An Exploration of Neural Network Modelling Options for the Upper River Ping, Thailand

This thesis reports results from a systematic experimental approach to evaluating aspects of the neural network modelling process to forecast river stage for a large, 23,600 km2 catchment in northern Thailand. The research is prompted by the absence of evidenced recommendations as to which of the array of input processes, validations and modelling procedures might be selected by a neural network forecaster.

The flood issue for forecasters at Chiang Mai derives from the monsoon rainfall, which leads to serious out-of-bank flooding two to four times a year. Data for stage and rainfall is limited as the instrumentation is sparse and the historical flood record is limited in length. Neural network forecasting models are potentially very powerful forecasters where the data are limited. The challenge of this catchment is to provide adequate forecasts from data for relatively few storm events using three stage gauges and one rain gauge. Previous studies have reported forecasts with lead times of up to 18 hours. Thus, one research driver is to extend this lead time to give more warning.

Eight input determination methods were systematically evaluated through thousands of model runs. The most successful method was found to be correlation and stepwise regression although the pattern was not consistent across all model runs.

Cloud radar imagery was available for a few storm events. Rainfall data from a network was not available so it was decided to explore the value of the raw cloud reflectivity data as a catchment-wide surrogate for rainfall, to enhance the data record and potentially improve the forecast. The limited number of events makes drawing conclusions difficult, but for one event the forecast lead time was extended to 24-30 hours. The modelling also indicates that for this catchment where the monsoon may come from the south west or the north east, the direction of storm travel is important, indicating that developing two neural network models may be more appropriate.

Internal model training and parameterisation tests suggest that future models should use Bayesian Regularization, and average across 50 runs. The number of hidden nodes should be less than the number input variables although for more complex problems, this was not necessarily the case. Ranges of normalisation made little difference. However, the minimum and maximum values used for normalisation appear to more important.

The strength of the conclusions to be drawn from this research was recognised from the start as being limited by the data, but the results suggest that neural networks are both helpful modelling processes and can provide valuable forecasts in catchments with extreme rainfall and limited hydrological data. The systematic investigation of the alternative input determination methods, algorithms and internal parameters has enabled guidance to be given on appropriate model structures.