Data-driven approaches for modelling of sub-critical coal-fired boiler

Abstract

Due to increasing shares of renewable electricity sources in the grid, thermal power plants need to operate in a more flexible manner in the future. This will involve more frequent startups, shutdowns, and load changes. A central part of a thermal power plant analysed in this study is the coal-fired boiler. In a previous study, a first-principle model of a sub-critical coalfired boiler has been developed and validated with operational data from a Polish power plant. Based on this model, this work aims to develop a computationally efficient and sufficiently accurate data-driven model that is easy to implement in new software. A selection of multi-output algorithms was first compared using nonoptimised parameters, with very few adaptations to the data set. Then, each algorithm had undergone three different optimisation routines to tune the hyper-parameters. The results of the nonoptimised models were compared with the optimised ones, and then compared to the reference first-principle model using the average Mean Absolute Percentage Error as a score. The methods used comprise six base learners and three algorithms using ensemble methods. The optimisation routines were based on the Powell conjugate direction method, Bayesian optimisation and evolutionary algorithm. All the data-driven models had shown a lower percentage error than the first principle model, and optimisation had resulted in improved prediction capacity for every base learner, but not for ensemble method-based algorithms.