Research and Application of a Lithology Prediction Method Based on Hybrid Neural Network and Multi-source Data Fusion

Lithology identification is critical for guiding well trajectory adjustments, optimizing drilling parameters, and ensuring wellbore stability, particularly under complex geological conditions. Traditional lithology identification methods rely on well log interpretation or empirical judgment, which often suffer from poor timeliness, incomplete coverage, and delayed responses, making them insufficient for the real-time, accurate formation sensing required in modern intelligent drilling. This paper proposes a lithology prediction model named CNN-LSTM-MDF, which combines a CNN-LSTM hybrid neural network and multi-source data fusion to predict formation lithology using multi-source data. CNN-LSTM-MDF leverages convolutional neural networks (CNN) to extract spatial features from drilling data and long short-term memory (LSTM) networks to capture temporal dependencies. In addition, lithological information from adjacent wells is incorporated as prior knowledge to enhance the accuracy and reliability of the predictions. Models were trained and tested using data from 10 wells in a specific block. After multiple rounds of optimization, it achieved high prediction accuracy. Experimental results show that the proposed method reached a lithology prediction accuracy of 90.32%, outperforming both traditional approaches and conventional machine learning models. By combining the strengths of CNN and LSTM, the method exhibits powerful nonlinear mapping and long-term memory capabilities, effectively capturing both spatial and temporal characteristics of drilling data. Combined with multi-source data fusion strategy, the accuracy and interpretability of the model are further improved. This study offers a novel solution for intelligent drilling, providing substantial practical value and strong potential for deployment in complex geological environments.