Design and implementation of machine learning algorithms for degradation estimation of Lithium-ion batteries and electrochemical capacitors
Abstract
Lithium-ion (Li-ion) batteries and electrochemical capacitors (EC) are the preferred
energy storage technology in stationary storage systems as well as other critical applications to the decarbonisation of our infrastructure including electrical vehicles,
busses, trains and planes. The gradual degradation of energy storage devices, however, is a major concern for the adoption of storage technology across a wide range
of industries. Li-ion batteries and EC cells continuously degrade with time. Their
degradation rate is highly dependent on their construction and the wide variation
in operating conditions. Given that a cell’s degradation is an unavoidable effect, it
is critical to understand its degradation and predict its lifetime to ensure resilience
and reliability.
First, the research in this thesis concentrates on understanding which factors
affect the degradation of Lithium-ion batteries. Whilst the operating temperature
is regarded as one of the highest stress factors accelerating battery degradation, the
effect of other stress factors such as discharge and charge current, charge cut-off
current and depth of discharge, is an understudied topic. The research in this thesis
describes a half-factorial design of experiment (16 test cases) consisting of a total
of 96 batteries from two manufactures. Two machine learning algorithms, random
forest and lasso regression are trained on the generated data and subsequently ranked
all five operational factors and their two-way interaction. The results indicate that
the two-way interaction effects of charge current and depth of discharge are in the
top 3 significant stress factors for the capacity fade in Li-ion batteries, which was
previously not commonly accepted knowledge in the battery literature.
Secondly, in this thesis, EC degradation at extreme temperatures is investigated
due to a lack of studies addressing EC cell operation outside manufacturer specified
temperature operating envelope. This is critical for EC operation in extreme environments such as drilling where temperatures can reach up to 200 °C. The research
first develops a design of experiment approach to generate data for EC degradation
under high-temperature conditions ranging from 80°C to 200 °C. The obtained data
is then used as input to a Gaussian process algorithm that estimates cell degradation in pre-specified temperature conditions with a mean absolute per cent error of
1.47%.
Finally, the thesis addresses the problem of state of health estimation of Li-ion
battery cells. State of health is a variable that characterises the condition of an
energy storage device throughout its lifetime and is typically measured as capacity fade or resistance increases. SOH estimation is critical to recognise a sudden degradation of a battery cell and greatly affects battery state of charge (SOC) calculation. Therefore, the research proposes a battery chemistry and design agnostic
machine learning pipeline. The method uses Li-ion battery charge curves as input and estimates degradation measured as capacity. This eliminates the need for
time-consuming and computationally expensive physics of failure and electrochemical models traditionally used for SOH estimation purposes. The pipeline operates
by passing incoming data streams through a hierarchical sequence of processing
steps to fuse them into a model. Each step of the pipeline has the goal of eliminating or minimising the typical disadvantages plaguing machine learning-based
algorithms to date. Namely, the pipeline engineers feature by summarising each
incoming charge curve data stream to one single value, thus becoming robust to
different data captures. In the case of limited training data, the pipeline introduces
adversarial examples, minimising overfitting. Additionally, the pipeline also associates a confidence interval with each SOH estimation and re-calibrates the models.
Results indicated that when deployed on Li-ion batteries subjected to a fast-charge
protocol, the pipeline achieved a mean absolute percent error of 0.45%.
Overall, the research in this thesis highlights the promise of combining machine
learning models with a design of experiment-based data generation phase for a better understating of degradation of complex dynamic electrochemical systems such
as lithium-ion battery and electrochemical capacitors. Furthermore, the proposed
methodology accurately estimates state of health of complex non-linear systems and
has the capability to scale to other energy storage devices in the future.