Improving information gain from optimization problems using artificial neural networks

Decisions in energy policy are influenced by the results from energy systems optimizations. Uncertainties regarding the input parameters of optimization problems, e.g. cost developments of technologies and resources in the future, may influence the optimization results in such a way, that an easy interpretation of results is not possible. The methodology presented herein aims to overcome the problem of uncertainties and to allow taking into account probability distributions (pd) for all input parameters while limiting the number of necessary optimizations to a minimum. This is achieved using design of experiment (DoE) to select the appropriate input parameter combinations to train an artificial neural network (ANN). The resulting ANN is then used to predict the optimization result for all possible input parameter combinations, which are then weighted with a pd according to user preferences. In this contribution, an explanation of the new methodology OPANN (optimization considering probabilities with artificial neural networks) and its application are presented. The information gained from a number of random or selected (e.g. scenario based) simulations is compared with the results following the DoE approach and the application of ANN and pds. The number of necessary simulations with the new methodology is then evaluated with regard to the applicability of the Monte Carlo method and stochastic optimization and the cost-benefit ratio for the considered methods at different numbers of runs of the original optimization problem is compared.