This paper discusses local composition soft sensors in a distillation column with regards to different operation modes, which are caused by effect of load disturbance such as feed composition change, wide operation regions, estimation of various product specification, etc. Linkage_based clustering method and PLS are used for the classification of various operation regions, modeling of local soft sensor respectively. The proposed method was illustrated using estimation of ethanol in a binary distillation column simulated in simulator, HYSYS^TM and toluene in an industrial splitter column within a wide operation window. In simulation, process data for sub-operation regions have been collected to reflect the changes of feed composition that is of frequent occurrence in real distillation column. Local PLS soft sensors were compared to global PLS, nonlinear PLS based on neural network and neural network. and shown better performance than other methods.