In order to take the advantages of hydrological similarity and ensemble extreme learning machine (ELM) to improveflood forecasting accuracy, we proposed a flood forecasting method for small- and medium-sized rivers, which is ensemble extremelearning machine based on similarity matching (SM-ELM). Firstly, we found the optimal ELM for each training sample from manyELMs. Then, we found k training samples that are most similar with the testing sample from the training set. At last, we used theoptimal ELMs of the k training samples to predict testing sample and combine the model outputs through weighted average strategy.We took the historical floods in the Chuanghua River Basin as study case to prove the validity of the proposed method. Theresults indicate that the ensemble ELM significantly improve the flood forecasting accuracy as compared with the single ELM. Theaccuracy of ensemble ELM is about 10%-15% higher than that of the single ELM in terms of root mean square error. Among thethree ensemble methods, SM-ELM can achieve higher and more stable forecasting accuracy with less number of ELMs.