The practical use of AI in EV battery technology

EV battery technology has advanced in recent years. However, several issues still exist; EV manufacturers face unique challenges in different stages, from design to battery material discovery, battery management systems and range estimation. Finding lasting solutions requires a holistic approach, such as AI for EV battery technology. AI techniques, including machine learning models (MLMs), are revolutionising the EV sector. This article will focus on practical AI applications in EV battery technology.  

Machine learning in EV battery research and development

Continuous EV battery research and development (R&D) is crucial. Batteries impact EV performance, safety and cost, and R&D teams are constantly exploring new battery materials to enhance performance and safety. However, this process is costly and time intensive, and R&D  teams now leverage AI techniques for battery material discovery, characterisation and manufacturing.

Machine learning (ML) is a stand out contributor to EV battery R&D, helping to simplify battery material discovery:

Step 1: R&D teams use ML with quantum mechanics (QM) and materials datasets like QM9 and the Chemical Space Project. This step visualises the trends of known materials in the database.

Step 2: ML models exploit the trends visualised in step 1 to predict the anticipated properties of new materials from other known properties.

One of the latest breakthroughs in EV battery technology is from the PNNL-Microsoft collaboration, where researchers used an AI program on Microsoft Azure to screen over 32 million materials for low-lithium batteries. They identified a new sodium variant with 70% less lithium than a conventional battery. This material could cut lithium usage in EV batteries by up to 70% and help reduce EVs’ price and environmental impact.

Machine learning in battery management

The battery management system (BMS) in an EV ensures safety, efficiency and longevity by managing and optimising battery performance. It should estimate the battery state of health (SOH) and state of charge (SOC) and predict future degradation accurately.

However traditional methods for battery estimation, which include model based techniques such as ECMs and EMs, can’t predict EV battery states accurately, and EVs don’t have enough processing power to support real-time monitoring and control. The best option for real-time EV applications is ML-based battery state estimation; ML algorithms for BMS can analyse real-time data such as current, temperature and voltage, ensuring precision in SOC and SOH estimations. The BMS then uses these estimates to optimise the battery performance.

Another challenge facing EV owners and manufacturers is battery degradation. The capacity of EV batteries drops by about 10% after around 6.5 years of service. Overcharging and over-discharging contribute to capacity decline, complicating remaining useful life (RUL) estimation and degradation monitoring. Fortunately, ML models can handle these complex tasks. ML considers overcharge and over-discharge cycles and predicts non-linear battery capacity degradation trajectories. This data allows the BMS to extend the battery’s lifespan by minimising overcharging, over-discharging and other factors.

AI-based algorithms for BMS in EVs

EV manufacturers use ML to develop the RUL prediction model, and each ML approach has benefits and drawbacks. AI-based algorithms offer several advantages for EV battery technology over traditional methods. However, not all ML techniques are created equal. This section reviews AI-based algorithms for BMS in EVs and analyses adoption challenges.

XGBoost: Perfect for regression tasks. In a recent study, XGBoost provided near-perfect RUL predictions and the lowest root mean squared error (RMSE). This robust ML algorithm can handle complex datasets and help BMSs optimise battery performance and EV operational efficiency. EVs can also use XGBoost with electrochemical impedance spectroscopy (EIS) to visualise how future usage protocols will impact the discharge capacity. In one study, EIS predicted the next cycle and longer term cell capacity with less than 10% test error. It also doesn’t require historical data from the cell’s cycling trajectory.

Long short-term memory (LSTM): LSTM is a type of recurrent neural network (RNN) that processes and predicts issues based on sequential data. The LSTM model can help EVs maximize BMS potential. In a recent study, researchers applied LSTM to standardised data to estimate battery charging voltage, scrutinising its performance based on predictions from popular predictive models. The LSTM model delivered accurate charging voltage predictions for BMS, enabling proactive SOC and SOH management. Future hybrid models can combine LSTM with traditional regression methods to provide advanced predictive capabilities.

Deep learning models: popular deep learning models include FNN, CNN and LSTM. Each has unique benefits and drawbacks. A recent study focused on building deep learning models to predict battery capacity accurately and to capture degradation’s impact on battery performance. The researchers used non-destructive techniques such as EIS to visualise aging mechanisms in Li-cells. Then, they applied FNN, CNN and LSTM to raw data downloaded from the BMS. LSTM outperformed the other deep learning models in RMSE evaluation.

Random Forest (RF): RF is a versatile ML algorithm. This model executes classification and regression tasks using decision trees. In a recent study, researchers applied a dataset from GitHub on AI technologies to identify the best ML algorithm for BMS. RF obtained a more accurate dataset than other models. It also outperformed alternatives in discharge prediction. The main challenge with RF is its computational complexity, which increases with the dataset size. This issue undermines RF’s effectiveness for real-time BMS tasks with strict reaction time constraints. Future RF implementations for BMS can leverage deep neural networks or federated learning to address computational challenges.

ML in range optimisation

One of the issues discouraging some consumers from switching to an EV is inaccurate range estimation (RE). Traditional RE methods predict the EV’s driving range based on past energy consumption, without considering other important factors such as environmental changes, road conditions or differences in driving behaviours. Some EVs overestimate the range by around 50%.

Misleading information also contributes to range anxiety facing many EV drivers, but AI can help address these issues. Here are two practical ML applications for EV range optimisation:

For you

Be part of something bigger, join BCS, The Chartered Institute for IT.

Hybrid ML models: improving energy efficiency and reducing driving time can increase EV driving range. Range optimisation during driving can help EVs achieve this goal. A recent study proposed an LSTM-DNN mixture model that considers the relation between the EV’s electric energy supply and its mechanical energy demand. The researchers trained this ML model to exploit real-time speed, acceleration, map and traffic data. It also considered SOC and environmental conditions. This LSTM-DNN model had a range prediction accuracy of 2-3 km in a 40-minute time window. Its continual learning method creates room for continuous iteration and improvement. For example, EV manufacturers can train the LSTM-DNN model on battery aging to ensure range prediction accuracy against future battery degradation.

Transfer learning method: RE requires substantial data for accurate energy consumption and driving range prediction. This reality creates a challenge for new EV models with limited data from trips in real-world environments. One way to address this issue is by adopting a new transfer learning method. The idea behind this concept is to construct a prediction model for new EVs based on prediction models for popular EVs. This system can use data from EV trips collected by applications on drivers’ smartphones and the GPS of the car’s position. Smartphone apps can also collect SOC data when drivers stop to charge their EVs. Using this data, a manufacturer with a new electric car can construct a data driven ML model for range prediction. Future EVs can access data from other EVs in real-time through intelligent transportation systems for accurate range prediction and optimisation.

Conclusion

EVs are green mobility solutions that can help curb CO₂ emissions and protect Earth — but first, the sector must overcome EV battery development challenges, from design to power management during operation. AI in EV battery technology can help address these issues. Adopting ML for EV battery R&D, BMS and range optimisation offer several benefits over traditional methods. They improve EV performance, safety, energy efficiency, and economic viability, enticing more eco-conscious consumers. AI for EV battery technology is the now and the future.