Synergy of feature-enhanced inputs and timexer for burn through point prediction in sintering processes

Abstract Accurate and timely prediction of burn through point (BTP) is essential for intelligent control in iron ore sintering. BTP is jointly affected by intrinsic temporal dynamics and external process variables, and their relative influence varies under different operating conditions. As a result, predictive performance often degrades when BTP changes rapidly or the process becomes strongly fluctuating. To address this issue, a statistical feature-enhanced TimeXer model, termed SF-TimeXer, is proposed. First, four domain-guided time-varying statistical features, namely difference, rolling mean, rolling standard deviation, and rolling slope, are constructed to characterize local change, trend, and volatility. These descriptors enrich the input information and are then compressed before entering the forecasting backbone, so that the backbone input dimensionality remains unchanged. Second, TimeXer distinguishes endogenous BTP dynamics from exogenous process-variable influences, allowing the model to better integrate intrinsic temporal evolution with external disturbances. Third, the feature-enhanced inputs are particularly compatible with TimeXer’s variate-wise embedding, which helps preserve trajectory-level information of each variable. Experiments on an industrial dataset show that statistical feature enhancement consistently improves multiple baselines, with the largest gain observed for TimeXer. On the test set, SF-TimeXer achieves a hit rate above 90% within a ±0.25 tolerance and an MSE of 0.0290, corresponding to a 47.0% reduction in MSE relative to the best baseline. These results indicate that combining domain-guided statistical descriptors with endogenous-exogenous attention is effective for BTP prediction in sintering processes.