Electric cars' global sales rose an astonishing 43% during 2020! That number includes both fully electric and plug-in hybrids. At the same time, the conventional automotive industry took a hard hit in 2020 due to the COVID-19 pandemic with a 15% decline in YoY sales. Expectations based on available projections are that EV (electric vehicle) sales will further increase in the upcoming years.
Nevertheless, according to Autolist's 2021 Electric Vehicle Survey, there are still several significant hurdles to overcome to achieve a wider EV adoption. Compared to traditional ICE (internal combustion engine) vehicles, electric vehicles exhibit:
Points 1, 3, and 4 are contributors to a common problem called range anxiety. EV range anxiety is the fear of running out of power on a journey and not finding a charging point. Studies show that driving range and a lack of charging infrastructure are the primary reasons people do not consider EVs when buying a new vehicle.
The range anxiety problem is further exacerbated by the remaining driving range (RDR) number calculated in electric vehicles, which drivers usually find pretty inaccurate. RDR is an estimation of how far you can get with the current battery charge. But, the energy consumption of an EV is not constant since it depends on the speed at which the car travels, the state of the roads, head/tailwinds, whether the roads are flat/undulating, and so on. Therefore, when the vehicle system reports an RDR of 200km, it can easily be anywhere in the range of 100 and 300km, or even more!
Typically being so inaccurate, the RDR becomes a part of the range anxiety problem, not a tool that may reduce anxiety. On the other hand, estimating RDR with much higher accuracy could dramatically improve the driver's confidence and thus alter the whole car industry towards electrification.
But is high-accuracy RDR estimation even possible? If yes, what could be the right approach towards the solution?
The basic idea to tackle the RDR estimation problem is to determine battery consumption for a specific route. Many scientific papers tried to address the problem with a physics-based model, taking into account the kinetic and potential energy change, rolling resistance, and drag (air resistance). The main problem with this approach lies in the factors that depend on the vehicle speed - kinetic energy and drag. We don't know the exact vehicle speed at every moment of the trip before the trip. Thus we cannot utilize a physics-based model without additional models estimating the vehicle speed.
Furthermore, many naïve implementations of the battery consumption model often employ a linear model with just a couple of factors influencing the battery consumption, like road segment length, the elevation difference, and accessory status (A/C, headlights, wipers, etc.). The problem with this approach is that neither is the linear model sufficient for describing all the relations, nor does it include many other relevant factors. Vehicle mass (including passengers), temperature, varying battery consumption depending on speed, typical traffic conditions for the particular time of day, road type (urban, suburban, highway), altitude which influences air density and therefore also drag, driving habits of different drivers, wind direction and speed - all influence RDR.
The more promising idea is to use machine learning methods and "teach" the RDR estimation system how an actual car and actual driver consume the energy in different circumstances. This approach alleviates all the problems mentioned above and allows for a personalized model for every vehicle or driver. Personalization in a car is not a new concept, and it makes a lot of sense. For example, driver profiles which remember seat and rear-view mirror positions are widely available for more than a decade. Similarly, we can create an intelligent system in a vehicle that can learn one's driving patterns and predict battery consumption more accurately depending on who is driving the car. This model keeps learning as one's habits or driving style evolve and vehicle and battery characteristics change.
The machine learning approach to RDR estimation opens, however, a number of questions. Would it be accurate enough? For how long do we need to train the system before it becomes acceptably accurate? Which set of sensors and other values do we have to take as inputs into the model?
To prove the feasibility of the machine learning approach, Mireo has developed an RDR estimation system using AI models for several production EV cars. We have leveraged Mireo SpaceTime, a connected car data platform for remote vehicles sensors' data acquisition and processing, to build and test RDR estimation models that turned out to be highly accurate and personalized.
What have we learned throughout the process? Well, for once, we don't need that much driving data to make a good prediction. Some 5,000 km is enough to develop a baseline (non-personalized) battery consumption model. Of course, the data has to cover driving in various circumstances. For example, it should cover different outside temperatures, road types, driving in the city, mountains, highways, longer and shorter trips, various battery charge (SoC) levels, etc.
As for the prediction accuracy, on 32 km long trips, we achieved a relative error of only 7%. It means that if the actual battery consumption were 10% of battery, we would typically make an error of 0.7%. The error is even smaller on the longer trips!
One of the EV models we tested was a Tesla Model 3, which uses a similar logic for determining the RDR. When a user enters the destination, the vehicle estimates the remaining battery percentage at the end of the trip. Our model's predictions were, in almost all the cases, very close to Tesla's.
But we actually took a step further there. Having to enter a destination to get a basic idea about your RDR may be cumbersome at times. This is why we calculate the maximum driving range in all directions and draw it on a map in the form of a reachable area. We also call this feature a spider map. One look at it reveals a whole lot more than just a number that is more often than not wildly inaccurate.
Genius Maps integration of EV battery consumption model – charging points and reachable area
Moreover, suppose your desired destination is outside your existing range. In that case, the navigation automatically routes the vehicle onto the most convenient charging stations along your route, so you can safely make it to your destination.
If all that sounds interesting, check out the Interactive Demo of Mireo's SpaceTime Connected Car Data Platform. The interactive demo presents the real-time processing and analyzing terabytes of data received from 200,000 vehicles. SpaceTime Connected Car Data platform offers all the necessary components to store and analyze real-time and historical data from millions of connected vehicles using an absurdly fast and completely customizable set of advanced analyses accessed by a standard SQL interface.
If you wish to hear more about how we can help you with EV battery consumption estimation, contact us.