Deep Neural Network Optimization Method And Its Application In Reservoir Porosity Prediction

As one of the important parameters of reservoir physical properties,the accurate prediction of porosity can provide a reliable basis for reservoir evaluation and is an important guide for the rational deployment of oil and gas field development plans.The traditional reservoir porosity prediction methods are costly,and the prediction accuracy is not ideal and the engineering applicability is poor due to the complex geological environment and measurement tools.Deep neural networks have been widely used in reservoir prediction to help improve accuracy and reliability due to their advanced performance in massive,high-dimensional and complex nonlinear data processing.However,deep neural networks still have problems such as difficult to obtain model connection weights,easy to fall into local optimum,and slow convergence speed,which lead to poor reservoir porosity prediction based on deep neural networks.To this end,this paper focuses on the model parameters and gradient descent optimization method of deep neural networks in combination with intelligent algorithms,and applies the optimized model to reservoir porosity prediction.The main research work of this paper is as follows.1.To address the problems of slow convergence and easy to fall into local optimum caused by random initialization of deep neural network weights,a shuffled frog leaping algorithm combining roulette and genetic coding improvement is proposed for optimizing the initial weights and thresholds of the network.Firstly,to solve the problem of weak optimization performance of traditional shuffled frog leaping algorithm due to poor individual interaction ability,a roulette mechanism is introduced in the sub-population division stage to eliminate non-elite individuals by calculating the selection probability of fitness value and improve the global optimization-seeking ability of the algorithm.Secondly,the global and local optimal solutions of the population and the boundary value information of the sub-population where the individuals are located are genetically encoded to avoid the algorithm from falling into the local optimum too early.Subsequently,the correctness and performance of finding the optimal solution of the improved algorithm are verified by comparison experiments with four population intelligence algorithms on twelve benchmark functions.Finally,the proposed algorithm is applied to obtain the optimal initial weights and thresholds,and its effectiveness and superiority for neural network parameter optimization is verified by a large number of comparison experiments.2.To address the problem that the traditional gradient descent algorithm in deep neural networks unable to find approximate optimal connection weights and the ineffectiveness of processing high-dimensional vectors,an Adam algorithm based on random block coordinates fusion with adaptive parameters and compound gradient improvement is proposed.Firstly,the adaptive parameters are used to calculate the degree of difference between the first-order momentum and the current gradient and correct the search direction to improve the gradient deviation caused by outliers,and then improve the search speed and accuracy of the algorithm.Secondly,the predicted gradient is introduced and the current gradient is fused with the first-order momentum to form a composite gradient,which determines a more accurate search direction and improves the global search capability.Finally,random block coordinates are used to determine the gradient update method by randomly selecting variables from the subset of parameters,so as to reduce the computational overhead as much as possible while ensuring the convergence accuracy.The performance of the improved Adam algorithm is verified by extensive comparison experiments on 2 standard classification datasets,and then it is applied to deep neural network gradient descent optimization,and the convergence speed and accuracy of the improved model are verified by extensive comparison experiments with 5neural network optimization algorithms.3.The above two algorithms are used to optimize the initial weights,thresholds and gradient descent of the deep neural network,respectively,to propose a porosity prediction model based on the optimized deep neural network.Firstly,the logging data are pre-processed by gray correlation analysis and deviation normalization.Subsequently,the porosity prediction of the workings is achieved based on the proposed model with the optimized natural gamma,density and other important parameters.The proposed model is compared with four neural network models using root mean square error and mean absolute error as evaluation indexes to verify that the proposed model has better results for porosity prediction.Finally,the reservoir porosity prediction platform based on the optimized deep neural network is designed,and the overall system design,detailed design and porosity prediction results and analysis are given.