Range anxiety has always been a major factor in defining a customer’s response to EVs. Range figures offered by manufacturers can be misleading as they do not assimilate varied traffic and weather conditions that influence the range. EV adoption faces significant challenges for both existing and potential customers as a result of inept range estimates and insufficient charging infrastructure. Addressing these issues, Intangles’ Co-founder and Head of Analytics, Aman Singh, writes for the Times of India. The article emphasises the role to be played by physics-informed AI in carrying out accurate range prediction and makes for a very insightful read.
The advent of Electric Vehicles (EVs) has been a mercurial process. In India particularly, the adoption of EVs is hampered by several distinct factors, the most significant being range anxiety. In simple terms, the fact that a vehicle can be left stranded in the middle of nowhere due to non-specific, erroneous Distance-to-Empty (DTE) estimates provided by OEM-integrated range prediction consoles, is what creates unease in the consumer base. An EV with a viable battery which has a life of about 8-10 years and offers a range of 150-200 miles on a single charge, faces tough competition from an ICE (Internal Combustion Engine) platform loaded with the latest in driver-assist and infotainment features, longer service life, full-tank range upwards of 300 miles in city conditions and an extensive energy-sourcing network. Usability issues include the shortage of charging infrastructure that peeves the user. An EV battery pack sources power for torque generation via motors, lighting instruments, HVAC, and driving aids such as ADAS, ESC and a wide array of media systems. The performance of these systems can vary based on ambient weather, traffic conditions and driving behaviour attributes. These factors collectively contribute to unstable range prediction models.
Range estimates published by the vehicle integrated battery management systems are found to have a tolerance of 10-15%. The irregularities in range estimates are amplified furthermore in the rugged usage conditions of the commercial vehicle industry. The inaccuracies stem from a combination of issues. Firstly, SOC estimation itself can have variance resulting from the use of legacy book-keeping workflows as opposed to more advanced adaptive or hybrid models. Secondly, Range Prediction Models are uninformed of performance dynamics resulting from changing weather, traffic and road conditions. On top of all this, fleets have their own notions around SOC thresholds below which vehicles should not be operated. If we consider EV solutions and concepts deployed in last-mile mobility, long haul transport and mining, the above-described issues can manifest in range biases as high as 30%. These hassles are causing aware users to commit to ICE vehicles over more sustainable EV solutions.
The aforementioned challenges present an opportunity for Telematics and cloud-based interpretation of data. These opportunities though are not without challenges. In an evolving ecosystem, there is very limited standardisation in the context of EV data bus topologies and communication protocols. Access to minimal signals around active flow current, measurements around voltage and impedance, is a struggle. Adding to the woes is the orchestration of data from a variety of controllers such as the BMS, TMC, BCU, ESC, HVAC & media controls.
So how do we get out of this conundrum? The answer lies in artificial intelligence. Data scientists are progressively developing machine learning workflows that run on physics informed algorithms. These algorithms can interpret and in some cases, infer system-level functions in the context of SOC estimation and trending SOC gradients over time. The development life cycle of such data models entails leveraging telematics for collecting real-time data on distances covered between charging sessions. Publicly available signals including normalised motor torque, wheel speed and battery pack temperatures are also logged. A real-world usage profile is simulated by running the vehicles in diverse geographies and weather conditions. Forecasting models can be trained to predict the aforementioned performance characteristics. Deep neural networks can be trained to derive SOC Rates (change in SOC with time and distance traversed) as a function of predicted performance parameters, biased against ambient forecasts around weather and traffic. These workflows can learn to tide over irregularities in SOC estimation and dynamically changing discharge rates.
Telemetry data and user experience both suggest that temperature is a crucial component that influences discharge characteristics. When the weather is chilly, the chemical kinetics of the battery are slower resulting in reduced current output. The denser air also impacts the aerodynamic profile of the vehicle. When the temperature is excessively low, tires can harden, impacting rolling resistance. In the course of very high temperatures, the reaction rates in the battery are faster, which means that the pack is expending too much current and not generating enough power. Moderate temperature conditions are most conducive to vehicle range and battery health. The behaviour of parasitic consumers such as the HVAC and lighting instruments can also be mapped against weather forecasts.
Machine learning solutions for range prediction integrated with ‘Fully Networked Vehicles’, leveraging multi-parametric forecasts, deep networks for interpreting physics and cloud-based forecasts for ambients such as weather and traffic, can consistently deliver accuracies of 94-96%. With the outgrowth of Industry 4.0, we are in for a roller coaster ride that will involve bumping into complications like range prediction and discovering solutions through technological breakthroughs.
Source: Times of India