ojcsv02i02p127-132

INTRODUCTION

Artificial neural networks are widely usedas an effective approach for handling non-linear andnoisy data, especially in situations where thephysical processes relationships are not fullyunderstood and they are also particularly well suitedto modelling complex systems on a real-time basis.Fuzzy logic is a generalisation of Boolean logicimplementing the concept of par tial truth oruncertainty, so within the fuzzy set theory anelement can have a gradual membership to differentsets. To describe system behaviour with fuzzy logic,you need to define fuzzy sets, fuzzy rules or socalled IF-THEN rules and apply a fuzzy inferencescheme. The generation of a fuzzy forecast model

Oriental Journal of Computer Science & Technology Vol. 2(2), 127-132 (2009)

Overview of the artificial neural networks and fuzzy logicapplications in operational hydrological forecasting systems

JITENDRA P. DAVE¹ and DIVYAKANT T. MEVA²

C.U. Shah College of Master of Computer Application,Wadhwan City, District Surendranagar, Gujarat (India).

(Received: July 25, 2009; Accepted: September 30, 2009)

ABSTRACT

Damage due to flooding has increase in many countries in the last years, and due to the globalclimate change, which is now recognized as a real threat, an increase in the occurrence of floodingevents and especially of flash flooding events is likely to continue into the future. In those conditionsand because building new flood defences structures for defending vulnerable areas has serious financialimplications, the timely forecasting of floods is becoming more important for flood defence and ingeneral for water management purposes. The complexity of natural systems and of hydrologicalprocesses that influence river levels evolutions make the traditional modelling approaches, based onmirroring natural processes with physically based equations very difficult. Despite the fact that in thelast decades the Operational Hydrological Forecasting Systems were significantly developed, becomingmore and more complex systems, ingesting and processing in real time a great amount of data fromautomated hydrometrical and meteorological stations networks and high resolution grided data fromradars and satellites, together with the use of distributed hydrological models, the warning and forecastsimprovements are not very significant, in many cases the performance of the new physically baseddistributed models being comparable with the “older” conceptual lumped models. The paper presentsan overview of some alternative and complementary modelling approaches, artificial neural networksand fuzzy logic systems, possible applications for the improvements of the Operational HydrologicalForecasting Systems, and presenting also some example of rainfall-runoff modelling implementations.

Key words: Artificial neural networks, fuzzy logic applications, forecasting systems.

can be based both on experts knowledge andhistorical data. In conclusion, both artificial neuralnetworks and fuzzy logic modelling systems offerthe potential for a more flexible, less assumptionapproach to hydrological processes, and they havealready been demonstrated as successfullysubstitutes for the classical rainfall – runoff models,and also as tools for the real time updating ofhydrological forecasting models and especially forthe multimodel approach. Keywords: hydrologicalforecasting model, artificial neural network, fuzzylogic, operational hydrological forecasting systems.

Operational hydrological forecasting systemsOn a global scale, floods account for over

65% of people affected by natural disasters and

128 Dave & Meva, Orient. J. Comp. Sci. & Technol., Vol. 2(2), 127-132 (2009)

they are the most damaging of all natural disasters.Better forecasting floods and with a larger lead time,is the main sustainable way of adapting to andmanaging such disasters. Operational hydrologicalorecasting systems, which link state of the rivercatchments, river discharges and water levels,recorded precipitations and weather forecasts, canbe used to respond to floods as they occur and toreduce their costs in term of lives, property andother damages. Current flood forecasting andwarning systems have several limitations, such as,insufficient lead-time to provide accurate floodwarnings, inadequate spatial and temporalresolution of the real-time rainfall observations andforecasts for flood producing storm, little integrationof different sources of forecast information.Moreover their ability in considering theuncertainties in estimating and forecastingprecipitation and flood discharges is very limited.

The desirable characteristics of a goodflood forecasting system are:

TimelinessThe lead-time is the time between making

a forecast of an event and its occurrence, if sufficientlead-time is available and the predictions areaccurate then evacuation, even of relatively largenumbers of people may be possible. The increaseof the lead-time is mainly limited by the availabilityof reliable quantitative precipitations forecasts, butalso can be limited by the hydrological models orforecasting methodologies that are implemented.

AccuracyIs usually related to the correctness the

forecasts of the magnitude and time of the floodpeak and of the resulting levels. In special situations,it may relate to the forecasts of the completehydrograph of the flood. The more accurate theforecast the better flood control/modification anddamage mitigation measures can be implemented.

