Dynamic neural networks with data assimilation

✍ Scribed by Henk van den Boogaard; Arthur Mynett

• Publisher: John Wiley and Sons
• Year: 2004
• Tongue: English
• Weight: 142 KB
• Volume: 18
• Category: Article
• ISSN: 0885-6087
• DOI: 10.1002/hyp.5579

✦ Synopsis

Neural networks (NNs) are often used as black-box techniques for the modelling of system relations. Standard NNs are static models, whereas in practice one often has to deal with dynamic systems or processes. In such cases, dynamic neural networks (DNNs) may be better suited. We will argue that the temporal processing in a DNN is highly related to that in dynamic physical-based models, and that this similarity provides the opportunity to include other important issues of conceptual dynamic models into an NN environment as well. In particular, we will propose to extend and upgrade NN models with data assimilation facilities. In hydrology one usually has to deal with complex dynamic systems, and lumped, aggregated, empirical, or fully black-box and data-driven techniques are often used for modelling. For this reason, the issues proposed here for upgrading of standard NNs may be particularly relevant for hydrology.