Readers of this blog have long seen it coming: Batteries are becoming a key component of our energy and mobility future. With this, the topic of battery performance and health becomes vital to the business models of companies around the world. Deepening our understanding of batteries has never been more important, yet at the same time it has never been more challenging to keep up with the rapid developments in material science, production processes, supply chain and business models.
To make informed decisions in an ever-accelerating battery world, companies need reliable data, from the electrochemical properties of batteries to their real-life behavior in the field. This article will give an overview of the field performance of batteries, and present the tools and methods that are available to collect, manage, and draw conclusions from battery field data.
1. Complexity overflow
Batteries are complex in several ways, including a multitude of cell chemistries, form factors, interconnection types, pack layouts, cooling strategies and BMS designs, to name just a few.
Adding to these technical configurations is the challenge of a multi-layered supply chain. As an example, system integrators typically have 3–5 battery suppliers, each of which makes completely different cells. Every 18 months, new product generations are launched with little to no connection to previous product lines. Add some widely changing usage patterns and lifetime expectations of 10 to 20 years into the mix and the title of this paragraph might start to make sense.
2. A few questions about batteries
While complexity is not a bad thing in itself, it complicates decision making and can lead to suboptimal outcomes. Every single system integrator and/or operator of battery systems is asking these questions:
Does my battery work as planned? As a crucial component of the new energy and mobility world, batteries need to be reliable. But how is this being determined? Tracking the performance of batteries is essential to ensuring that their intended benefits are realized and allowing for future improvements.
One issue we see far too often is battery modules not behaving as the data sheet says they should. A battery management system (BMS) might (seemingly for no reason) derate the charge current to much lower values than specified, leaving users with longer charging times or empty batteries.
In another case, a battery system that worked seamlessly in northern Europe might struggle to operate reliably in more hot and humid environments, even though the data sheet suggests otherwise.
Battery performance tracking brings transparency to the unpredictable reality of a battery operating system, unlocking tremendous value for predictive maintenance or improved system designs.
Are we running into warranty problems? For batteries to be economically and ecologically viable, a certain minimum service life is crucial. That’s why batteries almost always come with long-term warranties. But how valuable are these warranties? The largest part of the paperwork for battery warranties usually consists of complex guidelines specifying the conditions under which the battery must be operated for the warranty to remain in effect — like limiting the time a battery can be stored above a certain state of charge (e.g., not more than 4 hours per day above 95% state of charge [SOC]) or setting rules on how much warranty coverage the operator loses for each hour the battery is above a certain temperature. Of course, these guidelines are different for every battery supplier, and oftentimes even between products or product generations. As a result, for a surprising number of the batteries operated today, it is not clear if their warranty is still valid or even how this should be evaluated.
What’s the (resell) value of my batteries? New batteries come with a price tag, but what about used ones? Whether a private person wants to sell a used electric motorcycle or project developers are negotiating a complex asset deal over a 500 MWh BESS, we need to be able to determine the value of batteries that have already been used by others. This is especially true for all second-life business models, where batteries are meant to be used in stationary applications after their first life in electric cars.
All these questions have one thing in common: With better information on battery performance, they would be much less of a headache. But how can we get this information? Let’s dive straight into the technicalities.
3. Getting the data you need
To show the whole process of making sense of field data, we will follow a data stream from its origin at the individual battery cell all the way to the colorful bar chart presented to management at the end of a quarter (see Figure 3).
3.1 The data source
A unique aspect of lithium-ion battery (LIB) systems is that they are, by their very nature, IoT devices which generate a continuous stream of data. The main reason for this is safety — if a LIB is operated outside of its defined state, it can overheat and catch fire or even explode (the dreaded thermal runaway).
The electronic circuitry safeguarding LIB operation is called the BMS (battery management system). A typical BMS scans a battery cell’s voltage, current, and temperature at around 1 kHz, or 1,000 measurements per second.
The data generated, however, varies widely: From the number of temperature sensors installed in a battery module (typically 2–8) to the quality of the DC current sensors (usually between pretty bad and downright horrible) to the performance of the microcontroller (8/16/32-bit), no two BMS’s are directly comparable in their output.
In addition to the measured values, the BMS also calculates abstract values such as state-of-charge (SOC) and state-of-health (SOH). The quality of these values, however, is often less than perfect due to limited local computing capabilities and the lack of comparable data. After the data is generated by analog sensors on the battery and digitized in the BMS, it moves up to the system level, towards the energy management system (EMS).
