【摘要】 在石油和天然气的勘探开发中，人工智能模型被广泛应用于地层孔隙压力的预测。其中单一模型容易出现过拟合或预测结果不稳定的问题，在预测精度和泛化能力方面仍有提升空间。为了提高地层孔隙压力预测精度，基于深度学习技术，建立了CNN-Attn神经网络地层孔隙压力预测模型。优选了5种测井和随钻数据，使用Pearson相关系数法验证数据与地层孔隙压力的线性相关性。通过对一维CNN的结构进行优化，使模型能够有效捕捉数据的局部特征，并与自注意力机制结合，增强模型对全局依赖关系的捕捉能力，从而提高模型的表现力和理解能力。为了验证该模型的预测精度，对巴彦区块两口井进行预测，预测结果的平均绝对误差均小于1 MPa,均方根误差均小于1 MPa,平均相对误差均小于1.3%,决定系数均大于0.9,比BP、CNN和LSTM模型精度高。该模型提升了地层孔隙压力预测精度，并为钻井安全性提供了数据支持。更多还原

【Abstract】 In the exploration and exploitation of oil and gas, artificial intelligence models are extensively employed in the prediction of formation pore pressure. Among them, single models tend to encounter problems such as overfitting or unstable prediction outcomes, leaving room for improvement in aspects like prediction accuracy and generalization ability. To enhance the prediction accuracy of formation pore pressure, a CNN-Attn neural network-based formation pore pressure prediction model was established by virtue of deep learning technology. In this research, five types of logging and while-drilling data were optimally selected, and the linear correlation between the data and formation pore pressure was verified using the Pearson correlation coefficient method. Through the optimization of the structure of the one-dimensional CNN, the model can effectively capture the local characteristics of the data and, when combined with the self-attention mechanism, strengthen the model’s ability to capture global dependencies, thereby elevating the model’s expressiveness and comprehension. To validate the prediction accuracy of this model, two wells in the Bayan block were subjected to prediction. The average absolute errors of the prediction results were both less than 1 MPa, the root mean square errors were both less than 1 MPa, the average relative errors were both less than 1.3%, and the determination coefficients were both greater than 0.9, with higher accuracy compared to the BP, CNN, and LSTM models. This model has improved the prediction accuracy of formation pore pressure and provided data support for drilling safety.更多还原