Predict Oil Production from Geological and Petrophysical Data Before Hydraulic Fracturing Using an Improved Particle Swarm Optimization Based Least Squares Support Vector Machine

Abstract

Abstract For no natural productivity, hydraulic fracturing is widely used in unconventional reservoirs. Thus, oil production prediction before hydraulic fracturing is crucial. In general, it is usually obtained based on percolation theory and fracture geometrical model from engineering data. However, before hydraulic fracturing, the obtained data are mainly geological and petrophysical data. Hence, predicting oil production from them before hydraulic fracturing can make a lot of sense for guiding fracturing design. Although geological and petrophysical data have complex and nonlinear relation with oil production after hydraulic fracturing, least squares support vector machine (LSSVM) is very suitable for such high-dimensional and nonlinear problem with small datasets. As the performance of LSSVM highly depends on parameter selection, an improved LSSVR is designed based on particle swarm optimization (PSO). The selected study area is tight oil reservoirs of Chang 8 Formation, Triassic, Ordos Basin, China. In order to build a reliable data-driven oil production prediction model, a systematic formation evaluation workflow is proposed. Firstly, choose wells with similar hydraulic fracturing type in the study area, which is the base. Secondly, the oil production prediction problem is converted into a classification problem by dividing the oil production into four levels. Thirdly, the dataset of 24 geological, petrophysical and engineering parameters from 85 wells are selected and constructed, reflecting lithology, physical property, saturation, rock mechanics, thickness and facies. Fourthly, 8 parameters, including density, the difference of neutron and density porosity, permeability, water saturation, effective thickness, Poisson's ratio, interlayer stress difference and displacement amount, are optimized as sensitive ones by principal component analysis (PCA). Fifthly, the data from 85 wells are separated into two categories: training and testing data, according to the proportion of 70% and 30%, respectively. The former are used to train the model, while the latter are used for verification. Sixthly, the improved PSO based LSSVM and the LSSVM are trained and tested via the dataset respectively. The tested results present that the coincidence rate of the prediction of the two methods are 88.06% and 81.16% respectively. This work verifies that predicting oil production from geological and petrophysical data before hydraulic fracturing via an improved LSSVM is feasible and cost-effective, which can be important for guiding fracturing design and providing scientific basis for deployment of development program.