Generalization performance is one of the most important problems of intelligent approaches for parameters forecasting. This paper presents an ensemble of support vector regression (ESVR) which has better generalization performance than other intelligent approaches. To increase the diversity among individuals of ensemble, the paper proposes an individual generating approach based on clustering technique. Firstly, ESVR is used to classify training samples as several subclasses that are used to train different individuals with different kernel functions. The ensemble weights of individuals are determined by the generalization errors on the validation sets. ESVR is tested on the parameter drift data of gyroscope. By comparing single SVR, single neural network, neural network ensemble and combination approach with ESVR in generalization performance, the results reveal that ESVR has better generalization performance than other intelligent approaches in most cases.