Reducing EV range anxiety: How a simple AI model predicts port availability

The transition to electric vehicles (EVs) is accelerating globally, bringing with it the critical need for a reliable and robust charging infrastructure. While building out more physical charging stations is an important step, an equally important task is maximizing the efficiency of this infrastructure and minimizing "range anxiety”, a term used to describe an EV driver’s fear of running out of battery before reaching their destination or the nearest available charging station. These concerns led us to design an approach for EV routing that reduces range anxiety by integrating charging stations into the navigational route based on the battery level and destination.

This week we announced a new lightweight, highly efficient prediction model that can answer the core question, “What is the probability that an EV charging port will be available at a specific station a certain number of minutes from now?” We found that the most sophisticated model isn't always the best solution. By co-designing the model and the deployment infrastructure, we were able to create a highly effective prediction system based on a simple linear regression approach. This model’s simplicity is its strength, allowing it to rely on easily accessible features while still achieving performance improvements over a strong baseline. Our work demonstrates that combining intuitive real-world logic with machine learning can deliver significant operational and user experience benefits.

Our goal was to maximize predictive power while minimizing the feature set (i.e., the specific, measurable data points the model uses to make a prediction) to ensure speed and low-latency deployment. After testing various architectures, including a decision tree and a simple neural network, a straightforward linear regression model proved to be the most performant and robust for this specific task.

We trained the model using real-time availability data from charging networks to calculate the true number of available charging ports within a certain number of minutes from the current observation time using criteria for model features and weights. We uniformly sampled ports from two distinct regions (CA and Germany). Larger stations were more likely to be included in the training set because they see more traffic than isolated ports and more closely reflect real-world usage.

The model uses the hour of the day as a key piece of information (a "feature"). It treats each hour (or hour range) separately. For example, "9 AM" is one feature, and "5 PM" is another.

The "weights" are the specific numerical values that the linear regression algorithm learns during training. These numbers dictate how much each specific hour of the day affects the final prediction.

• A positive weight means that during that hour (e.g., 7:00 AM), ports tend to get occupied (the occupancy is increasing).
• A negative weight means that during that hour (e.g., 5:00 PM), ports tend to get freed up (the occupancy is decreasing).
• A zero or near-zero weight means that during that hour (e.g., 3:00 AM), there is little change in port status.

These “hour feature weights" are the model's learned coefficients that quantify the predictable rate of EV port occupancy change for every hour of the day. Essentially, the model learns to express the difference between the current number of available ports and the future number of available ports as a function of the hour feature weights.

The feature weights learned for each hour of the day are particularly insightful because they directly represent the rate at which port occupancy changes. As illustrated by the chart below, there are clear, predictable trends tied to driver schedules:

Note that the model only differentiates from the current state when the change rate is significant (e.g., rush hour) or the station is large (more ports amplify the predicted change), which are intuitively the correct times to issue an updated prediction.

Our evaluation was designed to be rigorous and representative of real-world usage. For both the 30-minute and 60-minute time horizons, we evaluated predictions on 100 randomly selected stations, sampling their occupancy status 48 times daily (every 30 minutes) for a full week.

The model was benchmarked against a remarkably strong baseline: the "Keep Current State" approach. This baseline simply assumes that the number of available ports a certain number of minutes (H) in the future will be exactly the same as the current number.

While simple, this baseline is very hard to beat, especially over short horizons. For example, our data showed that on the US East Coast, never more than 10% of ports change their availability state within a 30-minute block. Since most of the time the state doesn't change, the simplest prediction — no change — is correct most of the time, making the task of adding predictive value extremely difficult.

We focused on two key metrics to measure the model’s accuracy for predicting the exact number of free ports: mean squared error (MSE) and mean absolute error (MAE). A ratio of MSE/MAE ≥ 1 free port measures the accuracy of the most critical binary task for the user: “Will I find at least one free port (Yes/No)?”

The evaluation confirmed that the linear regression model provides crucial gains over the strong "Keep Current State" baseline, primarily by correctly identifying the infrequent, yet vital, moments of high occupancy turnover.

We sampled test instances from among stations with at least 6 ports with horizons of 30 to 60 minutes, a realistic set of cases for charging in urban environments. We evaluated the model for the task of predicting the availability of at least one port in a station. This evaluation focused on the station profile and time of day when the model would differentiate from the baseline, namely large stations at times of significant rates of change.

The table below presents the fraction of time in which we provide a wrong prediction (which is equivalent to the MAE for this problem) for the times of highest change (8am and 8pm).

In summary, deploying the regression model allows us to reduce the number of bad predictions by approximately 20% in morning peak times and by approximately 40% in evening peak times.

Further examinations revealed that while the shape of the change rate curve (when ports fill vs. when they empty) is similar across regions, the magnitude of the change is distinct enough to warrant separate models. For instance, training the model separately for regions like California and Germany yielded better performance than pooling all data together, suggesting that it’s necessary to account for unique regional EV usage patterns.

We have successfully developed and deployed a lightweight, linear regression model that effectively predicts EV charging port availability. By focusing on simplicity, speed, and co-designing the model with the existing infrastructure, we bypassed the complexity and latency associated with more detailed, but often unscalable, approaches.

The resulting model provides a crucial predictive advantage over a strong "Keep Current State" baseline, particularly during high-traffic periods. This capability translates directly into an improved user experience: reduced anxiety, smarter routing decisions, and a better overall experience that supports the continued growth of electric mobility. Future work will focus on extending the prediction horizons to provide even greater value for long-distance travel planning.

We thank our collaborators Achir Ramadhan, Sreenivas Gollapudi, Shubham Gupta, Ilya Eyzerman, and Ivan Kuznetsov.