Water-wall tube temperature prediction model for supercritical coal-fired boiler based on improved stochastic configuration network

Over-temperature of the boiler water-wall poses significant operational safety risks, particularly during deep peak shaving, making accurate prediction of the water-wall temperature crucial for boiler combustion regulation. The water-wall temperature prediction model using the hiking optimization algorithm-based stochastic configuration network (HOA-SCN) was proposed in this study. To create high-quality input features, Light gradient boosting machine (LightGBM) was utilized to evaluate feature variables with correlation. Considering the delay characteristics of boiler working medium, the delay time of the characteristic variables was estimated. To improve the efficiency of traditional SCN searching for weights and biases, to alleviate the ill-posed problem, and to improve the generalization and robustness, the SCN is improved with HOA algorithm and L2 Norm Regularization. Simulation experiments were conducted by extracting actual operating data from an in-service 660 MW coal-fired boiler. The results showed that the improved HOA-SCN achieves significantly better prediction than the conventional models, such as BP and Long Short-Term Memory (LSTM). The training time of HOA-SCN is 3.927 s, compared to 169.464 s and 13.868 s for BP and LSTM. Compared to the baseline model, the RMSE for HOA-SCN was only 2.06 °C, while the RMSE for SCN and Greedy-SCN was 4.484 °C and 3.628 °C.