Structural optimisation and input selection of an artificial neural network for river level prediction

SummaryAccurate river level prediction, necessary for reliable flood forecasting, is a difficult task due to the complexity and inherent nonlinearity of the catchment hydrological system. Although artificial neural networks (ANNs) offer advantages over mechanistic or conceptual hydrological models for river level prediction, their applicability is limited by the fact that each ANN has to be specifically optimised and trained for a particular prediction problem and suitable input vectors selected. A recently developed novel optimisation algorithm combining properties of simulated annealing and tabu search is used to arrive at an optimal ANN for the prediction of river levels 5 h in advance. The algorithm seeks to minimise the value of a cost function based on the complexity and performance of the ANN. This is done by removing inter-neuron connections and adjusting the weights of the remaining connections. The candidate inputs presented to the algorithm were: current values of river levels at the flood location and two upstream locations; the change in level over the previous 4 h at the flood point, mean sea level pressure (SLP) and the change in SLP over the previous 24 h. The optimisation removed 79% of the network connections and three of the candidate inputs, leaving the current levels at the two upstream locations and at the flood point as the only inputs. The optimised ANN was then trained using the standard backpropagation algorithm. This methodology produces an ANN of greatly reduced complexity albeit with a reduced performance compared to an unoptimised ANN trained with backpropagation only. However, it has the advantage of being generally applicable and represents an improvement over trial and error as a method of ANN structural optimisation and input selection. For this prediction problem, current levels at two upstream locations and at the flood point are the best predictors of the level at the flood point.