Development of a Neural Network Model for Dissolved Oxygen in the Tualatin River, Oregon

Dissolved oxygen concentrations in the lower reaches of the Tualatin River in northwest Oregon are the result of many processes. Temperature imposes a seasonal signal through the solubility of oxygen in water. Streamflow determines the travel time through the system and affects the amount of oxygen consumed via processes such as ammonia nitrification and the decomposition of organic material in the sediment and water column. Streamflow also affects the rate of oxygen exchange across the air/water interface. The available solar energy limits the photosynthetic production of oxygen by phytoplankton. Many of the processes that affect dissolved oxygen concentrations in the Tualatin River – solubility, sediment oxygen demand, photosynthesis, respiration, biochemical oxygen demand, and reaeration – are controlled to some extent by physical and meteorological factors such as streamflow, air temperature, and solar radiation. To test the extent of that control, an artificial neural network model was constructed to predict dissolved oxygen concentrations in the Tualatin River at the Oswego Dam using only air temperature, solar radiation, rainfall, and streamflow as inputs. The Oswego Dam is a low-head structure located on a bedrock sill 5.5 kilometers upstream from the river's mouth. Hourly dissolved oxygen concentrations have been collected there since 1991; the available dataset spans more than 10 years. Feedforward neural network modeling techniques, the most widely used type, were applied to this dataset. Data were segregated into calibration, verification, and test subsets. Two neural network models were constructed in series: the first model simulated daily mean dissolved oxygen concentrations, while the second superimposed any daily periodic signals. The final calibrated neural network models predicted the dissolved oxygen concent ration with acceptable accuracy, producing high correlations between measured and predicted values (correlation coefficient of 0.83, mean absolute error less than 0.9 milligrams per liter). By some measures, neural network model performance was better than that of a calibrated, mechanistic model of dissolved oxygen in the Tualatin River. As expected, however, dissolved oxygen concentrations affected by factors other than the physical and meteorological factors used as model inputs, such as large point-source ammonia releases, were not predicted well by the neural network model. Nevertheless, the neural network model demonstrated potential for use as a river management and forecasting tool to predict the effects of flow augmentation and near-term weather conditions on Tualatin River dissolved oxygen concentrations.