機器學習於鋰離子電池狀態估測與電壓行為預測之應用

Abstract

電池的荷電狀態(SoC)以及電壓行為(Voltage behavior)是電池管理系統中的兩個關鍵指標，其對於下游電池管理功能(例如功率狀態估測和電池平衡)尤其重要。在本論文的第一部份中，我們以機器學習的方法建立了一個能在動態負載和溫度變化的情形之下估算電池之荷電狀態的估算器。該估算器係由一系列長短期記憶類神經網路單元(LSTM Cell)堆疊而成，並以電池之電流、電壓、溫度作為輸入，以及電池當下之荷電狀態為輸出，並同時以兩組不同固定溫度下的動態負載數據訓練之，待訓練完畢後再以變溫之動態負載數據對其進行測試驗證， 以評估其性能。最後，我們成功開發出一能夠在極為複雜的操作條件下依然擁有97%的高精度荷電狀態估計器。 在本論文的第二部分裡，我們同樣以機器學習的方法開發了一能在動態負載和溫度變化情形之下預測下一時間點電池之電壓的預測器。此預測器同樣由長短期記憶類神經網路單元堆疊而來，並從當下以及過去二十九個時間步的過去資料中(共三十個時間步，包含電流、電壓、溫度等資訊)預測下一時刻的電池電壓。和上述之荷電狀態估測器類似，我們以定溫下的動態負載數據對預測器進行訓練、並以變溫動態資料測試之。結果，在變溫和動態負載條件下，預測器以小於0.1%的平均絕對百分比誤差(Mean absolute percentage error, MAPE)準確地估計出電池下一時間點的電壓。

The state of charge (SoC) and the voltage behavior of a battery is one of the key quantities in battery management systems and is especially important for downstream battery management functions such as state of power prediction and battery balancing. In the first part of this work, we built a model based on machine learning to estimate the SoC of batteries under dynamic loads and changing temperatures. The estimator is a stacked LSTM network with voltage, temperature, current as inputs and the SoC as its output. The model was first trained on the dynamic load data acquired at two different constant temperatures and was then tested on the dynamic load data obtained at varying temperatures. As a result, the final product is a SoC estimator that achieves a high accuracy of 97%. In the second part of this work, we developed a program that predicts the voltage of the battery at the next time step, which takes the current, voltage and temperature of the battery as input and the predicted battery voltage as output. This predictor was constructed by a many-to-one recurrent neural network with long short-term memory (LSTM) and is trained and tested similarly to the SoC estimator, only with the training data augmented from two sets into an even greater of five. As a result, under varying temperature and dynamic load conditions, the predictor accurately estimates the voltage for the next time step with a low mean absolute percentage error (MAPE) of less than 0.1%.