Abstract:
In recent years, the demand for hydrogen in refineries has been greatly increased. Rationally allocating the balance between hydrogen supply and consumption, and efficiently utilizing current hydrogen resources have important theoretical and practical industrial benefits. Research has demonstrated that XGBoost model showed excellent performance in many fields, but it has not been applied to hydrogen network industrial engineering. In this paper, a dynamic multi-output prediction model of the hydrogen network in a real-world refinery based on XGBoost model is studied. The dynamic multi-output prediction is carried out using the dynamic data of two indexes of the minimum fresh hydrogen consumption and the minimum hydrogen surplus. The fresh hydrogen refers to high purity hydrogen, and the dynamic data is obtained by solving linear model in our patent. The performance of the model is evaluated, with the aim to have a better reflection on the actual situation of the prediction error and to measure the deviation between the predicted and the true values. The MAE (mean absolute error) and the RMSPE (root-mean-square percent error) are selected as evaluation criteria; The prediction results are compared with those of BP (back propagation) neural network model, and good prediction results have been obtained. Finally, the influences of five types of operational parameter features on the output indexes are analyzed. The five types of operational parameter features are characterized by four reactor temperatures and recycle hydrogen, respectively. The diagram of the feature importance scores on the two output indexes is obtained. Based on the analysis of the diagram, the features that have the highest impact on the two output indexes of the model are obtained.