聯合卡爾曼濾波器與深度學習於鋰離子電池多狀態估計

Abstract

電池管理系統會對鋰離子電池進行狀態監測，以防止電池發生過充電、過放電或是熱失控，確保電動車運行的效率和安全性。在本研究中，我們提出了一種計算架構以估計電池的充電狀態（State of Charge, SoC）、健康狀態（State of Health, SoH）和表面溫度。該框架結合三種不同演算法，包含向量型遞迴最小平方法(Vector-type Recursive Least Squares, VRLS)、自適應擴展卡爾曼濾波器（Adaptive Extended Kalman Filter, AEKF）以及深度神經網路（Deep Neural Network, DNN）。此架構應用在駕駛循環測試中依序分為三個計算步驟。首先，VRLS利用動態變化的電池電壓和電流識別等效電路模型(equivalent circuit model, ECM)的參數。接著，AEKF負責運用ECM的狀態方程式來估計電池的SoC和SoH。最後，DNN則以電壓、電流、SoC以及ECM的參數作為輸入，估計電池的表面溫度。該架構的特點是僅需要電池電壓和電流的量測資訊，而不需要額外感測器，能更泛用於實際情況。實驗結果表明，在此框架下，SoC和SoH的平均估計誤差小於0.03，測試數據中電池表面溫度的平均估計誤差可以小於0.5°C。

Battery management systems monitor the state of lithium-ion batteries to prevent overcharge, over-discharge, and thermal runaway, ensuring the efficiency and safety of electric vehicles. In this study, we propose a calculation framework for estimating the battery state of charge (SoC), state of health (SoH), and surface temperature. The framework combines three different algorithms: vector-type recursive least squares (VRLS), adaptive extended Kalman filter (AEKF), and deep neural network (DNN). The proposed framework is applied in drive cycle tests and consists of three calculation steps. First, VRLS identifies the parameters of the equivalent circuit model (ECM) using the dynamic variations of battery voltage and current. Then, AEKF utilizes the state equations of the ECM to estimate the battery SoC and SoH. Finally, DNN takes voltage, current, SoC, and ECM parameters as inputs to estimate the battery surface temperature. The key feature of this framework is that it only requires measurements of battery voltage and current, eliminating the need for additional sensors and enhancing its applicability in practical scenarios. Experimental results demonstrate that the average estimation errors of SoC and SoH are below 0.03, and the average estimation error of the battery surface temperature in the testing data can be less than 0.5°C.