Research On Extreme Process Neural Net-work Models And Algorithms For Shale Oil Logging Evaluation

Recently,with the sustained and rapid development of economic,the people's oil demand is increasing.But the reserves of conventional oil and gas resources are not sufficient.The unconventional oil and gas resources represented by the shale oil and gas are increasingly becoming the global foucus nowdays.It becomes the important national oil and gas relay and reserve resources nowdays.The logging evaluation plays a crucial role in the process of shale oil exploration and development.Since the complexity of the evaluation of shale oil and gas,the existing evaluation methods have some shortcomings,such as poor adaptability and difficulty in modeling.Aimed at the three typical problems include multi mineral component retrieval,total organic carbon(TOC)prediction and lithology recognition,they are classified into three categories: nonlinear constrained optimization,nonlinear time varying system prediction,nonlinear time varying system diagnosis.In order to solve these problems,the extrem process neural network and quantum-inspired cuckoo search algorithm are proposed and researched in the paper.The relevant research is as follows:Firstly,Aiming at the nonlinear constrained optimization in multi mineral component retrieval,it combined with the theory of quantum computation,the quantum-inspired cuckoo search algorithm(BQCS)and the quantitatively orthogonal crossover quantum-inspired cuckoo search algorithm(DQQCS)are proposed in the paper.The individual in the two algorithms are all coded by quantum bit.The angle of quantum rotating gate is determined by Lévy flights.It expands the search scope of the population and improves the optimizing ability.In addition,DQQCS make the fine search in the local orthogonal region.The validity about two algorithms is verified by optimization problem of benchmark function.The two algorithms are also used to the training of the process neural network.The result shows that the training convergence and approximation ability are better than the gradient descent method based on orthogonal basis expansion.At last,the proposed algorithms are used to the problem about multi mineral component retrieval and achieve the better calculation accuracy than other evolutionary algorithms.Secondly,in order to improve the generalization ability of the process neural network,and aimed at total organic carbon(TOC)prediction based on logging curve,it proposed the prediction model based on extreme process neural network.Aimed at the problem that the process neural network has the slow learning speed and low training convergence,the fixed hide nodes extreme process neural network(FE-PNN)and fixed hide nodes optimized ex-treme process neural network(OFE-PNN)are proposed in the paper.The generalized inverse of hidden output matrix is calculated by SVD and the hidden ouput weights are fastly calculated by least square method.In order to enhance the training convergence about FE-PNN,OFE-PNN uses the quantum-inpried cuckoo search algorithm to optimize the hidden layer input parameters.At last,the OFE-PNN is applied to the TOC prediction and achieves better prediction accuracy than other artificial neural network and?logR method.In order to further increase TOC prediction accuracy and aimed at network structure,the incremental extreme process neural network(IE-PNN)and pruned extreme process neural network(PE-PNN)are proposed in the paper.The adaptive structure process neural network which has extreme learning mechanism is builded.IE-PNN fixes the parameters of the existing nodes and performs correlation analyzation;fast optimization of hidden layer input parameters and output weight calculation for the adding node.PE-PNN performs two pruning process by the QR decomposition and binary cuckoo serach algorithm.The practical application results show that the prediction accuracy of TOC has been further improved.In addition,aimed at lithology recognition based on logging curve,it proposed the diagnostic identification model based on extreme process neural network.In order to enhancing the classification and recognition ability,the integrated extreme ridgelet process neural network(AE-RPNN)is proposed.Firstly,the ridgelet function is taken as the active function in the process neuron.The learning algorithm includes the gradient descent training method and extreme learning method based on full rank decomposition computing generalized inverse.Secondly,it uses the AdaBoost to integrate the weak learning model which is extreme ridgelet process neural network.The classification and recognition ability of process neural network is enhanced.At last,the method's validity is verified by the work condition diagnosis of indicator diagram.The shale lithology identification result shows that the classification recognition effect about AE-RPNN is better than single model and AE-RPNN has the fast learning speed.Finally,the theoretical research results are applied to the problems about logging evaluation of shale oil in Damintun Sag Shahe Group,and the better results have been achieved in in practical application.It provides the new model and algorithms for solving the problem of shale oil logging evaluation.