Abstract: When dealing with softsensor modeling problem with lots of samples, the traditional least squares support vector machines (LS-SVM) algorithm has some shortcomings, such as, the complexity of model structure, the loss of sparseness, and the difficulty in selecting normalizing parameter and kernel parameter. In this paper, an improved algorithm is proposed to overcome the above drawbacks. The samples are firstly pre-processed and the similarities between samples are analyzed by computing their Euclidian distances. Moreover, one third of the original sample points is removed so that the SVM model structure is simplified and the computing speed is increased. A chaotic map is defined and the ergodicity of chaos is analyzed. Then, the proposed chaos optimization algorithm is used to optimize the LS-SVM model parameters so as to raise the fitting accuracy and enhance its generalization ability. Finally, the proposed method is applied to the soft sensor modeling of the acrylonitrile yield. It is shown from simulation results that this model has higher prediction precision and better generalization ability, and can satisfy the requirement of spot measurement.