Image-based deep neural network prediction of the heat output of a step-grate biomass boiler

This work investigates the usage of deep neural networks for predicting the thermal output of a 3 MW, grate-fired biomass boiler, based on routinely measured operating parameters and real-time flame imaging. It is hypothesized that flame imaging can provide information regarding the quasi-instantaneous state of combustion, therefore supplementing conventional measurements that generally produce lagging feedback. A deep neural network-based, continuous multistep-ahead prediction scheme was proposed and evaluated by using operational and image data collected through extensive campaigns. It was found that flame imaging increases the accuracy of predictions compared to those obtained by only using operational data. The complexity of biomass combustion was well captured by the proposed deep neural network; furthermore, the deep architecture produced better predictions than shallower ones. The proposed system can reliably predict output water temperatures with errors up to ±1 °C, up to approximately 30 min ahead of the current time.