Impact of electric vehicle charging demand on power distribution grid congestion

By 2045, most feeders in California need to be upgraded. These infrastructure upgrades depend on the extent the peak load on each feeder exceeds the capacity limit threshold. In Fig. 2, we show how feeders become increasingly stressed over time due to the growing EV charging demand. These overloads (colored red) generally start to appear in the population-dense areas such as the Bay Area, but are mostly less intense in the early 2020s—with the overload power below 25% of the existing feeder capacity. The early-overloaded feeders continue to become intensely stressed with the growth of EV charging demand while more feeders become overloaded with the expansion of EV uptake. By 2045, most feeders are severely loaded with the total load reaching nearly twice the current capacity. Neighboring feeders that are not yet overloaded (colored blue) tends to be very close to overloading as well—with less than 25% of capacity headroom remained.

The growth of overloaded distribution infrastructure with EV uptake over the years is shown in Fig. 3. The number of overload feeders (Fig. 3A) ramp up rapidly from mid-2020s to mid-2030s, with only 7% of feeders overloaded by EVs in 2025, but that grows to a total of 27% in 2030 and 50% in 2035. The growth slows down afterward, rising to 63% in 2040 and up to 67% in 2045. PG&E takes up most of the overloaded feeders in the 2020s, but SCE catches up after 2030. The total upgrade need (Fig. 3B) grows dramatically from 3.5 GW in 2030 to 25.4 GW in 2045. Although SCE has more feeders that need to be upgraded after 2030, PG&E has the highest requirements for capacity upgrades. This indicates that while the SCE territory has the widest range of overloading, the magnitude of overloading in the PG&E territory is generally the most severe.

The range of corresponding infrastructure upgrade costs is estimated using the 25th, 50th, and 75th percentiles of reported circuit upgrade cost per kW at different scales (Methods, Distribution Grid Data), and depicted in Fig. 3C. The range of circuit upgrade costs is quite broad, with the total cost by 2045 ranging from $6 billion to $20 billion. The total historical distribution grid capital cost of the 3 IOUs through 2022 is around $51 billion in total, with a split of around $18 billion from PG&E, $27 billion from SCE, and $6 billion from SDG&E (38). And the total operation and maintenance cost in 2022 is around $6 billion (38). The statewide EV-related upgrade cost in the next two decades is expected to be around 10 to 40% of the total existing distribution grid capital costs, or around 1 to 3 times the annual operation and maintenance costs. The increase in infrastructure costs puts an upward pressure on electricity rates, while the growth of EV load makes a downward pressure. Assuming a 7.5% rate of return (38) and considering only the distribution grid rate base, we estimate a negative net rate impact from EV growth for all three utilities: By 2045, the impact is around −0.006 $ kWh for PG&E, −0.011 $/kWh for SCE, and −0.062 $/kWh for SDG&E. This aligns with CPUC’s estimations (33). While electricity price may decrease, the overall bill of all consumers will still increase due to the growth of total electricity consumption, which will make up for the total upgrade costs. We note that the uncertainty could be even higher in actual future upgrade costs, since we are using near-term cost projections from the utilities and are not considering possible future change in the costs due to technology development or bottlenecks.

Among the three IOUs, PG&E accounts for the largest split of these costs in early years, and from 2035 SCE is estimated to need the highest upgrade costs. The difference in total upgrades and costs among different IOU territories are largely caused by the different numbers of customers they serve. As can be seen in Fig. 3D, the upgrade costs per customer in the three IOU territories are similar in scale, but the difference in their trends can still be observed. These differences are a strong indication that impacts to distribution grids are unlikely to be homogenous in other territories as well, and utilities may need to prepare for upgrades based on local conditions without relying on insights derived from other regions.

Despite these differences in the total upgrade needs and costs, some generalized trends can be observed in the relationship between capacity headroom left in each individual feeder and the possibility of overloading among the feeders with similar headroom, which is shown in Fig. 3E. In 2022, feeders with a headroom of over 50% of its capacity are safe from being overloaded by EV charging even with as high as 8% of EVs in total LDV stock. But when EV adoption grows to 25%, only the feeders with more than 80% headroom currently will be all safe. As EVs surpass 50% of the LDV population, only the feeders with over 90% headroom can completely avoid upgrade. But in general, the greater the headroom preserved in a feeder at present, the lower the possibility of requiring an upgrade in the future as the EV population expands. Even with a 95% EV adoption rate, among the feeders with over 55% headroom today, less than half of them will need an upgrade. Therefore, to reduce potential upgrade costs, a high-resolution forecast of EV charging load is essential to accurately determine which of the feeders are problematic.

