关键词: 油气井产量预测, 时域卷积神经网络, 多头注意力, 动态卷积, 自适应

Abstract:

The current poor performance of machine learning in predicting oil and gas well production is primarily due to conventional methods relying excessively on historical production data features, which results in predictions that largely reconstruct past information and struggle to predict new trends. These methods overlook other important time-series variables, such as the development stage of oil and gas wells, pressure, and water production, which affect production. To address these issues, this study proposed strategies associating pressure and water production with output and established a novel oil and gas well production prediction method based on a temporal attention and dynamic convolutional neural network (TADyC). This method used a temporal convolutional neural network as the base model and introduced multi-head attention and dynamic convolution mechanisms to capture long-term dependencies between different time steps in the input feature sequence and assign different weights to each time step. The dynamic convolution module dynamically generated convolution kernel parameters based on the output of the temporal attention module, thereby adapting to the input features across different production stages. The superiority of the TADyC-based oil and gas well production prediction model was demonstrated through validation using multiple real and complex well cases from the Anyue gas production area. The results showed that the proposed model achieved better prediction performance when tested on four randomly selected wells. Furthermore, visualization analysis of the attention and dynamic convolution weights revealed that the model could dynamically adjust the convolution kernel weights according to different development stages, particularly for the initial, transition, and decline stages of gas wells. By integrating the relationships between pressure, water production, and production at different development stages, the TADyC model can adaptively adjust its structure and parameters, thereby achieving accurate prediction of oil and gas well production.

Key words: oil and gas well production prediction, temporal convolutional network, multi-head attention, dynamic convolution, adaptivity