Automatic calibration of numerical model using artificial intelligence based techniques

Abstract

Many energy companies rely on natural resources to produce energy. They use advanced models to estimate how much of those resources they have access to, but if a model is to make an accurate estimation it needs to be accurately calibrated. There is little agreement in the science literature about what automatic calibration method is the best one to use on numerical model. The Shuffled Complex Evolution (SCE-UA) method is considered state of the art, and while it has been over 20 years since it was developed it is still in use both for commercial purposes and research. We compared the SCE-UA method to three other methods that can potentially be used for parameter optimization; Continuous Action Learning Automata(CALA), Genetic Algorithms(GA) and a Monte Carlo Scheme. We implemented and configured these methods to run an implementation of the HBV hydrological model. The purpose of this was to see if the SCE-UA method was still the best one to use compared to these more general methods. We designed a test protocol and an evaluation method to compare the methods on a level playing field. To be able to do this we had to research the characteristics of the methods and how to configure them to work with the HBV model. Our results conclusively showed the SCE-UA and Genetic Algorithm methods giving the most accurate and efficient results. However, both their results were so similar that we could not make a decisive conclusion of which one of them was the best. We concluded that with our evaluation and test procedures they produced roughly equal results. The CALA method came out worse than any of the other methods.