Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/68128
Title: | 機器學習用於預估鋰離子電池狀態及其電化學參數 On the Use of Machine Learning in Estimating the State and Electrochemical Parameters of Lithium-ion Battery |
Authors: | Chun-Shih Lin 林君實 |
Advisor: | 陳國慶(Kuo-Ching Chen) |
Keyword: | 鋰離子電池殘存電量,機器學習,轉移學習,老化參數,多任務學習, Lithium-ion battery remaining capacity,machine learning,transfer learning,aging parameters,multi-task learning, |
Publication Year : | 2020 |
Degree: | 碩士 |
Abstract: | 鋰電池是為現今最主要儲能方式之一,因此準確估算鋰離子電池之殘存電量及剩餘壽命是相當重要的,但此兩者資訊均無法經由直接觀察或測量而得。在鋰電池殘存電量部份,我們僅能透過可量測訊號(電壓、電流、溫度)估算而得,但可量測訊號與殘存電量之間存在非線性關係,不容易透過人工找出其對應關係,因此近年來機器學習已廣泛用於鋰離子電池模型建立。機器學習神經網路模型具有極佳的非線性擬合能力,且無須精確電化學模型,僅需數據即可自動從中學習有用知識,能準確反映出可量測訊號與殘存電量之關係。本研究又透過堆疊降噪自動編碼器(SDAE)訓練神經網路模型,使訓練出之模型有進一步提升,在固定常溫25°C之預測結果,平均絕對誤差(MAE)為1.17%,而由此完整訓練之模型,除了用以預測鋰電池之殘存電量之外,透過轉移學習也能幫助新模型之建立,減少新模型數據收集之時間及成本,且能提升新模型之精度。在鋰電池剩餘壽命部份,透過模擬軟體COMSOL之電化學模型,得出不同程度老化電池之放電曲線特徵與對應之電池老化參數,再由此些數據建立神經網路模型,僅需輸入放電曲線特徵即可估算其電池老化參數。透過多任務學習方式能再進一步提升模型估測精度,估測三個老化參數,平均絕對百分比誤差(MAPE)均不超過4%。 Lithium batteries are one of the most important energy storage methods today, so it is very important to accurately estimate the remaining capacity and remaining life of lithium ion batteries, but neither of these information can be obtained by direct observation or measurement. We can only estimate the measurable signal (voltage, current, temperature) of the remaining power of the lithium battery. However, there is a non-linear relationship between the measurable signal and the remaining power. It is not easy to find the corresponding relationship manually, so in recent years machine learning has been widely used to build lithium-ion battery models. Machine learning neural network models have excellent nonlinear fitting capabilities, and do not require accurate electrochemical models, only data is needed to automatically learn useful knowledge, which can accurately reflect the relationship between the measurement signal and the remaining power. In this study, the neural network model was trained by Stacked Denoising Autoencoders (SDAE) to further improve the trained model. The prediction result at a fixed normal temperature of 25°C has a Mean Absolute Error (MAE) of 1.17%. In addition to predicting the remaining capacity of the lithium battery, this fully trained model can also help the establishment of a new model through transfer learning, which can reduce the time and cost of data collection for the new model and improve the accuracy of the new model. In the remaining life of lithium batteries, through the electrochemical model of the simulation software COMSOL, the discharge curve characteristics and corresponding battery aging parameters of aging batteries of different degrees are obtained. From these data, a neural network model is established, and then input the characteristics of the discharge curve can estimate the battery aging parameters. The multi-task learning method can further improve the estimation accuracy of the model. The three aging parameters are estimated, and the average absolute percentage error (MAPE) does not exceed 4%. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/68128 |
DOI: | 10.6342/NTU202003794 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 應用力學研究所 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
U0001-1708202016263900.pdf Restricted Access | 3.95 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.