Xvolta: New Testing of Energy Storage using Machine Learning
Using Machine Learning to reduce testing time of energy storage systems by 96%.
Our story starts as we were reading Bill Gates Annual Letter where he mentions: “If you wanted to store enough electricity to run everything in your house for a week, you would need a huge battery and it would triple your electric bill.”
When we think energy storage, we think batteries but based on what Bill Gates is saying batteries will not be not adequate to meet energy demands and we need new, cheap ways to expand energy storage across our globe.
We’ve all heard of needing to move towards renewables. You might also know that renewables have become cheap and competitive with current fossil fuels.
So then what’s the thing holding us back from using them?
Well, the main problem is exactly what Bill Gates mentioned; storing energy, especially renewable turns out to be surprisingly hard and expensive.
The IRENA estimates that the world needs 150 GW of battery storage to meet the target of 45% of power generated from renewable sources by 2030. For context, current global energy storage capacity is ~8 GW. We’ll need to almost 20x our global energy storage capacity to afford renewables.
This is where we come in.
Xvolta is a project that is hoping to use Machine Learning to reduce testing times of energy storage devices, specificially supercapacitors, by 96% (from 3 months to just 3.5 days). This will significantly speed up the development and help expand energy storage accross the globe.
Before we get into how we do this, let’s take a step back, really understand the root problem we’re tackling and how we got there.
We need to improve energy storage.
Current utilities and energy storage systems are simply not enough to meet the changing demands of renewable power, such as solar and wind, and maintain 24/7 reliability.
On top of this, renewables sources like wind or solar are now cheaper than many old coal plants. In the case of both utility-scale solar and onshore wind power, this rate has dropped to about $40 per megawatt-hour, which is lower than the cost of building new power plants that burn natural gas or coal.
BUT the economics of storing energy is still an issue for the renewables industry. The cost of batteries continues to fall, but when a utility needs to store electricity for more than a few hours, it’s still hard for the cost of that storage to compete with fossil fuels.
It is predicted that the costs of storing energy are falling and could be $200 per kilowatt-hour in 2020, half of today’s price, and $160 per kilowatt-hour or less in 2025. As costs continue to decrease, we need new ways to think about energy storage expansion and ways to power our world.
The key things here are: Expansion and reduced cost in energy storage. Costs will continue to decrease but we need more people focusing on the expansion and scaling of promising energy storage systems.
This is why we’ve decided to focus on the root problem (testing) of expansion which slows the development of some of the most promising energy storage systems, specifically supercapacitors.
Current Energy Storage Methods.
… and why they aren’t enough.
The main energy storage methods being deployed right now include:
- Mechanical Storage: harnessing kinetic or gravitational energy to store electricity. It mainly includes but is not excluded to pumped hydropower which is creating large-scale reservoirs of energy with water.
- Chemical (Batteries): a range of electrochemical storage solutions, including advanced chemistry batteries, flow batteries, and capacitors.
- Thermal: capturing heat and cold to create energy on demand or offset energy needs.
- Electromagnetic: energy stored in the form of an electric field or a magnetic field. These commonly include supercapacitors or ultracapacitors (EDLCs) and superconducting magnetic energy storage (SMES).
If you’re curious to learn more about this, check out this talk by Robert Piconi that breaks down different methods to tackle energy storage.
There will pros and cons to every energy storage system depending on how it works, its energy/power density, etc. and of course the technology.
For example, the basic premise of thermal storage is to convert surplus electrical energy into heat or cold that can be used later. But underground thermal storage relies on geography, energy stored decreases with the time due to the heat losses, and other technologies can be very expensive. These are some of the main reasons why it hasn’t been widely adopted.
Pumped-storage hydropower (PSH) are large-scale energy storage plants that use gravitational force to generate electricity. They are by far the most popular form of energy storage in the US, where it accounts for 95% of utility-scale energy storage.
Every good also has a bad side, which is why pumped hydropower generation includes high initial capital cost and reliance on site-location.
Yet, despite the widespread use of PSH, in the past decade, the focus of technological advancement has been on battery storage, which (spoiler alert) doesn’t make too much sense.
Batteries can’t solve the energy-storage problem.
Most of the battery storage projects are for short-term energy storage and are not built to replace the traditional grid. Most of these facilities use lithium-ion batteries, which provide enough energy to shore up the local grid for approximately four hours or less.
Another promising technology are supercapacitors, but not as many people are working on them.
Supercapacitors are promising, here’s why
Batteries and Supercapacitors (SCs) do a similar job storing electricity but they do this in different ways. While batteries use chemistry, supercapacitors use static electricity (electrostatics) to store energy.
A supercapacitor is a device that stores energy in the electric field created between a pair of conductors on which equal but opposite electric charges have been placed. Through this, energy can flow quickly in and out of capacitors with extremely high efficiency.
