Laporkan Masalah

STATE OF HEALTH ESTIMATION OF LITHIUM ION BATTERY USING LIGHTWEIGHT PERCEPTRON EQUIVALENT CIRCUIT MODEL

Muhammad Dzaky Ashidqi, Dr.Eng. Ir. Adha Imam Cahyadi, S.T., M.Eng., IPM.; Ahmad Ataka Awwalur Rizqi., S.T., Ph.D.

2024 | Tesis | S2 Teknik Elektro

Penelitian mengenai State of Health (SoH) dan model degradasi baterai lithium telah secara masif diteliti dan dipelajari oleh banyak peneliti. Diantara beberapa metode yang dikembangkan diantaranya adalah metode berbasis data-driven. Namun demikian, penggunaan metode ini dalam memprediksi model degradasi memerlukan beban komputasi yang tinggi. Konsekuensinya, implementasinya dalam embedded system akan memerlukan biaya yang tinggi juga.

Untuk mengatasi permasalahan ini, sebuah metode diusulkan dengan berbasis model Equivalent Circuit Model (ECM) dan Artificial Neural Network (ANN) yang ringan. Baterai lithium dimodelkan dengan Equivalent Circuit Model (ECM) tipe the first order thevenin model. Dari ECM ini didapatkan beberapa parameter yaitu internal resistance, open circuit voltage (OCV) dan polarization voltage. Ketiga parameter ini kemudian diestimasi melalui model ANN dengan menggunakan dataset baterai lithium ferrous phospat (LFP) dengan 300 siklus data. Kemudian data internal resistance yang didapatkan dari estimasi menggunakan ANN digunakan untuk memprediksi degradasi baterai menggunakan metode linear least squares. Kemudian hasil prediksi degradasi kapasitas baterai tersebut digunakan untuk mengukur SoH baterai.

Hasil dari simulasi model menunjukkan bahwa model yang diusulkan dapat menghasilkan akurasi sebesar 93.1% dibandingkan dengan nilai aktual degradasi baterai. Akurasi model juga diuji untuk mengestimasi SoH di bawah kondisi kinerja baterai yang bervariasi. Pada variasi rating arus 0.5c, 1c dan 3c, model dapat mengestimasi SoH dengan eror yang minimum yaitu 3.67, 1.58, dan 1.12 berdasarkan perhitungan error menggunakan RMSE. Sedangkan pada kondisi depth of charging yang bervariasi dan juga initial SoC yang bervariasi model juga dapat menunjukkan performa yang baik dengan akurasi yang sama seperti pada kondisi charging 100%. Di sisi lain, beban komputasi dari metode ini juga sangat ringan dengan jumlah parameter yang minimum dimana hanya terdiri dari satu hidden layer. Dengan menggunakan big O analysis, total beban komputasi yang terukur adalah O(13) yang mana ini menunjukkan beban komputasi yang ringan. Metode yang diusulkan juga menunjukkan waktu komputasi yang cepat dengan hanya 499 ms waktu training per siklus baterai.

The research in the subject of battery state of health (SoH) and degradation models have extensively studied and investigated by researchers. A recent development in this field, pursued by numerous researchers, involves employing a data-driven approach, which has shown remarkable precision. However, the utilization of the data-driven technique in predicting degradation models requires high computational resources. Consequently, its implementation in an embedded system would require high cost and computational resource.

To overcome these problems, an alternative approach is suggested, involving degradation modeling that relies on an equivalent circuit model and a lightweight neural network. The battery is modeled on the equivalent circuit using the first order Thevenin model. From this equivalent circuit model, several parameters including internal resistance, open circuit voltage, and R-C voltage were obtained using a lightweight neural network model which uses 300 cycles of data of LFP battery acquired from the experiment. These parameters will be obtained by fitting the error between terminal voltage from dataset and model output voltage. By employing the linear least squares technique, the internal resistance obtained through neural network training, is fitted to degradation data obtained from experimental procedures. This facilitates the prediction of capacity loss for each cycle based on the internal resistance and the acquired degradation data. In the next step, SoH of the battery is measured based on capacity degradation in each cycle.

The results of the model simulation show that SoH estimation using proposed method can obtain 93,1?curacy compared to actual degradation. The accuracy of the model is also tested to estimate SoH under various condition. In various charging current rate with 0.5c, 1c and 3c charging rate, the model can estimate the SoH with minimum error: 3.67, 1.58 and 1.12 measured using RMSE. Meanwhile, in various depth of charging and various initial SoC charging process, the model also can provide good performance with similar accuracy as SoH estimation in 100?pth of charging. In other hand, the computational of this method is lightweight with small computational resources and minimum parameters from lightweight neural network model that consist only two neurons on one hidden layer. Using big O notation analysis, the total computational cost of this method is only O(13) which mean that the computational cost is extremely low. The model also can perform fast computational time with 499 ms training time per cycle.

Kata Kunci : SoH, LFP Battery, ECM, Neural Network

  1. S2-2024-489162-abstract.pdf  
  2. S2-2024-489162-bibliography.pdf  
  3. S2-2024-489162-tableofcontent.pdf  
  4. S2-2024-489162-title.pdf