Preprocessing methodology for time series: An industrial world application case study.

• Refining crude oil is a highly complex industrial process, subject to a large number of variables. • We propose a novel preprocessing methodology for obtaining quality data and extracting information from the data involved in the crude oil refining process. • The methodology incorporates dynamic knowledge, treatment of noise, reduction of the dimensionality, feature selection and introduction of slopes. • The proposal is validated through optimization of three state-of-the-art regressors: GB, RF and SVR. • This methodology along with SVR offer useful information for the expert of the refining process. This paper proposes a novel preprocessing methodology, framed within the field of time series forecasting. The aim is to get quality data and to extract information on the most important variables involved in a real-world crude oil refining process. To achieve this objective, the methodology incorporates the addition of dynamic knowledge, treatment of the noise present in the data, reduction of the dimensionality, feature selection and the introduction of slopes in the variables. Predictions are made for each step of the methodology and evaluated based on four measures: MAE, MSE, SMAPE and the delay in the prediction as compared to the original variable. The final solution is chosen based on these four measures and delivered to the experts so that they can optimize the crude oil refining process. [ABSTRACT FROM AUTHOR]