Fig 1.
Conceptual representation of the proposed HITL-MGA workflow and of the synthetic experiment we carry out in this study to test the workflow.
The theoretical framework we propose envisions that we collect the high-level system design preferences arising from a first iteration of an MGA design space with stakeholders. Then, we decode which technical features underlying such preferences and translate them into parameters that may inform and re-tune the MGA search for design options. The resulting updated design space is richer in design options that align with stakeholders’ high-level preferences, and it facilitates consensus formation. In our synthetic experiment, we replace stakeholder interaction with synthetic high-level stakeholder preferences, from which we derive five top-rated designs to initiate the HITL-MGA workflow. This way, we can test the outputs of the HITL-MGA workflow against the initial high-level preferences in a way that would not be possible if interacting with real-world stakeholders whose high-level preferences are typically unknown.
Fig 2.
Stylised representation of the SPORES algorithm, in its standard (a) and HITL (b) formulations.
Starting from the cost-optimal solution, the algorithm iteratively looks for additional feasible and only marginally costlier system designs in parallel batches. The standard formulation (a) foresees S parallel batches that generate alternative designs based on the sole objective of making them spatially and technologically distinct from previously found solutions. Each of these S batches targets decision variables pertaining to a specific energy sector, such as power, heating, mobility or synthetic fuels. At the same time, 2xM additional batches generate (in parallel) alternatives around the intensification of specific features of the design space, for instance, the very high or very low deployment of offshore wind power overall. In the HITL formulation (b), each parallel batch is informed by a high-level stakeholder preference M and intensifies one or more technical features that underlie such a preference while also targeting spatial and technological distinctiveness from previously found solutions. One additional batch (M+1) targets all the given high-level preferences simultaneously.
Fig 3.
Results of the automated decoding of distinctive technical features in each top-rated system design based on comparison with the statistical distribution (represented by strip plots) of technical system features across the entire design space.
In the figure, the reliance on each technical feature is normalised with respect to the maximum deployment of the given feature experienced across the entire design space. ‘Desired’ features are highlighted in blue, whilst ‘undesired’ ones are highlighted in red. In this case, the threshold for detecting a given feature as desired or undesired is set to a 15% deviation from the distribution’s mean before normalisation.
Fig 4.
Comparison of the original design space generated via standard MGA and the one generated with the HITL-MGA workflow in terms of percentage of designs matching the assumed high-level stakeholder preferences.
HITL-MGA results are shown both for the default case of a 15% threshold in the decoding procedure and for the case of a stricter (10%) threshold.
Fig 5.
Solution space generated by the original simple MGA and the HITL-MGA workflows from the perspective of technical (system-wide capacity deployment of a given technology) and social system features of stakeholder interest.
We show the designs (x-axis) by their performance on a given metric compared to the maximum experienced across both design spaces for the same metric (y-axis). To ensure comparability, results for each metric are re-scaled between 0 and 1 based on the minimum and maximum values experienced across either design space (an approach also known as min-max normalisation). In each design space, the designs highlighted in red are those that we define as ‘near-highest-consensus’. They are within 25% of the system design option that, across both design spaces, ensures the best performance across all synthetic high-level preferences (see sub-section 2.1).
Fig 6.
Spatial deployment of renewable energy generation capacity, system-wide deployment of electrolysers and degree of reliance on imports in the system’s primary energy supply.
The results are shown for two illustrative system design options, namely the options – in the original and in the human-trained (HITL-MGA) option spaces – with the best performance in terms of limited regional concentration of wind farms among the subset of overall best-performing, or ‘near-highest-consensus’, options. The geodata source to produce the maps is GISCO-Eurostat [24].