用于储量渗透率预测的高效梯度提升决策模型

渗透率预测本质上属于拟合问题,因此可用拟合模型进行解决。机器学习模型是解决拟合问题的利器,其中LightGBM (light gradient boosting machine)表现出色,为此选用该模型进行预测。然而,LightGBM预测性能受自变量的数量和性质影响较大,同时较多超参数的使用使其预测状态难以最优,为此采用MIV (mean impact value)算法和CD (coordinate descent)算法对模型进行改进。为验证提出模型的预测性能,以姬塬油田西部长8段致密砂岩储层为例进行研究。设计了三个实验分别对提出模型进行性能分析。根据实验结果发现MIV和CD的使用能提高LightGBM的预测性能,同时提出模型在预测上较常规混合机器学习模型表现更为高效。实验结果证明提出模型可在纯数据驱动下高效地预测渗透率,较经典物理模型更具有适用性和推广性。

Permeability Prediction using Improved Light Gradient Boosting Machine in Appraisal of Oil-Gas Reserves

Permeability prediction essentially can be regarded as an issue of fitting, thus fitting model naturally being accessible on the prediction. Under modern conditions, machine learning models preferably are employed in the fitting, and LightGBM (light gradient boosting machine) is universally viewed as one of the best solvers, so that this model is selected as a predictor to achieve permeability data. Nonetheless, the working performance of LightGBM will be seriously limited by the amount and the property of independent variables of learning samples, and simultaneously LightGBM has much difficulty to be in a best predicting status as many hyper-parameters are utilized in the modeling stage. In view of that circumstance, MIV (mean impact value) and CD (coordinate descent) algorithms are introduced as optimizers to LightGBM. To verify the predicting capability of the proposed hybrid model, tight sandstone reservoirs within the Chang 8 member of Jiyuan Oilfield are selected as study objectives to realize verification. Three experiments are well designed to validate the predicting capability of the proposed hybrid model, respectively. The experimental results manifest the integrations of MIV and CD are very significant to the improvement on the prediction of the proposed hybrid model, meanwhile the proposed hybrid model showing higher efficiency on the prediction compared to other competitive hybrid models. The results well demonstrate the proposed hybrid model is high-efficient to figure out the permeability data based on pure data-driven computing conditions, and compared to the classic physical models, presents to be more suitable and generable on the permeability prediction.