How analytics and AI can save our energy grid from electric vehicle overload

Leveraging advanced analytics and AI can unlock key insights, including learning which customers are attaching chargers, improving time-to-detection, increasing cost savings from load shifting, recognizing trends in attachment and identifying how many customers are charging during peak times.

Any model that attempts to interpret energy curves must account for events with a similar load profile. For example, large 220V appliances with sustained draw, such as a pool heater, can exhibit curves that might look similar to an EV charging cycle. It’s important to find signatures that fit an EV charging cycle, not just look for an amplitude spike.

A recent Smart Electric Power Alliance paper on EV charging cycles illustrates the varying usage spikes that can occur:

Not all customers charge the same way. Companies that are leading in EV detection have shared that any detection work they do must account for variability in behavior.

This includes everything from lifestyle differences (remote vs on-site work) to classes of cars with unique charging characteristics (size of vehicle and battery pack).

Daily habits and lifestyle variances have big impacts on a charging curve. Some customers may only charge as soon as they return from work, resulting in a clear, sustained spike in usage. Others may charge at bedtime, which could push the charging curve across days even though it’s one charging event. Others may connect the charger every time they return home throughout the day, creating multiple random, and often short, charging cycles. Utility companies need to be prepared for all scenarios.

Don’t fall into the fallacy of assuming a detection model that works well for a large urban area will work well for a smaller city or rural area. While still useful, they will likely be less effective.

Factors such as charging durations, times of day and predominant EV models may vary greatly by region.

There are various approaches in the Data & AI world to address the challenges with false-positives, behavior variability and local variability.

As previously described, false positives occur when energy spikes from non-EV sources trigger monitoring and detection algorithms. Data & AI capabilities can help prevent these scenarios by:

• Identifying and extracting patterns such as ramp time, frequency of usage and charging duration. This identification will likely happen over the course of weeks/months based on the amount of charging a customer performs.
• Ensuring similar load-heavy components and appliances have their characteristics and profiles accurately identified and provided to detection algorithms. Doing so allows for both positive and negative testing regarding EV detection.
• Using deep learning models such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to separate EV charging signals from other loads.

Energy usage varies significantly by user habits, charging frequency and time-of-day preferences. This makes pattern recognition more complex, and addressing behavioral variability can increase outcome confidence levels:

• Develop user personas based on charging patterns—daily commuter, weekend charger, traveler, etc. Leverage unsupervised learning such as Graph Neural Networks (GNN) and Wavelet Transform Clustering and density-based spatial clustering (DBSCAN) to segment users into behavioral groups.
• Implement adaptive machine learning models that adjust based on historical charging patterns. Apply long short-term memory (LSTM) or Bayesian Neural Networks (BNN) trained on energy consumption wavelets to identify and recognize trends. This also helps to address seasonal variations.
• Implement a continuous learning pipeline that refreshes and retrains models with new data each month. Real-time feedback loops, such as opt-in confirmations, can be used to fine-tune false positive rates as it relates to behavior.

There are significant variations in grid topology, infrastructure and regional EV adoption rates based on living locations—urban and rural, residential and commercial, and regulated and deregulated. Understanding this variability is crucial to appropriately interpreting analysis:

• Build and train models separately based on geographic region to accommodate local grid conditions. Consider leveraging transfer learning to adapt pre-trained models to new regions or areas where limited data may be available.
• Focus on integrating GIS data to build a spatial analytics layer that could be used to map EV charging density by area and circuit. This can then be overlayed with circuit configuration details to provide better overall capacity management.
• Use interpretable ML models such as SHAP and LIME to explain why a charge event was classified as an EV session. This should enable utility teams to validate AI insights and adjust detection rules based on local variations.

The pressure to invest in EV detection will only grow as EV buying trends continue. Utilities can benefit greatly by starting the foundational data and analytics projects that help them understand the impacts to the grid. By arming themselves with the right data and AI solutions, companies can stay ahead of the growth and ensure customer reliability.