Naslov
Neural networks as geomathematical tool in petroleum geology

Autori
Malvić, Tomislav

Izvornik
Autorove skripte i prezentacija

Vrsta, podvrsta
Ostale vrste radova, ostalo

Godina
2008

Ključne riječi
neural tools; Okoli field; Beničanci field; Kloštar field

Sažetak
Neural networks as geomathematical tool in petroleum geology Tomislav Malvić1, 2 1INA-Oil Industry Plc., Reservoir Engineering & Field Development, Zagreb, Croatia, tomislav.malvic@ina.hr 2Faculty of Mining, Geology and Petroleum Eng., Institute of Geology and Geol. Eng., Zagreb, Croatia Abstract: The neural network concept is derived from the architecture of human brains and chemo-biological connection among biological neurons. Such concept was transferred in theoretical model of artificial neurons that also combined different inputs (as signals), transformation algorithms (as hidden layers and activation function) and finally one output (carrying information). The electrical activation impulse in biological neuron is replaced, in the artificial neural system, by so called activation function. Petroleum geological department and INA in Zagreb started to study neural architecture in 2006. The set of petroleum data collected at the Croatian fields had been used for learning and testing of neural network using different software like Matlab, Neuro3, cVision and Statistica. Data from three Croatian fields are very detailed analysed using neural networks. The first analysed hydrocarbon field, in 2007, is Okoli located in the Sava depression. Clastic facies of Lower Pontian were predicted using neural network, using advanced learning algorithm RProp. The network is trained using log data (GR, R16", R64", PORE/T/W, SAND & SHALE logs). The real position of reservoir sandstones and marls is registered, and, using reduced dataset, facies positions are predicted on previously known lithologies (as test of neural approach). The network was over-trained, i.e. the marls are mostly replaced by sandstones. The useful prediction by RPORP algorithm could be reached with more than 90% probability of correct prediction (Okoli network reached 82.1%). The second target was the Beničanci field, located in the eastern part of the Drava depression. It is still one of five main oil reservoirs (Badenian breccia) in Croatia. That makes very meaningful to plan and perform a new set of geological reinterpretations and improvements of field geological model. Neural networks had been applied for seismic attribute processing and porosity prediction. Interpolated neural porosity maps strongly indicates on conclusion that neural estimation is more precisely tool for porosity prediction than previously applied inverse distance weighting and geostatistical methods. Finally, in 2007 and 2008, the prediction of the lithology and oil saturation in oil reservoirs at the Kloštar field is performed using Radial Basis Function and Multi Layer Perceptron networks with the high reliability. Although the input set was relatively small (SP, R16” and R64” logs) it was the most successful application of neural network in petroleum geology in Croatia up to now.

Izvorni jezik
Engleski

Znanstvena područja
Geologija

Napomena
Predavanje je održano kao pozvano za studentski odsjek Sveučilišta u Szegedu IAMG-a ('International Association of Mathematical Geology').

POVEZANOST RADA