Deep Learning technique for Model Dynamic System Identification and Forecasting

The mathematical modelling of a system is an activity as complex as it is useful for many fields of work and research. Technological development in areas such as machine learning opens the door to new approaches in data analysis and data knowledge extrapolation. What we propose in this work is a new approach in system identification that takes the advantage of the newest Deep Learning research for the virtualization of models from measurements. NeuralODEs are a new formulation of the classical Neural Network. NeuralODE set themselves the objective not of imitating the pattern of a system but really learning its dynamics from data. The aim of the work is to validate this framework for the generation of a surrogate model able to learn the set of ODEs that represent the model under study. We propose different test case with different dynamic to test the abilities and the limitations of this tool. Furthermore we test the NeuralODEs for different unseen initial condition condi??tions and input values. The framework studied shows a better response in the learning of systems with a content degree of difficulty, showing forward a good scalability. Research around increasing reliability for systems with greater complexity is an interesting challenge.