Optimisation procedures

The iterative steps of model calibration could be undertaken manually or automatically using an optimisation method or procedure. Effective automatic methods are preferred by some as they reduce subjectivity (Ndiritu and Daniel, 1999). However, where the modeller is adequately experienced with a given model, manual calibration could suffice (Ndiritu and Daniel, 1999). The use of optimisation techniques will allow the model to be calibrated satisfactorily for the range of catchments it is intended to represent such that there is knowledge on how the parameters are varying between catchments. Optimisation procedures therefore enhance the ability of the modeller to assess the predictive power of the model (Klemeš, 1986).

Manual optimisation generally involves modifying one or more parameter value and observing the effect on the fitting criteria between the observed and the simulated values and then repeating the exercise (Gorgens, 1983). However manual optimisation is time consuming and can be confusing particularly if there is a high degree of interaction between parameters (Eckhardt and Arnold, 2001; Madsen, 2000). The success of a manual calibration is said to depend heavily on the experience of the modeller and their knowledge of the model components and parameter interactions (Eckhardt and Arnold 2001).

Automatic optimisation means that parameters are evaluated using an automatic optimisating technique (Rosenbrock, 1960; Ibbitt and O’Donnell, 1971; James, 1972; Gupta and Sorooshian, 1985; Yapo et al., 1995; Ndiritu and Daniel, 1999). This procedure lessens reliance on the subjective judgement of the modeller and calibration can be significantly faster. According to Ndiritu and Daniel (1999), due to parameter correlations, it is often found that a unique global optimum parameter set does not exist and therefore recommend that the detection of parameter correlations should become an integral component of parameter identification and that an adequate level of optmisation should detect parameter correlations precisely. Since only one objective function can be used in some automatic optimisers, it may be necessary to carry out manual adjustments to parameter values to fine tune the model to achieve a better overall correspondence based on several criteria.

Eckhardt and Arnold (2001) suggest that since the high number of parameters in distributed models make special demands on the optimization process, there is need to develop a strategy of imposing constraints on parameters to limit the number of independently calibrated values.

In automatic optimization, calibration based on a single performance measure is often inadequate to properly identify the successful simulation of all the important characteristics of the system that are reflected in the observations (Madsen, 2000). This then has led to the development of automatic calibration strategies with the use of multi- objective functions (Madsen, 2000). These are known to be quite complicated but may be a step in the right direction to enhance the usefulness of automatic optimization routines.

3.3.3 Model validation

There is some confusion in the literature between model validation and verification and they have frequently been used to describe the same thing. There are essentially two checks that are needed in the application of hydrological models. The first is that the model algorithms and the way these are represented in the computer code are correct and

accurate (Stedinger and Taylor, 1982). This may include an assessment of whether the model is achieving a water balance (i.e. no water is being lost in the model). For the purposes of this document, this process is referred to as ‘model’ validation or verification.

Once the calibration of a model is completed, it becomes necessary to check that the calibrated parameter values can be used satisfactorily to simulate events, or climatic conditions, other than those used for the calibration (Sorooshian, 1983; Pirt and Bramley, 1985). This is referred to in this document as ‘parameter’ validation or verification, as it is the calibrated parameter set that is being assessed, rather than the model structure.

Another level of validation is to check perceived associations between model parameters and physical catchment characteristics. If these associations can be established and are then applied to different data sets from the calibrated catchments, then parameter validation would involve the assessment of the success of the transfer of the parameter relations to other catchments. This is vital to demonstrate whether the model can be used successfully to simulate flows in ungauged catchments.

3.4 PRACTICAL USE OF MODELS FOR WATER RESOURCE

ASSESSMENTS

The primary objective of hydrological modelling is quite often to generate a long representative time series of streamflow volumes from which water supply schemes and civil structures can be designed (Hughes, 1995). Sufficient knowledge of streamflow is necessary to aid in the efficient design and construction of bridges, conduits, dams, flood control structures, irrigation schemes, mutipurpose water supply schemes, transboundary water transfers and any water resources related infrastructure such as hydroelectric power stations. For efficient and dependable design decisions to be made, longer streamflow time series are required than are frequently available. Therefore flow time series have to be generated with sufficient accuracy through the use of hydrological models.

Many water resources problems are best tackled by testing the simulated performance of alternative designs in relation to representative sequences of hydrological input in the system. For the purpose of simulation, usually a hydrograph covering a period of 50 years of monthly river flows is required at each key point of interest in the system. Getting this length of reliable and continuous river flow data in most countries is difficult as streamflow observations are rather sparse and good records cover relatively short periods. In order to remedy the situation, the rainfall-runoff modelling approach is often used to make up for the missing data and satisfy the minimum record length requirement. Rainfall-runoff modelling often provides a solution because precipitation records normally extend back much further than flow records. This is due to the relative ease of making precipitation measurements as compared to setting up a discharge measuring station. Raingauges are easy to install with a much greater flexibility of choice of site which can be conveniently located near a centre of human activity thus facilitating operation and maintenance. River gauging stations on the other hand, must be located at hydraulically acceptable sites which are often remotely located, with gauge plates often being susceptible to damage or lack of maintenance. Gauging stations also require calibration through the establishment of rating curves which may require many years to adequately cover the range of stage variations. These requirements, added to the fact that resources in terms of both human and material costs are limited, has led to the situation where there are fewer river gauging stations with relatively shorter records while precipitation stations are numerous and have longer records (Shawinigan, 1993):

Hydrological models are therefore a useful tool to aid decision making in water resources assessments, planning and management (James, 1991). Specific applications may include: forecasting and predicting hydrologic phenomena; provision of sufficient information for engineering structural design, record extension, reservoir operation simulation, data in-filling and revision and the assessment of effects of land use changes or other catchment devlopments.