Improving neural network models of physical systems through dimensional analysis

Testing the applicability of neural networks as a gap-filling method using CH<sub>4</sub> flux data from high latitude wetlands

“…2, Table 2). This selection helped to prune the network by avoiding insignificant input data (Gunaratnam et al, 2003;Saxén and Pettersson, 2006) and to avoid cross-correlations (to some extent) between input variables, adding little extra information to the network. We have tested three approaches by using three different sets of input data: the most simple approach included solar radiation as a time of day indicator and the three seasonal fuzzy sets ("seasonal", Table 2), a second approach in which solar radiation has been removed and replaced by the time of day fuzzy sets ("diurnal", Table 2), while a third, a more thermo-hydrological approach is tested by integrating the lagged effect of precipitation and WTD which was applied to four out of six data sets ("lagged", Table 2).…”

“…Nevertheless, their application is time consuming, particularly in finding the appropriate input variables, the appropriate number of hidden layers, and neurons/nodes within these layers, as well as the choice of training and test data sets (data rows). Furthermore, the global minimum (Hammerstrom, 1993;Nguyen and Chan, 2004) is not unique and changes with each training run because every training run/repetition (a run includes many iterations) achieves different fitted weights and results (it is important to find a set of weights that processes data accurately enough for the application).…”

“…One of the drawbacks of neural networks is the nonuniqueness of the global minimum (Hammerstrom, 1993;Nguyen and Chan, 2004) which changes as each training run achieves different weights and results (it is important to find Fig. 2.…”

“…This can happen when data used to train the network is not representative enough for the entire observation span, if the number of hidden layers or neurons is not correct, if the global minimum is overshot or when the network learns the training pattern well but is underperforming during its validation (poor generalisation) (Jain et al, 1996;Gardner and Dorling, 1998;Nguyen and Chan, 2004;Wang et al, 2005;Saxén and Pettersson, 2006;Stathakis, 2009). To prevent this from happening, one can remove redundant input data (Gunaratnam et al, 2003;Saxén and Pettersson, 2006), reduce or increase the number of neurons in the network and use the appropriate generalisation such as early stopping (Hansen and Salamon, 1990;Amari et al, 1997;Svozil et al, 1997;Wang et al, 2005) or another stopping criteria (Günther and Fritsch, 2010;Fritsch and Günther, 2012).…”

Application of neural networks and fuzzy logic models to long-shore sediment transport

“…Chakraverty and Gupta [19] modeled the rainfall using artificial neural network (ANN) with different network configurations. More researches in water resources and hydraulics engineering based on NN, ANN, FL and ANFIS models are available in the literatures [20][21][22][23][24][25][26]. Among other research subjects in civil engineering affected by many parameters, which needs advanced methods for analyzing data, are the one that deals with sedimentation and local scour as well as ocean waves.…”