Short-term wind speed prediction is an indispensable part of the operation and control of wind energy power generation systems. However, many prediction models proposed by researchers did not preprocess the original data or consider the limitations of a single prediction model, resulting in poor prediction accuracy. Based on the no-negative constraint theory, this study uses five neural networks with advanced optimization algorithms and data preprocessing to obtain high-precision prediction results. Four experiments were designed to test the effectiveness of the proposed model and four analysis strategies were used to discuss the experimental results. The empirical study used wind speed data from China. The results show that the MAPE and Std performance indicators in the multi-step prediction of the hybrid model are lower than in other benchmark models; the proposed model is far superior to comparable models in terms of accuracy and stability.