Prediction of lithium polymer battery parameters by deep learning methods

As carbon emissions increase, the greenhouse gas effect occurs, and problems such as climate change and global warming arise. Vehicles that are equipped with internal combustion engines are among the factors that cause an increase in the carbon rate in the air. In order to prevent global warming and climate change, the use of electric vehicles has become important in areas such as transport, production, and travel. However, it is an important issue to use electric vehicles, hybrid electric vehicles, and robots in a healthier way with successful, reliable, and fast battery state of charge estimation, which has an important role in the battery management system. As the use of electric vehicles and robots has increased over time, the accurate determination of lithium-based battery parameters has become very important. In this study, a lithium polymer battery dataset, which has not been sufficiently studied in the literature, has been prepared in order to predict the battery state of charge used in electric vehicles and robots. A new experimental system was created to obtain the dataset by measuring the current, voltage, and temperature parameters of lithium polymer batteries. Battery state of charge estimation is performed successfully with deep learning methods, and new methods are proposed. By evaluating the structure of convolutional neural networks, intuition was obtained by examining the effects of hyper parameters such as batch size and fully connected layer neurons. Unlike deep learning models that require high computational load in electronic cards, the proposed method is based on four different fully connected layer neuron values and two different batch size values and obtained remarkable results. While determining the appropriate parameters for successful predictions, the proposed model was obtained by performing experiments on the optimizer, learning rate, number of fully connected layer neurons, and batch size values. In addition, the performance of cutting-edge deep learning methods and bidirectional long short-term memory methods was compared by determining hyper parameters using the trial-and-error method and the optimization method. The performance of these models is compared with deep learning training on a computer using the Python programming language on the Ubuntu operating system. It is thought that the proposed methods will be successfully applied in real life.