A novel ensemble learning framework based on a dynamic pruning strategy for steel tube outer diameter prediction in the hot rolling industry

Abstract

Precise prediction of cold-state outer diameter (OD) is essential for dimensional quality control in hot-rolled seamless steel tube production, yet remains challenging under multi-grade and multi-specification operation where process regimes change frequently, and prediction errors can become heavy-tailed. This study proposes an interpretable ensemble learning framework for OD prediction based on a stacking architecture, in which an instance-wise dynamic pruning strategy is introduced to deactivate locally uncompetitive base learners during inference using out-of-fold error evidence. Evaluations under progressively reduced training scales demonstrate consistent generalisation gains; at a 90% training size, the pruned ensemble reduces the test root mean square error from 1.1705 mm to 0.9551 mm with a test mean absolute error of 0.7534 mm and a coefficient of determination of 0.9972, indicating effective suppression of extreme deviations. Model transparency is established via Shapley Additive Explanations and partial dependence plots, which reveal a dominant sizing-stage speed-schedule structure and stage-resolved interaction patterns that are operationally meaningful for process monitoring. Validation at plant scale conducted on 15,300 production samples encompassing nine steel grades and five OD specifications demonstrates a stable error envelope centred near zero. Across all grades, the proportion of predictions within a tolerance of ±4 mm consistently exceeds 98.4%.