ﻣﻘﺎﯾﺴﻪ ﮐﺎرﮐﺮد ﺷﺒﮑﻪﻫﺎي ﻋﺼﺒﯽ ﻣﺮﺳﻮم ﺑﺮاي ﺑﺮآورد ﺗﺨﻠﺨﻞ در ﯾﮑﯽ از ﻣﯿﺪانﻫﺎي ﻧﻔﺘﯽ ﺟﻨﻮب ﺧﺎوري اﯾﺮان

Comparison of the function of conventional neural networks for estimating porosity in one of the southeastern Iranian oil fields

در ﺻﻨﻌﺖ ﻧﻔﺖ از ﻫﻮش ﻣﺼﻨﻮﻋﯽ ﺑﺮاي ﺷﻨﺎﺳﺎﯾﯽ رواﺑﻂ، ﺑﻬﯿﻨﻪﺳﺎزي، ﺑﺮآورد و ردهﺑﻨﺪي ﺗﺨﻠﺨﻞ ﺑﻬﺮهﮔﯿﺮي ﻣﯽﺷﻮد. ﯾﮑﯽ از ﻣﻬﻢﺗﺮﯾﻦ ﻣﺮاﺣﻞ ارزﯾﺎﺑﯽ ﭘﺎراﻣﺘﺮﻫﺎي ﭘﺘﺮوﻓﯿﺰﯾﮑﯽ ﻣﺨﺰن، ﺷﻨﺎﺳﺎﯾﯽ وﯾﮋﮔﯽﻫﺎي ﺗﺨﻠﺨﻞ اﺳﺖ. ﻫﺪف اﺻﻠﯽ اﯾﻦ ﭘﮋوﻫﺶ ﻣﻘﺎﯾﺴﻪ درﺳﺘﯽ و ﺗﻌﻤﯿﻢﭘﺬﯾﺮي ﺳﻪ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﭼﻨﺪ ﻻﯾﻪ ﭘﯿﺶﺧﻮر )MLFN(، ﺷﺒﮑﻪ ﺗﺎﺑﻊ ﺷﻌﺎع ﻣﺒﻨﺎ )RBFN( و ﺷﺒﮑﻪ ﻋﺼﺒﯽ اﺣﺘﻤﺎﻟﯽ )PNN( ﺑﺮاي ﺑﺮآورد ﺗﺨﻠﺨﻞ ﺑﺎ ﺑﻬﺮهﮔﯿﺮي از وﯾﮋﮔﯽﻫﺎي ﻟﺮزهاي اﺳﺖ. در اﯾﻦ راﺳﺘﺎ، دادهﻫﺎي زﻣﯿﻦﺷﻨﺎﺳﯽ 7 ﺣﻠﻘﻪ ﭼﺎه ﯾﮏ ﻣﯿﺪان ﻧﻔﺘﯽ ﻓﺮاﺳﺎﺣﻠﯽ ﻫﻨﺪﯾﺠﺎن در ﺷﻤﺎل ﺑﺎﺧﺘﺮي ﺣﻮﺿﻪ ﺧﻠﯿﺞ ﻓﺎرس ﻣﻮرد ارزﯾﺎﺑﯽ ﻗﺮارﮔﺮﻓﺖ. اﻣﭙﺪاﻧﺲ آﮐﻮﺳﺘﯿﮏ ﺑﺎ ﺑﻬﺮهﮔﯿﺮي از روش واروﻧﮕﯽ ﻣﺒﺘﻨﯽ ﺑﺮ ﻣﺪل ﺑﺮآورد ﺷﺪ و ﺳﭙﺲ ﺷﺒﮑﻪﻫﺎي ﻋﺼﺒﯽ ﯾﺎد ﺷﺪه ﺑﺎ ﺑﻬﺮهﮔﯿﺮي از وﯾﮋﮔﯽﻫﺎي ﻟﺮزهاي ﺑﻬﯿﻨﻪ ﻃﺮاﺣﯽ ﺷﺪه و ﺑﺎ روش رﮔﺮﺳﯿﻮن ﮔﺎم ﺑﻪ ﮔﺎم ﻣﻮرد ارزﯾﺎﺑﯽ ﻗﺮار ﮔﺮﻓﺘﻨﺪ. ﺳﺮاﻧﺠﺎم ﻣﺸﺨﺺ ﺷﺪ ﮐﻪ ﻣﺪل MLFN ﺑﺮاي ﺑﺮآورد ﺗﺨﻠﺨﻞ ﺧﻮب ﻋﻤﻞ ﻧﻤﯽﮐﻨﺪ. PNN از ﺑﻬﺘﺮﯾﻦ دﻗﺖ ﮐﺎرﮐﺮد در درونﯾﺎﺑﯽ ﺗﺨﻠﺨﻞ ﺑﺮﺧﻮردار اﺳﺖ، اﻣﺎ ﺗﻌﻤﯿﻢﭘﺬﯾﺮي RBFN ﺑﻬﺘﺮ اﺳﺖ.

In the oil industry, artificial intelligence is used to identify relationships, optimize, estimate and classify porosity. One of the most important steps in evaluating the petrophysical parameters of the reservoir is to identify the porosity properties. The main purpose of this study is to compare the accuracy and generalizability of three multilayer feed neural networks (MLFNs), radius base function networks (RBFNs) and probabilistic neural networks (PNNs) to estimate porosity using seismic properties. In this regard, geological data of 7 wells were evaluated from an offshore oil field in Hindijan in the northwest of the Persian Gulf basin. Acoustic impedance was estimated using model-based inversion method and then the mentioned neural networks were designed using optimal seismic properties and evaluated by stepwise regression method. Finally, it became clear that the MLFN model did not work well for estimating porosity. PNN has the best performance accuracy in porosity interpolation, but RBFN generalizability is better.