Boosting Electric Vehicle Performance: Hybrid Deep Neural Network Unleashes State of Charge Estimation

By Joel ScanlonReviewed by Susha Cheriyedath, M.Sc.Jul 13 2023

An article accepted for publication in the journal Energy proposed a novel approach for state of charge (SoC) estimation in electric vehicles (EVs) using a hybrid multi-layer deep neural network (HMDNN). This network utilized the mountain gazelle optimizer (MGO) as its training algorithm. The findings indicate that the neural network-based approach proposed exhibits superior accuracy and faster convergence compared to currently available state-of-art methods. As such, this approach can potentially enhance the efficiency of EV operation and extend the battery's lifespan.

​​​​​​​

Background

SoC pertains to the quantity of electrical energy that is contained within a battery at a specific point in time. The precise determination of SoC holds significant importance in various domains, such as battery management, forecasting electric vehicle range, and scheduling renewable energy systems.

In the realm of EVs, precise SoC estimation is critical in terms of forecasting the vehicle's range and guaranteeing that the driver possesses adequate energy to reach their intended destination, thereby enhancing the overall driving experience of EVs. The increasing adoption of EVs can be attributed to their potential to mitigate greenhouse gas emissions and reduce reliance on fossil fuels.

Historically, the estimation of SoC has been conducted through the utilization of mathematical models that rely on the electrochemical properties inherent to the battery. Nevertheless, it is important to note that these models can exhibit a high level of complexity and may not accurately represent the battery's behavior in real-world scenarios. This is particularly true when considering situations where the battery is old or has been subjected to other external factors that can potentially impact its overall performance.

Proposed technique

In the present study, a mathematical model of the MGO algorithm was formulated based on the fundamental principles that govern the community and grouping dynamics observed in the mountain gazelle. In the optimization procedure of the MGO algorithm, every gazelle (Xi) can associate with either a group of maternity herds, lonely males, or territorial males.

One of the three herds has the potential to produce offspring in the form of a young gazelle. The adult male gazelles within the herd's territory represent the most advantageous choice for MGO globally. Roughly 33% of the search population, comprising juvenile male bachelor gazelles who lack the necessary maturity or strength to engage in reproduction or assert dominance over females, will exhibit the lowest fitness level. This is because these gazelles are still in an immature stage and lack the ability to reproduce or establish dominance over female gazelles.

In addition to the MGO algorithm, this work aims to examine using deep neural networks (DNNs) to estimate  SoC for batteries. Instead of depending on the expertise in battery chemistry, the authors adopt a data-driven methodology where DNN is trained using a dataset of battery parameters that have been measured in controlled laboratory settings. The dataset comprises voltage, current, and temperature measurements, which are utilized for the estimation of SoC. Although the use of this technique necessitates meticulous sensor calibration and precision, it presents a viable alternative to model-centric methodologies that heavily depend on mathematical representations of battery performance.

Moreover, this study primarily examines two crucial hyperparameters that substantially influence the functioning of DNNs: the quantity of hidden layers and the weights and biases associated with these layers.

Major results

In each of the four drive cycle datasets, the proposed MGO-DNN consistently demonstrates a significant reduction in the cost function throughout the iterations, resulting in lower overall costs compared to standard state-of-art techniques like the mayfly-based DNN (MFA-DNN), the particle swarm optimization-based DNN (PSO-DNN), and the classical back propagation-based DNN (BP-DNN). The rapid decrease in the cost function achieved by MGO-DNN demonstrates the high efficacy of MGO in updating the weights and biases.

The BP-DNN algorithm experiences a persistent issue of becoming trapped in local minima during the training process. However, training the DNN and MFA-DNN models effectively minimizes the cost function. Hence, the efficacy of employing the MGO algorithm as a deep neural network training method for system-on-a-chip estimation has been verified.

The evaluation metrics employed in the proposed methodology to assess the performance of the testing process include normalized mean square error (NMSE), root mean square error (RMSE), mean absolute error (MAE), relative error (RE), and time. It exploits the inherent correlation between the system-on-chip and the voltage/current measurements of the EV battery, enabling precise real-time estimation of the SoC. This algorithm achieves an average NMSE of 0.1% and an average RMSE of 0.3% across all datasets.

Quantitative results utilizing the INR 1865020R LiNiMnCoO2/NMC Li-ion battery cell show the efficiency of the suggested methodology in calculating the SoC of a lithium-ion battery for an electric car. The Granger causality test also reveals the smallest discrepancy between the expected and real SoC. A more detailed examination of the model's performance is possible using several performance metrics.

Overall, the proposed model may be used to improve the efficiency and dependability of real-time monitoring and control systems for electric cars. More investigations might be needed to enhance the performance of the suggested model and investigate its potential for usage in other applications.