [31]. The typology of municipality is classified as either urban, sub-urban and rural, based on population density and local accessibility [32]. The locations of activities, such as shopping, education, etc., are extracted from OpenStreetMap [33] data.
The daily-activity model is comprised of four sub-models: a car ownership model; a mode choice model; a statistical matching model; and a plan generation model. With inputs of household income, household typology, public transit quality and municipal typology, the multinomial logit car ownership model determines if a household has 'no', 'one', 'two' or 'three or more' cars. With inputs including age, income, employment status, public transit quality, and number of cars in the household, for each agent a preferred transport mode for each travel leg is assigned in the mode choice model. A probabilistic statistical matching model best matches the behavioural patterns of the population agents with similar candidates in the Microcensus. Then, 'activity chains' which consist of a sequence of trips, each with departure and arrival times duration of activity, and mode of transport is generated for each agent in the plan generation model. Figure 4 compares the predicted departure times of synthetic agents to actual departure times from the Microcensus.

Traffic model
Rather than applying statistical, top-down driving cycles, Ener-Pol's traffic model is applied to simulate Switzerland as a whole for a typical working day with a high temporal resolution (1 s).
EnerPol's traffic model is written in C++ using CUDA SDK to accelerate parts of the code using a GPU [20]. For the mobility simulation, two main components are input: travel demand and supply. A road network represents the supply part and comprises links (unidirectional segments of streets) and nodes (intersection points of links). Each of these nodes has a georeferenced location, and each link has physical properties such as length, number of lanes or the speed limit. The supply part comprises daily-activity plans of all agents. For each driver, a travel leg contains a sequence of network links to follow to get from one activity location to another. The traffic propagation model, Figure 5, [21] uses queues to simulate congestion and the spillover effects of other traffic phenomena. In the model, each link (a span of a street from one intersection to another) is represented by two buffers (queues). [1]: The spatial buffer N l represents the physical capacity of a street, that is, how many vehicles can simultaneously be on the same segment of the road. [2]: The capacity buffer N f indicates the flow properties of a street, that is, how many vehicles can leave a street segment in a period of time (depending on speed limits, traffic lights, pedestrian crossings and other factors). The size of the buffers is calculated as follows: In this equation, L link is the physical length of a link, L vehicle is the average gross space occupied by a single vehicle, N lanes is the number of lanes, q is the flow capacity of a link (that is, 1,200 vehicles per hour), t cycle is the simulation time step (currently, 1 s), and t period is the duration of the period used to define q. When entering a spatial buffer, a vehicle must stay for at least seconds there, where v link is the allowed free speed of the link. The traffic propagation is conducted in two steps, for each link and for each node of the road network, for each simulated second: • Per-link step: a vehicle at the front of the spatial buffer is moved to the capacity buffer of the link if (1) it has spent t link of time in the spatial buffer, and (2) the capacity buffer has free space available. This operation is repeated at each link until it has vehicles and both conditions are satisfied. • Per-node step: a vehicle at the front of the capacity buffer of an upstream link is moved to the spatial buffer of a downstream link if [1] the link flow constraints are satisfied, and [2] the downstream link has free space in the spatial buffer. For example, if a link has a flow capacity of 3,600 vehicles per hour, then one vehicle at most can leave a link in a simulation step. This operation is repeated at each link until it has vehicles and both conditions are satisfied. As each node (intersection) can have multiple upstream links, each of the links is processed, and the order is uniformly defined at each simulation step at random and proportionally to the flow capacities of upstream links:

FIGURE 4
Departure times of synthetic agents throughout a day compared to survey FIGURE 5 Traffic queue model: [1] vehicle is propagated from spatial to capacity buffer, [2] and to downstream link. Adapted from [21] where p k is the probability of selecting the k-th upstream link, and q k and q i are flow capacities of the k-th and i-th links, respectively.
To summarise, the first step moves vehicles along the streets, and the second step moves vehicles across intersections. Therefore, if downstream links are full and congested, vehicles from upstream links cannot propagate further, and a spillover effect occurs.
As the speed and acceleration of an electric vehicle affect the power drawn from the battery, the vehicle's acceleration is accounted for in the model, as shown in Figure 6. The acceleration is considered each time a vehicle enters a new link, which also determines the minimum amount of time a vehicle should spend on the link. Therefore, the travel period along a link is split into three stages: • Normalisation stage: a vehicle attempts to normalise its speed according to the defined speed limit. • Cruise stage: a vehicle moves at a constant speed without acceleration. • Deceleration stage: a vehicle attempts to match the speed limit on the following link by breaking or keeping the current speed if the ensuing link's speed limit is not slower.
Thus, the following are required to evaluate the acceleration: v 0 , the velocity of a vehicle when entering a link [m/s]; v lim, the speed limit of the link [m/s]; v 1, the speed limit on the downstream link [m/s]; L, the length of the link [m]; a norm the acceler-ation/deceleration [m/s 2 ] used by vehicles. The value for a norm is currently set to 3 m/s 2 and is considered to be comfortable for drivers. In the algorithm, lines 2-9 represent the normalisation stage, where a driver attempts to adjust speed according to the link limits, if required. Lines 10-25 represent the deceleration stage if permitted by the normalisation distance from the previous step. Lines 12-13 check the distance required to accelerate and decelerate from the cruise speed. If acceleration or deceleration is impossible (lines [14][15][16], then the vehicle gradually changes speed from the previous normalised speed to the speed limit of the next link. Otherwise, (lines 19-20) the vehicle cruises. In a case where a vehicle enters a link with a speed that is too high (line 24), the speed is reduced to best match the downstream link limits. After the length required for each stage is calculated, the minimum travel time can be obtained (lines [26][27]. As EnerPol's traffic model implements a mesoscopic traffic model (no direct interaction of vehicles on a link is modelled), average link velocities and accelerations are used in the BEV power consumption model.

