Thermal power plants mainly use zirconia sensors and magnetic oxygen sensors to measure the oxygen content in the flue gas. Due to the heavy dust in the measurement environment, and corrosive media such as sulfide, etc. the measurement ring chamber is prone to blockage and the thermal element corrosion. Therefore it has poor stability, large measurement errors, and is prone to failure. Aiming at this situation, a measurement model based on genetic algorithm and neural network was proposed. According to the existing measurement points and mechanism analysis of the power plant, the auxiliary variables of the model were preliminarily selected, and the data was preprocessed before modeling. The Laida rule was used to remove the gross errors and the five-point three-time smoothing filter was used to remove the noise. Partial least squares was used for principal component analysis. The genetic algorithm was finally used to optimize the weight threshold of the neural network. A feedback neural network model based on the genetic algorithm to optimize the initial weight and threshold was constructed. The results show that the neural network based on genetic algorithm optimization has high accuracy in predicting the oxygen content of flue gas, and the convergence speed is fast.