关键词: 冷轧带材, 板形预报, 时间滞后补偿, 孤立森林算法, 极端梯度提升算法

Abstract: An ensemble learning method was employed to develop a high-accuracy, strong-generalization cold rolled strip flatness prediction approach. Firstly, a model training dataset was constructed based on industrial big data, which had large data scale and high degree of flatness diversity. The time lag compensation was employed to eliminate time asynchrony among data based on the relative position between rolling mill and flatness meter. The isolation forest algorithm of data mining technology was used to clean outliers in data, improving the quality of training data and model performance. Subsequently, an architecture of multi-channel flatness parallel prediction was constructed based on XGBoost algorithm. This architecture was trained using processed production dataset to obtain the cold rolled strip flatness prediction model(CCFD_M). Lastly, based on CCFD_M, a flatness channel optimization algorithm was proposed to eliminate the issue of “pseudo-flatness” in prediction results, and the practical version of cold rolled strip flatness prediction model(CCFD_OM) was obtained. After verification on test set, the prediction error indictors mean absolute error(MAE) and root mean square error(RMSE) of model CCFD_OM reach 0.4044I and 0.6816I, respectively. And the fitting performance indictor R2 reaches 0.83, which may meet the practical production requirements.

Key words: cold rolled strip, flatness prediction, time lag compensation, isolation forest algorithm, extreme gradient boosting（XGBoost） algorithm