Comparison of three updating schemes using artificial neural network in flow forecasting

Abstract

Three updating schemes using artificial neural network (ANN) in flow forecasting are compared in terms of model efficiency. The first is the ANN model in the simulation mode plus an autoregressive (AR) model. For the ANN model in the simulation model, the input includes the observed rainfall and the previously estimated discharges, while the AR model is used to forecast the flow simulation errors of the ANN model. The second one is the ANN model in the updating mode, i.e. the ANN model uses the observed discharge directly together with the observed rainfall as the input. In this scheme, the weights of the ANN model are obtained by optimisation and then kept fixed in the procedure of flow forecasting. The third one is also the ANN model in the updating mode; however, the weights of the ANN model are no longer fixed but updated at each time step by the backpropagation method using the latest forecast error of the ANN model. These three updating schemes are tested for flow forecasting on ten catchments and it is found that the third updating scheme is more effective than the other two in terms of their efficiency in flow forecasting. Moreover, compared to the first updating scheme, the third scheme is more parsimonious in terms of the number of parameters, since the latter does not need any additional correction model. In conclusion, this paper recommends the ANN model with the backpropagation method, which updates the weights of ANN at each time step according to the latest forecast error, for use in real-time flow forecasting.

Real-time flood forecasting by employing artificial neural network based model with zoning matching approach

“…The source for data training can be based on the same data type as the forecasting or it can come from several data types. Data types that are used as the source for data training are river flow (El-Shafie et al, 2008;Thirumalaiah and Deo, 1998;Shrestha et al, 2005), rainfall (Toth et al, 2000;Dawson et al, 2001), water level and rainfall (Alvisi et al, 2006;Campolo et al, 1999;Lihua et al, 2004;Boucher et al, 2010), water level and sea level pressure (Leahy et al, 2008), and flow, rainfall, temperature and snowmelt (Coulibaly et al, 2000).…”

A hybrid modelling approach for prediction of UCS of rock materials

“…As can be seen from the above, the ANN and ANFIS models outperform the existing empirical models; however, they may encounter some issues such as trapping in local minima [19][20][21]. Besides the ANN and the ANFIS, at present, the least-squares support vector machine (LSSVM) is one of the widely used AI techniques.…”

A robust recurrent ANFIS for modeling multi-step-ahead flood forecast of Three Gorges Reservoir in the Yangtze River

“…Data-driven techniques like artificial neural networks (ANNs) and fuzzy inference systems (FIS) have been widely applied with success to modeling runoff based on rainfall data in operational hydrology (Darras et al, 2015;Chen et al, 2014;Chen et al, 2013;Yang et al, 2013;De 2013;Mount et al, 2013;Li et al, 2009;Xiong et al, 2004). The FIS is 20 capable of coping with imprecision and uncertainty while ANNs have adaptive learning capabilities of being identified input-output patterns (Wandera et al, 2017;Khan 2016;Sun et al, 2016;Baghdadi et al, 2012;Gunnink et al, 2012;Lohani et al, 2011).…”