EV charging patterns at different locations can vary significantly, as can be seen in Fig. 4. Workplace charging mostly start around 8:00 in the morning, when people arrive at work; and home charging usually start after 18:00 when EV owners get back; the start time of public charging tend to be distributed a bit more in the middle, when people are running errands. For home charging, chargers that are installed at multifamily dwellings have more events that start earlier in the day than those in single-family homes. And within public charging, L2 and DC fast charging events also have quite different distributions of charging behavior.

The variation in electricity usage at different types of locations can have a strong impact on how the grid is strained by the load. Studies have shown that overnight home charging puts more burden on the bulk level of the power system than daytime workplace or public charging (6) but a similar comparison at the distribution level is still lacking. We simulate EV charging profiles by bootstrapping from empirical charging data (separately for each type of charging location), based on different trip purposes derived from a travel demand model. This allows us to observe the spatial distribution of different types of charging load and how they affect the local distribution network differently. Fig. 5A shows the split of total travel demand in California among trips going home, to work, or to public locations. Compared to the split of EV charging demand shown in Fig. 5B, it can be observed that the share of home- and public-charging demand is higher than that of travel demand. The reason is twofold: 1) People do not necessarily charge after each short-distance trip, and their choice of charging locations within a tour depends highly on their preference, which is modeled based on survey data in our analysis (Materials and Methods, EV Charging Load Simulation). According to the survey, the majority of people prefer to charge at home (39); 2) During long-distance trips, people may stop on the way to charge at public-charging stations, causing an increase in public-charging load.

We categorize the feeders by the charging load type that draws the most energy from each feeder. The map shown in Fig. 5C depicts the spatial distribution of the feeder categorization. The majority of feeders are dominated by either home-charging load or public-charging load. Very few feeders are dominated by workplace charging, due to both the relatively lower travel demand and the mostly lower power level of workplace chargers. These workplace charging feeders are never overloaded according to our results, so we only discuss the home- and public-charging feeders in the following sections. This implies a potential to reduce or even eliminate infrastructure congestion by encouraging workplace charging.

Home charging–dominated feeders are generally more stressed by the EV uptake than the public charging–dominated feeders. After 2025, the number of overloaded home-charging feeders are nearly twice as much the number of overloaded public-charging feeders. A similar ratio is estimated for the total capacity upgrade needs in home vs public-charging feeders. By 2045, home-charging feeders statewide will need around 16 GW capacity increase, with a corresponding cost of $3.9 billion to $12.9 billion; while public-charging feeders need 9 GW of upgrade, with a corresponding cost ranging from $2.2 billion to $7.1 billion.

The substantial difference at the aggregate level is caused by nuances in the overload mechanism between home- and public-charging feeders. We examine the overload intensity with the ratio of total load over feeder capacity, and evaluate the overload frequency with the proportion of number of overload hours within all hours. The distributions of these two indices within home-charging feeders and public-charging feeders respectively are shown in Fig. 6. Home-charging feeders tend to have more intense but less frequent overloading, and public-charging feeders generally have less intense but more frequent overloading. This is related to the difference in both spatial and temporal distributions of home-charging versus public-charging demand. Spatially, 57.8% of the feeders are dominated by home-charging load, while the amount of home-charging energy takes up only 44.4% of the total charging demand. This means that home-charging load is more spatially clustered than public-charging load, even though they have the similar amount of total demand. Temporally, home-charging behaviors are usually clustered in the middle of the night, while public charging tends to be more spread out throughout the day. This difference in overload hours will be discussed further in detail in the next section. The temporal and spatial clustering of home-charging demand indicates the potential of reducing capacity upgrade costs by shifting the home-charging load, which will be discussed further in Discussion and Conclusions.

The specific hours of the day that overload tends to occur varies across different charging locations. Fig. 7 shows the probability density distribution of overload hours on home charging–dominated feeders versus public charging–dominated feeders. On the feeders that have more home-charging demand, overload tends to happen more after 18:00, and overloading is most likely to occur at midnight; In contrast, for feeders that have more public-charging load, overloading is more likely to take place during the day than in the night, and the highest possibility for overloading occurs in late afternoon.

The hour that a feeder overload is a result of the synergy between baseload and EV load patterns. In Fig. 8 we present two case studies on the peak load day of August 2045 to explain this synergy. The graph on the left shows a typical feeder where the majority of EV charging load comes from home charging. The aggregate charging load pattern on this feeder has a high peak during the night and valley during the day. On the right, there is a typical public charging–dominated feeder, which is coupled with some workplace charging demand as well. It shows a cumulative charging pattern that peaks during the day. The composition of the EV charging load type is the result of local travel demand, which reflects the local facilities and type of buildings available, and is related to the land use or zoning of this area. The home charging–dominated feeder is likely located in a residential area, which explains its baseload pattern that has a peak in the evening around 18:00. The public charging–dominated feeder, on the other hand, is possibly from a commercial area, where the baseload peak occurs in the early afternoon. The combination of baseload and EV load pattern determines the time that overload tends to happen. On the residential feeder, overload happens during the night; and on the commercial feeder, most of the overload take place during the day.

