专注智能油藏储量预测的深度时空注意力模型

A deep spatio-temporal attention model focusing on intelligent reserve prediction

现有油藏储量预测方法的精度远不能满足实际应用的需求。受循环神经网络和注意力机制的启发,提出一种专注智能油藏储量预测的深度时空注意力模型。该模型通过时间注意力模型来捕获输入数据之间的关键信息,空间注意力模型捕获隐藏状态之间的关系紧密程度,能够缓解数据波动对预测结果的不利影响,从而大幅减小预测误差。结果表明,相比传统方法和已有的深度学习方法,该模型预测精度有显著提高,为今后油藏储量预测提供一种更优的选择。

The accuracy of existing reservoir reserve prediction methods is far from satisfying the needs of practical application. Inspired by recurrent neural network(RNN) and attention mechanism, this paper proposes a deep spatio-temporal attention model focusing on intelligent forecasting of reservoir reserves, in which the time attention model can capture the key information within the input data, and the spatial attention model can capture the link between the hidden states. Therefore, the adverse effect of data fluctuation can be mitigated on the prediction results to greatly reduce the prediction error. Experimental results on the real data of a large oil field show that compared with traditional methods and existing deep learning methods, the prediction accuracy of the model by this method is significantly improved. Our approach therefore provides a better choice for reserve prediction in the future.