Battery electric vehicle power consumption
A power-based BEV energy consumption model is integrated into the EnerPol framework. The salient formulas and parameters are described below, while the reader is referred to Fiori et al. [22] for more complete details.
The power at the wheels is calculated as follows: is acceleration; is the street slope; A f [m 2 ] is the frontal area; m [kg] represents vehicle mass and C D is the drag coefficient. The simulated velocity, acceleration and street slope of each vehicle are determined in the agent-based traffic simulation. Car properties, such as vehicle mass, frontal area and drag coefficient, are considered for each BEV owner and differentiated for 49 different models, including Tesla, BMW, Mercedes, VW and others. The ranges of these properties are summarised in Table 1.
The power at the electric motor assumes a driveline efficiency of η Driveline = 92%, with an electric motor efficiency of η ElectricMotor = 91%. More details can be found in [22,34]. A battery discharge efficiency of η Battery = 90% is assumed. The dependence of the auxiliary power consumption on ambient temperature is considered based on measurements from a Chevrolet Volt [35], as shown in Figure 7. This temperature dependence accounts for heating and cooling [35,36]. The ambient temperatures are derived from the mesoscale weather model, which is integrated into the EnerPol framework [37].

Distribution grid model
Distribution grid power flow simulations of the grid infrastructure in the canton of Zurich are conducted using pandapower FIGURE 7 Temperature-dependent BEV auxiliary power demand [38], which is based on the Newton-Raphson method [39]. Figure 8 shows the distribution grid and its link to the transmission grid.
The simulated grid includes 1,100 local and 8,700 regional distribution grid powerlines. The capacity, reactance, resistance and susceptance are specified for each line. The simulated local distribution grid is shown in Figure 9.
The different capacities of the local distribution grid, operated at 400 V, are shown. In this village, the powerline capacities in the same grid level differ up to a factor of 7.5. Hence, it is evident that each powerline's individual properties should be considered for an accurate assessment of the impact of BEVs on a local distribution grid. For each of the 398,000 customers in the canton of Zurich, annual power consumption and BEV charging are georeferenced and allocated to the closest network element. Hourly power demands are obtained by applying the measured temporal distribution of transformer power demand to the annual power demand of each georeferenced customer.
Up to 8.5% BEV share of total passenger cars is simulated in our future scenarios, which is much larger than the 2019 share of 0.6%. It is thought that the share of 8.5% will be reached after 2025. Within our charging scenarios, each car is charged daily with 11 kW until either the car leaves the charge or the battery is full. In the simulations, we distinguish between 'low-tariff 8 pm' and 'low-tariff 10 pm' charging as different low-tariff options might be used for peak shaving. Specifically, Elektrizitätswerke der Stadt Zürich offers a low tariff from 10 pm, and Elektrizitätswerke des Kantons Zürich offers a low-tariff option from 8 pm.

Agent-based BEV power consumption
The unsteady electric vehicle power consumption simulation is compared to measurements in a VW Golf conducted by the Swiss Federal Laboratories for Materials Science and Technology [40], Figure 10, for a summer trip with an ambient temperature of 22 • C. Figure 10 shows that the predicted battery SoC has good qualitative agreement with the measurement. After the 23min trip, the quantitative difference between simulation and measurement is 1%. Factors such as wind velocities and the use of the vehicles' radios or lights could account for this difference.
Agent-based traffic simulation results for the street slope, velocity and acceleration are fed into the electric vehicle power consumption model to calculate the power consumption and SoC. Three simulated trips for a Tesla Model S are shown in Figure 11. On the left y-axis, the SoC is shown in each plot. On the right y-axis, the altitude (top row of plots), velocity   (middle row) and acceleration (bottom row) are shown. The first column of plots displays a 'country road trip', which includes velocities from 50 to 100 km/h. Through this 50-km trip, the battery is discharged to 89%. The negative acceleration (deceleration) leads to recuperation and charging of the battery. High velocities and uphill driving at around 5 km decrease the SoC by 1%. The second column shows a mixed freeway and city trip, with the highest velocities around 105 km/h. During the 80-km trip, the battery is discharged by 20%. The last column presents a city trip with peak velocities of 40 km/h and rather small accelerations. This scenario would lead to a small battery discharge, but the altitude climbing from 400 to 650 m discharges the battery. This situation can be identified in the top-right plot from 15 to 32 km distance, where the battery is discharged from 97% to 90.5%. Only an agent-based model can consider all the details of the various trip characteristics for each individual agent. This modelling is conducted for all agents. Additionally, a 1% BEV share increase in the canton of Zurich corresponds to a 9 GWh annual increase in electric power demand.

