TL;DR
This paper tackles the challenge of accurately estimating a battery's State-of-Energy (SOE) and State-of-Health (SOH) in electric vehicles (EVs) by proposing a novel vehicle-cloud collaboration approach. By leveraging deep learning models on the cloud and an equivalent circuit model (ECM) on the vehicle, they achieve highly accurate joint estimation, crucial for improving EV range prediction and battery management.
What Problem Does This Paper Address?
Electric vehicles (EVs) face two major challenges related to their batteries: range anxiety due to uncertain driving range and the degradation of battery performance over time. Accurately estimating the State-of-Energy (SOE) is vital for predicting the remaining driving range, while estimating the State-of-Health (SOH) is crucial for understanding battery degradation and ensuring safety and reliability. However, accurately estimating both simultaneously is difficult due to the complex interplay between battery aging, varying environmental conditions, and the limited computational power of in-vehicle battery management systems (BMS). Existing methods often struggle with accuracy, require extensive parameter tuning, or cannot be implemented on-board. This paper aims to bridge this gap by proposing a vehicle-cloud collaboration framework that combines the strengths of both environments for robust joint SOE and SOH estimation.
Methods
The proposed approach utilizes a hybrid deep learning architecture within a vehicle-cloud collaboration framework.
- Vehicle-Side: An Equivalent Circuit Model (ECM) is employed for real-time battery state estimation. This model is parameterized using experimental data and optimized with a Particle Swarm Optimization (PSO) algorithm.
- Cloud-Side:
- SOH Estimation: A Bi-directional Long Short-Term Memory (Bi-LSTM) network is used for SOH prediction. This model leverages historical battery data and is optimized using Bayesian optimization to fine-tune its hyperparameters.
- SOE Estimation: A Convolutional Neural Network (CNN) combined with LSTM (CNN-LSTM) is employed for direct and nonlinear SOE mapping. This approach avoids explicit parameter identification and updating, making it suitable for complex operating conditions.
- Joint Estimation: The core innovation lies in the joint estimation strategy. The SOH prediction from the cloud-side Bi-LSTM model is used as feedback to correct and improve the SOE estimation performed by the CNN-LSTM model. This allows for a more accurate SOE estimation that accounts for battery degradation.
Datasets
The study utilizes the following datasets:
- NASA Ames Prognostics Center of Excellence (PCOE) Battery Data Set: Specifically, the Battery Aging ARC-FY08Q4 data set, including battery life test cycles B0005, B0006, B0007, and B0018. This dataset is used for training and validation of the proposed models.
- Custom Experimental Data: Data collected from a pulse discharge test of an 18650 lithium battery on a self-established experimental platform. This data was used for parameter identification of the ECM and OCV-SOE curve fitting.
- Dynamic Stress Test (DST) Data: A DST test was conducted on the custom battery test platform to simulate real EV operating conditions and further validate the proposed method's reliability.
Key Insights and Ideas
- Vehicle-Cloud Synergy: The paper convincingly demonstrates that combining on-board ECMs with cloud-based deep learning offers a powerful solution for complex battery state estimation, overcoming the limitations of each individual approach.
- SOH as a Crucial Feedback for SOE: The study highlights that accurately estimating SOH is not just important for battery health monitoring but is critical for improving the accuracy of SOE estimation, especially as batteries age. Without SOH correction, SOE estimates can become significantly inaccurate over time.
- Direct Mapping for SOE: Using CNN-LSTM for direct SOE mapping bypasses the need for complex parameter identification and updating, making the model more robust to varying operating conditions.
- Bayesian Optimization for Hyperparameter Tuning: Employing Bayesian optimization for hyperparameter tuning of the Bi-LSTM model is shown to be an efficient way to find optimal configurations, leading to better SOH prediction accuracy.
Key Results
- The proposed vehicle-cloud collaboration approach achieves a joint estimation error of SOE and SOH within 3%.
- The CNN-LSTM method for SOE prediction showed superior performance compared to a standard LSTM approach, with RMSE reaching 1.61% versus 2.46% for B0005.
- In the DST condition, the CNN-LSTM method achieved an RMSE of 1.10%, significantly outperforming the LSTM method's RMSE of 2.04%.
- The study found that SOE estimation accuracy drops sharply when SOH is unknown, especially after many charge/discharge cycles. With SOH prediction and correction, the accuracy is maintained well.
Limitations and Future Work
- Communication Problem: The paper explicitly states that the communication problem between the vehicle and the cloud is not considered in this study. Future work will address this by designing control strategies for communication signals.
- Parameter Optimization: The authors suggest that the CNN-LSTM parameters proposed in this paper can be further optimized.
- Algorithm of Device-Cloud Fusion: The current algorithm's dependence on communication signals is a limitation. Future work will focus on designing the communication strategy for the device-cloud fusion.
Summary updated