A Multi-Condition Pulverized-Coal Flame Stability Evaluation Method Based on an Image-Level Adversarial Autoencoder and Adaptive Updating

To support carbon-neutrality targets, high penetration of wind and photovoltaic power has increased peak-shaving requirements for coal-fired units, leading to frequent boiler load variations and combustion instability. An adaptive framework for pulverized-coal flame stability evaluation under multiple operating conditions was developed based on a self-built 10 kW combustion test rig. First, an AE-Res-CBAM-SSIM-GAN architecture was proposed for latent representation learning and flame image reconstruction. In this architecture, autoencoder (AE) reconstruction was enhanced by residual feature propagation, convolutional block attention module (CBAM)-based attention, structural similarity index measure (SSIM) constraints, and generative adversarial network (GAN)-based adversarial distribution constraints. Compared with the original AE, the average mean squared error (MSE) was reduced from 6.26 to 3.60, while the peak signal-to-noise ratio (PSNR) and SSIM were increased from 40.48 dB to 42.72 dB and from 0.879 to 0.959, respectively. Second, an adaptive updating strategy was established in the latent feature space. During single-condition incremental updating, the silhouette coefficient increased from 0.4034 to 0.5841, and the number of noise points decreased to 0, indicating that a stable clustering structure was established for newly introduced conditions. Finally, a static–dynamic fused stability factor was constructed by integrating flue gas parameters, spectral features, and dynamic flame image features. The comprehensive stability factor increased from 0.00 to 0.89 with increasing primary air ratio, and combustion stability under different O2, N2, and CO2 atmospheres was effectively differentiated. The framework enables interpretable and adaptive assessment of pulverized-coal flame stability.