Intelligent Predictive Models for Water resources Engineering

Abstract

Fresh water is considered to be one of the most important resources for humans and the environment. Due to the increase in population and the currently unsustainable usage of this limited resource, more attention is needed in the management of water resources. Advanced computational methods can help in attaining a better understanding of all aspects of water. Indeed, a better understanding of water resources requires a vast knowledge of a wide variety of fields such as atmospheric science, geology, hydrology, hydraulics and mathematics etc. To assist in this process computing techniques have been widely applied in water resources engineering problems. An artificial neural network (ANN) has been applied to solve many engineering problems since the 1980s. However, there are still many engineering fields that have the potential to benefit from ANN, such as water resource engineering. In the present research two important applications; time-series prediction and function estimation for water resource engineering are investigated. Within water engineering the prediction of river discharge is important. The results can be used for many purposes including flooding management, risk assessment and saving lives. New techniques are always being sought to improve the accuracy of predictions. In the first part of this research a neural network model was developed as a tool for time-series prediction to forecast water flow discharge of Fitzroy River near Rockhampton in central Queensland. A feed-forward back-propagation network was selected to predict the daily time-series of the Fitzroy Rivers’ discharge at The Gap station, Queensland. The data was derived from the Queensland Department of Natural Resources and Mines. The two developed ANN models are investigated and compared after many trials with a number of inputs, outputs, hidden layers, learning rate and transfer functions. The final model uses the flow data for 15 successive days and then predicts the discharge for the next 4 days. The results show that an accurate prediction was obtained during flood events. The advantage of the ANN model, when compared to other numerical models, is that it only uses the historical data of the discharge from this particular river. Thus it is free of the need for other data such as rainfall data, topography of the area and stream sections. In addition, after the ANN was trained, a very fast prediction was obtained. Consequently, this model can be used as a real-time tool for flow forecasting in the Fitzroy River. Similar models could be developed, based on the structure of this ANN model, for any river in Australia and in the world. Another interesting problem in water resource engineering is groundwater dynamics that occur near the coast. Indeed, a knowledge of groundwater dynamics in coastal aquifers is important for understanding sediment transport processes in the swash zone; shoreline stability; the design of coastal structures close to beaches; water quality in closed coastal lakes and lagoons; the operation of dune sewage disposal and domestic water supply. Analytical methods or numerical models have been used to predict this groundwater table fluctuation due to tides, waves and precipitation etc. In the present study ANN is adopted to simulate groundwater table fluctuations. In the study a multilayer feed-forward neural network model has been developed and trained using a back-propagation algorithm. The training data was based on field measurements (KANG et al., 1994a) from five different locations down the east coast of Australia. The data included information on watertable, tide elevation, beach slopes and hydraulic conductivity at each beach. The results from the developed model show that the artificial neural network model is very successful in terms of the prediction of a target that is dependent on a number of variables. Sensitivity analysis was undertaken which confirmed that a variation in tide elevation is the most important parameter to use for simulating groundwater flow in coastal aquifers. In contrast the low number of training data available for hydraulic conductivity and beach slope did not have a significant effect on the prediction of groundwater table fluctuations in this model. Thus, to improve the accuracy of prediction for the developed model, more data should be collected.