LITHIUM-ION BATTERY HEALTH ESTIMATION USING SUPERVISED MACHINE LEARNING BASED ON INCREMENTAL CAPACITY ANALYSIS

• Main Author: Maulanal Haq, Muhamad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86778
• Institution: Institut Teknologi Bandung

Battery is an energy storage device and its use is currently projected to increase in application along with the increasing use of electric vehicles, especially in Indonesia. Batteries need to be equipped with a Battery Management System (BMS) as a system that monitors their performance. BMS is an important battery management system in monitoring and maximizing battery performance, one of the function is monitoring the battery health or state-of-health (SOH) of the battery. The state-of-health (SOH) of a battery is the condition of its current capacity compared to the capacity in its new condition, which decreases as the battery degrades due to usage and exposure to environmental conditions. Estimating the state of health of a battery is a challenge, especially in the selection of methods and the characteristics of each battery. One of them is the estimation of the battery health of several different battery materials. In this study, it is intended to be able to estimate the health condition of batteries in several types of battery materials. To achieve this goal, the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology is used as an estimator development flow consisting of several stages. The first stage is the business understanding, it was determined that the goal to be achieved was to estimate SOH with one material dataset, which could then be applied to other materials. It is also intended that the estimator development only requires minimum computational cost. At the data understanding stage, several material data were determined to be the object and limitation of the research. This battery data is obtained from research data that has been published by the researcher, namely Lithium Nickel-Cobalt-Alumunium-Oxide (NCA) battery data from Aachen University, Lithium Iron Phosphate (LFP) battery data from Berkeley University, and Lithium Cobalt Oxide (LCO) battery data from Toyota Research University - MIT - Stanford, each of them has a different nominal capacity and voltage. In the data preparation stage, there are conversion, compression, feature creation, feature selection, and data cleaning steps are carried out so that the collected data can be used as training data, test data, and validation data for the SOH estimator. The feature creation process is done by combining several statistical features from voltage, current, and temperature. Then Incremental Capacity Analysis (ICA) is also used to obtain some features obtained from charging or discharging methods with constant current. From the input features obtained, feature selection is carried out by finding correlations with the target feature: SOH, to get features that have a high correlation, and provide the results of three main features derived from the ICA process. After data preparation, the modeling process is carried out using three supervised machine learning methods, namely XGBoost, SVR, and Stacking. The training data set for making estimators is 70% NCA material data, with 30% data as test data. Experiments were also conducted using the addition of features that have a lower correlation than those obtained in the ICA process. In addition to making the SOH estimator, the relationship between the health condition or SOH feature with the peak value of ICA and the number of cycles is also tested to see the process of health condition decline. The last stage is the evaluation stage where the estimator results that have been trained based on 70% NCA data with error metrics such as RMSE, MAE, and MAPE. The estimator model will be evaluated using 30% NCA data, and validation data from other different material data, namely LFP and LCO to see the ability to predict other materials. Based on experiments conducted by comparing the use of supervised machine learning: XGBoost, SVR, and Stacking. It was found that the Stacking method can provide a lower error than the standard XGBoost and SVR models with an RMSE of 0.6030 against the test data. Experimental results show that the combination of features obtained from the ICA process which are ICA peak value, voltage at peak value, and peak area, is a feature with a high correlation with SOH features, and by using only one training data, the NCA material dataset. The results of evaluating the estimator model with LFP and LCO validation data show that the SOH estimator made from supervised machine learning: Stacking, is able to generalize SOH predictions to other material data, LFP with the lowest error value RMSE at 1.0476%, MAE at 0.7676%, MAPE at 0.0084%, and LCO with the lowest error RMSE at 0.7317%, MAE at 0.4601%, MAPE at 0.0047%. In addition, it was found that the relationship between the decrease in ICA peak and the decrease in health condition (KK) of the NCA and LFP material data. Keywords: battery, BMS, estimator, lithium, ICA, SOH, supervised machine learning, LCO, LFP, NCA, stacking. ?