Explorative application of discrete Bayesian networks as surrogate models for energy systems analysis

This work investigates data-based discrete Bayesian Belief Networks (BBNs) as surrogate energy system models for result analysis and interactive analyses, such as what-if analyses. A simplified version of the Swiss TIMES (STEM) model, referred to as STEM-lite, is used for demonstration. A method to optimize the BBN model is devised, based on performance metrics related to the accuracy of the BBN predictions, calculated over data records unseen by the BBN in the training phase. Further validation of the BBN on a set of seven scenarios yielded an average relative error below 2 %, suggesting adequate performance as surrogate model. The application of the surrogate BBN model is demonstrated to highlight its benefits, which include enabling interactive analysis (supported by the visualization of key variables, their relationships and interactions), fast and intuitive uncertainty propagation, and support for goal-driven analysis (backward reasoning from outcomes to the inputs that produce these outcomes). The surrogate BBN presented here was developed to elaborate the methods for constructing, validating, and using BBN models for energy systems analysis and to demonstrate the benefits of such a model; at this stage, this model is not intended for energy systems and economics policy discussions. For practical applications, future work is needed to reduce the number of data records to construct the BBN, to introduce the option to treat the time dependence of the input variables, and to allow for larger BBN models (involving more variables) that reflect the increasing complexity of energy systems. • We present methods for building, validating, and using BBNs as surrogate models. • With average relative error below 2 %, the BBN model is an adequate model surrogate. • Benefits: interactive analysis, intuitive uncertainty propagation, goal-driven analysis.