In order to improve the prediction accuracy of effectiveness of weapon equipment system-of-systems which trains on highdimensional and noisy small samples, a hybrid model based on stacked denoising autoencoder (SDA) and support vector regression (SVR) is proposed. By taking the advantage of SDA, the method extracts common features autonomously on related but different domain data. This hybrid model is pretrained by using a large number of source domain data. Then the hybrid model is transferred as prior knowledge on target domain. By transferring these prior knowledge, the hybrid model is finetuned on highdimensional and noisy small target domain data, making up for the defects of traditional SVR. In a certain battle scenario, the model with the simulation data produced by simulation testbed is validated. Experimental results demonstrate the effectiveness of the proposed model.