3.2 The gatekeeper
Batteries are almost always part of a larger system — be it a smartphone, an electric vehicle, or a storage container next to a wind farm. For this system to work, information about the battery needs to be transferred between the BMS and the EMS, which is done through the so-called SCADA (Supervisory Control and Data Acquisition). But since there is no need to transfer thousands of data points every second, the BMS only passes on a predefined subset of the raw data.
The kind and amount of data transferred varies widely from battery to battery. Two of the main distinctive features are:
- Granularity of data points
- Sampling of data points
Granularity of data points Some BMS’s pass the values of all installed sensors to the SCADA. This means one voltage per cell, one current for every parallel string and all available temperatures.
Other BMS, however, pass only the total voltage of a battery module (e.g. 51.2 V) as well as the average, minimum, and maximum cell voltage in that module (e.g. 3.92 V, 3.88 V, 3.95 V) on to the SCADA. If multiple temperature sensors are installed, the same aggregations may be applied.
In addition, BMS may share up to 80 additional signals such as locally-estimated SOC and SOH values, current limits, error flags and support voltages. However, this does not necessarily occur all the time, and companies generally use different naming and normalization conventions.
Sampling of data points Since a sampling frequency of 1,000 measurements per second is usually not required and certainly not practical for most battery applications, some form of time aggregation needs to be done by the BMS before data is passed on to the SCADA.The two most common ways to do this are by 1) passing on snapshot values (“every thousandth value, as it is”) or 2) applying a rolling mean to the data stream which is then sampled at a lower frequency (e.g. 1 s or 1 min). It goes without saying that a direct comparison of these two methods of data capture will not lead to meaningful results.
At ACCURE we found that, in order to achieve meaningful results regarding safety, performance, and aging, a minimum viable dataset needs the MIN/MAX/AVG values of every battery module with a sampling rate of no less than 1 reading per minute. More data always helps, of course, but if handled in a smart way, even seemingly limited amounts of data can go a long way.
Another important thing to consider are the measurements that you don’t have — like self-consumption of the BMS, which can silently drain the battery without a single amp-hour being counted.
With all values on the bus, let’s head for the cloud.
3.3 The transmitter
The SCADA communicates with the cloud through Wi-Fi or LTE/GSM. Depending on the use-case, data transfer can be continuous (e.g. stationary battery storage systems with a solid internet connection) or batched (e.g. electric ferries in French Polynesia that log into the harbor Wi-Fi once in a while). The key here is robustness. If you install a data logger in an electric boat on the other side of the world, it must be maintenance free, even under unexpected conditions — imagine maniacally switching the power and internet connection on and off while spraying salt water on everything. Good logging solutions will stoically reboot and get to work, no matter what.
Data volume optimization is crucial in every IoT application to save bandwidth and keep server costs in check. So how much data are we talking about for battery field data analytics? Here is a real-life example:
A single battery module sampling at a 1 s time resolution generates around 400 MB of data every month.
For a 50 MWh battery storage unit (which is on the smaller side of new installations in the US) this adds up to about 25 TB every year.
Happy computing!
Battery analytics are a true “Big Data’’ use case that will drive any local Matlab / Simulink setup to its limits and its users into insanity. In the above example, a workstation with 64 GB RAM could handle just under one day of operation data at a time. At ACCURE, we are currently approaching 1 petabyte of time-series data on our platform, with much more to come.
3.4 Stop! Quality check
By now you probably feel slightly uncomfortable with all the chaos we have allowed into our database while tracking battery field operation. Data cleaning manages this complexity, unifies formats, detects, and, if possible, corrects measurement errors. While there are some amazing methods and tools available to support us in the process of data cleaning — anomaly and outlier detection, for example, are areas where machine learning can really shine — the most significant part of the process is transitioning battery expert knowledge into functional SQL transformations and Python preprocessing scripts.
Before even beginning to analyze the data, you may want to ask yourself the following questions:
- What does it really mean if the current measurement shows 2.384 A, given its sensor type and data aggregation level?
- How can we homogenize values from systems with different time resolutions in a storage efficient way?
- How do we link the metadata, such as cell chemistry, nominal capacity, form factor or current rating, with the time-series data?
- Does this value spark joy (i.e., generate useful insights) or can we get rid of it altogether to reduce cost and complexity?
In all seriousness, data cleaning is probably the most important and the most underrated part of the whole process of battery analytics. Scientists love to rush to the “exciting” parts where they can test their ideas on real data or benchmark different approaches published by prestigious research groups against each other. But if the quality of the input data is muddled, the results of even the most advanced methods are worthless.
