River flow results from the interplay of numerous variables for which
quantitative information is not easily available. The aim of this pape
r was to develop a river-flow forecasting model based only on rainfall
and runoff information. The proposed model has been implemented with
the combined utilization of two methods: a neural network method, whic
h takes into account the non-linearity of the rainfall-runoff relation
ship and an adaptative technique, the Kalman filter, allowing real tim
e correction of estimates. The rainfall-runoff relationship has been m
odelled on two rivers in northern France. The weekly and daily time st
ep models gave satisfying forecasts, even for different lead times.