NOx produced during biomass combustion in thermal power plants causes serious environmental pollution. Accurate prediction of NOx emissions is crucial to reducing environmental pollution. The NOx emission prediction model established based on traditional data-driven methods does not extract deep feature information sufficiently and has poor robustness. To address the existing problems, a hybrid prediction model for NOx emissions based on flame imaging, SADAE-MSViL, is proposed. First, a self-attention mechanism is introduced into the adversarial denoising autoencoder to extract deep features of the image and effectively remove noise interference. Secondly, a multi-scale feature fusion mechanism with a combination of scales of 8 and 16 is designed to fully capture the flame frequency domain information of image blocks at different scales. Finally, by improving Linformer and integrating the gated low-rank attention mechanism, the NOx emission prediction accuracy is improved while ensuring the operating efficiency of the model. Experimental results show that the R2 of the model reaches 0.98 and the RMSE is 3.0. The prediction accuracy is better than other models, showing high robustness and reliability.