An online coal calorific value prediction method through combustion radiation monitoring for industrial boilers was presented. Multiband combustion radiation signals in visible, infrared, and ultraviolet ranges were monitored. Multi-scale variables were extracted from the signals as flame radiation features in time and frequency domains. Principal component analysis (PCA) was used to eliminate information redundancy and noise disturbance. Correlation between obtained key principal components and the coal calorific value was established by training a support vector regression (SVR) model. Particle swarm optimization (PSO) and genetic analysis (GA) methods were used to search for the best SVR construction parameters. Performance test results showed that the optimized PCA+SVR model-based coal calorific value prediction results had a mean error of 110.9 kcal/kg relative to the lab analysis results, while the standard deviation (STD) was 151.9 kcal/kg. In order to reveal dynamic correlations among the multi-scale feature variables, dynamical principle component analysis (DPCA) was further used. Good consistence was obtained between the coal calorific values predicted by the DPCA+SVR model and the lab analysis results. The mean absolute error and the STD of the coal calorific values predicted by the DPCA+SVR model were diminished to 98.0 kcal/kg and 129.4 kcal/kg, respectively. The presented online coal calorific value monitoring system is highlighted by its low cost, easy installation, robustness to harsh application environment, and is expected to supply meaningful data for coal fired combustion process diagnostics and help realize more reliable regulation of the combustion process.