基于图神经网络的注采井网产油量预测方法OA

Traditional neural network-based oil production forecasting methods fail to account for the mutual influence between injection and production wells.To address this limitation,this study proposed a graph neural network(GNN)-based oil production forecasting method for injection and production well patterns.The method integrates dilated convolutional neural networks(dilated CNNs)with causal convolutional neural networks(causal CNNs)to extract production dynamic features.An adaptive inter-well connectivity matrix was constructed based on reservoir geological characteristics,and a graph diffusion neural network(GDNN)was employed to capture inter-well connectivity information,thereby enhancing the forecasting capability of the GNN under injection and production well pattern conditions.By using production data from a well group comprising nine production wells and five injection wells in a block of the Daqing Oilfield as a case study,the hyperparameters of the GNN-based model were optimized,and the accuracy of the prediction results was validated.The results indicate that compared with traditional neural network models,the proposed GNN model significantly reduces prediction errors,with the mean absolute percentage error(MAPE)decreasing by 32.431%and the normalized deviation(ND)decreasing by 7.785%.The model generates an optimized inter-well connectivity matrix based on historical production data,ensuring the matrix aligns better with actual reservoir geological properties.This optimized matrix can subsequently assist in future oil production forecasting.Furthermore,optimizing hyperparameters,such as the maximum diffusion order,effectively improves prediction accuracy,confirming the necessity of hyperparameter optimization for GNN models.Verified with field production data,the proposed method demonstrates high prediction accuracy and strong robustness,indicating its applicability for oil production forecasting in actual oilfields under water injection development.