基於Elman神經網路演算法的電動車鋰電池電量狀態估測與電路實現;Estimation and Circuit Implementation of Lithium Battery State of Charge in Electric Vehicles Based on the Elman Neural Network Algorithm

NCU Institutional Repository > 資訊電機學院 > 電機工程研究所 > 博碩士論文 > Item 987654321/93621

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/93621

題名:	基於Elman神經網路演算法的電動車鋰電池電量狀態估測與電路實現;Estimation and Circuit Implementation of Lithium Battery State of Charge in Electric Vehicles Based on the Elman Neural Network Algorithm
作者:	劉秉鈞;Liu, Bing-Jun
貢獻者:	電機工程學系
關鍵詞:	電量狀態估測
日期:	2024-01-24
上傳時間:	2024-09-19 17:22:30 (UTC+8)
出版者:	國立中央大學
摘要:	電池充電狀態 (State-Of-Charge; SOC)的估算在電池管理系統 (Battery Management System; BMS)中佔有極重要的角色，其估算精度對於系統安全有密切關係。然而隨著深度學習及神經網路的發展迅速，透過神經網路搜索估算電池狀態特徵的方式來取代傳統估算方式，變成接下來電池狀態估算的發展態勢，本論文依據神經網路模型之特色及適用架構，基於 Elman神經網路 [1-4]作為神經網路估算系統之模型主軸，規劃基於神經網路演算架構估算電池 SOC，系統規劃中將估測系統分割成參數訓練模(Training Part)與即時估測模組 (Real-Time Estimate Part)兩個部分，透過參數訓練模組依據電池過往充放電資料找尋電池特徵參數，利用更新電池特徵參數的方式更新即時估測模組中的電路，實現深度學習晶片規劃。於即時估測模組中，在實際電池運作時，因為電池負載變動所造成的電池電壓電流波動過大，進而導致電池 SOC估算精度下降，為避免估算誤差本論文於晶片中規劃三階權重適配器來微調權重以提升即時估測精度，克服估測誤差。本文以電動車的動態壓力測試 (DST)、聯邦城市駕駛週期表 (FUDS)與電動車真實行車資料 D3 2014 BMW i3 BEV(SOC 90% - 10%)、 D3 2014 BMW i3 BEV(SOC 56.8% - 9.9%)進行系統模擬參數訓練與晶片設計規劃。其晶片設計部分以 TSMC-40nm製程為基底規劃，並利用 SMIMS Veri Enterprise Xilinx FPGA驗證其功能，並達成深度學習演算法估算電池 SOC之目的。;The estimation of State-Of-Charge (SOC) in Battery Management Systems (BMS) plays a crucial role, and its accuracy is closely related to system safety. However, with the rapid development of deep learning and neural networks, there is a shift towards utilizing neural networks to search for battery state features, replacing traditional estimation methods. This trend becomes the forefront of future developments in battery state estimation. This paper, based on the characteristics and applicable architecture of neural network models, adopts the Elman neural network [1-4] as the central model for the neural network estimation system. The system is designed to estimate battery SOC based on the neural network algorithm framework. In the system design, the estimation system is divided into two parts: Parameter Training Module (Training Part) and Real-Time Estimate Module (Real-Time Estimate Part). The Parameter Training Module utilizes past charge-discharge data to identify battery feature parameters. The Real-Time Estimate Module updates the circuit in real-time based on the updated feature parameters, achieving deep learning chip planning. In the Real-Time Estimate Module, during actual battery operation, the significant fluctuation in battery voltage and current caused by dynamic changes in battery load leads to a decrease in SOC estimation accuracy. To mitigate estimation errors, this paper incorporates a three-stage weight adapter in the chip to fine-tune weights and enhance real-time estimation accuracy, overcoming estimation inaccuracies. The system simulation for parameter training and chip design planning is conducted using dynamic stress testing (DST), Federal Urban Driving Schedule (FUDS), and real driving data from a 2014 BMW i3 BEV (SOC 90% - 10%) and a 2014 BMW i3 BEV (SOC 56.8% - 9.9%). The chip design is based on the TSMC-40nm process, and its functionality is verified using the SMIMS Veri Enterprise Xilinx FPGA, successfully achieving the goal of deep learning algorithm-based battery SOC estimation.
顯示於類別:	[電機工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	14	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....