How Tesla dramatically improved driving range prediction |AI may defeat EV Range Anxiety - Part 3

Make my ease of ownership better, and I'll consider an EV.

This quote by Mary Barra, the CEO of General Motors, summarises the entire consumer perspective on electric vehicles (EVs) and their willingness to own one.

They (customers) start to lose range anxiety at about 300 miles of range.

On the others side, it couldn’t better describe the general public's perspective towards range anxiety.

Range anxiety is still an issue without robust charging infrastructure; therefore, accurate range prediction is the cornerstone of every EV trip. The driver must know how far the car can travel.

Mathematics gives you hope that every problem has a solution

Recently, we've spoken about traditional EV range prediction methods, all of which happen to be physics-based models and, consequently, can be up to 30% inaccurate. Such models rely on mathematical formulas to describe how much the various variables, such as car speed, terrain shape, air resistance, and others, discharge the battery, thus limiting the EV car range.

Accurate battery consumption estimation is the cornerstone of robust EV range prediction.

The naive approach to modeling battery consumption with physics formulas fails despite including more and more metrics, and the physics battery consumption models (i.e., formulas) fit only to a certain extent. The harsh reality is:

• Some contributing factors will easily be omitted
• It is difficult to determine to which extent each input affects battery consumption
• Physics models are not adaptable to variable and evolving factors such as driver's driving styles, battery degradation, etc.
• Physics models don't react to contributing factors that cannot be directly measured by car sensors (e.g., car air resistance coefficient).

Estimate the EV Driving Range with a continuous refinement

Which approach could yield far better and more consistent results than physics-based battery discharge models?

Suppose we could somehow learn how the car battery discharges on one journey.

In that case, we could have a mechanism to estimate the battery consumption on a similar one.

Furthermore, suppose we could find a technique to continuously update our knowledge on battery consumption on all future journeys.

In that case, we'd get a robust and sustainable instrument for estimating battery consumption on any kind of journey.

Machine learning algorithms (ML) do precisely that - learn how an actual car driven by an actual driver consumes energy in different circumstances and learns how much each input journey feature affects battery consumption. You can think of it as how a child learns through constant experiences and replication.

Combining machine learning as base elements, we can build an artificial intelligence (AI) battery consumption algorithm that will, with great precision, be able to estimate the required energy on any journey and, more importantly, continuously refine itself.

Speaking in terms of ML, the goal should be to:

• model the real-world drives and predict energy continuously accurately
• adapt to the new conditions, even to those not present in the learning phase, such as personal driving styles
• adapt to hidden (directly unmeasurable) factors, such as the car's air resistance mentioned above

The art of ML is to select the features that are significantly correlated with the output battery consumption and then train the model to find the connection between those two - to determine the contributions of each feature in the overall energy consumption.

If our AI model can learn how the EV car battery discharges on a Milan - Rome trip, it would know to estimate the battery consumption for all subsequent Milan - Rome journeys. However, the model wouldn't out-of-the-box know to predict the battery consumption on a Madrid - Barcelona trip.

The secret ingredient to help the model would be to break down the Milan - Rome trip into smaller distinctive parts and to build its knowledge about energy consumption on each trip part. Estimating the battery consumption on any new journey would be almost equivalent to collecting the energy consumption on all distinctive comprising parts.

Training phase

Initially, ML algorithms learn and establish the connections between car sensor data and energy consumption based on pre-recorded vehicle journeys (training set). The more diversity in the learning journey data, ML can discover better connections. Therefore, the resulting algorithm can better predict energy consumption on future real-world drives.

Re-training phase

The based model produced in the training phase continues learning from the upcoming trips and updates the connections between input features (car sensors) and energy consumption, thus adapting to the new driving conditions.

Due to its nature, the machine learning approach provides the following:

• The high level of personalization and adaption to a particular driver's driving style
• Consistent energy consumption prediction results
• Continuous improvement and adaptation to new and unknown journey types

Building the AI model

Use historical driving patterns to build an AI model for battery consumption for a specific route and trained on driver's habits

As previously said, the art of ML is the selection of features that highly correlate with the output battery consumption. Journey features range from car speed, temperature, air conditioning, and heating usage to terrain slope and road curvatures.

An interesting example of a feature typically not included in the model is the wipers on/off signal. Although the wipers consume energy, their contribution to the total energy consumption is entirely neglectable.

To discover the highly correlating journey features, we first start with as many features as we have at our disposal and then aim to retain the high-contributing ones.

To calculate each feature's contribution, we must be able to retrieve those from the journey dataset exceptionally quickly. Fast data retrieval and rapid hypothesis testing reduce the time to market and speed up the re-training of the model.

For this precise reason, we store the training journeys dataset in a proprietary connected car platform Mireo SpaceTime which particularly excels in split-second journey data retrieval. Namely, car trip data are highly non-uniformly distributed in terms of space and time, and off-the-shelf databases fail at scale. Therefore, we had to develop a highly specific storage and analytical engine - Mireo SpaceTime.

Using historical driving patterns to build an AI model for battery consumption - for a specific route and trained on driver's habits

Personalized EV range estimation - the results

Once the base model is built, it's ready to be used in in-car navigation systems, where it will continue to improve as it adapts to new, unseen journeys and driving styles.

The base model becomes extremely precise after only 500 kilometers driven.

We tested the model for battery consumption and range prediction on several EV cars, including Tesla Model 3, and achieved highly consistent and precise results.

• On most of the 100+ km journeys, the difference of our prediction of the remaining battery percentage at the end of a trip and the actual battery percentage never exceeded 1% of total battery capacity.
• Moreover, the prediction error was at most 4% of the total battery capacity.

What do EV owners expect?

Unlike gasoline car drivers, EV drivers on longer journeys simply cannot rely on finding a free and operating charger nearby. The sparse charging network, long charging times, and moderate EV range don't allow that level of convenience. And the carmakers that actively work to alleviate these problems for their drivers are precisely those with the biggest market share.

Three things are required to ease the ownership of an EV at the moment. Those are the up-to-date map of chargers, accurate range prediction, and GPS navigation, and should preferably come in an integrated system fashion.

The following blog post will illustrate why embedding those three hard requirements is the go-to way to mitigate the range anxiety of today's EV drivers and how to achieve it.