MAISY AI Agent-Based EV Models

AI Provides Unique Insights on Grid-EV Hourly Load Challenges

Why AI Models Are Preferred to Traditional Statistical Models for EV Forecasting

AI analysis overcomes limitations of traditional fixed parameter forecasting methodologies like regression analysis. Regression model parameters reflect the relative impact of different influences; however, they provide a relatively crude extrapolation approach to forecasting compared to the granularity and flexibility of non-parametric AI modeling approaches.

For example, an AI KNN nearest-neighbor agent-agent based model reflects EV ownership and hourly load charging impacts for individual households (i.e., agents) based on demographics, income, commuting and other factors identified as influential in ownership and EV use. These AI models apply machine-learning using a training database to define model relationships. The KNN methodology is described as non-parametric because it does not make any prior assumptions on the data or relationships within the data. Regression analysis requires pre-estimation specification of model relationships.

The non-parametric nature of AI model forecasts can identify ownership and commuting patterns missed by the more rigid parametric structure of traditional statistical models.

Paper Summary

Most electric utilities are struggling to understand how the rapidly growing number of EVs (electric vehicles) will impact their distribution systems. Our previous paper presented an AI EV ownership model and 2030 EV forecasts for individual Rhode Island ZIP codes. This paper describes an extension to the AI model to include hourly commuter charging kW demands for households, ZIPs and neighborhoods within ZIP code areas.

The AI model, which applies information on more than seven million actual US households including their commuting characteristics, is applicable to any ZIP-defined geographic area in the US including utility service and metro areas, states as well as neighborhoods and census tracts within.

MAISY AI Agent-based EV Modeling

2030 Rhode Island (RI) EV ownership forecast. The first table below shows the rapid growth of recent RI EV ownership, especially in recent years with approximately 50% annual increase in EVs over each of the three years from 2021 through 2023. The 2023 EV household saturation is 2.6%.

The AI agent-based modelis intuitively appealing; behavior of each household (i.e., agent) in a statistically representative sample of actual households is modeled providing a forecast of the entire population. Increases in EV ownership and charging load profiles are determined as follows:

EV Ownership Forecasts. The AI process determines the probability of EV ownership for each household. The EV purchase probability is estimated with an AI KNN “nearest neighbor” machine learning algorithm that draws on a separate household/EV (HH/EV) ownership “training” database of more than 26,000 individual households to calculate the EV ownership probability for each individual household. Increase in ownership over time is modeled by including households whose ownership probability is greater than the forecast threshold.

EV Charging Load Profile Forecasts. Commuting data for each household including commuting time, distance and arrival times are applied to model EV charging requirements for each hour after arrival home from work. Load profiles are calculated and applied for each household whose ownership probability exceeds the ownership forecast threshold.

Whole Building 8760 Hourly Loads Including EV Charging. MAISY models provide whole-building 8,760 hourly loads for each individual household prior to EV charging. These models, model results and their applications are documented in detail at https://maisy.com/menergy.htm . Hourly EV charging kW loads are added to whole building loads to provide information on the relative impact that EV charging has on current transformer, feeder and substation loads.

Example Commuting EV Charging KW Load Forecasts

The graphic below illustrates three distinct insights provided by MAISY AI EV charging forecasts. This example applies data from the MAISY AI model for ZIP code 02806 in Rhode Island.

Insight 1. Household EV Charging. The chart on the right top shows unmanaged EV charging profiles for a sample of individual households in ZIP 02806. The chart illustrates the diversity across individual households in timing and loads required to charge EVs after the evening commute. The diversity in charging timing and required charging loads results in an average that is considerably less than the maximum 7+ kW charge for a single level 2 in-home charge. As indicated in Insight 2, the ZIP 02806 maximum EV charging load average is less than 0.5 kW.

Insight 2. Household EV Charging. The chart on the right top shows unmanaged EV charging profiles for a sample of individual households in ZIP 02806. The chart illustrates the diversity across individual households in timing and loads required to charge EVs after the evening commute. The diversity in charging timing and required charging loads results in an average that is considerably less than the maximum 7+ kW charge for a single level 2 in-home charge. As indicated in Insight 2, the ZIP 02806 maximum EV charging load average is less than 0.5 kW. EV Charging kW Loads

Insight 3. ZIP Average Whole Building Charging Diversity. The second chart on the middle right shows the average 2030 forecast impact on whole building EV load charging demands for all households in ZIP 02806. The average increase in whole building residential loads is about 0.37 kW which is an increase in overall whole building loads of about 25 – 35 percent. Whole building loads include AC and other electricity use so this 0.37 average EV kW impact relative to the total whole building impacts varies by day. Load diversity along with the relatively modest 2030 increase in EVs (8 percent) in this ZIP limits the total impact of EV charging well below the maximum 7+kW impact of individual level 2 chargers.

Insight 4. Neighborhood Hotspots. With a forecast 2030 EV saturation of 18 percent and EV charging load diversity, ZIP 02806 would not seem to be at the top of EV charging concerns. However, drilling down on neighborhood characteristics identifies a potential EV charging issue for a nearly 600 household customer segment reflecting more affluent neighborhoods, identified here as households with incomes greater than $200,000 and dwelling unit square feet greater than 3,000.

EV ownership and charging demands for this utility customer segment is significantly different than the average 02806 household as indicated in the third chart on the bottom right. These affluent neighborhoods reflect an EV increase in whole house demand of nearly 1 kw which translates to an increase of about 50 – 100 percent of August whole building hourly loads depending on daily AC weather impacts. This level of unaddressed household demand can cause local voltage sags, reduce transformer life and even cause transformer failure.

Takeaways

AI models provide more granularity and flexibility than fixed parameter models. The non-parametric nature of AI model forecasts can identify ownership and commuting patterns missed by the more rigid parametric structure of traditional statistical models.
AI grid-edge EV load models provide new insights on future grid reliability issues with the ability to merge, integrate and process data from a variety of data sources in ways that maintain data integrity and provide statistically reliable planning tools.
Utility geographic distribution threats resulting from EV charging can be anticipated by identifying ZIP areas, census tracts and neighborhoods most likely to gain a large number of EVs and EV charging demands added to existing whole building loads.
AI model forecasts of EV charging impacts can be used to evaluate utility programs designed to mitigate the new EV charging demand including time of use rates, managed charging and vehicle-to-grid initiatives.

MAISY AI agent-based Models

MAISY AI agent-based models apply actual income, demographic, commuting and other data of individual households across the US to forecast electricity use, 8760 hourly kW loads, EV ownership and charging profiles. Models are available for any ZIP defined area (utility service area, state), neighborhood or census tract for any year. Model software and data are embedded in an Excel workbook providing easy access to input and output data along with user options to specify alternative forecast scenarios.