Long-horizon prediction of NOx emissions and boiler performance for operational optimization using a guided rollout hybrid forecasting network

Accurate long-horizon forecasting of NOx emissions and key boiler performance metrics may help improve operational control in coal-fired power plants. This study proposes a guided rollout hybrid forecasting network (GRHFNet) to predict the long-horizon behavior of NOx and O2 concentrations and the gas temperature (Tg,exit) at the boiler exit. A total of 84,360 samples collected at 1-min intervals over 69 days from a 500 MWe coal-fired boiler were preprocessed. The data were then categorized into three groups: operational variables, performance metrics, and target variables. GRHFNet consists of two components: a guidance model (Model G) and a rollout forecasting model (Model R). Model G generates a horizon-wide trajectory from performance-metric histories, while Model R performs autoregressive one-step-ahead predictions using operational inputs and target histories. Based on cross-correlation time-delay analysis, the input sequence length and forecasting horizon were set to 30 min and 20 min, respectively, to capture control-induced dynamic responses. Various deep learning architectures were evaluated to improve prediction accuracy. The final GRHFNet employs a gated recurrent unit (GRU)–Transformer. Over the 20-min horizon, average R2 values of 0.945, 0.946, and 0.993 were achieved for NOx, O2, and Tg,exit, respectively. GRHFNet was further coupled with grey wolf optimization (GWO) to optimize operational variables. NOx emissions were reduced while maintaining boiler performance, suggesting its potential as a practical optimization tool for large-scale thermal power systems.