//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>
E-mobility is advancing at an increasing pace, emerging as the reference model in the automotive sector. If electronics play a fundamental role in the architecture of electric vehicles, artificial intelligence (AI) is becoming an essential component—especially for monitoring and managing the state of charge (SOC) and state of health (SOH) of an EV’s battery.
AI technologies are currently used to optimally manage various factors crucial to the operation of EVs, such as powertrains, battery management and fast charging. AI will also play an essential role in the future, with the spread of vehicle-to-grid technology, wherein EVs will exchange electricity with the grid using a bidirectional energy transfer. Furthermore, AI is contributing to the search for new materials for battery manufacturing and the identification of new recycling techniques.
More importantly, however, the use of AI in EVs will enable users to monitor and predict the SOC and SOH of an EV’s battery, which is essential for prolonging battery life and increasing vehicle range.
Battery stacks based on lithium-ion (Li-ion) cells are used in many applications, including EVs and hybrid vehicles, renewable energy storage systems, and energy storage on the grid.
In these applications, it is important to measure both the SOC of the cells, which is an indication of the available residual capacity, and the SOH, which represents a measure of the battery’s ability to store and deliver electrical energy, compared with a new battery.
At any given time, the SOC can be computed according to the following formula, where Qrem is the remaining capacity in the battery and Qmax is the maximum (or rated) capacity as specified by the manufacturer:
You can also calculate the SOH with the formula below, where Qnom is the current nominal capacity and QnomBOL is the nominal capacity at the beginning of life:
With time, and after several charging/discharging cycles, aging and material degradation can cause the battery’s SOH to fall below its initial value.
As a result, the estimation of these parameters is not an easy task, as it depends on several factors and requires high accuracy to improve system reliability, performance and the battery life cycle. Besides traditional estimation methods, such as coulomb counting, voltage method, and Kalman filter, an AI-based approach could offer an efficient alternative.
One key function of battery management systems (BMSes) is accurate SOC estimation, which can help increase system performance and dependability, as well as the battery’s duration.
In fact, an accurate SOC assessment can prevent unexpected system interruptions and keep batteries from being overcharged and overdischarged, which could harm the internal structure of the batteries permanently. However, it is difficult to precisely predict the SOC under diverse operating conditions, because both battery charge and discharge involve intricate chemical and physical processes.
The standard method for determining SOC is to accurately measure the coulombs and current going in and out of the cell stack under all operating situations, as well as the individual cell voltages. To create a precise SOC estimate, this data is combined with previously loaded cell-pack data for the precise cells being observed.
This method, also known as ampere-hour counting and current integration, requires additional information, such as the cell temperature, the cell age, and whether the cell was charging or discharging at the time of the measurements.
At a given time, you can assess the SOC with the following equation, where SOC(t0) is the initial SOC, Cnom is the nominal (rated) capacity, Ib is the battery current, and Iloss is the current due to loss reactions:
Once an SOC has been assessed, it is up to the system to maintain the SOC throughout subsequent operations, effectively keeping track of the coulombs that enter and exit the cells. Not knowing the starting SOC to a precise-enough state and other issues, such as cell self-discharge and leakage effects, can compromise this approach’s accuracy.
Using a controlled discharge test, it is possible to determine a battery’s SOC, or its remaining capacity. The voltage approach uses the battery’s well-known discharge curve (voltage versus SOC) to translate a reading of the battery voltage into the corresponding SOC value.
By employing a lookup table of the battery’s open-circuit voltage versus temperature and correcting the voltage reading by a correction term proportional to the battery current, it is possible to increase the accuracy of this method. The voltage approach, however, is challenging to apply, as the batteries must have a constant voltage range. Moreover, the discharge test typically involves a subsequent recharge, making it too time-consuming to be considered in most applications.
Because the Kalman filter is an algorithm for estimating the internal states of any dynamic system, it may also be used to determine a battery’s SOC. For both state-observation and state-prediction issues, the Kalman filter offers a recursive solution to optimal linear filtering. This method inherently offers dynamic error boundaries on its own state estimates, unlike previous estimating techniques.
The Kalman filter calculates the values of the desired unknown parameters (like SOC) and provides error bounds on the estimations by modeling the battery system to include them in its state description. When providing real-time predictions, it then transforms into a model-based state-estimate technique that makes use of an error-correction mechanism.
The extended Kalman filter, which boosts the capability of real-time SOH estimates, can be utilized when the battery system is nonlinear and a step of linearization is required. The Kalman filter, however, needs a precise parameter identification and an appropriate battery model. It also requires substantial computer capability and precise startup.
A BMS is an essential part of any EV, as it oversees management and protection of the battery pack. The operations performed by BMSes include cell balancing, which makes the SOC and voltage of each cell equal, SOC and SOH estimation and fault management.
All the previously mentioned functions might be done using AI, which requires those same machines to be properly trained with some datasets. AI-based prediction provides several advantages over traditional methods, like higher accuracy and faster response.
To make BMSes more intelligent, the first step is to collect a huge amount of data that can be used to train the AI algorithm (normally some kind of neural network). Once the algorithm has been trained, a BMS will be able to predict SOC and SOH, protect the battery pack, perform cell-balancing operations and detect faults (such as overvoltage, overcurrent and overtemperature) before they can damage the battery.
To properly identify the battery parameters, a model, such as the battery equivalent-circuit model (ECM), shall be used. ECM provides an electrical representation of the charging and discharging characteristics of the battery using a set of resistors and capacitors with a voltage source and current, as shown in Figure 1.
The dynamic performance of the cell over its useful voltage ranges can be assessed using hybrid pulse-power characterization (HPPC) tests, which also provide accurate identification of the battery’s parameters and reliable battery models for the training of AI algorithms.
A plot of HPPC data at one SOC and one temperature is shown in Figure 2. The most comprehensive set of HPPC data includes a curve for each SOC, as well as curves for various temperatures, charging rates and after-pulses that discharge and charge.
A comprehensive set of this kind might have several hundred of these curves from various samples. Such information is required by a battery’s ECM to appropriately model the cell’s performance in various scenarios.
To train AI algorithms (neural networks) to predict SOC and SOH, data collected from HPPC testing can be used as input datasets. Because the number of training data affects the accuracy of the system, once the application has been trained with a large amount of data, it may make accurate predictions at various drive cycles.
Similarly, for fault management, the initial step is to gather a sizable amount of data by generating numerous errors under various operating conditions. This dataset can then be used to train the AI algorithms.
It should be noted that fault management can effectively take advantage of deep-learning (DL) techniques, which (as the name itself implies) corresponds to the innermost level of AI. An artificial neural network (ANN), which is a type of neural network used in DL, allows the algorithm to train itself. When applied to BMS fault management, ANN can predict potentially hazardous conditions with great accuracy and very low latency.
An example of an open-access dataset is the SicWell Dataset, as shown in Figure 3 below, available on IEEEDataPort.
For modeling and troubleshooting purposes, the SiCWell Dataset includes data on Li-ion batteries used in battery EVs. In this experiment, Li-ion pouch bag cells of the automotive grade are cycled with current profiles appropriate for EVs, including cell-aging scenarios, current ripple evaluation, cell cycles and cell checkups.
Read the full article here