Boiler combustion system is a typical dynamic system with many variables, strong coupling, large-lag, and multiple input/output. It is very difficult to build a combustion system model that conforms to the actual working conditions. This paper presents a new modeling method of boiler combustion system based on bidirectional threshold cycle unit (Bi-GRU), and establishes the training model of combustion system under variable load (low, medium and high load)conditions. At the same time, gradient lifting decision tree (GBDT) is used to reduce the dimension of input characteristic matrix. GBDT model can evaluate the weight of input features under different loads and outputs, and can identify the feature with the largest weight proportion on the basis of retaining the original physical meaning of the feature. The feature selection model based on GBDT can not only reduce the original input dimension, but also provide theoretical guidance for the subsequent combustion control strategy. The calculation results of actual operation data show that the new combustion system model established by Bi-GRU and GBDT can accurately reflect the dynamic changes of main steam flow, main steam pressure and NOx emission under different loads. Compared with the traditional recurrent neural network (RNN) model, the accuracy and performance of the new model in this paper are significantly improved, and the structure is simple and the amount of calculation is small.