储层参数是储层的重要特征，它们的大小影响着油气勘探开发进程。在实际工程中，储层参数反演问题有两种形式：一种是利用测井曲线和取芯所得的储层参数建模，用测井曲线反演储层参数；一种是利用空间统计学方法建模，对物性参数在储层层面上进行插值预测。

首先，论文介绍了相关基础理论知识。第一，研究了测井曲线、储层物性参数的基本概念和它们之间的联系；第二，研究了论文所需的基本模型方法：主成分分析、K-means聚类、生成对抗网络、长短时记忆网络、克里金插值法、BP神经网络、粒子群算法等，这些基础理论为后文的模型建立、参数优化起着重要的作用；第三，研究了模型评价指标，用来对模型的优良进行判断。

其次，论文研究了利用测井曲线反演储层参数，提出一种融合领域信息的LSTM-CGAN（​​Long Short Term Memory-Conditional Generative Adversarial Network​）储层参数反演模型。这个问题早期常采用统计学方法，例如多元回归方法，但由于地层的非均质性，测井曲线与储层参数之间有复杂的非线性关系，传统统计学不能准确有效地反演。论文采用主成分分析进行特征筛选，用K-means聚类进行井群划分来强化邻域信息，将LSTM引入到CGAN框架，进行储层参数孔隙度、渗透率反演方法研究。将新方法应用于实际数据中，相较于线性回归、全连接神经网络、支持向量机等传统机器学习模型，孔隙度、渗透率的四类井群平均误差分别比传统最优模型降低24.7%、53.9%，结果表明在储层参数反演中新方法有优异的性能。

最后，论文研究了储层参数的层面插值方法，提出一种基于协同实验的PSO-BP（Particle Swarm Optimization-Backpropagation​）网络的插值建模方法，在二维层面上利用协同变量对储层参数插值预测。这个问题早期常采用地质统计学方法，例如克里金插值法，但其存在对数据分布依赖性高、计算复杂度高等问题。论文首先将粒子群优化算法引入BP神经网络，优化初始权值阈值，缓解了传统BP神经网络易陷入局部极值、过拟合等缺陷。其次，将孔隙度、渗透率及其组合作为协同变量，对目标储层参数插值预测，对比各模型预测误差及可视化效果，得出最优模型。将新方法应用于实际数据中，研究结果表明，在插值储层参数净毛比时，引入渗透率作为协同变量的插值结果明显优于未引入协同变量的情况，插值结果的均方误差降低6.4%。与传统克里金插值法对比，均方误差降低21.8%。这个研究为储层参数的层面插值提供了更具潜力的解决方案。

综合来看，论文研究的两种反演问题有很好的关联性，层层递进。首先利用测井曲线反演不同深度的储层物性参数，形成单井纵向参数序列；进而利用不同井在相同储层层面的预测结果，实现井间储层参数的层面插值预测，两者综合为储层三维地质建模提供了理论和实践的基础。

Reservoir parameters are important characteristics of reservoirs, and their magnitude affects the process of oil and gas exploration and development. In practical engineering, there are two forms of reservoir parameter inversion problems: one is modeling the reservoir parameters obtained from logging curves and coring, and inverting the reservoir parameters with logging curves; and the other is modeling the physical parameters using spatial statistical methods to interpolate the physical parameters to predict them at the reservoir level.

Firstly, the thesis studied the related basic theoretical knowledge. First, the basic concepts of logging curves, reservoir physical parameters and the connection between them are studied; second, the basic modeling methods needed for the thesis are studied: principal component analysis, K-means clustering, generative adversarial network, long and short-term memory network, Kriging interpolation, BP neural network, particle swarm algorithm and so on, which are basic theories that play an important role for the model building and parameter optimization in the later paper; third , the model evaluation index is studied, which is used to judge the excellence of the model.

Secondly, the paper studies the inversion of reservoir parameters using logging curves, and proposes a LSTM-CGAN (Long Short Term Memory-Conditional Generative Adversarial Network) reservoir parameter inversion model that integrates domain information. Statistical methods, such as multiple regression, were often used in the early stage of this problem, but due to the non-homogeneity of the formation and the complex non-linear relationship between logging curves and reservoir parameters, the traditional statistics can not be accurately and effectively inverted. The paper adopts principal component analysis for feature screening, K-means clustering for well group division to strengthen the neighborhood information, and introduces LSTM into the CGAN framework for the research of inversion methods of reservoir parameters porosity and permeability. Applying the new method to actual data, compared with traditional machine learning models such as linear regression, fully connected neural network, and support vector machine, the average errors of the four types of well clusters for porosity and permeability are reduced by 24.7% and 53.9%, respectively, compared with the traditional optimal model, and the results show that the new method has excellent performance in reservoir parameter inversion.

Finally, the paper investigates the level interpolation method of reservoir parameters, and proposes an interpolation modeling method based on PSO-BP (Particle Swarm Optimization-Backpropagation) network of synergistic experiments, which interpolates and predicts the reservoir parameters using synergistic variables at the two-dimensional level. This problem often used geostatistical methods, such as Kriging interpolation, in the early stage, but it has the problems of high dependence on data distribution and high computational complexity. The paper firstly introduces the particle swarm optimization algorithm into the BP neural network to optimize the initial weight threshold, which alleviates the defects of the traditional BP neural network that is easy to fall into the local extreme value and overfitting. Secondly, porosity, permeability and their combinations are used as synergistic variables to interpolate the prediction of target reservoir parameters, and the optimal model is derived by comparing the prediction error and visualization effect of each model. Applying the new method to actual data, the results show that when interpolating the net gross ratio of reservoir parameters, the interpolation result with permeability introduced as a covariate is significantly better than that without the introduction of covariate, and the mean-square error of the interpolation result is reduced by 6.4%. Compared with the traditional kriging interpolation method, the mean square error is reduced by 21.8%. This study provides a more potential solution for the level interpolation of reservoir parameters.

Taken together, the two inversion problems studied in the paper have a good correlation with the layers. Firstly, the inversion of reservoir physical parameters at different depths using logging curves forms the longitudinal parameter sequence of a single well; and then the prediction results of different wells at the same reservoir level are used to realize the level interpolation prediction of reservoir parameters between wells, and the combination of the two provides the theoretical and practical basis for the three-dimensional geological modeling of reservoirs.