Tight Oil Well Productivity Prediction Model Based on Neural Network

Productivity prediction has always been an important part of reservoir development, and tight reservoirs need accurate and efficient productivity prediction models. Due to the complexity of the tight oil reservoir, the data obtained by the detection instrument need to extract data features at a deeper level. Using the Pearson correlation coefficient and partial correlation coefficient to analyze the main control of productivity factors, eight characteristic parameters of volume coefficient, water saturation, density, effective thickness, skin factor, shale content, porosity, and effective permeability were obtained, and the specific oil production index was used as the target parameter. Two sample structures of pure static parameters and dynamic and static parameters (shale content, effective permeability, porosity, water saturation, and density as dynamic parameters, volume coefficient, skin factor, and effective thickness as static parameters) were created, and corresponding model structures (BP (Backpropagation), neural network model, and LSTM-BP (Long Short-Term Memory Backpropagation) neural network model) were designed to compare the prediction effects of models under different sample structures. The mean absolute error, root mean square error, mean relative percentage error, and coefficient of determination were used to evaluate the model results. The LSTM-BP neural network was used to predict the production capacity of the test set. The results showed that the average absolute error was 0.07, the root mean square error was 0.10, the average absolute percentage error was 21%, and the coefficient of determination was 0.97. Using wells in the WZ area for testing, the LSTM-BP model’s predictions are evenly distributed on both sides of the 45° line, separating the predicted values from actual values, with errors from the line being relatively small. In contrast, the BP model and analytical method are unable to achieve such an even distribution around the line. Experiments show that the LSTM-BP neural network model can effectively extract dynamic parameter features and has a stronger generalization ability.