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 overly relying on historical production data features, which results in predictions that largely resemble the recombination of past information, failing to forecast new trends. These methods overlook other important time-series variables, such as the development stage of the oil and gas well, pressure, and water production, which affect production. To address this issue, this study proposes a strategy for associating pressure and water production with output and establishes a novel oil and gas well production forecasting method based on a temporal attention dynamic convolutional neural network (TADyC). The method uses a temporal convolutional neural network as the base model and introduces multi-head attention and dynamic convolution mechanisms to capture the long-term dependencies between different time steps in the input feature sequence, assigning different weights to each time step. The dynamic convolution module can dynamically generate convolution kernel parameters based on the output of the temporal attention module, thus adapting to the input features of different production stages. Through verification with real-world, complex multi-well cases from the Anyue gas production area, the superiority of the TADyC-based oil and gas well production forecasting model is demonstrated. The study shows that the proposed model performs better when tested on four randomly selected wells. Further, through the visualization analysis of attention and dynamic convolution weights, it is found that the model can dynamically adjust the convolution kernel weights according to different development stages, particularly for the initial, transition, and decline stages of the wells. By combining the relationships between pressure, water production, and yield at different development stages, the temporal attention dynamic convolution neural network model can adaptively adjust its structure and parameters, thus achieving accurate oil and gas well production forecasting.

Key words: Oil and gas well production prediction, Temporal Convolutional Neural Network, Multi-head attention, Dynamic convolution, Adaptivity