Despite the general trend of difference discussed above, it is worth noting that the overloading time in each specific distribution network is more complicated than a residential versus commercial categorization. Most of the feeders in our analysis are a mix of home, public, and workplace charging load, so to determine the possible overloading hours, the local electricity consumer composition needs to be examined carefully.

To test the sensitivity of our results to some key parameters and assumptions, we run the model under several extreme scenarios, results of which can be seen in Table 1. Improving EV energy efficiency from their current average efficiency to the highest observed efficiency can decrease the total capacity upgrade needs by 12%, while at the current lowest observed EV efficiencies, the total upgrade requirements would increase by 17%. Spatial shift of charging loads has the potential to reduce capacity upgrade needs. Allocating more charging events to the locations with more feeder capacity headroom left can reduce infrastructure upgrade costs by around 10%. And reducing the probability of home charging from 86 to 56% has a similar effect.

Overall, the sensitivity of our results remains within 20% of our baseline analysis under most of the scenarios, except for the extreme cases regarding DC fast charging. Removing all DC fast charging would result in a 12% drop in total number of feeders that need upgrade, and a 23% decrease in total capacity upgrade needs. On the other hand, if all public-charging events are conducted by DC fast charging, while the number of feeders that needs to be upgraded is only increased by 15%, the size of upgrade needs more than doubles due to the higher charging power demands. This implies that the adoption and usage of DC fast chargers need to be planned and managed carefully.

In 2016, the CPUC required utilities to perform ICA for the distribution system. The ICA maps from PG&E, SCE, and SDG&E provide publicly available data on the information of the distribution network down to the circuit level, including network pattern, feeder capacities, as well as hourly load profiles per feeder. The ICA data contains both thermal and voltage load allowances for their distribution circuit segments across each of the three major utilities in California. These capacities are calculated by running iterative steady-state power flow simulations at each node in the distribution system model: after adding existing baseload, incremental load is added per iteration to determine the maximum amount of load that can be penetrated without triggering thermal or voltage criteria violations (40). These load allowances are provided on an hourly basis by the maximum and minimum load days observed in each month of the year. Due to data availability issues of SCE, we use Grid Need Assessment (GNA) data for the feeder capacities in their territory. Unlike ICA data that is resolved by hour, the GNA data provides a single capacity rating value for each feeder, which is obtained by performing power system analysis on a single annual loading peak, and does not account for fluctuations from hour to hour (41). In our analysis, we adopt the ICA data in baseline year 2022, on the peak load day of each month (that is, the day with the highest baseload), when overload is most likely to happen. We then add the projected EV charging load to the baseload on each feeder to check whether the total load exceeds the feeder capacity, and the extent to which they are exceeded at each hour. The maximum overload observed on each feeder within each year is considered as the capacity upgrade need of this feeder.

PG&E provides data on their estimates of distribution grid upgrade projects and the corresponding investments in their Distribution Investment Deferral Framework (DIDF) map. Elmallah et al. (31) calculated the per-kW upgrade cost of each project, and divided these data into different scales to address the economies of scale (cost per-kW decrease with the increase of project scale). The authors use the 25th, 50th, and 75th percentiles of the per-kW costs within each scale group as references to calculate future grid upgrade costs. We apply the same methodology to calculate expected feeder upgrade costs according to our results of projected feeder capacity upgrade needs.

We use the CSTDM together with the EV Toolbox (42) to determine the trips that are made by EVs. CSTDM simulates all the trips that take place in California in one typical weekday, including the origin and destination traffic analysis zones (TAZs), the distance of each trip, the household that this traveler belongs to, which TAZ is the household in, as well as the purpose of each trip (whether it is going home, to work, or to other public use locations like dining, shopping, school, recreation, etc.). To determine which of these trips are made by EVs, we adopt the EV Toolbox to project future EV adoption until 2045. The EV Toolbox uses diffusion of innovations model to estimate the number of households that will own at least one EV in the future in each census tract, considering social-economic factors of the households and retirement of existing vehicles (37, 43). With the share of EV households in each area, we are able to sample the households in CSTDM that are expected to own EVs at each year and determine the EV trips in each zone from CSTDM considering spatial heterogeneity.

We calibrate the model to ensure that EV sales growth aligns with the Zero-Emission Vehicle standard in California (2), which requires up to 100% EVs within all new light-duty vehicles sold in 2035. With this assumption, we project a state-wide EV adoption of around 6.6 million by 2030—which is more than 30% above the 5 million goal to be reached by 2030 and over 24 million EVs on the road by 2045.

