Using Artificial Neural Networks to Fill-in Missing Annual Peak Flows

The objectives of this project were to: 1) develop Artificial Neural Networks (ANNs) models to fill-in missing data from the peak annual flowrate records for the Santa Clara River watershed, and 2) compare the ANN results with results from linear regression. The purpose of this work was to provide estimates of the missing data so that a frequency analysis would be more accurate in predicting various flood events. The Santa Clara River Watershed in Southern California has a total drainage area of approximately 1,630 square miles. Peak annual flow data was available for nine stations throughout the watershed since 1933. Eight of the nine stations had some missing data. Each station was modeled and inputs to the ANN model consisted of: peak flow from nearby stations, precipitation data, and temporal data. Inputs were studied to determine their importance to modeling success and either kept or discarded. Sensitivity analysis was conducted to help determine the input parameters: one to four neighboring station peak flows, 10-day precipitation data, and the year. While it is easy to understand the explicit relationship between missing streamflow and neighboring station peak flows as well as precipitation data, the relationship between missing streamflow and the year is implicit. This might reflect the changes of watershed characteristics (landuse, regulated dams, etc.) over time and possibly the long-term change of weather patterns. Model characteristics (number of nodes and layers, transfer functions, data pre-processing methods, number of epochs, etc.) were also studied to optimize the ability of the ANN to learn relationships between the inputs and the peak flow. In general: the models performed well with one to four neighboring station peak flows, 10-days precipitation data, and the year; and it was common for testing results to be within about 20% of the target. Linear regression using the same data sets was also performed. The ANN models had a sum of the squared error value 2 to 400 times less than the respective linear regression models.