Multi-model data fusion for hydrological forecasting

This paper outlines some simple data fusion strategies for continuous river level forecasting where data fusion is defined as the amalgamation of info rmation from different data sources. The objective of data fusion is to pro vide a better solution than could otherwise be achieved from the use of sin gle-source data alone. In this paper, the simplest data-in/data-out fusion architecture was used to combine neural network, fuzzy logic, statistical, and persistence forecasts using four different experimental strategies to p roduce a single predicted output. In the first two experiments, mean and me dian values were calculated from the individual forecasts and used as the f inal forecasts. These types of approaches can be effective when the individ ual model residuals follow a consistent pattern of over and under predictio n. In the other two experiments, amalgamation was performed with a neural n etwork, which provided a more flexible solution based on function approxima tion. The four individual model outputs were input to a one hidden layer, f eed-forward network that had been trained to produce a single final forecas t. The second network was similar to the first, except that differenced val ues were used as inputs and outputs. These various data fusion strategies w ere implemented using hydrological data for the River Ouse gauge at Skelton , above York, in Northern England. Neither the mean nor the median produced improved results, whereas the two neural network data fusion approaches pr oduced substantial gains with respect to their single solution components. The potential to obtain more accurate forecasts using data fusion methodolo gies could therefore have significant implications for the design and const ruction of automated flood forecasting and flood warning systems. (C) 2001 Elsevier Science Ltd. All rights reserved.