Research On Optimization Of Quantitative Analysis Of Oil Content In Oilfield Reinjection Water Based On Neural Network Model

In the process of oilfield-produced water treatment for reinjection,several water quality parameters usually need to be monitored to ensure the water quality of injection water meets the standard and ensure the efficient extraction of crude oil,among which oil content is a key indicator.Research on real-time in-situ monitoring technology of oil content in reinjection water is the key to developing green oilfields and smart oilfields.Spectroscopy,as a fast and non-destructive contaminant detection method,is also the core method to realize online monitoring of oil content in reinjection water.At present,when using the spectral information of reinjection water to quantitatively analyze its oil content,traditional chemometric methods are usually used to establish quantitative calibration models.The quantitative model’s prediction accuracy is limited due to the influence of spectral band overlap and spectral noise,as well as the modeling method’s ignoring of the non-linear relationship between spectral information and oil content.To address the problems of the current quantitative oil content detection model of reinjection water,this paper conducts a quantitative study of oil content neural network under the selection of UV and visible transmission spectral features of reinjection water.Successive Projections Algorithm(SPA),Competitive Adaptive Reweighted Sampling method(CARS),and Genetic Algorithm(GA)are introduced for After that,BP,CNN,and ResNet neural networks are trained to build the quantitative oil content prediction model using full spectrum and feature spectrum,respectively.The structure and hyperparameters of the three neural network models are optimized,and the effects of turbidity variables on the quantitative results of the models are discussed on the basis of the best models.The main research contents and results are as follows.1.The SPA,CARS,and GA methods were used to reconstruct the characteristic spectra of the reinjection water.Taking oilfield reinjection water as the research object and scanning its transmission spectrum in the range of 190–900 nm,the characteristic spectra were selected by the above three methods.77,46,and 179 characteristic wavelengths were extracted from 356 full-spectrum wavelength points by SPA,CARS,and GA models,respectively,and the characteristic wavelengths were mainly located in the visible region,which is an important region for the quantitative analysis of the oil content of reinjection water.2.To investigate the quantitative prediction effect of the PLS quantitative model on the basis of the characteristic spectrum of reinjection water.A PLS quantitative model of oil content based on characteristic spectra was developed to evaluate the quantitative effect with three parameters: r,RMSE,and MAE.GA-PLS was the best quantitative model with a test set correlation coefficient of 0.9102 and an RMSEP of3.15.The PLS modeling approach ignores the nonlinear relationship between spectral variables and oil content,resulting in limited prediction accuracy for this quantitative model.3.To investigate the quantitative prediction effects of BP,CNN,and ResNet neural networks on the basis of feature spectra.The three neural network models were trained with full and feature spectrums,and r,RMSE,and MAE were used to evaluate the quantitative prediction accuracy of the models.The GA-ResNet model was the best quantitative model with a correlation coefficient of 0.9978 and a MAE of 0.44 for the prediction set.The quantitative results of the model were quite accurate.4.The effect of turbidity variables on the quantitative accuracy of the model was investigated.Turbidity variables are stochastic in improving the prediction accuracy of quantitative models,and considering the cost of data used for modeling,it is not recommended to train quantitative models using turbidity.