Item:OSW4a6ab634411a4916a50ff318e41dfc0e: Difference between revisions

David Howey is Professor of Engineering Science at the University of Oxford, UK. He has an MEng from Cambridge University and PhD from Imperial College London. His research group focuses on energy storage engineering, where he has co-authored 120+ peer-reviewed journal and conference articles, and several patents, and is the recipient of grant funding from UKRI, the EU, Faraday Institution, and industry. He currently leads a £1M EPSRC project on grid integration of energy storage, and leads workstreams within the Faraday Institution’s £6M Multiscale Modelling project and two €5M+ EU projects (‘IntelLiGent’ and ‘DigiBatt’). Howey is organiser of the award-winning Oxford Battery Modelling Symposium, and a member of the editorial boards of IEEE Transactions on Industrial Informatics and Oxford Open Energy, plus the Strategic Leadership Group for the Ayrton Challenge on Energy Storage. He is also co-founder of Oxford-based spin-out company, Brill Power Ltd., who raised $10M in 2022

Okay. So here's an outline of the talk. I'll give some context, motivation, and just general background on our group. And then really the three main sections I want to dive into a bit more detail. The first one is on modelling, which I'll hopefully give you a bit of context for why I think it's important, and how it can have an impact. And then towards the end of that, I'll talk about how we parametrise models, which is a really important challenge, I think. Second, major topics on degradation and lifetime. Again, I'll try to place that in context and then talk a bit about some of the work we've done. And then the final one is on control. So once we've got all this information, we've done some modelling, what do we do? What decisions do we make? What actions do we take? And again, I'll try and show some examples of how we can use this information to make batteries perform better.

Okay. I'm sure this audience knows the context pretty well, so I'll just go through this quickly, but we know that lithium-ion demand is increasing rapidly. Cost has been dropping, astronomical amounts by 97% from the mid-90s to just a few years ago, and it's continuing to decrease. And this has really caused an acceleration in the rollout of batteries, not just for consumer electronics as it was in the early days, but these days, obviously for electric vehicles, grid storage and so on.

So the bit where we fit in is maybe to say, "Okay, how can we extend the life as long as possible and even know what it is going to be for a particular usage application?"

So what is systems engineering? I guess it's quite a broad topic within engineering, but from a batteries point of view, I suppose what we're trying to say is that we want to look after batteries better. Batteries are a bit like people. They don't want to be too hot. They're mechanically fragile and unfortunately they degrade over time.

And so you could think of these kind of three areas. For example, it's key places where the stuff you put around the battery, the thermal management system, the mechanical containment and the usage electronics and sort of measurements and control systems can make a real difference between a system performing really well and not performing quite so well. So these are the kinds of topics that I'll be talking about.

Just to give you a little bit of background on the situation that we have here in Oxford University. We're really lucky to have great heritage. This was the place where John Goodenough and team back in 1980 actually developed the lithium-ion cathodes material. There's a plaque on a wall down the road that commemorates this.

In my group, we are interested in a variety of things, but mainly we're interested in modelling, we're interested in diagnosing health and performance from data. We do quite a bit of testing mainly of lifetime, and then we are interested in how we control batteries to make them perform better. And this is just a photo from our spinout company that Simon mentioned in the introduction, Brill Power, which builds highly modularized battery systems with a lot of active management of balancing and so on.

And we've been lucky to have several successes in the last few years. I can't list them all, but I couldn't help listing a few of them, if you'll indulge me for a minute.

And these days, we're thinking quite a lot about grid energy storage and the interaction between battery lifetime and revenue streams that we could get from grid storage. And we were very pleased to be involved with this large project just down the road from here in Oxford, which is a very large 50 MW/50 MWh transmission-connected grid battery as part of Energy Superhub Oxford.

Okay, so hopefully that gives you at least five minutes of background. And what I want to do now is just go through in a bit more detail some of the work that we've done in these three areas.

And then we have other colleagues who go down to this level as well. And then up on the right-hand side of this, I think there are interesting conversations around long-term investment decisions, circular economy, and that kind of thing as well.

And modelling can help with all of these things. We can use modelling to get properties from structures, we can get performance from properties, and then eventually start making decisions around that.

We can measure voltage, current, temperature, but what we want to know is things like state of charge, which is quite hard to get directly, and state of health: how big is the capacity of the battery, and so on and so forth. Ideally, we want to do this in situ, non-invasively, without interrupting normal operation. So estimation is a big part of model impact as well.

