Regional Risk Prediction Method Based on Deep Learning and Multi-source Data

Drilling in complex formations in new exploration areas poses the challenge of accurately predicting downhole engineering risks before drilling. To address this issue, researchers are exploring the use of intelligent algorithms to analyze the complex relationship between multi-source data and underground engineering risks. However, due to the limited number of drillings in these areas, there are few risk samples, which can result in insufficient generalization ability of the training model and poor prediction effect. To overcome these challenges, this paper introduces a quantitative evaluation method for drilling well engineering risk, which enables the construction of a complex underground risk probability profile rich in geological-engineering information. This risk profile provides reliable risk samples for subsequent model training. Additionally, the paper proposes the concept of virtual wells and its deployment method. The LSTM deep learning model is used to mine the quantitative relationship between multi-source data, such as seismic interval velocity, well logging, and rock mechanics parameters, and the downhole risk probability profile. To achieve a quantitative prediction of underground engineering risk probability profiles of virtual wells, relevant parameters of virtual wells are calculated using Depth Adjustment and Kriging interpolation method. The example calculation demonstrates that the addition of virtual wells can significantly improve the regional drilling engineering risk understanding compared to the 3D drilling engineering risk body constructed only based on wells. The prediction accuracy of engineering risks in unexplored areas can be increased by up to 23.6%.