Below a further elaboration of the modelling phases is given.
The analysis and modelling phase leads to a conceptual model. During this phase the system in focus is defined. A system is a limited part of the real world that will be analysed. Outside the system is the environment, which might influence the system, but cannot be controlled. Building a conceptual model requires the definition of parameters (dependent and independent), their possible values, as well as the numerical relationships between those. In addition, for optimisation models, a criteria function must be defined. Overall, this phase should lead to the formulation of a mathematical model of the problem.
Taking the Balmorel model as an example, this phase has numerous important decisions that must be made. This includes the overall delimitation of the type of questions to ask as e.g. illustrated in Section 1.4, the choice of model type, and how best to represent the different model elements. An example of this is the choice of time resolution and between deterministic and stochastic programming. Chapter 4 describes these examples in more detail.
The computer programming and implementation phase is the transfer of the mathematical model into a computerised model. The model and solution method can be programmed from scratch but often some modelling/simulation languages are used that permit a near mathematical formulation of problems and efficient solution by accompanying optimisation or simulation software.
However, sometimes the analyst must implement methods for solution, as no existing ones fit with the model of the problem. Some model types are easily solved while others cannot—even with the technology and knowledge of today—be solved to optimality. Instead the method implemented must be able to find good, but not necessary optimal solutions to the models.
In relation to the Balmorel model, it was chosen to use the algebraic modelling language GAMS, see Brooke, Kendrick, and Meeraus (1989), and using commercial solvers to solve the linear programming problem formulated in this, as very efficient solvers for linear programming problems are available. This was to assure that focus could be on the model formulation and not on implementing the model or solution methods. See further the discussion in Section 4.3.
For the stochastic model that was built, a solution algorithm was developed as no existing code for efficient solution existed.
Finally, the experimentation phase addresses the computations made using the computerised model. Looking at problems in general, this can be computations to verify whether a hypothesis holds or simulating the consequences of taking different actions. For decision-problems, the results of the model run should indicate which decisions to take given the problem in view.
Many experiments have been made with the Balmorel model during the development stage. Some of the experiments made have been for analysing a specific problem.
Typically of these is the one presented in Paper E. But mainly experiments have been made to test the validity of the model and to assist in the decision-making during the analysis and modelling phase. This is part of the important issue during model development known as model Validation and Verification (V&V). As seen in Figure 16 the V&V is part of all the phases of the modelling process.
In order to trust the model results, it has to be determined whether the underlying assumptions and mechanics of the model are sound and the data used is acceptably accurate. Validation of a model ensures this. Verification on the other hand is to check that the formulated mathematical model and the dataset to be used are implemented and ran correctly on the computer. In general, verification is testing whether the model is implemented as intended while validation is whether the intended model is the right model for the problem.
The various V&V steps in the process of modelling can be presented as follows.
The data validation concerns whether data is available for use, reliable, consistent, and up-to-date. Relevant issues to consider include how the data is measured and what assumptions that have been made. Also, the validation should ensure that the transformation of data from the ones collected to the ones required by the model has been done correctly.
The conceptual model validation is addressing whether the choices during the analysis and modelling phase are sound. The design choices may be backed by earlier experiences. If no such are available, the choices should be based on theoretical considerations or empirical analyses, to make sure that the assumptions, the defined model elements, and their relations are sound compared with the answers to be found.
The computerised model verification concerns whether the conceptual model has been correctly transformed into a computer program. This includes the implementation of any method, whether optimisation or simulation, used for analysing it.
The operational validation is dealing with the results of the whole model. Basically, it is tested whether the overall model results behave like the real world does. In general, this can be analysed by either historical or model comparison. In the historical comparison, input similar to those found in the real world should result in a behaviour reasonably close to the one observed in reality. Model comparison may be done if there is no real world data to compare with but also to see the performance of the computerised model developed in comparison with similar existing models.
V&V is actions performed by the analyst during the model development. However, it is not enough that the model developer trusts the results, as the decision-maker, who is to use its results for assistance when taking decisions, should share this belief. This leads to the discussion of model evaluation in the next section, where the acceptance of the model as seen by the decision-maker is addressed. This also includes an overview of literature relevant for both V&V and model evaluation.