Answers are grounded in this paper and cite their sources.
Coal-fired power units increasingly operate under deep peak regulation, so accurate real-time prediction of flue-gas CO2 concentration is needed for combustion optimization and for feedforward control of post-combustion carbon capture. Hardware Continuous Emission Monitoring Systems suffer from measurement delay, while existing data-driven models are limited by two properties of furnace signals that are seldom addressed jointly: strong non-stationarity with distribution shift between lookback and prediction windows, and the superposition of a load-driven macroscopic trend with combustion-induced local fluctuations. This paper couples data analysis with model design in an end-to-end framework for multi-scale CO2 concentration forecasting. A multi-stage feature selection procedure combines Pearson correlation screening, gradient-boosted feature importance and Shapley additive explanations to obtain an interpretable, compact input set; the retained variables are then fed to a dual-branch network that models the macroscopic trend and the local fluctuations with structures matched to their physical origin, supported by interactive convolution and adaptive spectral filtering. On operational data from four units at three sampling resolutions, the model attains the best or second-best value in all 36 metric comparisons, the best value in 23 of them, and an inference latency of 5.28 ms, compatible with power-plant control system deployment.
No takes yet. Share an insight, caveat, or question.
Xia et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: