A method for remaining discharge time prediction of lithium-ion batteries under dynamic uncertainty

Abstract

Summary Remaining discharge time of the battery system in electric vehicles relates strongly to the decision‐making of driving. Subjected to the various uncertainties, such as modeling uncertainty, state estimation uncertainty, and future load uncertainty, the accuracy and reliability of the remaining discharge time prediction reduce, which will lead to range anxiety. A stochastic framework based on the state‐of‐charge estimation and prediction strategy is proposed to predict remaining discharge time against the uncertainty. Firstly, the equivalent circuit model is established to model the dynamic behavior of the battery. Through the analysis of different fitting functions of open circuit voltage and state‐of‐charge, the Akaike information criterion is introduced to select the best function. Secondly, the coestimator is employed to bound the influence of the model parameter uncertainty and noise uncertainty. Finally, the uncertainties are quantified by a probabilistic method, and a Monte Carlo–based stochastic prediction strategy is proposed to predict the probability distribution of remaining discharge time under dynamic uncertainty. Experimental results demonstrate that the proposed prediction framework is of great effectiveness as it can provide an accurate remaining discharge time interval under dynamic uncertainty, which helps to promote the energy‐saving driving and overcome the range anxiety.

State of health prediction model based on internal resistance

“…However, the diffusion coefficient of active substances inside the current collector was inversely proportional to the DOD. An increase in the DOD reduced the amount of electrolyte, lithium ions, and other active substances, which resulted in an increased battery IR 26‐30 . The aforementioned condition may have increased the battery's IR and degraded its performance.…”

Remaining energy estimation for lithium-ion batteries via Gaussian mixture and Markov models for future load prediction

“…Conventional model-based RDE or EoDT prediction methods are all sensitive to noises, initial conditions as well as the model uncertainties. To address this challenge, adaptive RDE estimation methods via Kalman filters (KFs) and particle filters (PFs) are proposed [8,11,14,15,16,17,18,19]. Although these filters mostly focus on the state of energy (SoE) estimation rather than the RDE itself, but they are still significantly useful for vehicle and range estimation applications.…”

“…To consider the uncertainty of the future conditions, probabilistic and stochastic algorithms are proposed. In [11] the battery future load is characterised through a Gaussian process and in [8], it is addressed through a Markov model with the states of the minimum and maximum of the load in a fixed length window of the historical data. Both methods address the uncertainties in the predicted results but the first approach is still unable to address the transient behaviour of load, and the second approach is still either over estimating or under estimating loads because of its based on the maximum and minimum load value for Markov model states.…”

“…Three main definitions can address the EoDT, (i) the time when the SoC reaches a predetermined lower limit [8], (ii) the time when the terminal voltage reaches a lower threshold cut-off voltage [9], (iii) the time when the battery temperature hits a lower or upper limit. While the temperature and voltage cut-off limits are normally set via battery cell manufacturer's recommendation, the SoC limit may be selected dependent to the use case and the preferred safety and reserved margins [8,10,11,12]. Based on either definitions of EoDT, remaining discharge time (RDT) is the defined as RDTt = EoDTt -t.…”

Two Layer Markov Model for Prediction of Future Load and End of Discharge Time of Batteries

“…In [8], battery RDT prediction is performed by assuming that future load is mean value of the historical load. Future load characterization based on Gaussian distribution with the mean and variance of the historical load is presented in [9]. A Markov process with two states of the minimum and maximum load obtained from the historical data within a fixed length window is used for representing the future scenarios in [2].…”