Using Artificial Neural Networks and Genetic Programming in rainfall/runoff modeling   [ANN] [NN] [GP]

by

Drécourt, J.-P. and Drecourt, J.-P.

Literature search on Evolutionary ComputationBBase ©1999-2013, Rasmus K. Ursem
     Home · Search · Adv. search · Authors · Login · Add entries   Webmaster
Note to authors: Please submit your bibliography and contact information - online papers are more frequently cited.

Info: 3rd DHI Software Conference \& DHI Software Courses (Conference proceedings), 1999
Keywords:genetic algorithms, genetic programming
Abstract:
The main problem in rainfall/runoff modeling is to obtain data about the catchment with sufficient accuracy. Since self-learning tools only need knowledge about rainfall and runoff, they can offer a good alternative to classical model. The present study focuses on Lindenborg, a Danish catchment situated in the northern part of Jutland, between Hobro and Ålborg. It is characterized by high groundwater contribution and thus a very persistent flow regime. The tools used were artificial neural networks (ANN) [ANN] [NN] and genetic programming (GP). [GP] The purpose was to compare the efficiency of these tools with a classic lumped model (NAM) and a naïve prediction (i.e. the runoff does not change between one day and the next one). The study with GP was oriented in two directions : the prediction of the runoff, and the prediction of the variation in the runoff. In both cases GP was given the rainfall and runoff of the past days, and it was assumed that the rainfall was predicted without any error for the target day. Each strategy has its own advantages. Predicting the variation is considered to be closer to the relationships given by physics, whereas predicting the runoff takes in account the large auto-correlation of the runoff time series. [TS] Since it is difficult to predict the upper boundary of runoff, the ANN worked exclusively with the time variation. The variation in runoff is less likely to saturate the network than the runoff itself, especially in this catchment where the dynamics are relatively slow. Therefore, the sensitivity of the prediction is increased. Time lag recurrent network (TLRN) were used for this study as they allow to take in account smoothed version of the past time series, [TS] both in the input and the hidden layers. The comparison of the different models was based on the Pearson coefficient of correlation, which gives a good overview of the performance of the prediction. This study was realized in relationship with the Department of Hydrodynamics and Water Resources of DTU as a special course for the Master of Science in Environmental Engineering.
Notes:
http://www.dhi.dk/softcon/index.htm
URL(s):(G)zipped postscript
MS word

Review item:

Mark as doublet (will be reviewed)

Print entry




BibTex:
@InProceedings{drecourt:1999uANNGPrrm,
  author =       "J-P. Drécourt",
  title =        "Using Artificial Neural Networks and Genetic
                 Programming in rainfall/runoff modeling",
  booktitle =    "3rd DHI Software Conference \& DHI Software Courses",
  year =         "1999",
  address =      "Helsingør, Denmark",
  month =        "7-11 " # jun,
  organisation = "Danish Hydraulic Institute",
  keywords =     "genetic algorithms, genetic programming",
  URL =          "http://www.dhi.dk/softcon/abstract/102.doc",
  abstract =     "The main problem in rainfall/runoff modeling is to
                 obtain data about the catchment with sufficient
                 accuracy. Since self-learning tools only need knowledge
                 about rainfall and runoff, they can offer a good
                 alternative to classical model. The present study
                 focuses on Lindenborg, a Danish catchment situated in
                 the northern part of Jutland, between Hobro and Ålborg.
                 It is characterized by high groundwater contribution
                 and thus a very persistent flow regime. The tools used
                 were artificial neural networks (ANN) and genetic
                 programming (GP). The purpose was to compare the
                 efficiency of these tools with a classic lumped model
                 (NAM) and a naïve prediction (i.e. the runoff does not
                 change between one day and the next one).

                 The study with GP was oriented in two directions : the
                 prediction of the runoff, and the prediction of the
                 variation in the runoff. In both cases GP was given the
                 rainfall and runoff of the past days, and it was
                 assumed that the rainfall was predicted without any
                 error for the target day. Each strategy has its own
                 advantages. Predicting the variation is considered to
                 be closer to the relationships given by physics,
                 whereas predicting the runoff takes in account the
                 large auto-correlation of the runoff time series.

                 Since it is difficult to predict the upper boundary of
                 runoff, the ANN worked exclusively with the time
                 variation. The variation in runoff is less likely to
                 saturate the network than the runoff itself, especially
                 in this catchment where the dynamics are relatively
                 slow. Therefore, the sensitivity of the prediction is
                 increased. Time lag recurrent network (TLRN) were used
                 for this study as they allow to take in account
                 smoothed version of the past time series, both in the
                 input and the hidden layers.

                 The comparison of the different models was based on the
                 Pearson coefficient of correlation, which gives a good
                 overview of the performance of the prediction.

                 This study was realized in relationship with the
                 Department of Hydrodynamics and Water Resources of DTU
                 as a special course for the Master of Science in
                 Environmental Engineering.",
  notes =        "http://www.dhi.dk/softcon/index.htm",
}