以深度學習方法估測鋰電池健康度

Abstract

在現今注重綠色能源的氣氛下，電動車也隨之蓬勃發展，而電池作為電動車主要提供能源的核心裝置，因此關於電池各類特性響應需要著實監控。 電池經過多次使用後，各參數如：電池容量、內阻會有顯著的差異並影響能量見控制上有影響，因此建立具備電池老化估測之模型，有助於改善控制策略以達到節省能量，並知曉是否將之汰換。本論文循環充放電實驗以接近實車的測試模式為主，其考量不同充放電情況和使用次數對電池老化的影響，並運用TensorFlow™開源軟體庫建立人工智慧模型，預測出電池健康度。本論文將實驗數據分為訓練資料與測試資料，經過深層神經網路(Deep Neural Network)訓練後，並估測出不同實驗方式之結果，且最佳估測結果為方均誤差小於0.1。本論文另一電池模型為遞歸神經網路(Recurrent Neural Network)，能記憶之前時刻的資料，電池老化與時間序列有些許關聯，也能精準地預測SOH。

In development of Battery models are vital for the development of electric vehicles. The battery life change the battery response. As a result, the model, which has battery degradation factors helps improve the efficiency of other control algorithms and maintain the safe usage of the battery. In this thesis, use charge-discharge cycle that is similar to a vehicle testing pattern and TensorFlow™ to build artificial neural network of battery cell which can predict the battery life. The battery model considers different parameters like patterns of charge current and the number of cycle which have an influence on the result of prediction. We have training data that is for a Deep Neural Network and test data that is identified for the model accuracy. The best result is that it can predict state of health within 0.1 of mean square error. In this thesis, the other model that is Recurrent Neural Network has the ability of remembering previous data. It has also good performance on estimating SOH.