Research On Two Kinds Of Parameter Prediction Problems Based On Recurrent Neural Network

In recent years,Recurrent Neural Network(RNN)has been widely used in the field of Scientific Computing.This paper focuses on the parameter prediction of numerical reservoir simulation and parametric solver of linear algebraic system based on RNN.For the reservoir parameter characteristic datasets generated by simulating the fracture-vuggy three-phase black oil model with the numerical reservoir simulator KarstSim,firstly,the structured supervised datasets is constructed by data preprocessing methods such as normalization and sliding window method.Then,several prediction models were constructed based on common RNN such as SRN,LSTM,and GRU,and the relative optimal model LSTM was established by comparing different models and different loss functions.Then,two subsets with weaker multi-scale are selected from the original data set,and the corresponding prediction models are established.Numerical experiments show that the prediction performance is getting better and better with the weakening of multi-scale properties,which shows that the multi-scale of datasets is one of the important reasons affecting the prediction performance.Furthermore,we also convert the Timestep feature contained in the above feature datasets into physical time features who has better smoothness.So we obtain a prediction model with higher prediction efficiency.For the defects of unbalanced number of training epochs and excessive number of total training epochs in the training process of weighted Jacobi acceleration model with equidistant distribution weight sequence,two optimization problems are proposed.About the optimization problem 1,firstly,a Binary Search Algorithm for generating weight step sequence is designed,so that the number of training epochs of the accelerated model based on the sequence is balanced.Then,the corresponding supervised datasets is constructed by using data preprocessing methods,and several prediction models are constructed based on SRN,LSTM and GRU,among which GRU model has the best prediction performance.Furthermore,a prediction model with better comprehensive performance is obtained by converting the weight step sequence into weight sequence data.About the optimization problem 2,firstly,a weight sequence generation algorithm is designed,which make the training epochs segmented and the number of total epochs as small as possible.Then,by using the generated weight sequence data and the corresponding supervised dataset,the prediction model based on LSTM is constructed.Numerical experiments show that,compared with equidistant weight sequence,the new prediction weight sequence can effectively reduce the total number of training epochs.Furthermore,the number of total training epoch is further reduced by using extrapolation technology for the smaller weights.