Electric cars, or EVs, are becoming more popular. We need more charging places for them. Building more stations is good. Making current stations work better is also important. This helps EV drivers feel less worried about running out of battery. This worry is called "range anxiety." We made a smart way to plan EV trips. It uses charging stations and your battery level to help you reach your goal.
This week, we shared a new, small AI model. It answers an important question: How likely is it that a charging spot will be free at a station in a few minutes? We learned that the most complex model is not always the best. By working closely on the model and how it runs, we made a very good prediction system. It uses a simple math idea called linear regression. This model is strong because it uses easy-to-get information. It works better than other simple ways. Our work shows that using simple real-world ideas with AI can help a lot.
Make the AI Model
We wanted the model to guess very well. We also wanted it to use only a few types of data. This makes it fast and easy to use. We tried different ways to build the model. We tested a decision tree and a simple neural network. A simple linear regression model worked best. It was also strong for this job.
We taught the model using real data about charging spots. This data showed how many ports were free at different times. We looked at ports in California and Germany. We included more data from bigger charging stations. These stations have more cars, so they show how people really use them.
What Data to Use
The model uses the time of day. This is a key piece of information. It looks at each hour as separate. For example, "9 AM" is one piece of information. "5 PM" is another.
How Data Affects Predictions
The "weights" are numbers. The linear regression system learns these numbers. They show how much each hour of the day changes the guess.
- A positive weight means that during that hour (like 7:00 AM), more ports get used.
- A negative weight means that during that hour (like 5:00 PM), more ports become free.
- A zero or small weight means that during that hour (like 3:00 AM), the port status does not change much.
These "hour weights" are what the model learns. They show how much the number of EV ports changes each hour. The model learns to show the difference between now and later. It uses the hour weights to do this.
The weights learned for each hour are very helpful. They show how the number of ports changes. Look at the chart below. It shows clear patterns based on when drivers use cars:
These show how port use changes each hour for the next 30 minutes.
These show how port use changes each hour for the next 60 minutes.
The model only gives a new guess when the change is big. This happens during busy times or at large stations. This makes sense for giving users helpful updates.
Testing the Model
We tested the model very carefully. We used real-world driving patterns. We looked at stations for one week. We checked them 48 times each day. This was for 30-minute and 60-minute future times.
We compared our model to a simple method. This simple method is called "Keep Current State." It just guesses that the number of free ports will stay the same.
This simple guess is hard to beat. Especially for short times. For example, in the US, only about 10% of ports change their status in 30 minutes. So, the simple guess is often right. It is hard for a model to do much better.
We measured how well the model guessed the number of free ports. We used two main checks: MSE and MAE. We also checked if the model could say Yes or No to this question: "Will I find at least one free port?" This is the most important question for drivers.
Results
The tests showed that our model is better than the "Keep Current State" guess. It helps most when ports are busy. It finds the times when many ports become free or used.
We looked at stations with at least 6 ports. We checked times from 30 to 60 minutes ahead. This is common for city driving. We checked if the model could predict if at least one port was free. We focused on when the model was different from the simple guess. This was for big stations during busy times.
The table below shows when we made a wrong guess. It shows this for busy times like 8 am and 8 pm. The number shown is like the MAE for this problem.
This table compares how often we guess wrong about finding a free port in 30 to 60 minutes.
Using our model means we make about 20% fewer bad guesses in the morning. We make about 40% fewer bad guesses in the evening.
Differences in Areas
We also saw that how ports fill and empty is similar in different places. But the number of ports that change is different. So, we need different models for different areas. For example, using a separate model for California and Germany worked better than using all data together. This shows we need to consider how people use EVs in each place.
The End
We made a simple AI model using linear regression. It is good at guessing when EV charging ports will be free. We made it simple and fast. We built it with the systems we already had. This is better than using complex methods that are slow.
Our model is much better than just guessing the number of ports stays the same. It works best when many people are charging. This helps drivers feel less worried. It helps them plan better trips. It makes driving an EV a better experience. This helps more people choose electric cars. We plan to make the model predict even further into the future. This will help more with long trips.