Capacitors have many advantages over batteries: they weigh less, degrade less because of little to no chemical reactions and they can be charged and discharged unlimited times without wearing out. SCs are 95% more efficient than the batteries which are 60–80% efficient under full load conditions.
The biggest benefit is that they have a higher power density. For example, lithium-ion, polymer, lead-acid batteries have different power density, from 1000 Wh per kg to 2000 Wh per kg. But SC power density varies from 2500 Wh per kg to 45000 Wh per kg.
As the costs of SCs continue to decrease, the main things holding us back from widely using SC are low energy density and slow development. As for energy density, lots of development in material science using carbon nanomaterials is happening to improve electrolytes and electrodes used. However, there aren’t enough people working on speeding up the development and testing of supercaps.
Supercapacitors: working principles
If you’re wondering how supercapacitors work, here’s a quick run-down on the most important components of this technology:
- Electrodes: electrodes are thin coatings that are applied to a conductive, metallic current collector typically made of porous, spongy material where the ions cling.
- Electrolytes: electrolytes consist of a solvent and dissolved chemicals that dissociate into positive and negative ions, making the electrolyte electrically conductive. The more ions the electrolyte contains, the better its conductivity. The electrolyte is important as it determines the capacitor’s characteristics: its operating voltage, temperature range, ESR and capacitance.
- Energy Capacity: determines the amount of energy a supercapacitor can help. This is determined by the capacitance and the operating voltage. These factors strongly depend on the electrode and electrolyte used in the system.
- Operating Voltage: The operating voltage is usually specified as the window of voltages (max and minimum) that you can operate a supercapacitor at.
- Capacitance: Capacitance can be described as the ability of a supercapacitor to collect and store energy in the form of an electrical charge.
- Internal resistance (or ESR): a measure of resistance within a supercapacitor cell.
The Problem: Lengthy Ageing Testing Process
Currently, testing for ageing in supercapacitors are done manually in labs. This can take upwards from several months to years and significantly slow down the production and output of supercapacitors.
Testing for ageing basically translates to testing for degradation of supercapcitors. This is a crucial test to preform so manufactures can evaluate under what conditions will the supercapacitor reach the point of no return when integrated within a certain application.
The key parameters that showcase the degradation of the supercapacitor are a decrease in capacitance levels and an increase in ESR levels. Ideally, a supercapacitor should have high capacitance levels and low ESR levels, however, ageing reverses that causing degradation.
The causes of ageing come from the key variables (these drive the capacitance and ESR levels of a supercapacitor):
- operating voltage
Identifying how certain configurations of temperature and operating voltage cause the acceleration of ageing is a process that takes upwards of 2 months or more to figure out.
Key Parameters: for causing ageing
Capacitance is a key measurement for knowing how well a supercapacitor can perform in different environments and over time, hence why it is a key parameter for ageing testing. Often a supercapacitor has “aged” once it has reduced in 20% of its capacitance. This is also called depth of discharge. As the percentage of discharge increases, the cycle numbers significantly decrease.
This is measured during discharge with a constant current source from its rated voltage to half its rated voltage.
Equivalent Series Resistance (ESR)
Charging/discharging a supercapacitor is connected to the movement of ions (which carry the charges) in the electrolyte across the separator to the electrodes and into their porous structure. Losses occur during this movement that can be measured as the internal resistance or ESR.
The ESR is time-dependent and increases during charge/discharge.
There are various different ways to test for and measure ESR (for ex., the Arbin method being a common one).
Operating Voltage and Temperature impact on Capacitance, ESR
Capacitance and Temp:
- The influence of voltage on the performance when held at rated voltage (highest voltage) and a lower voltage at its maximum rated environmental temperature.
- The graph shows a best case scenario, often the relationship between capacitance and temperature is not linear since it’s dependent on many other factors.
ESR and Temp:
- The influence on ESR when held at rated voltage and a lower voltage at its maximum rated environmental temperature.
- Test until we reach a 100%-140% increase in the ESR.
- Often supercapacitors held at lower temps last longer but at low-temperature extremes will increase the internal resistance of the cell. This is not a linear relationship, which is why the effect of temperature on a supercapacitor can vary a lot depending on many factors.
Our Solution: using ML to accelerate prediction of ageing
We present two Machine Learning models; ANNs and Bayesian Non-Parametrics Time-Series to predict the effect of various variables such as voltage, current, temperature on ESR/Capacitance.
Our goal is to accurately predict the time it takes to reach a 20% reduction of capacitance and a 100% increase in the equivalent series resistance (ESR).
With these models using only 4% of the lifetime data (or 30 cells) we can cut the time it takes to test aging by 96% from 3 months to 3.5 days. These models have an accuracy of 95% but by setting parameters, & training with more data for longer, we can continually increase this accuracy to near 99%.
