Colloquia

Samenvatting

In the last decades the interest in and the need for battery technology has risen. This new interest can be explained by the global shift from fossil fuels to renewable energy, or more specifically, by electric vehicles slowly phasing out fossil fuel powered vehicles. Although increasingly popular, electric vehicles still have limited driving range which can be improved by further optimization of battery design. To better understand internal battery processes, and hence enhance design, one has to develop the corresponding physics-based mathematical models, i.e. virtual phantoms, that are of predictive type, and can emulate battery behaviour in different environments. The typical example of such a model for the emulation of the electro-chemical processes of a Lithium-Ion battery is the pseudo-2-dimensional model (P2D model), the model considered in this thesis. The physics-based equations coming from a homogenisation principle are solved by a finite element approach and novel type of iterative algorithm. However, such an approach is shown to be computationally expensive. Therefore, in this thesis the novel approach to solving the physics-based equations is suggested. By use of the first principles and deep learning approximation of the battery state, the physics-based equations are solved in a purely machine learning setting without use of any other type of numerical approximation such as finite element methods.  The comparison results show that NN is able to mimic battery behaviour with reasonable accuracy and for a lower computational load compared to a classical approach. Finally, different NN architectures and training variations have been tested and evaluated.