Integrated ApproachesIntegrated Approachesfor Runoff Forecastingfor Runoff Forecasting
Ashu JainAshu JainDepartment of Civil Engineering Department of Civil Engineering
Indian Institute of Technology KanpurIndian Institute of Technology KanpurKanpur-UP, INDIAKanpur-UP, INDIA
OutlineOutline
• Hydrologic CycleHydrologic Cycle
• Global Water FactsGlobal Water Facts
• Indian Scenario & Possible SolutionsIndian Scenario & Possible Solutions
• Rainfall-Runoff ModellingRainfall-Runoff Modelling
• Existing ApproachesExisting Approaches
• Integrated Approaches (3)Integrated Approaches (3)
• ConclusionsConclusions
Hydrologic CycleHydrologic Cycle
(Source: http://saturn.geog.umb.edu/wdripps/Hydrology/Hydrology%20Fall%202004/precipitation.ppt)
Global Water FactsGlobal Water Facts
• Total water – 1386 Million Kilometer^3Total water – 1386 Million Kilometer^3
• 97% in oceans & 1% on land is saline97% in oceans & 1% on land is saline
• => only 35 MKm3 on land is fresh=> only 35 MKm3 on land is fresh
• Of which 25 MKm3 is solidOf which 25 MKm3 is solid
• Only 10 MKm3 is fresh liquid waterOnly 10 MKm3 is fresh liquid water
• Availability is CONSTANTAvailability is CONSTANT
• Water Demands are INCREASING (2050!)Water Demands are INCREASING (2050!)
• Optimal use of existing WR is neededOptimal use of existing WR is needed
Indian ScenarioIndian Scenario
Water availability in India Water availability in India is highly uneven with is highly uneven with respect to both respect to both spacespace and and timetime
Indian ScenarioIndian Scenario
Indian ScenarioIndian Scenario
Kanpur ScenarioKanpur Scenario
Dainik Jagran: 2 May 2007Dainik Jagran: 2 May 2007
Indian ScenarioIndian Scenario
• We depend on rainfall for meeting most of our We depend on rainfall for meeting most of our water requirementswater requirements
• Most of the rainfall in majority of the country is Most of the rainfall in majority of the country is concentrated in monsoon season (June-September)concentrated in monsoon season (June-September)
• The uneven spatio-temporal distribution of water The uneven spatio-temporal distribution of water and uncertain nature of rainfall patterns call for and uncertain nature of rainfall patterns call for innovative methods for water utilization and innovative methods for water utilization and forecastingforecasting
Possible SolutionsPossible Solutions
Solutions of water problems in India Solutions of water problems in India lie in its root causeslie in its root causes
Space => InterlinkingSpace => Interlinking
Time => Rainwater HarvestingTime => Rainwater Harvesting
Possible SolutionsPossible Solutions
Other solutions includeOther solutions include
• Optimal Management of Existing WROptimal Management of Existing WR
• Runoff ForecastingRunoff Forecasting
• Technological AdvancementsTechnological Advancements
• Innovative Integrated ApproachesInnovative Integrated Approaches
Runoff ConceptsRunoff Concepts
• Amount of water at any time Amount of water at any time measured in m3/sec at any location measured in m3/sec at any location in a river is called runoff.in a river is called runoff.
• A graph showing runoff as a A graph showing runoff as a function of time is called a runoff function of time is called a runoff hydrograph.hydrograph.
