Application of stacked bidirectional LSTM neural networks in reservoir porosity prediction

Porosity directly affects the fluid storage and transportation capacity of reservoirs, and is an important parameter for portraying the physical properties of oil and gas reservoirs. However, existing porosity prediction methods have obvious limitations. Traditional petrophysical modeling and cross-plot method struggle to characterize nonlinear relationships under complex geological conditions. Machine learning methods such as support vector machines (SVM) and BP neural networks rely on manual feature extraction, resulting in limited generalization ability. These methods show significant accuracy degradation in noisy data, with insufficient noise immunity. For this reason, this study uses petrophysical modeling to generate large-scale and diverse synthetic data sets of elastic and anisotropic parameters based on field logging data. And an enhanced stacked bi-directional long short-term memory (S-BiLSTM) model is proposed, which uses stacking of bi-directional LSTM layers to enhance feature extraction and prediction capabilities. Validation is carried out using seismic field-measured data from an oil field in southwest China, and the results show that the method has high accuracy and robustness in porosity distribution prediction. Specifically, the prediction accuracy of the S-BiLSTM model reaches 92.19%, which is 7.81% higher than that of the LSTM model and 14.06% higher than that of the RNN model. It also performs well under noisy data. At a signal-to-noise ratio of 10dB its accuracy is 82.81% surpassing RNN at 59.38% and LSTM at 68.75%. When the signal-to-noise ratio drops to 1dB its accuracy remains 76.12% significantly higher than RNN at 44.81% and LSTM at 53.5%. This study explores the potential of deep learning in the prediction of reservoir physical properties parameters, which has certain application value for oil and gas exploration and development.