Development of data-driven method for capacity estimation and prognosis for lithium-ion batteries

• Main Author: Koul, Akhilesh
• Other Authors: Xu Yan
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/137313
• Institution: Nanyang Technological University

With an ongoing transition from traditional energy sources to renewable energy sources, which inherently are intermittent in nature, the electrical energy storage is becoming more and more important for managing energy production and demand. Due to this, lithium-ion batteries have emerged as the key technologies in the field of energy storages. However, to ensure that adequate safety and better performance, there is a need for health monitoring of the current battery state and parameters such as capacity, state-of-health, remaining useful life. Monitoring of these parameters ensures that the batteries are being used efficiently. Also, to maximise battery life, such health parameter monitoring opens up the possibility of optimisation of battery usage. Capacity, which quantifies the available energy in a fully charged Li-ion battery, is an vital index that can be used in interfering the state-of-health or the remaining useful life. This work explores the data-driven method that can be used in estimating the current capacity and forecasting the capacity trend for the future charge cycles of the battery whose internal health parameters are difficult to gauge. The features used to develop data-driven capacity degradation model are obtained from voltage, current and time measurements observed during the charging phase of the battery, which is operated under the constant current - constant voltage charging protocol. With the developed model, capacity degradation can be estimated in respect with cyclic ageing of the battery. Stochastic gradient boosting regression (SGBR) ensemble with an autoregressive integrated moving average (ARIMA) is used for capacity estimation and prognosis. Features obtained are used to train respective SGBR models with a target value as the actual capacity or true capacity obtained using coulomb counting method from consecutive discharge cycle. For prognosis, ARIMA models are developed to forecast the features for future unobserved cycles using observed features and used as an input feature in another SGBR model to provide the predicted capacity for the unobserved cycles with the confidence interval. In actual operation, batteries are seldom fully charged/discharged, therefore during online capacity estimation, not all the features will be available. To solve the issue of data unavailability during partial charge/discharge, the presented method does not require the full range of measurements for prediction. Instead, it uses the sets of time window, i.e. measurements belonging to different voltage and current ranges, and since the features are obtained during charging of the battery, it does not affect the normal working on the battery. Fuzzy intelligent system is implemented to deal with the missing value cases in the runtime phase. In addition to this, the proposed method also presents a method for missing value imputation in case one of the voltages or current feature set is not observed using one-step forecasting via the ARIMA model. The results are presented using three independent experimental dataset provided by the National Aeronautics and Space Administration(NASA) Prognostic Center and Center for Advanced Life Cycle Engineering (CALCE) Battery Research Group based in the University of Maryland, in which various experiments were performed on a lithium-ion battery. The results demonstrate the effectiveness and accuracy of the proposed framework for battery capacity estimation and prognosis.