A Runoff HydrographA Runoff Hydrograph
Runoff ConceptsRunoff Concepts
Runoff at any time depends onRunoff at any time depends on
• Catchment characteristicsCatchment characteristics
• Storm characteristicsStorm characteristics
• Climatic characteristicsClimatic characteristics
• Geo-morphological characteristicsGeo-morphological characteristics
Rainfall Runoff ModellingRainfall Runoff Modelling
• Physical processes involved in Physical processes involved in hydrologic cycle hydrologic cycle – Extremely complexExtremely complex– DynamicDynamic– Non-linearNon-linear– Fragmented Fragmented
• Not clearly understood Not clearly understood • Very difficult to modelVery difficult to model
Rainfall Runoff ModelsRainfall Runoff Models
Conceptual or DeterministicConceptual or DeterministicSystems Theoretic or Black Box TypeSystems Theoretic or Black Box Type
RegressionRegressionTime SeriesTime SeriesANNsANNs
IntegratedIntegrated
Integrated R-R ModelsIntegrated R-R Models
• Innovative Integrated approachesInnovative Integrated approaches–Conceptual + ANNConceptual + ANN
–Decomposition + AggregationDecomposition + Aggregation
–Time Series + ANNTime Series + ANN……
IntegratedIntegratedRainfall-Runoff Rainfall-Runoff
Model-1Model-1
Conceptual + ANNConceptual + ANN Conceptual ModelConceptual Model
Conceptual + ANNConceptual + ANN ANN/Black Box ModelANN/Black Box Model
Conceptual + ANNConceptual + ANN
An An integrated/hybridintegrated/hybrid model capable of model capable of exploiting the advantages of exploiting the advantages of conceptual and ANN techniques may conceptual and ANN techniques may be able to provide superior be able to provide superior performance in runoff forecasting.performance in runoff forecasting.
Conceptual + ANNConceptual + ANN
Data Employed: Kentucky RiverData Employed: Kentucky River
• Spatially aggregated daily rainfall (mm) Spatially aggregated daily rainfall (mm)
• Average daily river flow (m3/s)Average daily river flow (m3/s)
• Total length of data – 26 yearsTotal length of data – 26 years
• First 13 years for training/calibrationFirst 13 years for training/calibration
• Next 13 years for testing/validationNext 13 years for testing/validation
Integrated R-R Model-1Integrated R-R Model-1
• Conceptual:Conceptual: Base flow, infiltration, continuous soil Base flow, infiltration, continuous soil moisture accounting, and the evapotranspiration moisture accounting, and the evapotranspiration processes are modelled using conceptual/ processes are modelled using conceptual/ deterministic techniquesdeterministic techniques
• ANN:ANN: Complex, dynamic, and non-linear nature of Complex, dynamic, and non-linear nature of the process of transformation of effective rainfalls the process of transformation of effective rainfalls into runoff in a watershed are modelled using ANNsinto runoff in a watershed are modelled using ANNs
• Training:Training: ANN training is carried out using GA. ANN training is carried out using GA.
Integrated R-R Model-1 ResultsIntegrated R-R Model-1 Results
Model AARE R During Training Conceptual 23.57 0.9363 ANN 54.45 0.9770 Integrated 21.58 0.9773 During Testing Conceptual 24.68 0.9332 ANN 66.78 0.9700 Integrated 23.09 0.9704
Integrated R-R Model-1 ResultsIntegrated R-R Model-1 Results
Observed and Predicted Runoff in 1986 (Dry Year)Observed and Predicted Runoff in 1986 (Dry Year)
ANN Model Results (Summer)ANN Model Results (Summer)
Integrated Model-1 Results (Summer)Integrated Model-1 Results (Summer)
IntegratedIntegrated Rainfall-Runoff Rainfall-Runoff
Model-2Model-2
Decomposition + AggregationDecomposition + Aggregation
Figure 1: Decomposition of a Flow Hydrograph
R1
R2
F1
F2
F3
Time
Flo
w
Integrated Model-2 DetailsIntegrated Model-2 DetailsTable 1: Details of Neural Network Models
