采用神经网络与贝叶斯反演法自动优化 乙酸化学反应动力学模型

In view of the low efficiency and the uncertainty for accurate measurement in the combustion kinetic model developing process, an intelligent modeling method based on experimental data was proposed in the work. The Markov chain Monte Carlo and Artificial Neural Network (ANN) methods were used for accelerating the Bayesian approach for solutions; the combustion kinetic model for acetic acid was automatically optimized based on the measured laminar flame speed. The simulated results show that the uncertainty of ANN test sets can be reduced to 1 % by increasing the sample size and hidden layers, which meets the requirement of the surrogate kinetic model. ANN can produce a large number of samples and effectively reduce uncertainty and improve convergence efficiency during MCMC sampling. Being constrained by measured laminar flame speed, the posterior probability of the rate coefficient of reaction CH2C02 H+ H; CH2CO + H20 has changed a lot compared with the prior probability. The average rate coefficient increases with a factor of 10 and the uncertainty also decreases. The acetic acid laminar flame speed is the most sensitive when subject to the above reaction. In addition, the modified combustion kinetic model shows accurate prediction on the measured laminar flame speeds at different temperatures and much smaller uncertainty of the model being constrained. The research results provide reference to the combustion kinetic model in practical engineering application.