We constructed neural network models for nine selected Canadian rivers to see if we could predict monthly surface runoff. We used as inputs current and previous monthly climate variables, such as mean, minimum and maximum temperatures, total rain and snow amounts. Our predicted output was monthly runoff for the basins. The data selected to develop the models were collected between 1980 and 1994. Model predictions made using the training data set showed that the model had good predictive capacity in seven of nine cases. However, in order to better test the models, we also used the models on three years of input data collected in the period between 1939 and 1978. We found the neural network models to provide reasonable predictions for six of the nine sites. Basin disturbances and data sets unrepresentative of basin conditions were the main reasons for model failure.
展开▼
