Machine Learning Interfaces for Optimal Design and Control of Solar Thermal Systems in Process Industry

No Thumbnail Available



Journal Title

Journal ISSN

Volume Title


Addis Ababa University


The trend of introducing solar thermal systems (STSs) in process industries has resulted in a new energy paradigm– an interactive platform where there are economic benefits and motivations to address sustainable development. On the other hand, this paradigm has also introduced fluctuations and uncertainties not previously seen on the energy system, and is challenging the industry. Accordingly, we are observing increasing need for robust design and control solutions that will facilitate the smooth operation and cost competitiveness of the industrial solar thermal system (ISTS). The possibility for developing such a solution exists, but only if the necessity to explicitly model, which might not work well, and also increase computational complexity in the ISTS, is removed. In this dissertation work, a machine learning (ML) approach is followed for design and control optimization of ISTSs, and is leveraged for two goals. First it is used as a multi-modelling tool for developing heterogeneous optimization interfaces, using stochastic and generic models. These interfaces are intended to be simple but are not simpler in order to simultaneously address both scalability-tractability tradeoffs and model inefficiencies of conventional methods. Afterwards, ML enabled linking up of these interfaces as building blocks for realizing a modular optimization framework, and of integrating different layers of functionalities. As a result, the ML approach allowed disaggregated modelling of several similar technologies (and processes) as well as parameterizing of their inputs and local condition differently. Using this method, it was also possible to represent distributed energy resources (DERs) and their additional capabilities of interactions. Furthermore, it allowed replication of simulation experiments with the same model and at varying scale levels. These are essential features that cannot be offered by conventional methods, and can be used to improve synergy and unlock the potentials of DERs in ISTSs. Following the ML approach, some important findings were made. Firstly, the solutions of the optimal design problems were scalable and tractable. This feature facilitates operation-based designs of STSs according to the specific requirements of process(s), heat distribution networks or existing thermal plant in industry. The approach also allowed the testing of an improved optimal control strategy, while at the same time, enabling controller tuning or model calibration. This capability is used to adapt an empirical solar radiation model, to serve as an efficient and low-cost sensor that can be integrated to ISTSs in real-time. However, due to the scope and limitation of the dissertation, these relevant findings provide mainly key design and control strategies and points of discussion instead of benchmarked results. Therefore, it is particularly desirable if further research could confirm these findings.



Design optimization., optimal control, model calibration, machine learning, solar thermal system, process industry