| Unconventional petroleum resources,such as shale oil and shale gas,are currently regarded as essential resources in the face of depleting conventional hydrocarbon resources.To determine accurately the potential of an unconventional shale reservoir requires in-depth knowledge of the distribution of the organic geochemical properties.In line with this,the accurate determination of the hydrocarbon potential of shale rock is critical and relies in part,on the total organic carbon TOC content.The most organic geochemical examination of organic matter is the Rock-Eval Pyrolysis technique.Laboratory pyrolysis of organic matter in the rocks attempts to track variations in the subsurface and is a helpful tool for identifying organic matter.This technique was adapted as a major laboratory examination of the organic matter of stratigraphic units spanning the Mihambia,Nondwa,and Mbuo formations in this research.Although direct measurement is the most accurate way to obtain a representative value,at the same time,it is a little difficult and expensive if carried out on all wells.In the absence of such facilities,other approaches such as analytical solutions and empirical correlations are used.However,there exist challenges when determining these parameters from analytical techniques such as the relationship between well log response and TOC values.For the case of TOC,analytical techniques assume linear relations such as resistivity and porosity logs which always end up on poor correlation.The utilization of Artificial Intelligence(AI)keeps on developing in ubiquity inside petroleum geosciences in perspective on regularly developing intricacy and extent of accessible subsurface information data.The use of AI in the prediction of geochemical logs of organic shale formation is turning out to be a perpetually regular spot as more specialists embrace this technique for displaying and prediction purposes.Typical approaches for artificial intelligence include Artificial Neural Networks(ANN)and Machine learning models amongst others which are run both in supervised and unsupervised modes.Group Method of Data Handling(GMDH)algorithm is a selforganizing technique that in computation intelligence can solve certain practical limitations.Therefore,as a means to apply novel approaches for the purposes of source rock evaluation and prediction of total organic carbon(TOC),the following studies were conducted.1.The first study conducted a detailed analysis of the Triassic-Jurassic of the Mandawa Basin,South-Eastern Tanzania to determine source rock potentiality based on Group Method of Data Handling(GMDH)neural network,Machine learning,and Geochemical using well logs Data.The overall analysis results revealed that the evaluated source rocks mainly yielded organic matter contents(TOC)ranging from 0.5 to 8.7 wt%;rated between fair and very good potential source rocks.The source rocks are mainly kerogens type I,II,III,and mixed Types II/III,with a predominance of Type II,III,and mixed Types II/III.Based on pyrolysis data(Highly variable HI = 13 to 1 000 mg/g TOC,(Tmax 417 to 473 o C,OI = 16 to 225 mg/g,PI = 0.02 to 0.47,)For the case TOC intelligent models,GMDH performed excellent with least error of 0.0381 and 0.1627 for MSE of and MAE respectively during training.This reveals the importance of adopting GMDH and machine learning in the evaluation and prediction of organic content which will help in the exploration and exploitation of hydrocarbon reserves economically.2.The second study applied and assessed for the first time the applicability of the generalized structure of the Group Method of Data Handling(g-GMDH)as a novel neural network model in predicting total organic carbon(TOC)from well logs.The study used bulk density,sonic travel time,deep lateral resistivity log,gamma-ray,spontaneous potential,neutron porosity well logs as input variables to predict TOC,of the Nondwa,Mbuo,and Mihambia Formations in the Triassic to mid-Jurassic of the Mandawa Basin found in South-East Tanzania.The TOC prediction results indicated that the g-GMDH TOC model trained well while generalizing better across the testing data than both ANN and Δlog R.Specifically,g-GMDH provided TOC testing predictions having the least error values of 0.40 and 0.45 for MSE and MAE as compared to 1.27 and 0.81,0.68 and 0.7,1.4 and 0.89 obtained by backpropagation neural network(BPNN),radial basis function neural network(RBFNN),and Δlog R,respectively.Hence,the improved generalization performance of g-GMDH makes it a novel form of a neural network for TOC prediction.The proposed model was further adopted to predict the geochemical data and determine the source rock quality for the East Lika-1 well which has no core data.3.In the third study,Total Organic Carbon(TOC)was predicted using the enhanced group method of data handling(GMDH)based on the modified Levenberg Marquardt technique.The study used natural gamma-ray,formation bulk density,limited effective porosity,neutron porosity,sonic travel time,and deep lateral resistivity log well logs as input variables.The GMDH-LM TOC model was developed with a high level of flexibility in each node and connectivity configuration.The modified GMDH-LM network is not limited to contiguous layers and becomes fully optimized structurally and parametrically.Results show that an enhanced method has a reasonable reduction in processing time with high accuracy.Compared to conventional GMDH,Random Forest(RF),and ANN of backpropagation neural networks(BPNN),the GMDH-LM used 30% less computation time and performed excellently during training with the least error values of 0.019 and 0.0105 for RMSE and MAE.Likewise,good results were observed during testing,obtaining the least error values of 0.0279 and 0.0198 for RMSE and MAE respectively.The modified generalization performance of GMDH-LM makes it an improved form of GMDH and can be adopted as an improved alternative in predicting Total Organic Carbon(TOC).Based on the above work,this study believes that based on different processing methods of geochemical and logging data,the TOC prediction model is established by using the neural network technology of d GMDH,which can predict the TOC of source rocks quantitatively in real-time,save time and obtain good prediction results.The main innovations of this study are the following two aspects:(1)it is confirmed that the generalized structure of the Group Method of Data Handling(g-GMDH)based on logging data as a new neural network model has good applicability for predicting total organic carbon(TOC)of source rocks,which is better than ANN and Δlog R.It has a better prediction effect and can become a new method of neural network technology for TOC prediction of source rocks.(2)The enhanced group method of data handling(GMDH)based on the modified Levenberg Marquardt technique is established to predict total organic carbon(TOC).It not only saves the calculation time but also provides good prediction performance and less error.It can be used as an improved alternative method to predict the total organic carbon(TOC)of source rocks.These technologies will have great application value for the prediction of total organic carbon(TOC)of organic shale rock and determining the source rock quality for other wells which has no core data.Developments of g-GMDH and GMDH-LM techniques for total organic carbon(TOC)model have a great contribution to the petroleum industry because it overcomes challenges facing standard ANN and ML such as BPNN,RF,and RBNN,which may be used in the prediction of total organic carbon(TOC)of the other basins,and also is important progress in the application of artificial intelligence technology in the petroleum industry. |
PDF Full Text Request