【摘要】：鋰離子電池具有較高的能量密度與功率密度,被廣泛用于電動(dòng)汽車(chē)(Electric vehicle,EV)儲(chǔ)能。電池管理系統(tǒng)(Battery management system,BMS)是電動(dòng)車(chē)關(guān)鍵技術(shù)之一,BMS對(duì)電池組的能量?jī)?yōu)化管理以及對(duì)電池的有效保護(hù)均依賴于電池荷電狀態(tài)(State of charge, SOC)的準(zhǔn)確估計(jì)。SOC不能被直接測(cè)量,傳統(tǒng)SOC估計(jì)方法存在參數(shù)選取困難、準(zhǔn)確性低、通用性差等缺點(diǎn)。近年來(lái)以機(jī)器學(xué)習(xí)為代表的數(shù)據(jù)驅(qū)動(dòng)建模方法蓬勃發(fā)展,此類方法通常有著良好的非線性逼近能力,且具有較好的可移植性。電動(dòng)車(chē)運(yùn)行過(guò)程中所產(chǎn)生的大量狀態(tài)數(shù)據(jù)也給數(shù)據(jù)驅(qū)動(dòng)建模方法提供了有利條件。然而,數(shù)據(jù)驅(qū)動(dòng)建模方法受制于數(shù)據(jù)自身的可靠程度,而車(chē)載數(shù)據(jù)采集設(shè)備有限的精度以及電動(dòng)車(chē)的復(fù)雜的工作環(huán)境使得數(shù)據(jù)中的噪聲、離群點(diǎn)無(wú)可避免。且由于數(shù)據(jù)驅(qū)動(dòng)模型基于“黑箱”假設(shè),難以分析誤差來(lái)源,因而可靠性難以保證。為了減少SOC估計(jì)誤差對(duì)BMS系統(tǒng)所造成的負(fù)面影響,考慮到數(shù)據(jù)不可靠性的多個(gè)來(lái)源并且結(jié)合機(jī)器學(xué)習(xí)方法自身的特點(diǎn),本文從多個(gè)角度開(kāi)展研究,以降低不可靠數(shù)據(jù)對(duì)模型產(chǎn)生的不利影響,并且通過(guò)估計(jì)所得到SOC的后驗(yàn)概率,降低BMS誤動(dòng)作風(fēng)險(xiǎn)。首先,考慮到最小二乘支持向量機(jī)(Least square support vector machine, LSSVM)具有良好的泛化性能,第二章中將最小二乘支持向量機(jī)回歸方法應(yīng)用于SOC估計(jì)問(wèn)題,并通過(guò)粒子群算法提高模型精度、優(yōu)化運(yùn)算效率。利用磷酸鐵鋰(LiFePO4,LFP)單電池恒定電流放電實(shí)驗(yàn)比較了幾種常用的核函數(shù)在SOC估計(jì)中的實(shí)際效果,并通過(guò)動(dòng)力電池組交替充放電實(shí)驗(yàn),模擬電動(dòng)車(chē)動(dòng)力電池工作狀態(tài),驗(yàn)證了算法的有效性,并分析了該方法存在的不足。其次,為降低數(shù)據(jù)噪聲對(duì)回歸模型所產(chǎn)生的影響,首次提出了一種加權(quán)模糊支持向量機(jī)回歸方法,根據(jù)樣本受噪聲污染的程度動(dòng)態(tài)調(diào)整權(quán)重,以提高估計(jì)的可靠性。由于不同數(shù)據(jù)特征中噪聲相對(duì)獨(dú)立,為了綜合評(píng)估多維數(shù)據(jù)樣本的置信程度,提出了一種基于T-S模糊模型的權(quán)值函數(shù)。此外,為度量不同特征量與估計(jì)結(jié)果的相關(guān)性,首次提出了一種夾角非線性相關(guān)性度量方法�；谪S田COMS電動(dòng)汽車(chē)實(shí)驗(yàn)平臺(tái),通過(guò)實(shí)測(cè)數(shù)據(jù)驗(yàn)證了所提出方法的有效性。接著,針對(duì)荷電狀態(tài)回歸問(wèn)題的中異方差、非平穩(wěn)性,建立了一種靈活的混合高斯過(guò)程(Gaussian mixture regression,GMR)荷電狀態(tài)估計(jì)方法。該方法通過(guò)多高斯分量線性組合的方式從實(shí)測(cè)數(shù)據(jù)中抽取數(shù)據(jù)信息,利用高斯過(guò)程回歸(Gaussian process regression,GPR)對(duì)每個(gè)子高斯分量進(jìn)行估計(jì),最后將子系統(tǒng)的估計(jì)結(jié)果加權(quán)求和以獲得最終SOC估計(jì)結(jié)果。針對(duì)混合高斯模型(Gaussian mixture model,GMM)的優(yōu)化問(wèn)題提出了一種進(jìn)化期望最大化算法(Evolutionary expectation maximum,EEM),數(shù)據(jù)分布較為集中的高斯分量通過(guò)對(duì)附近數(shù)據(jù)分布稀疏的高斯分量吞噬實(shí)現(xiàn)進(jìn)化,不斷迭代最終獲得最優(yōu)模型數(shù)量以及超參數(shù)。此外,首次提出了一種Pearson非線性相關(guān)性特征選擇算法,以降低模型過(guò)擬合風(fēng)險(xiǎn)。最后,深入研究了離群點(diǎn)對(duì)支持向量機(jī)、神經(jīng)網(wǎng)絡(luò)以及高斯過(guò)程回歸模型所造成的影響,針對(duì)高維數(shù)據(jù)中的離群點(diǎn)問(wèn)題,首次提出了一種加權(quán)高斯過(guò)程回歸方法(Weighted Gaussian process regression,WGPR),分別從權(quán)重空間以及函數(shù)空間的角度進(jìn)行推導(dǎo)。此外,提出了一種改進(jìn)的密度離群檢測(cè)方法,避免了高維數(shù)據(jù)離群檢測(cè)中性能劣化的問(wèn)題。實(shí)驗(yàn)中將所提出的算法分別用于解決荷電狀態(tài)估計(jì)問(wèn)題以及光伏發(fā)電系統(tǒng)短期功率估計(jì)問(wèn)題。
[Abstract]:The lithium ion battery has higher energy density and power density and is widely used for electric vehicle (EV) energy storage. The battery management system (BMS) is one of the key technologies of the electric vehicle, and the energy-optimized management of the BMS and the effective protection of the battery depend on the accurate estimation of the state of charge (SOC). the SOC cannot be directly measured, and the traditional SOC estimation method has the disadvantages of difficulty in selecting parameters, low accuracy, poor universality and the like. In recent years, the data-driven modeling method, which is represented by machine learning, is developing vigorously, and the method has good non-linear approximation ability and has good portability. The large number of state data generated during the operation of the electric vehicle also provides favorable conditions for the data-driven modeling method. However, the data-driven modeling method is limited by the reliability of the data itself, and the limited accuracy of the on-board data acquisition device and the complex working environment of the electric vehicle make the noise in the data, and the outliers are inevitable. and because the data driving model is based on the 鈥渂lack box鈥，

本文編號(hào)：2374488