Model 1: ANN
The ANN leverages two key pieces of data: the duty cycle and equivalent series resistance in order to make predictions for cycle life. The reason why an ANN is well suited for this problem is that neural networks are really proficient in making and drawing conclusions from data that isn’t linearly correlated. These two pieces of data are fed through a three-layer artificial neural network and a single output is returned.
ANN is a set of connected neurons organized in layers:
- input layer: brings the initial data into the system for further processing by subsequent layers of artificial neurons.
- hidden layer: a layer in between input layers and output layers, where artificial neurons take in a set of weighted inputs and produce an output through an activation function.
- output layer: the last layer of neurons that produces given outputs for the program. In this case the amount of cycles before it drops below a certain threshold.
Our data inputs would be Duty cycle and Equivalent series resistance. These are helpful for predicting final cycle life.
How the model works
Output & Results
We can do a cycle life prediction leveraging 4% of the data, at a 95% accuracy. We can predict the number of cycles before it drops below a certain voltage/esr/capacitance threshold.
The graph shows the Predicted specific cycle life over 3 months. It predicted cycle life based on the discharge curve of roughly 4% of the average cycle life data with a 95% accuracy.
This showcases, that if life cycle life testing of supercapacitors takes 3 months, we can cut the time down to 3.5 days.
Further, this model can also be used to predict Capacitance and ESR (this example was showcasing it being used for Cycle Life) but in this case, the inputs you would use would have to differ. These inputs would be similar to the ones used for the Bayesian NP model.
Model 2: Bayesian NP Time-Series Model
A Bayesian model is a statistical model that uses probability (or likelihood) to represent all uncertainty within the model. This can be uncertainty regarding the output or regarding the input (aka parameters) to the model.
Bayesian Models can use past and current data to make predictions about the future. They can also compare the performance of two different series of a different type for the same time duration. They are often used in various different forecasting & clustering applications for this reason.
They start out with a prior which is an initial belief they have and based on likelihood and data that you feed the model; it creates a posterior which is the updated belief:
Our model is similar to this expect it is a non-parametric model which means it can input an infinite amount of parameters and relationships between them based on time.
Our data inputs would be ranges of voltage, temperature and current over time (to predict) + capacitance and ESR over time (to validate).
How the model works
We would input our parameters into the input layer. These would just be a vector of inputs. These then get put through the hidden layers where the prediction happens.
Within the hidden layers, somethings called ‘load patterns’ are analyzed. These are essentially just time-series of current, voltage and temperature data between two capacitance and ESR measurements.
The changes of all the input data (current, temperature, voltage) over the time-series is then used to predict the effect on capacitance over that time-series.
On this graph, Q represents the capacitance at that time-series. We can see how it decreases as you continue measuring it. The red “cross” represents the new capacitance based on the changes or behaviour of voltage, current, and temperature at that time-series. This is measured until we reach a 20% reduction in capacitance:
Output & Results
Results show a relative accuracy on mean capacitance predictions that is within 5% of the actual values. This is around 95% accuracy of prediction.
Recap: the problem and our solution
By accelerating the prediction of ageing in supercapacitors, we’re confident that we can help supercapacitor companies produce 26x more supercapacitors.
This is considering that currently, it takes ~3 months to run through one batch → we can reduce this down to 3.5 days per batch.
We do this with:
- An ANN model that using Duty Cycle, and ESR data of just 4% of entire test data can predict Cycle Life with an accuracy of 95%.
- A Bayesian NP Time-Series model that using ranges of voltage, temperature and current over time (to predict) + capacitance and ESR over time (to validate) data can predict final ESR & Capacitance.
If we’re producing more supercapacitors, we can help widely deploy this as an energy storage system around the world. This is with the hopes that we can create a renewable future with cheap and scalable energy storage systems.
We’ve had the opportunity to speak with some of the top people in this field to validate our solution, here’s what they said:
- “It takes us an upwards of 2 months or even more to test lifetime and cycle life. If we can come up with an ML model to test for ageing and save us the time, this would be very helpful” -Head of Cell Development at Skeleton Tech
- “This is definitely a problem when developing supercapacitors and an interesting solution. It would be great if we can build this..” -Head of Technology at Zapgo
- “These kind of models for quicker estimation are absolutely welcome to make the overall system engineering more efficient and save time/cost”-Engineer at Tesla
We’re currently working on getting the data, cleaning the data and training our models.
We’d appreciate your help:
If you are connected to people at any of the following areas, we’d really appreciate an intro or connection:
- If you have any connections with people at top supercapacitor companies such as Maxwell (Tesla), Panasonic, Skeleton, etc.
- If you have or know of someone who has supercapacitor data (ranges of voltage, temperature and current over time + capacitance and ESR over time) that we can train our models on.
- If you’re connected with any of these funds or groups (Bill and Melinda Gates Foundation, Breakthrough Energy Ventures (BEV), MIT Solve + others!)
Make sure to follow our page at Xvolta to stay posted on future articles about our solution!