ReliabilityCan be associated with accuracy, but is

related to the overall long-term reliability of the floodforecasting system, and not just to the accuracy ofa forecast for a particular flood. Usually the long-term reliability of the system can be assessed byits performance in two respects. It should always

forecast a flood when one occurs and it should notforecast floods when one doesn’t occur. Thereliability, like accuracy, affects the confidence indeciding on response measures, as the publicperception of warnings messages may be veryimportant. Forecasts require both data collectionand modelling. The amounts of data and thecomplexity of the modelling necessary to achievespecific targets of lead-time, accuracy and reliabilityvary from catchment to catchment, and there is anatural conflict between the desire for greaterforecast leadtime and greater accuracy andreliability (usually the warning messages are basedon model simulations that take into account just therecorded precipitations and not the forecastedprecipitations). Generally the longer the lead timesthe less accurate and reliable are the forecasts offlood magnitude, location and timing.

Operational hydrological forecasting systemshave (or could have) the following componentsData acquisition systems

Is the basic component for an operationalsystem, and the data type and availability havemajor implications on the modelling part of thesystem.

Rainfall forecasts modelsIs the most important part for the forecast

lead-time increase. Unfortunately the present resultsof the numerical meteorological models are notenough accurate for the hydrological forecastsapplications. On short term, the solution could be acombination with the short future scan estimationsbased on radar information.

Rainfall-runoff forecasts modelsThe possible approach extend from the

extremely simple forecast relations, event typemodel, through conceptual semi-distributed models,which are still the most used models in operational,to complex physically based models.

Flood routing and flood plain modelsThe hydrological routing methods are still

extensively used, but the general direction is to useappropriate hydraulic models, which take intoaccount the river geometry, and allow reasonableestimations of flood maps.

129Dave & Meva, Orient. J. Comp. Sci. & Technol., Vol. 2(2), 127-132 (2009)

Flood impact analysis componentIf flood maps are available, flood impact

analysis could be finally obtained by superimposingflood maps with GIS georeferenced spatial data onconstructions, traffic, agriculture, etc.

Artificial neural networks modelsThe field of neural networks has a history

of five decades but has found solid application onlyin the past decade, and the field is still developingrapidly. Neural networks are composed of manysimple elements operating in parallel. Theseelements are inspired by biological nervous systems.The network function is determined largely by theconnections between elements.

Neural networks have been trained toperform complex functions in various fields ofapplication including pattern recognition,forecasting, identification, classification, speech,vision and control systems. Artificial neural networkscan be characterised most adequately ascomputational models with particular propertiessuch as the ability to adapt or learn, to generalise,or to cluster or organise data, and which operationis based on parallel processing.

An artificial network consists of a pool ofsimple processing units which communicate bysending signals to each other over a large numberof weighted connections.

A set of major aspects can be distinguished´ a set of processing units (neurons);´ a state of activation for every unit, which also

determines the output of the unit;´ connections between the units, generally

each connection is defined by a weight whichdetermines the effect which the signal of oneunit has on other unit;

´ a propagation rule, which determines theeffective input of a unit from its externalinputs;

´ an activation function, which determines thenew level of activation based on the effectiveinput and the current state;

´ an external input or offset for each unit;´ a neural network architecture;´ a training method.

Within neural systems it is useful todistinguish three types of units: input units whichmay receive data from outside the system, outputunits which send data out of the system and hiddenunits whose input and output signals remain withinthe system itself.

The neuron model and the architecturesof a neural network describe how a networktransforms its input into an output. Both the neuronmodel and the network architecture each placelimitations on what a particular model can compute.The way a network computes its output must beunderstood before training methods for the networkcan be explained.

Fig. 1: Typical representation of an artificialneuron model and for a full connected feedforward neural network, with two hidden layers

Backpropagation neural networks havebeen successfully used as a basic tool for theconstruction of effective forecasting schemes.Different neural network architectures have beenproposed as the basis of forecasting methods. Theone hidden layer feedforward network is gainingbroad acceptance due to its simplicity andexpressiveness power. This net can be thought ofas a nonlinear autoregressive (AR) model. Themathematical relationship, established through thesigmoid processing elements of the network, isessentially nonlinear.