Distribution grid
The simulated grid consists of regional and local distribution grids that have more than 9,800 powerlines, as shown in Figure 8. The power flow simulations are conducted for the entire grid, with an assumption of phase equilibrium. As a reference, the grid is simulated without BEV charging. Figure 12 shows the simulated load distribution in the regional distribution grid. Out of 8,700 powerlines, one line is operated above 50% of its capacity, at 63% average load. This shows that the system is neither over-nor under-designed. Indeed, the maximum difference between the hourly peak and the annual average confirms the design is satisfactory. Forty-two powerlines are operated for at least one hour above 50% of their capacity, and a maximum delta of 22% is found between the annual average and the hourly   Powerline hourly peak load of the regional (left) and local (right) distribution grid for BEV shares in the low-tariff 8 pm case peak. A similar analysis for the local distribution grid is shown in Figure 13.
A comparison with the regional distribution grid reveals that both the annual average and fluctuating loads are higher in the lowest grid level. Therefore, 11 of 1,100 powerlines are operated on average above 50% of their capacity, with a maximum average of 80.8%. In comparison to the fluctuating loads of grid level 5 in Figure 12, the fluctuations are up to 47% in the local distribution grid. This reference simulation indicates that as BEV penetration increases, more operational challenges are anticipated at the lowest-level distribution grid.
Below we present outcomes from agent-based traffic simulations coupled to power flow simulations for the entire distribution grid through the unsteady BEV power consumption model. Note that this is a future simulation that does not reflect today's grid operation and assumes simultaneous BEV charging with neighbourhood clustering effects in different low-tariff options. Figure 14 compares the impact of BEV penetration on powerline loading in the low-tariff 8 pm case.
In this case, each car is charged daily at 8 pm. Figure 14 shows that in this uncontrolled charging scenario, the highest loads in the regional distribution grid remain below rated capacity. How-ever, hourly peak loads are increased by up to 132% in the local distribution grid. The reasons for operation above 100% include clustering effects of future BEV ownership scenarios and simultaneous, uncontrolled BEV charging. Figure 15 compares the impact on powerline loads for 8.5% BEV penetration in the 8 pm and 10 pm low-tariff charging cases. Clearly, the loads are the largest for the 'low-tariff 8 pm' case, whilst loads are decreased in the 'low-tariff 10 pm' case. In the 'low-tariff 10 pm' case, ten powerlines of the local distribution grid are still loaded above rated capacity, but the peak line load is decreased from 132% to 122%.
These charging scenarios show that overloads can occur in the distribution grid, especially in the local grid at 400 V. However, both the temporal and spatial characteristics of the peak loads must be assessed to identify which elements face the highest risk. Figure 16 shows an anonymised future scenario map of peak powerline loads, shown for the 8 pm case with an 8.5% BEV share. The figure also reveals that overloads occur not only at the last connections but also along powerlines that are below main roads. Note that powerlines below main roads may be costlier to upgrade than the connections at the end-users.

FIGURE 15
Powerline hourly peak load of the regional (left) and local (right) distribution grid comparing low-tariff cases FIGURE 16 Future scenario map of georeferenced peak powerline loads and BEV owners in a local distribution grid. Simulated with an 8.5% BEV share in the low-tariff 8 pm case Powerlines below main roads can cause overloads because of the superposition of multiple end-consumer demands, including BEV chargers, which are fed through the main powerlines. The determination of which powerlines tend to be overloaded depends on the local grid architecture, BEV ownership and charging behaviour.

SUMMARY AND CONCLUSION
In this work, we demonstrate a novel physical bottom-up methodology to assess the impact of BEV charging on real and large-scale distribution grids that include thousands of buses and powerlines. The charging scenarios show that overloads result from BEV charging and that the lowest voltage infrastructure is the most critical. Hence, it is recommended that distribution system operators simulate their entire grid coupled with agent-based traffic and power demand models, as a 'reference' grid architecture may not accurately represent other local distribution grid architectures, customer behaviour and BEV ownership. Since hourly peak loads can spike up to 50% more than annual average loads, it is crucial to assess the operation of the distribution grid for all hours of the year and not just for 'representative' worst-case loads. Our future scenarios incorporate BEV ownership neighbourhood clustering effects and uncontrolled simultaneous BEV charging, revealing that critical loads were located mostly in underground cables below main roads in the local distribution grid and at branches with multiple BEVs. Consequently, distribution system operators should focus on demand time-shift measurements to decrease peak loads in these cables. In the canton of Zurich, an option that could decrease the cost of building of new infrastructure would be to incentivise BEV charging to around 10pm, as this change decreased the hourly peak load in our future scenarios from 132% to 122%.