Abstract：
The accurate prediction of the produced water quality is an important basis for evaluation of the treatment effect of
the produced water at the oilfield joint station, which can provide a scientific basis for early warning of water quality. In the tradi
tional method, we can see that the prediction of the oilfield produced water quality is mainly based on the experience of experts,
however, there is no doubt that this method has a strong personal subjectivity so it is difficult to reach an accurate prediction
of the quality of the produced water. There is also a part of existing studies to measure whether the produced water quality is
up to the relevant standard. However, this method has the disadvantage of taking a long time so that it is not conducive for the
efficient development of on-site work. Now there is a part of the existing research with the help of machine learning algorithms,
but the situation of data noise and data non-linearity is not fully considered in these methods. In response to the above problems,
a novel method for water quality prediction is put forward in this paper, which is based on the combination of the two-layer
decomposition method and the modified support vector machine (SVM) algorithm. Through the two-layer decomposition method
put forward above, the redundant noise in the prediction process can be eliminated effectively, and at the same time the major
features of the original data can be extracted. The method of stratified sampling is used to divide the original dataset so as to
avoid the sample deviation caused by the method of traditional random sampling. A modified particle swarm algorithm is applied
to optimize the parameters of the SVM so that the global convergence ability can be improved by this algorithm. On the basis of
the four cases of the Zhuangxi oil production plant joint station, the prediction accuracy of this method is evaluated in the light of
three evaluation indexes: the relative error, the average absolute percentage error and the determination coefficient. On the basis
of the average values of these 4 cases on the three indicators are -0.38
%
, 5.23
%
and 0.82
%
, respectively. Compared with the
existing mainstream machine learning algorithms, we can see that the method in this paper has higher prediction accuracy.