And just to show an example of this, this is typically what's done actually by electrical engineers, control engineers. So, in something like a battery management system (BMS), we would typically take a model. The model would, for example, take a measured current, predict a voltage, maybe a temperature, then we'd compare that to the actual voltage, get some kind of error, and we'd feed that back and use that kind of negative feedback to drive the model towards what we hope is the real-world state.

We can predict lifetime, which is important for warranties and so on. We can understand integration issues such as losses and voltage ranges, and so on and so forth. And there's many other things that are not mentioned in this slide, but I just wanted to give you a feel for some of the areas where modelling makes an impact.

One of the things that I think it's important for me to say, and, you know, I hope many of us appreciate this, but there isn't a single model which solves all of your problems. I think sometimes in the literature in this space, I see people write things like, "The Newman model, the P2D model, is more accurate than a circuit model," or whatever. That may be so, but actually, I think I'd flip it around and say there are different models for different purposes, and it depends what you want to do.  

And people get uptight about the fact that electrochemical models and circuit models are different, but actually, I don't think they're very different to each other. I think the key difference between, say, this model here and the SPM, for example, is that you're modelling two electrodes separately. There's a diffusion PDE in there, and maybe the properties are a bit more physical, if you're lucky. But actually, I think this model here is reasonably close to some of the stuff in this bottom right corner.

So if you're interested in playing with models, if you're new to modelling or you just want to have a go, I would encourage you to have a look at this website, [https://pybamm.org PyBaMM.org]. This is open-source software. It's well documented and you can run some of it directly in the browser. So go and have a look at that.

One of the things I like to keep in the back of my mind, it's easy to get sucked down the rabbit hole of modelling. But actually, we should remember that most electrochemical devices, on a very basic level, we can think of them as a voltage source and a resistor. The voltage source, if you like, is representing the thermodynamics, and the resistors are representing all the losses. And this framework, although very simplistic, it actually maps quite well to different devices.

So, for example, I won't go through these in detail, but for a supercapacitor, fuel cell, etc., the voltage source is defined through these equations. In some cases, it varies with state of charge. For a battery, for example, fuel cell, thermodynamic voltage is basically fixed, ignoring pressure and temperature variations. And then the resistance, if you like, represents all of the other stuff that causes a voltage drop when we start to pull current.

And a battery is a good example. Here we've got ohmic losses, or ionic, and then kinetics reaction, and then some diffusion and stuff. So you might be thinking it's way too simple, but actually, let me try and convince you that it's a good starting point, at least.

In a good battery, if we take like a well-designed commercial cell that is matched to the application that you have chosen, I would argue that the open-circuit voltage should be much larger than the overpotentials. We don't want the voltage drop across this resistor to be huge, because that's going to generate a lot of heat and so on and so forth. So in a good battery, this should dominate.

Now, if we care about things like heating and temperature distribution and degradation, then of course, we need to worry about that. But let's not lose the wood for the trees, as we say. The open-circuit voltage is critical.

So, at this point, I want to just take a quick aside to say I think there are still lots of interesting issues in modelling to be worked on and important issues. I've just mentioned how OCV is critical. One of the things that really messes up OCV, if you like, from a systems point of view, is voltage hysteresis. So this is the first important issue.

Now, there's discussions we could have about what's causing that. And is it so-called dissipative or not dissipative hysteresis? But from a practical point of view, you're definitely going to see this on the timescales that battery management systems care about. So putting these into models is important. And I know Greg Plett has done lots of nice work on that. So, if you're interested, go and have a look at what he's done. In our group, we're doing experiments and modelling around this stuff at the moment.

So that's important issue number one. Important issue number two is that I mentioned earlier that people, in my opinion, are maybe too obsessed with the P2D model as the end of the road. Maybe all of the P2D model emphasis on local transport across the cell is a red herring in some situations. And I want to suggest to you that one situation where it might be more important to look at something else is when you have a large-format pouch cell.

And then we've done quite a bit of modelling of this setup using kind of coupled simplified electrochemical thermal models. And then we've fitted those models to experimental data, and you can see we can get pretty good fits. And you can see that what happens is you get this hotspot kind of forming towards the tabs, which gradually moves down, and you can also see the kind of breathing effect as the cell charges and discharges related to the entropy term. So this is pulsing at 4C with 100-second pulses at 30% SOC.