Review of artificial neural networks and fuzzylogic applications in hydrological forecasting

Especially over the last decade, neuralnetwork-based flood forecast models have beenincreasingly used in hydrology. Usually, the inputdata of the network are composed by pastmeasurements of flows and rainfalls, and eventuallythe basin state could be assessed analyzing the


rainfall occurred on a certain time window beforethe flood event. The most used neural networks typeis the feedforward network trained withbackpropagation family algorithms. In someapplications are implemented neural network thatuses recurrence, i.e. outputs at some level are fedback in as inputs. This allows the net to better trainon time dependent data. Examples of this areJordan, Elman, Real-Time Recurrent Learning(RTRL) networks, or Time Delay Neural Network(TDNN), which uses not only the current input datavalues (time t) but also the previous several inputsets (t-1, t-2. etc.).

In general, the fuzzy models used forhydrological forecasting are based on a fuzzy rulebase describing the hydrological behaviour of theriver basin. Expert knowledge about specificdischarge situations combined with precipitationinformation and soil moisture conditions can betransformed directly in IF ... THEN ... rules, usinglinguistic entities like discharge=low ANDprecipitation=high AND soil moisture=high and thusbuilding up an initial rule base. Optimizationprocedures are then used to adapt the rule baseon the basis of data from former flood events inorder to achieve an optimal forecast model.

Such kind of modelling systems hasalready successfully been used for operationalhydrological forecasts elaboration in river basin withdifferent sizes and with different lead times. Thecomputational effort of forecasts with different timehorizons requires only few seconds on a standardPC. Another recently developed computingtechnique is the neurofuzzy approach, which is acombination of a fuzzy computing approach andan artificial neural network technique.

This approach is becoming one of themajor areas of interest because it gets the benefitsof neural networks as well as of fuzzy logic systemsand it removes the individual disadvantages bycombining them on the common features. Neuralnetworks and Fuzzy logic have some commonfeatures such as distributed representation ofknowledge, model-free estimation, ability to handledata with uncertainty and imprecision etc. Fuzzylogic has tolerance for imprecision of data, whileneural networks have tolerance for noisy data. A

neural network’s learning capability provides a goodway to adjust expert’s knowledge and itautomatically generates additional fuzzy rules andmembership functions to meet cer tainspecifications. This reduces the design time andcost. On the other hand, the fuzzy logic approachpossibly enhances the generalization capability ofa neural network by providing more reliable outputwhen extrapolation is needed beyond the limits ofthe training data. The neuro-fuzzy system consistsof the components of a conventional fuzzy systemexcept that computations at each stage is performedby a layer of hidden neurons and the neuralnetwork’s learning capacity is provided to enhancethe system knowledge.

The system contains at least the followingthree different layers:

Fuzzification LayerIn a fuzzification layer each neuron

represents an input membership function of theantecedent of a fuzzy rule.

Fuzzy Rule LayerIn a fuzzy inference layer fuzzy rules are

fired and the value at the end of each rule representsthe initial weight of the rule, and will be adjusted toits appropriate level at the end of training.

Defuzzification LayerIn the defuzzification layer each neuron

represents a consequent proposition and itsmembership function can be implemented bycombining one or two sigmoid functions and linearfunctions. The weight of each output link hererepresents the centre of gravity of each outputmembership function of the consequent and istrainable. After getting the corresponding output theadjustment is made in the connection weights andthe membership functions in order to compensatethe error and produce a new control signal.

Different successfully applications, ofthose different computing techniques, have beenreported by a large number of researchers,especially in the last decade. Between theapplications that are related to the hydrologicalforecasting activity we can mention:´ Rainfall-runoff modelling;

131Dave & Meva, Orient. J. Comp. Sci. & Technol., Vol. 2(2), 127-132 (2009)

´ Flood routing;´ Directly forecasts of water levels;´ Link different individual forecasting models

into a single forecasting system multimodel/ consensus / combination r iver flowforecasting);

´ Forecasting river ice jams break-up;´ Modeling of the flood forecasting uncertainty;´ Quantitative precipitations forecasts;´ River flow forecasts updating technique;´ Control and optimization of reservoir

operations;´ Replicating complex hydraulics or rainfall-

runoff models;´ Non-linear, non-unique stage-discharge

relationship.

CONCLUSIONS

After the general review of neural networksand fuzzy logic modelling approachescharacteristics and applications we can summarizesthe following conclusions:´ Both neural networks and fuzzy logic models

are a convenient way to map anndimensional input space to an m-dimensional output space, being especiallyuseful when the relation between input andoutput space variables is not well-known,except the fact that is a nonlinear relation.

