Soft Sensor Model Of Oxygen Content In Bagasse Boiler Based On Neural Network And Support Vector Machine

Oxygen content is an important factor to ensure optimization control of the bagasse boiler combustion system. The ratio of wind and fuel of bagasse boiler combustion system can be adjusted timely and effectively through monitoring the oxygen content, which reduces heat loss and improves efficiency, so that the boiler combustion can be optimized.Currently, bagasse boiler system mainly uses thermal-magnetic oxygen sensor and zirconia oxygen sensor to measure the oxygen content of the flue gas. But these oxygen sensors are inaccuracy, expensive, short-lift, and the lag is large in the process of measure, which are not conducive to online real-time monitoring in the boiler combustion process.Aiming at these problems, this paper is based on the characteristics of oxygen content of bagasse boiler flue gas, the relationship of the various factors, the common soft-sensor model, and the basic knowledge of data processing. It adopts a common neural networks and support vector machine method to establish soft sensor modeling of oxygen content of bagasse boiler.At the first, the collected data is analyzed and preprocessed, then the soft sensor modeling is established based on BP neural network. But since the prediction error data is large, and the generalization ability is poor, the Elman neural network is used to improve the method. The method can improve the prediction accuracy effectively, and convergence easier. However, due to the problem of instability of neural networks and local minimum, the paper decide to use support vector machine regression (SVR) method for modeling. The method adopts the squared of training error to replace the slack variable, but the calculation is too large, which causes longer training time. In order to avoid solving quadratic programming problems, improving the training speed, least squares support vector machine (LS-SVR) is used. In addition, due to the LS-SVR choose penalty factor C and the Gaussian kernel σ, SVR robustness and relaxation will be lost, which causes negative impact on the model. Particle swarm optimization (PSO) is used to optimize penalty factor C and the Gaussian kernel σ, which can be called soft measurement model of PSO-LS-SVR.Finally, via comparing the prediction error of data which come from SMPT-1000 platform and a sugar refinery actual field, which show the PSO-LS-SVR soft measurement model was able to achieve a good prediction effect, and it also meet the industry requirement.