What does this kind of exercise tell us? Well, it's quite interesting. I think it tells us a number of different things. The first thing we learn is that in this particular case, a large-scale measurement like the surface temperature, under these experimental conditions using a thermal camera, actually, the shape of that measurement relates strongly to our assumptions about diffusion in the solid.

So macro temperature variations are important, particularly in larger cells, and maybe something we should be focusing on a bit more using simplified electrochemical models.

The next topic I want to talk about, just to bring us into the end of this section on modelling, thanks for sticking with me, is this need to parameterise models in general. And one of the things I think is an active area of research in many groups around the world, including ours, is building tools to help parameterise models. This can be done in a couple of different ways. We can either rip batteries apart and measure stuff, or we can take data from full cells and try and fit it with models. And in practice, we'll probably do a combination of both those things.

The beauty of this, by the way, is a well-known thing within control engineering. It's called system identification. So we're not claiming to have invented it. But the beauty of this is we've got a very flexible framework for trying lots of different models, lots of different optimisers, lots of different cost functions.

To show you some of the things that you could do, here are a couple of examples. One is parameterising model parameters. In this case, it's kind of a toy example, but we're just fitting a discharge curve. So we add some noise to the discharge curve, and then we want to estimate two parameters: the negative particle diffusivity and contact resistance using the single particle model. You can see here, we can plot a cost function by brute force, showing that there's a kind of sweet spot where you get the best fit between the model and the data. This allows you to visualise that and then try out lots of different approaches to finding the cost function. You can even try different cost functions if you want, which is what's shown on the right.

But perhaps more excitingly, we realised if you're building a tool that allows you to fit models to data, you can also tweak parameters for design objectives by changing the cost function. So, say instead of "I want the model to fit the data," the cost function becomes "maximise the energy density at 1C." Here's an example where we take the single particle model with electrolyte (SPMe), constant 1C discharge, and we want to optimise the gravimetric energy density by changing some geometric parameters.

In this case, we actually have four geometric parameters. So we say that we are allowed to change the electrode thicknesses plus the active material volume fraction and a representative particle radius. Sorry, five parameters. Six if you count pairs. What we see—it becomes hard to visualise once we have six things changing—but you can sort of slice through the landscape. What you can see is that there's always a kind of Goldilocks region, if you like, where there's a sweet spot where the combinations of parameters really work, in this particular case for a 1C discharge.

Interestingly, we also get insights into why it doesn't work when it doesn't work. So, for example, up here, maybe there are transport limitations, and over here, maybe the electrodes are unbalanced with respect to each other. So I think this is a really exciting new area to look at.

One side point about impedance—we've done a lot of work on impedance. I don't have a huge amount of time to go into it, but I think it's a really cool tool for parameterising models efficiently. Just to give you an example from some work a few years ago, this is the use of EIS data at low frequency. So we're actually going from about 1 Hz down to microhertz to estimate diffusion timescales. So that's like R² over D.

At the moment, we're doing lots of work on new extensions to normal EIS. One area of interest is the so-called non-stationary EIS. This is where the system is actually traveling, say, in temperature or SOC, and we want to extract the whole family of curves with one set of measurements. The other one is non-linear EIS. This is where we inject a high-amplitude sinusoid and excite harmonics, and then these harmonics help us to identify parameters better in models. The real expert on this is Noêl Halleman, so do look up his work if you're interested in this topic.

Lifetime is important. I think we all appreciate this, but let's just ground it in some reality. If you're an electric car owner, you want to know how long your car is going to last. How much will it be worth in five years' time? Should you sell it before the warranty runs out, for example? If you're investing in a large battery project, you want to know what the return on investment is. If you're doing maintenance on off-grid systems, you need to know how many spares to order. So lifetime modelling is important for these kinds of reasons.

A particular issue that's been talked about, including in this paper on kneepoints, is this idea that there would be an acceleration of the capacity fade curve in later life. So everything's fine at the beginning. We degrade along a reasonably shallow curve, but then something happens later on, and the battery starts to drop off a cliff. This has been shown in quite a few datasets, but I think there's still quite a lot of research to be done to understand why this is happening and what particular interactions of complex pathways are causing it. There are lots of things that can interact, which are shown in this picture here, leading to the kneepoints. So that's an important area of research.

