According to the various components characteristics of the bridge deflection, the non-linear relationship of temperature and the temperature deflection effect was established. In order to improve the regressive ability to fit the temperature deflection effect, a multiple least square support vector machine (M-LS-SVM) regressive model was presented. The subtractive clustering was adopted to divide the input space into several sub-spaces, and sub-models were built by LS-SVM in every sub-space. In order to minimize the severe correlation among sub-models and to improve the accuracy and robustness of the model, the sub-models were combined by principal components regression (PCR). Experimental and analytical results show that the method can separate the temperature effect from deflection monitoring signals and provide the basis data for damage detection from the long-term monitoring signals.