´ Artificial neural networks are widely used asan effective approach for handling non-linearand noisy data, especially in situations wherethe physical processes relationships are notfully understood and they are also particularlywell suited to modelling complex systems on

a real-time basis.´ Fuzzy logic is a very powerful tool for dealing

quickly and efficiently with imprecision andnonlinearity.

´ The generation of a fuzzy forecast model canbe based both on experts knowledge andhistorical data.

´ Both artificial neural networks and fuzzy logicmodelling systems offer the potential for amore flexible, less assumption approach tohydrological processes, and they havealready been demonstrated as successfullysubstitutes for the classical rainfall – runoffmodels, and also as tools for the real timeupdating of hydrological forecasting modelsand especially for the multimodel approach.

In the next 3 years, the national DESWATdecisional and informational integrated nationalsystem for management in case of disasters projectwill be implemented. The hydrological modelling andforecasting system of this project will implement thefollowing modelling component:´ National Weather Service River Forecasting

System (NWSRFS - USA);´ NOAH – LIS distr ibuted modelling

component;´ TOPLATS distributed model, for some

specific catchments;´ FLASH FLOOD GUIDANCE component.

After the first year modelling systemimplementation, we intend to investigate thepossibility of adding into the system of modellingcomponent using the neural network and fuzzy logicapproaches.

1. Carling A., Introducing Neural Network.SIGMA PRESS - Wilmslow, United Kingdom(1992).

2. Michael A. Arbib, The Handbook of BrainTheory and Neural Networks. THE MITPRESS, Massachusetts Institute ofTechnology, United States of America (1995).

3. Toderean G., Costeiu M., Giurgiu M., Artificialneural networks, Blue Press, Cluj-Napoca.

REFERENCES

Romania (1995).4. Dumitrescu D., Hariton C., Neural networks,

theory and applications, Ed. Teora,Bucharest, Romania (1996).

5. French, M.N., Krajewski, W.F. andCuykendall, R.R., Rainfall forecasting inspace and time using a neural network,Journal of Hydrology, 137: 1-31 (1992).

6. Govindaraju, R.S., Anmala, J., Nedunuri, K.V.


and Koelliker, J.K., Use of neural networksin predicting the nonlinear rainfall-runoffrelationships for small watersheds, paperpresented at the European GeophysicalSociety General Assembly, Hamburg, 1995and abstracted in the conferenceproceedings (1995).

7. Liong, S.Y., Nguyen, V.T.V., Chan, W.T. andChia, Y.S., Regional Estimation of Floods forUngaged Catchments with Neural Networks.In Cheong, H-F., Shankar (1994).

8. Maier, H.R. and Dandy, G.C., The use ofartificial neural networks for the predictionof water quality parameters, WaterResources Research, 32(4): 1013-1022(1996).

9. Wilby, R.L., Abrahart, R.J., Dawson, C.W.,Detection of conceptual model rainfall-runoffprocesses inside an artificial neural network,Hydrological Sciences Journal, 48(2): 163-181 (2003).

10. Zekai, Sen, 2004 : Fuzzy logic and systemmodels in water sciences, Turkish WaterFoundation,Istanbul.

11. M.Stuber, P. Gemmar, M. Greving, Machinesupported Development of Fuzzy – Flood

Forecast Systems, European Conference onAdvances in Flood Research, Potsdam, PIKReport Nr. 65 (2000).

12. C. Mahabir, F.E. Hicks, A. Robinson Fayek,Forecasting Ice Jam Risk at Fort McMurray,AB, using Fuzzy Logic, Proceedings of the16th IAHR Intrnational Symposium on Ice,Dunedin, New Zealand (2002).

13. Linda See, Robert J. Abrahart, StanOpenshaw, An Integrated Neuro-Fuzzy-Statistical Approach to HydrologicalModelling, School of Geography, Universityof Leeds, UK (1998).

14. S. Alvisi, G. Mascellani, M. Franchini, A.Bardossy, Water level forecasting throughfuzzy logic and artificial neural networkapproaches, published in Hydrology andEarth System Science Discussions, EGU(2005).

15. P.C. Nayak, K.P. Sudheer, D.M. Rangan, K.S.Ramasastri, Short-term flood forecastingwith a neurofuzzy model, Water ResourcesResearch, 41 (2005).

16. Marius Matreata, National Institute ofHydrology and Water Management,Bucharest, ROMANIA.

ojcsv02i02p127-132

Documents