Page 104 - EER-2-1

Explora: Environment
and Resource Artificial neural networks

Table 7. Comparison of the proposed model with related work
Study Model type Prediction accuracy (%)
This work ANN-based electrothermal model Voltage: <2 error
Temperature: <3 error
Tu et al. 58 A physics-based model with machine learning Voltage: 3 error
Temperature: 4 error
Pour 59 Empirical thermal model Temperature: 2.5 error
Zhang et al. 60 Data-driven electrical model Voltage: 1.5 error
Abbreviation: ANN: Artificial neural network.

between actual and simulated data, remained below 3%, 3.5. Coupled electrothermal aging model
indicating satisfactory accuracy in estimating cell lifetime The primary objective of this study is to present a
under static conditions (Figure 8B). While static profile complete cell-level electrothermal model coupled with
validation provides valuable insights, a more comprehensive an ANN-lifetime model. This approach offers two
validation approach involves using dynamic profiles that significant advantages: it uses a simple neural network
differ significantly from the training data and utilizing a to estimate the cell’s lifetime, providing a fast-response
dynamic discharge current pattern, further demonstrating solution to the computational complexity challenge
the model’s robustness and predictive capabilities. typically associated with mathematical models, and it

3.4.2. Dynamic profile validation can predict cell voltage and temperature based on the
Ideally, validation requires using a completely independent cell’s cycle life. To validate the complete model’s voltage
and temperature estimations, experimental data from
set of long-term aging data, which differs from the
training data in key variables (temperature, current, etc.) the discharging phase (C/2) of RPTs at different cycle
points (500, 1000, and 1500 FECs) were used for two
or behavior. Therefore, to complete the lifetime model conditions: (i) condition II involved a 1C discharge,
validation, a separate testing condition was employed for 0.5C charge, 100% DoD, and 15 min of rest time,
model validation, involving a dynamic discharge current and (ii) condition IV included a 4C discharge, 0.5C
pattern rather than the constant current used in the other charge, 100% DoD, and 60 min of rest time. The results
conditions (Table 2). Specifically, the cycling profile, (Figures 10 and 11) demonstrate highly accurate voltage
in this case, includes: charge at C/2 (or 1.5A); a 1 h rest predictions with a maximum RMSE of <1% (Table 6).
period; dynamic discharge using the WLTC until the cell is Temperature estimations exhibit minor deviations,
depleted; and repeats for 100 cycles.
particularly in the middle of the discharge phase, but
Three cells were tested under these validation the overall error remains below 3%. The combination of
conditions, as previously described in the experimental a semi-empirical approach with ANN offers a powerful
section, with only the best-performing sample being method for predicting battery performance and lifetime.
utilized for validation. The remaining test conditions This hybrid approach leverages the strengths of both
were used to build and train the FNN model, Figure 9A model-based and data-driven techniques, potentially
illustrates a comparison between the measured data and providing more accurate and efficient predictions for
the simulated lifespan, plotting capacity retention (%) battery management systems.
against the number of cycles for 700 FECs. The FNN
model’s estimates closely align with the experimental 4. Conclusion
results. Despite some minor deviations at the 200 and In this study, a complete cell-level electrothermal model
300 FEC marks, the FNN accurately captures the capacity coupled with an ANN-lifetime model was proposed
retention trend across the full 700 cycles. In addition, for Sony 3 Ah 18650-type lithium-ion cells. First, a
the error metric—calculated as the absolute difference comprehensive study of the cells was performed, providing
between real and simulated data—is presented in valuable insights into their electrothermal characteristics
Figure 9B. The deviation remains below 2%, indicating and aging behavior but also the necessary model
that the model can simulate the accelerated profile with parameters for the development of the model.
satisfactory accuracy. This highlights the FNN model’s
robustness in predicting the cell’s lifespan, making it a The electrothermal characterization revealed
reliable tool for optimization algorithm development. important trends in IR, OCV, and capacity across different

Volume 2 Issue 1 (2025) 11 doi: 10.36922/eer.7228