If we think about the lifetime challenge in general, I think it's a two-level problem. The first level is diagnosing what the health of a battery is. In a real application, it's difficult to measure. To do a full measurement, you basically have to fully charge and discharge a system. That's not always practical. So knowing just where we are now—that's a challenge.

The other part of the problem is lifetime modelling—predicting what is the lifetime going to do in the future, given a certain set of assumptions about usage. From that, we can start making decisions. Let's talk a little bit about both of these.

First of all, let's talk about lifetime modelling. There are a number of ways of doing this, with physical models and electrochemical models, the kind of DFN stuff that I talked about a few slides ago. The basic idea here is that we would take something like a simplified electrochemical model. This is just an example showing a single slice through a single layer in an 18-65 cell. Then we would hypothesise a mechanism for degradation. It can be whatever you like—a popular one, for example, is solvent diffusion and then SEI growth. Then you can work out what the equations for that are, implement them in a tool like PyBaMM, and build a kind of virtual battery cycler to see how this is going to degrade under different situations.

Challenges with that are where do we get the parameters from? Do we actually understand what's going on? And there are many more. In later life, also, other things—interactions with different mechanisms—become a really tricky problem.

Some people, including us, have said, "Okay, that's very hard. Maybe we can just use data-driven modelling." In our group, we've used data-driven approaches and machine-learning approaches for some years, both to do diagnostics (that’s health estimation now) and prediction of future life. For example, there are some papers here about this where you just take a few voltage measurements and come up with a capacity estimate, and then some other papers where we predict future life based on correlations between capacity fade.

A big challenge with this is that we don't really have rich enough datasets. To generalise, you need not just large datasets so you can deal with cell-to-cell variability, but rich enough datasets, because so many things can change—temperatures, SOC ranges, chemistries, and so forth. If you take a model, train it on some data (such as this blue curve), and then predict it on a different test condition, maybe a different application, you're probably not going to get a good result. So we need this kind of richness of data. We need more data. If you're working on experiments, please release the data so that we can all see it. Maybe field data can help, and we've done some work on this.

To show you for a couple of slides what we've done with field data, we've done some work on using data from off-grid solar systems, mainly in Sub-Saharan Africa, where we have measurements of voltage, current, and temperature. We've worked closely with a company who has a lot of lead-acid battery systems in the field and is gradually switching to lithium iron phosphate.

We did this work on a thousand batteries. You can see some of the resistance trajectories that we extracted from the field. These have all been calibrated now. We combined these with some features that we extracted directly from the data, such as calendar age, charge throughput since the beginning of life, mean temperature since the beginning of life, and, in this case, the company has labels. Systems get taken to a workshop if they fail, so we could separately train a classifier to say, "This system has been independently tested and has failed or not." We were able to build a kind of life model that can predict whether a system is going to fail or not up to about two months in advance.

This graph shows the accuracy of that prediction under different assumptions. In the best-case scenario, it's about 85%, so it's not perfect. 50% here is flipping a coin, but it's about 20% higher than the benchmark, which doesn't do this kind of calibration step I talked about earlier. The cool thing about this is, first, we can give some useful information to an end user to say, "Look, the probability of failure on these 23 batteries is high." Then they can get the right stock to order. Secondly, we also get an ageing model out of this process, so we can look at what the classifier thinks is important. For example, in this data, obviously the resistance, which we fed in as an input, is important, but also the mean temperature and mean voltage are key factors impacting the life.

We are currently working on rolling this out to lithium-ion systems. This is just some example using NCA data. In this case, we can use this approach to predict capacity and resistance. We also get, in this case, fits of the resistance as a function of SOC and age, and also circuit parameters in an RC circuit model.

Once we have insight into how well or badly a battery is performing, the question becomes: what actions can we take? To me, this is really exciting because this is where we can start to actually show real-world benefits.

Imagine a situation where a battery manufacturer tracks data all the way through from components to systems that they build—pack diagnostics, BMS (Battery Management Systems), and so on—all the way through applications to end of life and recycling. They could then use later-life performance to influence earlier decisions. This would close the loop, creating a feedback system that spans the entire lifecycle of the battery.  

There are, of course, challenges with this vision. This is happening, potentially 10, 15, or 25 years later than when the cell was made, and by that stage, the technology may have moved on. But we can imagine smaller feedback loops within this process, and that's what I want to focus on now.

