| Abstract [eng] |
Lithium-ion batteries have a high energy density, making them appealing for applications in portable electronics and, thanks to features such as fast charging, long discharge cycles, and low self-discharge, in electric vehicles and even in electricity grid infrastructures. However, lithium-ion batteries face the problem of battery ageing. Battery degradation can lead to failures in electronic systems and machinery, leading to losses for companies, consumers and even life-threatening risks. Modelling and monitoring of battery ageing is, therefore, a relevant topic. An analysis of recent scientific literature reveals a variety of models applied for aging monitoring of lithium-ion batteries. Models of different complexity are used to simulate battery processes, depending on the available knowledge of the battery. Machine learning-based models are computationally simple and only require easily measurable battery performance data, making them suitable for use in battery management systems. These models provide estimates of battery states relevant to ageing monitoring. During the study, an equivalent circuit model is analyzed to assess the state of charge and state of health of a battery. The model is defined by a state-space model whose parameter distributions can be updated using Bayesian methods as more battery consumption data is collected. The model parameters include the parametric expression of the dependence function between open circuit voltage and state of charge f_OCV(SOC). This dependence function is an important characteristic for determining battery states, making the estimation of the model parameters a key research task. To estimate the model parameters, an adaptive Particle Markov Chain Monte Carlo algorithm is developed, which is able to universally adapt to various types of battery data and can efficiently approximate the posterior distributions of the parameters. Experiments on real battery data show that the shape of the dependence function is approximated by a 13th-degree polynomial with two logarithmic terms to replicate the ends of the curve. The number of particles used in filtering was found to influence computation time, thus the minimum number of particles required to obtain results was determined to be 400. A value of 0.86 for the Markov chain jump factor τ was found optimal to cover the parameter space of the model. The adaptive Particle Markov Chain Monte Carlo algorithm was used to estimate the states of two real batteries. For both batteries, the estimate of the state of charge is obtained by calculating the open circuit voltage using measurements of current and output voltage as well as the mean of the posterior distribution of the battery's internal resistance R, and by interpolating the inverse of f_OCV(SOC). The mean value of 0,6789 for the EVE LF173 battery of the state of health indicates strong degradation. For EVE LF105, the mean value of 0,9924 of the state of health means that the battery is similar to a newly manufactured battery, and no significant degradation is observed. The economic value of assessing battery states is justified using cost-benefit analysis. Three cases of lithium-ion battery production are compared: production from virgin materials, production from recycled materials, and battery repurposing for second-life applications. The economic analysis on a global scale shows that battery recycling and second-life repurposing cases are more cost-effective than the production of batteries from virgin materials, provided that the investment for the necessary infrastructure is limited to $51.51 billion and $54.36 billion respectively. A cost-benefit analysis over 5 years with a discount rate of 0.1 estimates that the net present value of the battery second-life repurposing case is greater than the net present value of the recycling case by $2.85 billion. |
