Modelling, in real time, hydrodynamic behavior of multi-phase fluid flow in transitory phase in pipe, comprises series of neuron networks

摘要

A number of neuron networks (Estra, Edisp, Eint) are constructed, each dedicated to different fluid flow regimes. A probability neuron network (RNProba) is constructed to evaluate at all times the probability that the flow in the pipe corresponds to each of the flow regimes, and the results from the different neuron networks are combined weighted by the different probabilities. The process takes into account the operating conditions fixed over a certain number of structural parameters defined relative to the pipe, and an assembly of defined physical dimensions, with fixed ranges of variation for the said parameters and dimensions, using neuron networks with inputs for the parameters and dimensions and outputs producing the results required to estimate the hydrodynamic behavior, and at least one intermediate layer. The neuron networks are determined iteratively to adjust themselves from starter base values with predefined tables connecting different values obtained for the output data to corresponding values of input data. At least three neuron networks are constructed, dedicated respectively to stratified, dispersed and intermediate flow regimes. The probabilities are calculated for the fluid flow in the pipe to correspond to each of these regimes and the results are combined linearly. When the database is sufficiently detailed to distinguish sub-regimes inside the same flow regime, a neuron network for probability (RN Proba) is constructed to evaluate the probabilities of each flow regime at any moment.