For the CO concentration in the decomposition furnace, a CNN-LSTM-Attention based CO prediction model is established. The decomposition furnace outlet temperature, coal feeding amount, raw material feeding amount, oxygen concentration, and waste flow rate are used as the characteristic auxiliary variables of the prediction model, and these variables are preprocessed and time series matched, and trained to obtain the CO prediction concentration through model processing. For the traditional predictive control of the outlet temperature of the GPC decomposer, a multi-operating condition optimization control strategy for the decomposer considering the CO concentration prediction model was investigated. The input of the model is the amount of coal fed and the output is the decomposer outlet temperature, the prediction model of the decomposer outlet temperature is described by the ARMAX model, and the control method adopts the GPC algorithm, but the CO concentration working condition switch is added on the basis of which there are two different control methods for the CO concentration in the normal and abnormal conditions. The simulation results show that the CNN-LSTM-Attention model predicts better and is better than the models LSTM, CNN-LSTM and LSTM-Attention, and the decomposer outlet temperature model considering CO concentration also has a better control effect.