File Download
Supplementary

Citations:
 Scopus: 0
 Appears in Collections:
Conference Paper: Rainfallrunoff modelling using genetic programming
Title  Rainfallrunoff modelling using genetic programming 

Authors  
Keywords  Datadriven models Evolutionary algorithms Genetic programming Rainfallrunoff modelling 
Issue Date  2005 
Publisher  Modelling and Simulation Society of Australia & New Zealand. 
Citation  The 2005 International Congress on Modelling and Simulation (MODSIM05), Melbourne, VIC., Australia, 1215 December 2005. In Conference Proceedings, 2005, p. 18411847 How to Cite? 
Abstract  The problem of accurately determining river flows from rainfall, evaporation and other factors, occupies an important place in hydrology. The rainfallrunoff process is believed to be highly nonlinear, time varying, spatially distributed and not easily described by simple models. Practitioners in water resources have embraced datadriven modelling approaches enthusiastically, as they are perceived to overcome some of the difficulties associated with physicsbased approaches. Such approaches have proved to be an effective and efficient way to model the rainfallrunoff process in situations where enough data on physical characteristics of catchment is not available or when it is essential to predict the flow in the shortest possible time to enable sufficient time for notification and evacuation procedures. In the recent past, an evolutionary based datadriven modelling approach, genetic programming (GP) has been used for rainfallrunoff modelling. In this study, GP has been applied for predicting the runoff from three catchments  a small steepsloped catchment in Hong Kong (Hok Tau catchment) and two relatively bigger catchments located in the southern part of China (Shanqiao and Shuntian catchments). For the runoff predictions in Hok Tau catchment, the performance of the datadriven technique was not very satisfactory. This catchment, being a very steepsloped catchment, has high peak discharge magnitudes with steep rising and recession limbs, which the GP models are unable to capture. This catchment being a small one with an area of about 5 km2 has a time of concentration of about 3045 minutes, but the time interval of the available data is one day, which seems to be another reason for GP's inability to capture the complex rainfall to runoff transformation on this catchment. Using a dataset of smaller time interval, the datadriven model should perform better. A key advantage of GP as compared to traditional modelling approaches is that it does not assume any a priori functional form of the solution. For instance, in a typical regression method, the model structure is specified in advance (which is in general difficult to do) and the model coefficients are determined. For neural networks, the time consuming task of initially defining the network structure has to be undertaken and then the coefficients (weights) are found by the learning algorithm. On the other hand, in GP, the building blocks (the input and target variables and the function set) are defined initially, and the learning method subsequently finds both the optimal structure of the model and its coefficients. Moreover, since GP evolves an equation or formula relating the input and output variables, a major advantage of the GP approach is its automatic ability to select input variables that contribute beneficially to the model and disregard those that do not. GP can thus reduce substantially the dimensionality of the input variables. In GP, as in any datadriven prediction model, the selection of appropriate model inputs is extremely important. This is especially so when lagged input variables are also used. Inclusion of irrelevant inputs leads to poor model accuracy and creation of complex models, which are more difficult to interpret as compared to simpler ones. Thus, for the remaining two catchments, an attempt is made to use the evolutionary search capabilities of GP for selecting the significant input variables. These variables, indicated as significant by GP are then used as inputs for the actual predictions. In contrast to the not so satisfactory performance by the GP models for predicting the runoff from Hok Tau catchment, their performance for the other two catchments is quite satisfactory, as the GP models are able to capture the peaks quite well and the goodnessoffit measures are also acceptable. These results indicate that GP can be used as a viable alternative for rainfallrunoff modelling, and the analytical form of the evolved equations facilitate easy interpretation. In this study, the GP evolved models are used for selection of significant variables influencing the rainfall to runoff transformation. 
Description  Conference Theme: Advances and Applications for Management and Decision Making 
Persistent Identifier  http://hdl.handle.net/10722/238593 
ISBN 
DC Field  Value  Language 