To simulate EV charging load by feeder, we need to determine where the EVs will choose to charge during the EV trips. CSTDM classifies the trips into short- distance (generally below 300 miles) and long-distance (up to 700 miles), and we analyze these two classes separately.

For short-distance trips, CSTDM provides the information of whether several trips belong to a tour. So we assume that people would charge after a certain trip in each tour, but not necessary after each trip in a tour. We categorize the trip purposes into home, public, and workplace, as the main types of charging location. Then we bootstrap people’s choice of charging location from the distribution provided by the eVMT survey (39), which involves 7,979 EV owners and gives the proportion of households that choose different combinations of charging locations (only charge at home, charge at workplace or public, charge at all three types of locations, etc.).

For long-distance trips, we assume that people might charge in the middle of a trip around every 100 miles. Given the origin and destination of each trip, the trip route can be estimated based on major corridors with Djikstra’s algorithm. We then assume that public charging would take place at the intersection of the routes, and the charging type at destination is determined by the purpose of the whole trip. We are then able to determine the locations of the EV charging events for both short and long-distance trips. The charging demand of each event equals the corresponding split of travel distance within the tour or trip.

Next, we allocate each charging event to feeders throughout the state by disaggregating charging events at the TAZ level to a census block level, so that each zone is small enough to be mapped to only one feeder. For home-charging events, the spatial disaggregation is based on population (44) while public and workplace charging events are based on the number of jobs (45).

Finally, we generate the charging profile for each charging event using bootstrap simulation. The empirical EV charging data that we use include a total of 6,458,576 charging session records from: 1) charging network providers including EVgo, BTC POWER, and ChargePoint; 2) charging infrastructure installed by utilities; 3) the eVMT logger data, which is collected by installing data loggers on a total of 300 EVs in multiple utility areas, to track their travel and charging behaviors (46). The records range from 2011 to 2023 and include information of the start and end time of each charging session, as well as the charger level and charger location. More information of the datasets can be found in Table 2. We first divide the datasets into pools of home, public, and workplace charging based on the charging locations. Then, for each charging event that we derived from CSTDM, we calculate the energy consumption corresponding to its travel distance assuming an efficiency of 3 miles/kWh (47). The energy consumption of the trip, however, is not necessarily the exact amount of energy that an EV end up charging during the following charging event—it is possible that the battery is not full at the beginning of the trip, causing the EV to charge more than the energy consumed during the trip; or that the parking time is not long enough, causing the EV to charge less. To cover this uncertainty, we divide both the CSTDM trips and the empirical charging sessions into multiple bins of charging energy: 0 to 5, 5 to 10, 10 to 15, 15 to 20, 20 to 30, 30 to 50, 50 to 80, and above 80 kWh. For each charging event that we derived from CSTDM, we bootstrap a charging session from the corresponding trip purpose/charging location pool (home, work, public) and energy bin in the empirical dataset. Weighted bootstrap is conducted for public L2 vs DC charging—based on the numbers of L2 and DC fast chargers (48); and multifamily home vs single-family home-charging—based on the numbers of single-family units and multifamily units in California (44). Based on the spatial distribution of trip purposes and travel distances given by the CSTDM, we generate the specific charging time, energy, and power level of each charging event from the empirical data. Finally, we add up the charging power of all the charging events on each feeder by hour and obtain EV charging profile by feeder.

We test the sensitivity of our results under several extreme scenarios. For the EV energy efficiency, we test the cases where all EVs are assumed to have the highest (0.22 kWh/mile) or lowest (0.47 kWh/mile) efficiency in the current battery EV models (47). To test the impact of varying charger power, we test the extreme cases of 0 and 100% DC fast charging within public charging by adjusting the weight in bootstrapping when sampling charging profiles from empirical charging records.

As described in the previous section, in the baseline scenario, people’s choice of charging location is derived from the eVMT survey, which reveals that 86% of the households would charge at home (39). However, it is possible that as EV adoption ramp up in the future, the proportion of households that performs home charging would decrease, as more and more consumers without access to home charging choose purchase EV. To simulate the other end of the spectrum, we test the scenario where the proportion of households that would choose home charging is equal to the share of single family units (56%) in California (44), while keeping the relative ratios of the other charging choices the same as eVMT survey results. This scenario also tests the potential effect of shifting home charging to public and workplace charging. Another scenario that tests the sensitivity to charging location is by changing the method of allocating charging events from TAZ level to block level. Instead of allocating by the ratio of population or number of jobs, we allocate the events by the ratio of feeder capacity headroom left in each block.

Y.L. and A.J. designed research; performed research; contributed new reagents/analytic tools; analyzed data; and wrote the paper.

The authors declare no competing interest.