When using support vector regression to predict building energy consumption, since the energy influence factors are quite abundant and complex, the features associated with the statistical model could be in large quantity. This paper focuses in feature selection for the purpose of reducing model complexity without sacrificing performance. The optimal features are selected by their feasibility of obtaining and the evaluation of two filter methods. We test the selected subset on three datasets and train support vector regression with two different kernels: radial basis function and polynomial function. Extensive experiments show that the proposed method can select valid feature subset which guarantees the model accuracy and reduces the computational time.