The combustion system of power plant boilers is a typical multi-variable dynamic system

which is strongly coupled

large-lag

and with multiple input/output signals. Therefore

it is difficult to construct a model of the combustion system that is close to the reality. A new modeling method for the boiler combustion sytem is proposed based on bidirectional gated recurrent unit(Bi-GRU)

and the training models for the combsution system under varying load conditions(low load

medium load and high load) are established. Meanwhile

the gradient boosting decision tree(GBDT) is adopted to reduce the input features matrix dimensions. The GBDT model can evaluate the weight of the inputs feature at different outputs and loads

and identify the most informative features on the basis of preserving the original physical meaning. The feature selection model by GBDT can not only reduce the original input dimension but also contribute to theoretical guidance for the subsequent combustion control strategy. The actual operation data calculation results show that

the new combustion system model established by Bi-GRU and GBDT can accurately reflect the output(main steam flow

main steam pressure and NO_x emission)dynamic change with different loads. Compared with the conventional recirculating neural network(RNN) model

the new model proposed above has significant higher accuracy and performance

and simpler structure and less computation time.