One of the simplest actions we can take is to adjust the way the battery management system controls the voltage limits to extend life. For example, in off-grid solar home systems, we’ve seen that there are high users, moderate users, and infrequent users or systems that are turned off. In all these cases, the systems are generally oversized—the battery capacity is larger than most people need, even for high-use cases. Different degradation rates are experienced by different use cases.

A similar logic applies to lithium iron phosphate (LFP) systems. For instance, my home battery has solar panels installed, and the software currently doesn’t allow me to control the upper state-of-charge limit. In the summer, the system charges up to 100% by midday and sits there until 6, 7, or 8 PM. If I could lower the upper voltage limit or adaptively adjust the SOC limits according to usage and weather forecasts, it could significantly impact the battery’s life because LFP systems degrade more quickly at higher voltages.

For larger grid battery systems, there’s a concept called energy arbitrage. This involves charging the battery when electricity prices are low and discharging it when prices are high to make a profit based on the price difference. However, we need to ensure that the revenue generated from this strategy exceeds the cost of aging the battery during these cycles.

However, if we had full control over the charge and discharge strategies, the approach could be much more nuanced. For example, if we exclude aging effects, we might aggressively charge and discharge the battery whenever there’s a small fluctuation in electricity prices. This approach maximizes short-term profit but accelerates battery degradation. Including a lifetime model in the optimization process changes the strategy. It reduces the number of cycles per day and focuses on maximizing revenue while minimizing degradation.

Companies working in this space often formulate this as an optimization problem. The objective is to maximize or minimize a cost function based on electricity prices, power limits, and time, while also considering the cost of aging. Constraints such as the finite power and energy levels of the battery are incorporated into the model.

The results were intriguing. In the first case, aggressive charge-discharge strategies resulted in significant revenue but caused rapid degradation. In the second case, penalizing the cost per cycle reduced cycling intensity but left economic value on the table. The third case, with the physics-based model, resulted in a more nuanced strategy, prioritizing long-term revenue by reducing degradation.

We took these three scenarios and tested them on six real cells. The cells were cycled using the optimized strategies for one year, and the results showed clear differences. The physics-based model outperformed the others by generating more revenue per percentage of lost capacity. It also followed a different degradation trajectory, demonstrating that complex models can yield practical benefits.

The key takeaway is that simple throttling of cycling doesn’t necessarily change the trajectory of degradation. It’s crucial to consider the impact of voltage and temperature. While simple models are easier to implement, they leave economic value behind.  

Implementing this approach in real systems requires tight coordination between the battery management system and the energy management system. This is challenging because these systems are often made by different companies. Vertical integration could help address this issue, but it remains a significant hurdle.

To summarize, control strategies can significantly extend the life of batteries and improve their economic value. By incorporating aging models into optimization frameworks, we can design more intelligent charge-discharge strategies that balance revenue generation with long-term performance. However, practical implementation requires overcoming integration challenges and ensuring robust communication between system components.

Thank you for sticking with me. I’ll now summarize what we’ve covered and leave you with some final thoughts.

The core idea behind systems engineering for batteries is to squeeze more out of the batteries we already have. This complements efforts to design better batteries, as both approaches are necessary for advancing the field. Throughout the talk, I’ve shared how we use battery modeling to support design, estimation, and control.

If you’re interested in exploring modeling further, I recommend looking into PyBaMM and PyBOP, two open-source tools that facilitate battery modeling and parameterization.

We discussed the importance of understanding battery lifetime, both for diagnostics and for making predictions about future performance. Battery lifetime modeling impacts warranties, maintenance, and investments.

* Collaborative efforts are needed to release experimental and field data to advance the field and enable generalizable models.

Finally, we explored how control strategies can be used to maximize battery performance while minimizing degradation. Control strategies bridge the gap between theoretical insights and real-world applications, delivering tangible benefits.

* Practical implementation requires tight integration between battery management systems and energy management systems, which is a significant challenge in many real-world setups.

None of this work would have been possible without the contributions of my research group, collaborators, and funders. A special thank you to Simon for his ongoing collaboration and support, as well as to all of you for taking the time to attend this seminar. I hope you found it informative and thought-provoking.

I’d be delighted to hear your questions and engage in further discussions. Batteries are a fascinating and critical area of research, and there’s still so much to learn. Let’s work together to advance the field and create more sustainable, efficient energy storage systems.

Thank you.