An Integrated Method of Data Pre-processing in Support Vector Regression for the Quality Prediction of Treated Gas-oil

The accuracy of the measured data is very important for better quality prediction by soft sensors. In order to determine the quality of the treated gas oil، a soft sensor is designed. The soft sensor design is based on a new machine learning technique called support vector regression (SVR). An integrated technique was developed for data preprocessing. In this technique، wavelet analysis and vector quantization were being used sequentially for random error elimination، data compression، and unusual data omitting. Different methods of wavelet analysis were used to remove the random errors and the best method was selected. Random errors were deleted using Harr and Daubechies basis function where Rigrsure، Minimaxi، Heursure، and Sqtwolog were the threshold algorithms. The results showed that the db4 basis function with Rigrsure threshold algorithms provided the best results for error removal. AARE and RMSE for this method was better than the other types of wavelet functions. Additionally، the results of SVR training based on the pilot plant data showed AARE of 0. 053. This showed the high accuracy of the SVR model for predicting treated gas oil quality.