Prediction method for the formation drillability based on improved Bi-LSTM

Rock drillability serves as a comprehensive parameter reflecting the difficulty of rock fragmentation during drilling. To address the challenges, this study introduces a Bi-LSTM and Stacked Autoencoder (SAE) hybrid model for predicting the drillability of deep undrilled formations. First, the SAE, composed of multiple autoencoders, is utilized to extract the features from the drilling data. The output from the last hidden layer of the SAE is then fed as input to the Bi-LSTM model to capture temporal dependencies in the drilling data and minimize their influence on prediction accuracy, thereby establishing the SAE-BiLSTM formation drillability prediction model. Subsequently, the training time and predictive performance of the SAE-BiLSTM hybrid model are compared with those of Backpropagation (BP), Support Vector Machine (SVM), Random Forest (RF), and a standalone LSTM model. It indicates that the proposed SAE-BiLSTM model achieves the shortest training time and the smallest error between predicted and actual measured values. Specifically, the model demonstrates a Root Mean Square Error (RMSE) of 0.2265, a Mean Absolute Percentage Error (MAPE) of 5.21, and a Coefficient of Determination (R 2 ) of 0.9781. Compared to other models, the SAE-BiLSTM achieves the lowest RMSE and MAPE, along with the highest R 2 , indicating superior prediction accuracy. This method offers a novel approach for predicting formation parameters and helps address common limitations of conventional prediction techniques, such as low efficiency and inadequate accuracy.