Google Unveils Simple AI Model to Predict EV Charging Port Availability

Google Research has announced that a lightweight artificial intelligence (AI) model has been developed to predict the probability of an electric vehicle (EV) charging port being available at a specific station within the next few minutes. The model has been designed to reduce range anxiety among EV users and to improve the efficiency of global charging infrastructure.

Range anxiety the fear of running out of battery before reaching a charging point has continued to be considered a major barrier to EV adoption. While the expansion of physical charging stations is ongoing, the intelligent use of existing infrastructure has been viewed as equally important for improving the driving experience. As part of this effort, Google’s new model has been created to forecast future port availability so that EV routing can be planned more reliably.

According to Google Research, the model has been developed using a simple linear regression approach after several architectures, including decision trees and neural networks, were tested. The linear model was found to be the most effective, offering low latency, high robustness, and reliable performance, even though more complex models were available. The approach has also been co-designed with Google’s deployment infrastructure to ensure real-world scalability.

The hour of the day has been used as the primary feature for the model. Each hour has been treated as a distinct indicator of charging activity, enabling the model to learn clear usage patterns. The model’s learned coefficients, or “weights”, have represented how port occupancy typically changes. Positive weights have indicated hours when ports tend to fill up, while negative weights have shown times when ports are more likely to be freed.

Training has been carried out using real-time availability data from EV charging networks in California and Germany. Larger stations have been sampled more frequently, as they reflect higher traffic and more realistic charging behaviour. The model has been built to predict availability for short time horizons such as 30 or 60 minutes.

To evaluate performance, data from 100 stations was used, and occupancy was sampled 48 times each day over a full week. The model was compared with a strong baseline that assumed no change in port status. This baseline, known as the “Keep Current State” approach, has been difficult to surpass because port availability often remains unchanged in short time intervals.

Despite this, Google’s model has delivered notable improvements. During high-traffic periods such as 8 a.m. and 8 p.m., the number of incorrect predictions has been reduced by approximately 20 percent and 40 percent respectively. The gains have been particularly significant at larger stations and during hours with high occupancy turnover.

Further analysis has shown that the model’s effectiveness has varied by region. Although usage patterns have followed similar daily shapes across markets, the magnitude of changes has differed. Separate models for California and Germany have therefore produced better results than a combined model, indicating that regional EV behaviour must be considered during training.

Google has stated that the model’s simplicity has enabled it to be deployed quickly and efficiently, improving EV navigation and offering drivers more reliable guidance. The company has added that the prediction system has the potential to reduce waiting times at stations, minimize uncertainty during travel, and support broader EV adoption. Future work is expected to focus on extending prediction horizons to assist long-distance drivers.