Neuronale Netzwerke zur Permeabilitätsbestimmung in Rotliegendsandsteinen

摘要

In der vorliegenden Arbeit wurden mit Hilfe von Logs und unter Verwendung eines neuronalen Netzwerkes synthetische Permeabilitätswerte für Rotliegendsandsteine bestimmt. Der Vergleich zwischen synthetischen und am Kern bestimmter Permeabilitäten ergibt eine gute Übereinstimmung. Zur Verfügung standen Bohrungen mit bohrlochkorrigierten Logs und Kernpermeabilitäten eines produktiven Gebietes im Norddeutschen Becken. Kernpermeabilitäten und Loggrößen wurden mit Hilfe eines neuronalen Netzwerkes verknüpft. Dabei bilden die Logs die Eingangsgrößen des Netzwerkes und die Kernpermeabilitäten die zu lernenden Sollwerte. Nach erfolgreichen Tests können anschließend in ungekernten Intervallen mit Hilfe von Logs und trainiertem Netzwerk synthetische Permeabilitäten bestimmt werden. Für die oben beschriebene Problemstellung ist das »Backpropagation Netzwerk« geeignet.%This study investigates the applicability of using Neural Networks to determine the permeability from wireline logs in Rotliegend sandstones in the gas province of the N-Ger-man Basin. Borehole corrected deep and shallow resistivity, sonic, bulk density, neutron porosity and log-porosity logs were used as input for specially designed Neural Networks. In the training phase, Neural Networks were designed using core permeabilities as output. In the application phase synthetic permeabilities were predicted and compared with core permeabilities from a verification data set. Analyses of the results showed that a Back-propagation Network with one hidden layer consisting of 15 nodes and using a sigmoid transfer function was best suited to solve this problem. Among the tested wireline log combinations, the selection of the log suite log-porosity/resistivity-deep/resistivity shallow resulted in the best match between predicted and measured core permeability. The comparison of synthetic permeabilities derived from logs by means of Neural Networks with core permeabilities can be summarized as follows: 1. the trained Network predicted the permeabilities on a verification data set which were not included in the training accurately with a mean error of 0.39 decades (layers of all thickness included). The permeabilities cover the range between 0.001 to 800 mD. 2. synthetic permeabilities are in close agreement with core permeabilities for layers more than 50 cm in thickness. 3. for layers less than 50 cm thickness the log response shows lower resolution due to the thin layer problem. Furthermore, the calibration of logs by cores in thin layers was sometimes problematic due to depth shift and plug sampling. 4. for low permeability ranges ( < 0.01 mD), deviations occured which might also reflect the problem of low permeability determination in plugged cores. 5. an interesting result was that permeabilities of > 1 mD were predicted for thin layers ( < 1 m) in uncored intervals. This information might be used for frac-treatment.