A real-time steam flow prediction framework based on two-stage feature selection and adaptive matching

Accurate short-term steam-flow forecasting is critical for stable operation of municipal solid-waste incineration (MSWI) plants, yet it is challenged by strong nonlinearity and rapid dynamics from combustion and equipment adjustments. This study contributes by RTF, a real-time online forecasting framework that combines a two-stage feature selection pipeline with an adaptive model matching mechanism. First, a filter step (Spearman) identifies candidate predictors and removes highly collinear variables; second, an embedded linear selection on reconstructed features selects the most informative historical input. A model library of offline Scaleformer–Autoformer (AS) models and lightweight online Ridge regressors is then exploited via adaptive time matching to select the best predictor for each local segment. On real industrial data from an MSWI plant, RTF achieves substantial improvements while remaining computationally efficient: training time is reduced from 8.6096 s to 0.1023 s after feature selection. Across 1–5 min horizons, RTF achieves MSE/MAE values from 0.0100/0.0769 (1 min) to 0.0499/0.1749 (5 min), outperforming Autoformer by at least 9.6% in MSE. These results indicate that RTF balances accuracy, responsiveness, and resource use, making it suitable for real-time monitoring and control in complex waste-to-energy systems.