A proposed methodology using artificial intelligence to estimate the state of charge for batteries of electric vehicle

• Autor(a) principal: SANTOS, Monalisa Cristina Moura dos
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFPE
• Texto Completo: https://repositorio.ufpe.br/handle/123456789/39127
• topic: Engenharia de Produção; Veículos elétricos; Estado da carga; Machine learning; Prognóstico; Health management

In order to decrease the emission of greenhouse gases and propose alternatives to the environmental effect of it, the development and improvement of “green technologies” have received special attention due to their utility to prevent the impacts caused by those gases. Thus, electric vehicles (EVs) were, also, an important advancement in this area. To work, the EVs need a reliable battery source and, for most EVs, a lithium-ion battery is used as a power source. Some advantages of lithium-ion batteries are high specific energy density, high cycle life, and low self-discharge. In the context of Prognostic and Health Management (PHM), estimation of the SOC (State of Charge) – which is the remaining charge within the battery and is defined as the ratio of the residual capacity of the battery to its nominal capacity – based on data-driven methods (e.g. Machine Learning – ML, Deep Neural Networks – DNN) and data storage (e.g. Big Data) has come as a suitable alternative to identify patterns in its degradation over time, also being much less time-consuming than physics of failure (e.g. coulomb counting and open circuit approaches) methods, which needs full discharging to estimate SOC. In this work, a methodology using DNN and Machine Learning (ML) algorithms is proposed to predict battery SOC. At first, the input – current and voltage – and the output – SOC – each given in the form of time series, are replicated using Maximum Entropy Bootstrap (MEB), a sampling technique used with non-stationary time series- this technique is used to further compute confidence interval of the remaining time until the next recharge. Afterward, the input dataset is processed using a windowing model as the pre-processing step; this processed dataset is used to train a DNN model. For purposes of comparison, the data is also fed into an ML model, with each replication training the model. Following the training phase, the predicted SOC, for both the DNN and ML model, is filtered by an Unscented Kalman Filter (UKF), which processes the predicted SOC time series in terms of its mean and covariance. Then, the remaining time until the next recharge is computed and compared with the real discharge time. Finally, the confidence interval of the remaining time until the next discharge is calculated for the DNN and ML models. Analyzing the results, the DNN model, which is performed by the Multi-Layer Perceptron, has better results compared with the other applied methods – Support Vector Machines, Random Forest and XGBoost – with lower root mean squared error results and percentage errors for the remaining time until the next discharge – for both non and postprocessed results. These results are achieved due to the complexity of the DNN model. However, further analysis in terms of the number of layers for the DNN method needs to be operated. For the Random Forest and XGBoost methods, which obtain the worst results, they are, generally applied for classification tasks, explaining the observed results.