【摘要】 复杂岩性井段对密度测井数据精度要求很高，传统的计算模型难以满足此类高精度要求，为此提出利用机器学习密度预测模型提高密度测井曲线的精度。首先，使用蒙特卡罗模拟双探测器密度测井仪器，获取不同地层的密度测井数据用以训练和测试机器学习模型。考虑到密度预测模型的过拟合问题和不同模型的密度补偿性能，基于麻雀搜索算法（Sparrow Search Algorithm,SSA）分别改进XGBoost、支持向量回归（Support Vector Regression,SVR）、随机森林回归（Random Forest Regression,RFR）和长短记忆网络（Long Short-Term Memory,LSTM）等模型，进而提出了SSA-XGBoost、SSA-SVR、SSA-RFR和SSA-LSTM密度预测模型。然后，使用量化评价指标和泰勒图模型对比分析各个模型的预测性能，分析了不同预测模型对实际密度测井数据的预测效果。结果表明：SSA-XGBoost模型的预测精度高于传统脊-肋图模型和其他机器学习模型，其预测地层密度的误差为0.017 4 g·cm^-3，远低于传统脊-肋图法的0.028 4 g·cm^-3，因此在实际密度测井数据处理中具有广阔的应用前景。更多还原

【Abstract】 [Background] Complex lithology well sections require high precision in density well logging data whilst traditional computational models are difficult to meet this high precision requirement. [Purpose] This study aims to improve the precision of density logging curves utilizing machine learning regression prediction models. [Methods]Firstly, Monte Carlo N-Particle transport code（MCNP） was utilized to obtain stratigraphic data of varying density of dual-detector density logging tool instrument to validate the predictive effectiveness of the model. Then, sparrow search algorithm（SSA） was adopted to enhance XGBoost model, resulting in the development of the SSA-XGBoost density prediction model. Subsequently, the parameters of support vector regression（SVR）, random forest regression（RFR）, and long short-term memory（LSTM） were optimized by employing the SSA to construct the SSA-SVR, SSARFR, and SSA-LSTM models to predict the simulated formation density, and quantitative evaluation metrics and Taylor diagram models were applied to the comparison and analysis of the predictive performance of each model.Finally, the performance of different prediction models was evaluated on actual density logging data. [Results]Results of the comparative analysis and processing of actual well density logging data with various models show that the SSA-XGBoost model exhibits smaller errors between predicted and actual density and its error in predicting formation density is 0.017 4 g·cm^-3, which is much lower than the traditional spine-ribs plots error of 0.028 4 g·cm^-3.[Conclusions] The SSA-XGBoost model demonstrates higher predictive accuracy than traditional spine-ribs plots and other models, showing great potential for applications in the processing of actual density logging data.更多还原