Battery ageing estimation with Artificial Intelligence

With the gradual introduction of electric vehicles (e-vehicles) in the market, climate change could be reduced. The goal is to make these vehicles more efficient than traditional fuel ones in such a way that in the future they could completely replace the old technology. One of the main components of an electric vehicle is the rechargeable battery, and lithium-ion ones have been proven as the most efficient available, nowadays. A lot of research has been done to boost performances of this type of battery especially in terms of energy consumption or in other words how many kilometers the vehicle is able to travel with respect to one single full charge. However, one problem that lithium-ion batteries face is that, for this technology, it does not exist an equivalent of a fuel level measurement unit; therefore, it is only possible to estimate the remaining charge. If this estimation is poor or wrong, it leads to several problems, like over-charges and over-discharges, which severely impact on the average battery lifetime. The goal of this thesis is to develop neural networks for estimating the state of charge (SoC) and the state of health (SoH) of lithium ion (Li-ion) cells used in automotive applications. To this purpose, different neural networks are trained and tested to obtain a good SoC and SoH estimator; both synthetic and real data of Yuasa's LEV50 battery cells have been used for training and testing the neural models. The architecture of these neural networks is composed of a Nonlinear AutoRegressive with eXogenous inputs (NARX) that predicts the SoC and in cascade a convolutional neural network (CNN) that estimates the SoH. Two versions of this network are presented: the first one is composed of a NARX followed by a unidimensional CNN (NARX + CNN1D) while the second one uses, in its convolutional part, a bi-dimensional CNN in combination with a bidirectional long short term memory (NARX + CNN2D biLSTM). The inputs are the cell terminal voltage, current and surface temperature. A 10-fold cross validation was applied to NARX+CNN1D and NARX+CNN2D biLSTM using exclusively synthetic data; both networks achieve an average accuracy of 95%. Using the same networks, training was repeated on a hybrid data set composed of synthetic and measured data. The test accuracy on the first model reaches 97% while in the second one 98%. This work concludes with the training and test of the NARX+CNN1D and NARX+CNN2D biLSTM using only measured data from laboratory, achieving respectively an accuracy of 97% and 98%.