[Improved generalized regression neural network for quantitative analysis of crude oil density by gas chromatography]

In the field of oil and gas exploration and development, the quick identification of reservoir crude oil properties has a guiding significance for engineers and technicians. Geochemical logging technology is a conventional method to evaluate the properties of crude oil in reservoirs, and it can provide professional knowledge for comprehensive evaluation of reservoirs. In this study, the principles of rock pyrolysis and gas chromatographic analyses in geochemical logging are studied. Moreover, a new method for quantitative analysis of crude oil density by chromatogram is proposed. Combined with the division standard of crude oil property density, the properties of reservoir crude oil can be quickly evaluated. In the experiment, first, the chromatogram was standardized and normalized using computer image processing software. The curve characteristic law of rock pyrolysis gas chromatogram was analyzed, and the corresponding characteristic parameter extraction method was proposed. The chromatogram was converted into a characteristic parameter matrix. Second, three types of artificial intelligence prediction and classification models were studied. The latest meta-heuristic optimization algorithm (sparrow search optimization algorithm) was used to optimize the hyperparameters of the generalized regression neural network, and the accuracy and convergence speed of the model were improved. To study the influence of different positions of rock samples on the experimental results, two groups of samples were utilized: cuttings samples and wall core samples. Based on a comprehensive comparison of the prediction results of the three models, it was found that the generalized regression neural network prediction model optimized by sparrow search algorithm provided the best effect, being a stable model, with small prediction density error, and strong generalization ability. The prediction error coincidence rate (absolute error < 0.02) of this model for cuttings and wall core samples was 95% and 100%, respectively. The root mean square errors were 0.0079 and 0.0069 respectively. The classification accuracy of crude oil properties was 95%. The analysis of the two groups of parallel experimental data indicated that the rock samples from the wall center can more accurately reflect the crude oil properties of the reservoir. Therefore, the method proposed in this study can provide reliable data support for reservoir comprehensive evaluation and on-site construction.