Rockefeller, RockefellerRockefellerRockefellerBah, BubacarrBubacarrBahZimmermann, Hans GeorgHans GeorgZimmermannMarivate, VukosiVukosiMarivate2024-05-272024-05-272023https://publica.fraunhofer.de/handle/publica/46879810.1109/ICECET58911.2023.103893202-s2.0-85187289097Wind power prediction is an attempt to model a complex dynamical system that is high dimensional, nonlinear, and only partially observable. The outcomes of such forecasts are important for the operation of the blades of the wind turbines, for the short-term stability of the electrical grid, and for the long-term planning of electrical power generation. Historical consistent neural networks (HCNNs) are a type of neural networks which were originally designed to model complex systems as described above. They were applied successfully in the financial environment, which shares the complexity feature as highlighted above. Because of the mathematical similarity, we would like to transfer the insights from these studies to model the physics of wind dynamics. HCNNs fulfill universal approximation property, give a causal description between past and future and allow a reconstruction of unobserved variables. The modeling of the wind power depends on a priori set-up of the network architecture together with learning from data. Therefore, a preprocessing stage is needed to transform the wind measurements into a formulation that suits to the learning framework. Our HCNNs approach formulates a causal understanding of the wind system, opposite to a pure phenomenological data analysis which does not allow the reconstruction of unobserved variables. In our paper, we show the success of this approach on South African wind data.encomplex dynamical systemsdeep learninghistorical consistent neural networksrenewable energystate space modelsconference paper