Feature Selection With Missing Data For Metallurgical Coal Gas System Based On Mutual Information

The gas storage level of steel enterprises is affected by many production and consumption users in the gas system, which brings great difficulties to build the gas system model. By feature selection we can determine the main users affecting the gas storage level, get rid of the users with little or no effect on the storage level. It can greatly decrease the complexity of the gas system modeling. Furthermore, the change of the gas storage level can be accurately estimated, so it makes an important contribution to realize the optimized scheduling of the gas system.For the missing data and delay phenomenon generated in the real industrial data, this paper proposes a feature selection method based on mutual information for the gas system of steel enterprise. This method consists of two parts:1)A mutual information estimation by using k nearest neighbors is adopted to directly calculate the mutual information value between the production and consumption gas system users and the gas storage level with missing data. Because of the distribution of gas pipeline network, the change of units gas flow may be reflected in the gas holder level after time delay. Starting from the gas system actual situation, we put forward a feature selection with time delay.2) To evaluate the feature selection for the system model’s influence, the feature selection is evaluated by using least squares support vector machine (LSSVM) at the complete data set after missing data imputation.To verify the effectiveness of the proposed method in this paper, the gas system in a domestic steel enterprise is the research object. The actual data of the blast furnace gas system in this plant is employed for the two gas holder simulation experiments with different missing proportion. Compared with the feature selection after data imputation and none feature selection, the prediction accuracy of validation experiments shows the validity of this method.