ارزﯾﺎﺑﯽ ﻋﻤﻠﮑﺮد ﺷﺒﮑﻪﻫﺎي ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﺗﻠﻔﯿﻖ ﺷﺪه ﺑﺎ اﻟﮕﻮرﯾﺘﻢ ﻫﺎي ﻓﺮااﺑﺘﮑﺎري وال و ﻣﻮرﭼﮕﺎن در ﺗﺨﻤﯿﻦ ﻧﺮخ ﻧﻔﻮذ ﺣﻔﺎري و ﻣﻘﺎﯾﺴﻪ ﺑﺎ ﺷﺒﮑﻪ ﻫﺎي ﻋﺼﺒﯽ ﺳﺎده و ﻣﺪل ﻫﺎي رﯾﺎﺿﯽ ﻣﺮﺳﻮم

Evaluation of the performance of artificial neural networks integrated with whale optimization and ant colony optimization algorithms in estimating the drilling rate of penetration and compare with simple neural networks and mathematical conventional models

ﺗﺨﻤﯿﻦ ﻧﺮخ ﻧﻔﻮذ )ROP( در ﯾﮏ ﻓﺮاﯾﻨﺪ ﺣﻔﺎري از آن ﺟﻬﺖ ﮐﻪ ﺳﺒﺐ اﻧﺘﺨﺎب ﺑﻬﯿﻨﻪ ﭘﺎراﻣﺘﺮﻫﺎي ﺣﻔﺎري و ﮐﺎﻫﺶ ﻫﺰﯾﻨﻪﻫﺎي ﻣﺼﺮﻓﯽ ﻋﻤﻠﯿﺎت ﻣﯽﺷﻮد ﺑﺴﯿﺎر ﺣﺎﺋﺰ اﻫﻤﯿﺖ اﺳﺖ. ﻫﺪف اﺻﻠﯽ از اﯾﻦ ﻣﻘﺎﻟﻪ، ﻣﺪﻟﺴﺎزي و ﺗﺨﻤﯿﻦ ROP ﺑﺎ اﺳﺘﻔﺎده از ﺷﺒﮑﻪﻫﺎي ﻋﺼﺒﯽ ﭘﺮﺳﭙﺘﺮون ﭼﻨﺪ ﻻﯾﻪ ﺑﻬﯿﻨﻪ ﺷﺪه ﺑﺎ اﻟﮕﻮرﯾﺘﻢ وال )WOA-MLPNN(، ﺷﺒﮑﻪﻫﺎي ﻋﺼﺒﯽ ﺑﻬﯿﻨﻪ ﺷﺪه ﺑﺎ اﻟﮕﻮرﯾﺘﻢ ﻣﻮرﭼﮕﺎن )ACO-MLPNN(، ﺷﺒﮑﻪﻫﺎي ﻋﺼﺒﯽ ﭘﺲ اﻧﺘﺸﺎر ﺧﻄﺎ )BP-MLPNN( و دو ﻣﺪل رﯾﺎﺿﯽ ﺷﺎﻣﻞ ﻣﺪل ﺑﻮرﮔﻮان و ﯾﺎﻧﮓ )BYM( و ﻣﺪل ﺑﯿﻨﮕﻬﺎم ﻣﯽﺑﺎﺷﺪ. دادهﻫﺎي ﻣﻮرد ﻧﯿﺎز ﺑﺮاي ﺗﻮﺳﻌﻪ ﻣﺪلﻫﺎ، از واﺣﺪ ﻧﻤﻮدارﮔﯿﺮي ﮔﻞ و ﮔﺰارﺷﺎت ﭘﺎﯾﺎﻧﯽ ﺳﻪ ﭼﺎه ﺣﻔﺎري ﺷﺪه در ﯾﮏ ﻣﯿﺪان ﻧﻔﺘﯽ واﻗﻊ در ﺟﻨﻮب ﻏﺮﺑﯽ اﯾﺮان ﺟﻤﻊ اوري ﺷﺪه اﺳﺖ، ﮐﻪ ﻧﺨﺴﺖ ﺑﻪ ﻣﻨﻈﻮر ﺣﺬف ﻧﻘﺎط ﺧﺎرج از ﻣﺤﺪوده و ﮐﺎﻫﺶ ﻧﻮﯾﺰ ﭘﯿﺶ ﭘﺮدازش ﺷﺪﻧﺪ. در اداﻣﻪ، از اﻃﻼﻋﺎت ﻣﻘﻄﻊ 12,25 اﯾﻨﭻ دو ﺣﻠﻘﻪ ﭼﺎه ﮐﻪ ﺷﺎﻣﻞ ﯾﮏ ﺗﻮاﻟﯽ ﻣﺸﺎﺑﻪ از ﺳﺎزﻧﺪهﻫﺎي ﺣﻔﺎري ﺷﺪه ﻣﯽﺑﺎﺷﻨﺪ ﺑﻪ ﻣﻨﻈﻮر آﻣﻮزش و آزﻣﺎﯾﺶ ﻣﺪلﻫﺎ اﺳﺘﻔﺎده ﮔﺮدﯾﺪ و ﺳﭙﺲ ﻣﺪلﻫﺎي ﺗﻮﻟﯿﺪ ﺷﺪه، ﺗﻮﺳﻂ اﻃﻼﻋﺎت ﭼﺎه ﺳﻮم ﻣﻮرد اﻋﺘﺒﺎر ﺳﻨﺠﯽ ﻗﺮار ﮔﺮﻓﺘﻨﺪ. در ﭘﺎﯾﺎن، ﻋﻤﻠﮑﺮد ﻣﺪلﻫﺎ ﺑﻮﺳﯿﻠﻪ ﺷﺎﺧﺺﻫﺎي آﻣﺎري و اﺑﺰارﻫﺎي ﮔﺮاﻓﯿﮑﯽ ﻣﺨﺘﻠﻔﯽ ﻣﻮرد ارزﯾﺎﺑﯽ ﻗﺮار ﮔﺮﻓﺖ. ﻧﺘﺎﯾﺞ اﯾﻦ ﻣﻄﺎﻟﻌﻪ ﻧﺸﺎن داد ﮐﻪ روشﻫﺎي آﻣﻮزش ﻣﺎﺷﯿﻦ ﻧﺴﺒﺖ ﺑﻪ ﻣﺪلﻫﺎي رﯾﺎﺿﯽ ﻣﺮﺳﻮم ﺑﺴﯿﺎر دﻗﯿﻘﺘﺮ ﻣﯿﺒﺎﺷﻨﺪ. ﻫﻤﭽﻨﯿﻦ، ﺑﺮرﺳﯽﻫﺎي ﺑﯿﺸﺘﺮ ﺛﺎﺑﺖ ﮐﺮد ﮐﻪ ﻣﺪل WOA-MLPNN ﺑﺎ ﻣﻘﺎدﯾﺮ AAPRE ﺑﺮاﺑﺮ 5,48 ،3,19 و 9,31 ﺑﻪ ﺗﺮﺗﯿﺐ ﺑﺮاي ﺳﻪ ﺑﺨﺶ آﻣﻮزش، آزﻣﺎﯾﺶ و اﻋﺘﺒﺎر ﺳﻨﺠﯽ ﺑﺎﻻﺗﺮﯾﻦ ﻋﻤﻠﮑﺮد را ﻧﺴﺒﺖ ﺑﻪ ﺳﺎﯾﺮ ﻣﺪلﻫﺎ دارا ﻣﯽﺑﺎﺷﺪ.

Rate of penetration (ROP) estimation in a drilling process is very important because it leads to the optimal selection of drilling parameters and reduction of the operating costs. The main purpose of this paper is to modeling and estimating ROP using optimized multilayer perceptron neural network with whale optimization algorithm (WOA-MLPNN), optimized multilayer perceptron neural network with ant colony optimization algorithm (ACO-MLPNN), back propagation multilayer perceptron neural network (BP-MLPNN) and two mathematical models including Bourgoyne and Young model (BYM) and Bingham model. The data required for development of the models were collected from the mud logging unit and the final reports of three drilled wells in an oil field located in southwestern Iran, which were first pre-processed to remove outliers and reduce noise. In the following, 12.25” hole-section information of two wells containing a similar sequence of drilled formations was used to train and test the models, and then the generated models were validated by the third well information. In the end, the performance of models was evaluated by statistical indicators and various graphical tools. The results of this study showed that the machine learning methods are much more accurate than conventional mathematical models. Also, more detailed studies showed that the WOA-MLPNN model with AAPRE values of 3.19, 5.48 and 9.31 for the three sections of training, testing and validation, respectively, has the highest performance compared to other models.