With the proliferation of renewable energy generations, power conversion systems (PCSs) are becoming much more complex; it is becoming challenging to search all possible power conversion architectures (PCAs) and find the best optimisation in terms of different objectives. Therefore, this study investigates a systematic approach to construct and evaluate PCAs using graph theory. First, the components in PCSs are graphically modelled as either nodes or edges. Then, a generalised PCA deduction methodology is proposed, and all possible PCAs can be mathematically deduced by modifying elements in adjacency matrices. For a fuel cell (FC) generation system, 45 possible PCAs are found with the proposed method. Furthermore, an evaluation methodology based on graph theory is proposed. The performance indices of the deduced PCAs, including costs, efficiency, and reliability, are calculated. Then, an optimisation approach is applied to finding the best architecture compromise, where the one with the shortest distance to the ideal architecture is considered the best architecture compromise. For the FC demo system, with the proposed assessment methodology, the best architecture compromise (dc-bus structure) is found among 45 possible architectures. Finally, the experimental platform, which adopts the dc-bus optimised architecture, is built and experimental results validate the architecture selection.