Optimization of Rock Mechanical Properties Prediction Model Based on Block Database

Rock mechanical properties are complex but essential for reservoir stimulation. Both experimental and empirical equation method have inevitable defects, and artificial neural network model (ANN) shows great potential in rock mechanics. This paper established a preliminary rock database from the Keshen block in Tarim Basin and improved clustering algorithm to assess the similarity between wells, creating sub-databases and realizing data sharing among similar wells. 18 parameters are used to predict ten rock mechanical properties and corresponding model inputs are optimized by correlation and sensitivity analysis separately. The results show that more accurate predictions can be obtained using optimized sample data and model inputs. The average predicted loss value using all 31 wells (221 samples) was 7.28%, while the value was reduced to 5.4% after clustering analysis using seven wells (95 samples). The average loss value can be further reduced to 4.3% using the optimized model inputs. As for rock peak stress, the predicted loss value is 2.5%, the determination coefficient (R-2) is 0.983, and the root mean square error (RMSE) is 0.03. By optimizing training method, improving training sample quality, and optimizing model inputs, the problem of overfitting can be alleviated and a more reliable ANN model can be obtained.