Artificial neural networks (ANNs) are effective tools in forecasting runoff in river because of their power capability in mapping in-output relations. However, the traditional ANNs based on back - propagation training algorithm need improvement because they have shortcomings in long training times and prone in falling into local optimum points. Therefore, 3 algorithms were used to train the ANNs-mean linear particle swarm optimization (ML-PSO) method, original particle swarm optimization (PSO) method and BP method. Their global optimization capabilities were first tested by using the 3 standard mathematical functions, and the ANNs based on the 3 training algorithms were applied in runoff forecasting to test their performances. The results show that among the 3 algorithms, the ML-PSO algorithm is the fastest and most robust one in finding global optimum, and it also is the most accurate one in forecasting runoff.