dc.contributor.author  Jayawardena, AW   
dc.contributor.author  Muttil, N   
dc.contributor.author  Fernando, TMKG   
dc.date.accessioned  20170220T01:01:51Z   
dc.date.available  20170220T01:01:51Z   
dc.date.issued  2005   
dc.identifier.citation  The 2005 International Congress on Modelling and Simulation (MODSIM05), Melbourne, VIC., Australia, 1215 December 2005. In Conference Proceedings, 2005, p. 18411847   
dc.identifier.isbn  9780975840009   
dc.identifier.uri  http://hdl.handle.net/10722/238593   
dc.description  Conference Theme: Advances and Applications for Management and Decision Making   
dc.description.abstract  The problem of accurately determining river flows from rainfall, evaporation and other factors, occupies an important place in hydrology. The rainfallrunoff process is believed to be highly nonlinear, time varying, spatially distributed and not easily described by simple models. Practitioners in water resources have embraced datadriven modelling approaches enthusiastically, as they are perceived to overcome some of the difficulties associated with physicsbased approaches. Such approaches have proved to be an effective and efficient way to model the rainfallrunoff process in situations where enough data on physical characteristics of catchment is not available or when it is essential to predict the flow in the shortest possible time to enable sufficient time for notification and evacuation procedures. In the recent past, an evolutionary based datadriven modelling approach, genetic programming (GP) has been used for rainfallrunoff modelling. In this study, GP has been applied for predicting the runoff from three catchments  a small steepsloped catchment in Hong Kong (Hok Tau catchment) and two relatively bigger catchments located in the southern part of China (Shanqiao and Shuntian catchments). For the runoff predictions in Hok Tau catchment, the performance of the datadriven technique was not very satisfactory. This catchment, being a very steepsloped catchment, has high peak discharge magnitudes with steep rising and recession limbs, which the GP models are unable to capture. This catchment being a small one with an area of about 5 km2 has a time of concentration of about 3045 minutes, but the time interval of the available data is one day, which seems to be another reason for GP's inability to capture the complex rainfall to runoff transformation on this catchment. Using a dataset of smaller time interval, the datadriven model should perform better. A key advantage of GP as compared to traditional modelling approaches is that it does not assume any a priori functional form of the solution. For instance, in a typical regression method, the model structure is specified in advance (which is in general difficult to do) and the model coefficients are determined. For neural networks, the time consuming task of initially defining the network structure has to be undertaken and then the coefficients (weights) are found by the learning algorithm. On the other hand, in GP, the building blocks (the input and target variables and the function set) are defined initially, and the learning method subsequently finds both the optimal structure of the model and its coefficients. Moreover, since GP evolves an equation or formula relating the input and output variables, a major advantage of the GP approach is its automatic ability to select input variables that contribute beneficially to the model and disregard those that do not. GP can thus reduce substantially the dimensionality of the input variables. In GP, as in any datadriven prediction model, the selection of appropriate model inputs is extremely important. This is especially so when lagged input variables are also used. Inclusion of irrelevant inputs leads to poor model accuracy and creation of complex models, which are more difficult to interpret as compared to simpler ones. Thus, for the remaining two catchments, an attempt is made to use the evolutionary search capabilities of GP for selecting the significant input variables. These variables, indicated as significant by GP are then used as inputs for the actual predictions. In contrast to the not so satisfactory performance by the GP models for predicting the runoff from Hok Tau catchment, their performance for the other two catchments is quite satisfactory, as the GP models are able to capture the peaks quite well and the goodnessoffit measures are also acceptable. These results indicate that GP can be used as a viable alternative for rainfallrunoff modelling, and the analytical form of the evolved equations facilitate easy interpretation. In this study, the GP evolved models are used for selection of significant variables influencing the rainfall to runoff transformation.   
dc.language  eng   
dc.publisher  Modelling and Simulation Society of Australia & New Zealand.   
dc.relation.ispartof  International Congress on Modelling and Simulation, Modelling and Simulation, MODSIM05 Proceedings   
dc.rights  This work is licensed under a Creative Commons AttributionNonCommercialNoDerivatives 4.0 International License.   
dc.subject  Datadriven models   
dc.subject  Evolutionary algorithms   
dc.subject  Genetic programming   
dc.subject  Rainfallrunoff modelling   
dc.title  Rainfallrunoff modelling using genetic programming   
dc.type  Conference_Paper   
dc.identifier.email  Jayawardena, AW: hrecjaw@hkucc.hku.hk   
dc.identifier.email  Muttil, N: nitinm@hkucc.hku.hk   
dc.description.nature  postprint   
dc.identifier.scopus  eid_2s2.068949124303   
dc.identifier.hkuros  123679   
dc.identifier.spage  1841   
dc.identifier.epage  1847   
dc.publisher.place  Australia   
dc.customcontrol.immutable  sml 170220   