A Method for Reservoir Permeability Prediction Based on Small Sample Data
A single hidden layer feedforward neural network(SLFNN)is designed to realize nonlinear regression in order to overcome the shortcomings of complicated processing steps in reservoir permeability calculation by Timur/Coates and SDR formulas.The SLFNN contains a hidden layer with a nonlinear activation function,two linear fully connected layers and a dropout layer.In order to prevent the learning process from falling into local optimum and over-fitting caused by small sample data set,Adam optimizer,ReLU activation function,Kaiming He's uniform distribution weight initialization method and cosine annealing hot restart learning rate adjustment algorithm are used.The number of neurons in the hidden layer,initial learning rate,and inactivation probability of neurons in the dropout layer are determined by using the 5-fold cross validation method based on the small sample data composed of the NMR logging and core from different layers of four production wells from A to D in an oilfield as the training set and validation set.Finally,taking the data of Well E in the same block as the test set,the four models of SLFNN,RFR,SVR and XGBR are used to compare the permeability prediction results.The experimental results show that the MAE and R2 of the SLFNN model are better than those of the other three models under the test set,which indicates that the SLFNN model is effective for the prediction of reservoir permeability.
nuclear magnetic resonance loggingreservoir permeabilityKaiming He weight initializationmodel evaluationcorrelation coefficient
国家科技期刊平台
NSTL
万方数据
