Adaptive weighted XGBoost optimization with multifeature fusion for industrial data prediction

Abstract

Traditional machine learning-based data prediction is core to industrial intelligent operation, equipment monitoring and energy scheduling. With the widespread application of sensors and industrial Internet, industrial scenarios continuously generate massive high-dimensional, noisy and multicollinear data, raising higher demands for prediction model accuracy and generalization[1]. As an efficient ensemble learning algorithm, XGBoost is widely used in regression prediction due to its superior nonlinear fitting, robustness and computational efficiency[2]. However, traditional XGBoost has apparent limitations in industrial prediction. Industrial high-dimensional data contains numerous irrelevant and collinear redundant features that interfere with model training and degrade prediction performance[3]. Besides, its fixed feature weight allocation and static learning rate fail to adapt to complex data distribution changes, causing slow convergence and local optimum problems[4]. To solve these problems, this paper proposes an integrated optimization pipeline for XGBoost, A dual-criterion screening strategy combining Pearson correlation coefficient and VIF is constructed to remove invalid low-correlation and multicollinear features[5]. An adaptive weighting mechanism based on tree splitting gain is designed to dynamically adjust feature contributions[6]. Meanwhile, an exponential decay dynamic learning rate is adopted to accelerate early convergence and stabilize late optimization[7] . Comparative experiments, ablation experiments and noise robustness tests are conducted on public datasets. The results show that the proposed method reduces MAE by 12.3% and RMSE by 10.8%, with the R² reaching 0.976, outperforming SVR, Random Forest, LightGBM and original XGBoost. Ablation experiments validate the effectiveness of each improved module, and robustness tests confirm its strong anti-noise capability and generalization[8] . This method can provide reliable technical support for industrial time-series prediction and intelligent early warning