________________________________________________________________________________________________ Model Portion Architecture Number Statistics Input Variables
of Data ( x , σ ) ________________________________________________________________________________________________ Model-I 5-4-1 4747 (146.7, 238.8) P(t), P(t-1), P(t-2), Q(t-1), and Q(t-2) Model-II Rising 5-4-1 1783 (233.5, 330.3) P(t), P(t-1), P(t-2), Q(t-1), and Q(t-2) Falling 3-3-1 2963 (94.4, 135.7) P(t), Q(t-1), and Q(t-2) Model-III Rising 5-4-1 1783 (233.5, 330.3) P(t), P(t-1), P(t-2), Q(t-1), and Q(t-2) Falling Recession 2963 (94.4, 135.7) Q(t-1), and Q(t-2) Model-IV Rising 5-4-1 1783 (233.5, 330.3) P(t), P(t-1), P(t-2), Q(t-1), and Q(t-2) Falling-I 3-3-1 1189 (198.5, 164.4) P(t), Q(t-1), and Q(t-2) Falling-II Recession 1774 (25.3, 20.1) Q(t-1), and Q(t-2) Model-V Rising-I Inverse Recession 182 (8.2, 2.1) Q(t-1), and Q(t-2) Rising-II 5-4-1 1601 (259.0, 339.4) P(t), P(t-1), P(t-2), Q(t-1), and Q(t-2) Falling-I 3-3-1 1189 (198.5, 164.4) P(t), Q(t-1), and Q(t-2) Falling-II Recession 1774 (25.3, 20.1) Q(t-1), and Q(t-2) SOM(3) High 5-4-1 693 (537.8, 384.2) P(t), P(t-1), P(t-2), Q(t-1), and Q(t-2) Medium 3-3-1 1061 (195.5, 127.6) P(t), Q(t-1), and Q(t-2) Low 4-3-1 2993 (38.8, 50.9) P(t), P(t-1), Q(t-1), and Q(t-2) SOM(4) High 5-4-1 409 (678.9, 426.3) P(t), P(t-1), P(t-2), Q(t-1), and Q(t-2) Medium-I 4-3-1 704 (280.4, 157.4) P(t), P(t-1), Q(t-1), and Q(t-2) Medium-II 3-3-1 1089 (136.7, 104.4) P(t), Q(t-1), and Q(t-2) Low 3-3-1 2545 (28.4, 34.3) P(t), Q(t-1), and Q(t-2)
Integrated Model-2 ResultsIntegrated Model-2 Results
Model AARE R AARE R During Training During Testing
Model-I 54.97 0.9770 65.71 0.9700 Model-II 61.28 0.9764 72.28 0.9696 Model-III 31.66 0.9607 36.45 0.9571 Model-IV 31.90 0.9777 39.56 0.9684 Model-V 23.85 0.9780 21.63 0.9678
Scatter Plot from Model-VScatter Plot from Model-V
Results-Model-V: Drought Year 1988Results-Model-V: Drought Year 1988
IntegratedIntegratedRainfall-Runoff Rainfall-Runoff
Model-3Model-3
Time Series + ANNTime Series + ANN
• Basic Steps in Time Series ModellingBasic Steps in Time Series Modelling– DetrendingDetrending– DeseasonalizationDeseasonalization– Auto-correlationAuto-correlation
• ANN modelling involves presenting raw ANN modelling involves presenting raw data as inputsdata as inputs
• Time series steps can be carried out Time series steps can be carried out before presenting data to ANN as inputs.before presenting data to ANN as inputs.
Time Series + ANNTime Series + ANN
• ANN1 – Raw DataANN1 – Raw Data
• ANN2 – Detrended DataANN2 – Detrended Data
• ANN3 – Detrended and ANN3 – Detrended and Deseasonalized Data Deseasonalized Data
Time Series + ANNTime Series + ANN
Data EmployedData Employed• Monthly runoff from Colorado River @ Monthly runoff from Colorado River @
Lees Ferry, USA for 62 yearsLees Ferry, USA for 62 years
• Past four months lagPast four months lag
• 50 Years for training50 Years for training
• 12 years for testing12 years for testing
Time Series + ANNTime Series + ANN
Lag 2 Results Lag 4 Results
AARE R AARE R
Time Series 92.78 0.48 88.52 0.51 ANN1 44.51 0.62 44.01 0.68 ANN2 19.55 0.77 17.67 0.80 ANN3 12.55 0.86 9.62 0.89
ConclusionsConclusions
• Runoff forecasting is important for efficient Runoff forecasting is important for efficient management of existing water resources.management of existing water resources.
• An individual modelling technique provides An individual modelling technique provides reasonable accuracy in runoff forecasting.reasonable accuracy in runoff forecasting.
• Neural network based solutions can be Neural network based solutions can be better than those obtained using better than those obtained using conventional methods.conventional methods.
ConclusionsConclusions
• Integrated modelling approaches have the Integrated modelling approaches have the potential for producing higher accuracy in potential for producing higher accuracy in runoff forecasts.runoff forecasts.
• Innovative integrated approaches dependent Innovative integrated approaches dependent on the nature of problem are needed in order on the nature of problem are needed in order to develop hybrid forecast models capable to develop hybrid forecast models capable of exploiting the strengths of the available of exploiting the strengths of the available individual techniques.individual techniques.
Thank YouThank You