3.5 Applying some battery magic
Once all the data is clean and made accessible, the fun part can finally start — at least if you are a battery scientist. While lab tests are a great way to learn about the fundamentals of batteries, their scope is always limited in breadth. Or, put more bluntly, since battery testing is horribly expensive and time consuming, it is usually kept to the minimum necessary, and sometimes even less than that.
Field data is different. Once deployed, the batteries do their own testing under the most relevant load profiles in existence: Reality. By extracting open-circuit voltages and complex impedances from the field data for as many operating conditions as there are, we can build battery models (or digital twins, if you are feeling fancy) of the batteries. These models allow us, among other things, to predict future safety, performance and aging. Some of the models suitable for this goal are listed below.
Equivalent-Circuit-Models (ECM) use standard circuit components to model the electrical behavior of batteries. A simple example of an ECM is a voltage source, a resistor, an inductance, and several RC branches in series. The values for all elements are taken from multi-dimensional look-up tables depending on the batteries’ temperature, SOC or SOH.
Machine-Learning Models (MLM), which exchange scripted battery knowledge for data-driven insight, is the new kid on the block and has taken the industry by storm. Since 2020 alone, there have been thousands of publications on MLM, promising better lifetime predictions based on less data or improved SOC estimation from field data.
The process of model-based analytics allows you to reverse-engineer the most important aspects of batteries, such as aging mechanisms (Is it material or lithium inventory loss? Is there lithium plating?) and predict battery performance. In a perfect world, our models would operate in a closed-loop system that continuously optimizes batteries for safety, performance, and lifetime. In reality, this optimization process is usually not fully integrated, although companies have been moving in this direction.
3.6 Dashboarding
A common challenge for decision makers in the battery space is that managers and their engineers often need fundamentally different things — colorful bar charts vs. time series data that can be easily tweaked or enriched with a linear regression at a single mouse click. With all the data in one place you can easily do both: Connect business intelligence tools like Tableau, PowerBI or Looker to allow business analysts to dig into the (high-level) data and make pretty reports for management, while data scientists connect their Jupyter Notebook or Matlab environment directly to the data warehouse.
4. Results and outlook
Now that we have put all the pieces together, I’d like to discuss an example of how analysis of battery field data can solve real-life problems. For confidentiality reasons, I constructed a situation that mimics an actual event we have dealt with at ACCURE Battery Intelligence earlier this year. I will treat the case as “real” going forward.
ACCURE Battery Intelligence was contacted by a European bus operator that experienced a mysterious problem: On days when temperatures were low in the winter, the range of some of their battery electric buses dropped to extremely low levels. The reduced range restricted their operation and eventually left passengers stranded and freezing at the bus stop. Not good. Initially, everyone thought the range was reduced because the vehicles needed more electricity in the winter for heating. But THAT much? It seemed unrealistic. So we connected our S3 buckets, normalized the operational data (charging and discharging), integrated information from the technical data sheets and finally ran a few very simple queries. The results are displayed in Figure 7.
The analysis shown in Figure 7 solved the mystery: At low temperatures, the BMS limited the maximum charging current to extremely low values (orange dots), leading to batteries that could no longer recharge as scheduled. Our analysis empowered the bus operator to demand a correction from their supplier, which quickly promised improvements via an over-the-air update. Two weeks later, we keenly monitored how the charging behavior of the batteries had changed following the update (grey dots). Although the charge rates still came in below specifications in many cases, the overall performance was much better and the bus operator could return to its schedule.
However, one question remains: Why was the battery supplier so conservative in their charge limits at low temperatures in the first place? Was it just a mistake or are these cells potentially prone to lithium plating? Would an increase in charge current (even within the data sheet specification) lead to potentially dangerous side effects? To be sure, we proposed running mechanistic algorithms on the battery systems, which analyze, among other factors, the changing of the OCV to look for hints of plating and warn the operator if certain thresholds are met. So far, no precursors have been found, but we are ready for the next winter.
5. Get in touch
If the topic of online battery diagnostics resonates with you, I’d be super happy to chat — connect with me on LinkedIn or shoot me an email.
About the Author
Dr. Kai-Philipp Kairies is a scientist and entrepreneur focusing on innovative energy storage solutions. Since 2009, he has worked as a battery researcher and consultant in Germany, Singapore, and California. Together with his team at ACCURE Battery Intelligence, Kai-Philipp supports international companies in understanding and improving the safety and lifetime performance of their batteries using advanced data analytics.
Interested in publishing in BatteryBits? Apply at this link to become a contributor.
Join our discussion with other battery professionals at the community slack space Battery Street!
