摘要
Long-term monthly flow forecasts are essential for decision making in a river basin system. Many studies have already been reported on monthly as well as seasonal forecast using artificial neural networks (ANNs). This study demonstrates that monthly forecasts can be significantly improved if the input variables in ANN models are chosen with due consideration, even if number of training patterns are less. Monthly forecast models up to 12-month lead-time have been developed for Mississippi River in USA. It is seen that direct forecast with only antecedent flows as inputs does not improve the result. It is better to develop individual models for each month separately with information from previous years for the same month. Further, the forecast is found to significantly improve if the difference in predicted and actual flows is also included as one of the input variables (i.e. error updating), particularly where there is a clearly observed pattern in the historical information.