Model linking

Model linking typically involves the exchange of a limited number of outputs between the models and iterating the linked model system until a convergence criterion of some kind has been reached. The linking exercise may also include some harmonisation, but this is rarely exhaustive.

While providing clear benefits for analysis, model linking also implies trade-offs and creates new problems to be solved. The more technical considerations range from deciding e.g. how to treat inconsistencies in temporal, spatial and variable definitions, system boundaries (including significant background assumptions) and model overlaps, different foresights for the modelled decisions, level of harmonisation, convergence criteria and technical implementation of the data exchange. There often are no easy, obvious solutions to many of these, and several of such problems are likely to be present in most linking exercises. As compromises will thus need to be made, understanding the implications of the specific ones made is of key importance. 

This leads to the second area of trade-offs; the changing interpretation of the model outcomes. Individual models have been built internally consistent, following e.g., a specific foresight and decision-making rationale that can be fairly easily explained also outside the model context. For the linked model system this is not necessarily true, as the model linking often breaks the internally consistent decision-making rationale of individual models. This is further complicated by the compromises made in the linking, as they will also feed into the results.

As part of IAM COMPACT, we have produced a blueprint for model linking in energy-climate-economy modelling science, including technical and conceptual challenges and best practices, published as a peer-reviewed paper (here).

In that, we we propose a general checklist that can be used for making initial decisions about linking models. As no two models—and thus no two linking tasks—are identical, this checklist is intentionally general: the questions must be addressed and the recommendations interpreted within the context of the specific linking task at hand. When relevant, we acknowledge different requirements and considerations in case the establishment of model links is strategic and long-term rather than a circumstantial, one-off activity—acknowledging that the former case requires additional considerations and guardrails. With that said, the one-off suggestions are also a minimum baseline for the strategic linking recommendation, when no new advice is offered for the same topic in the latter. We also refrain here from discussing the specific linking approaches (hard linking, bi-directional and unidirectional soft linking, etc.), as these have been previously discussed extensively.

The recommendations included in this checklist assume that the trade-offs associated with the model linking process have been considered and deemed beneficial. This means that the level of robustness required for the linking process has been considered, to avoid compromising the integrity of the results or the benefits that can be achieved; this is especially important for long-term/strategic model linking, which requires considering the level of added complexity, the increase in opaqueness of model dynamics, and the volume of applications that truly need the explicit representation of the cross-model dynamics. Not linking the models should be kept as an option, as the above conditions will not be fulfilled for many model linking possibilities.

Linking across temporal and spatial scales

• If the aim of the linking is not to increase the spatial or temporal granularity of the analysis, and there is an option to link models of similar temporal and/or spatial scope or granularity, then practitioners should prioritise linking such models.
• Model linking explicitly aiming to increase the spatial or temporal granularity of the analysis requires significantly more robust approaches than if the core purpose of the linking exercise is to simply overcome the spatial and temporal differences that exist in two models linked for other purposes.
• Upscaling and downscaling approaches (e.g. clustering) can be useful for moving data from one model to another – but with limitations (see below).
• If models with different foresight assumptions are linked, attention should be paid to the design and process of the data exchange.
• [Strategic long-term applications] Limitations to upscaling and downscaling due to important relative positions of data points (in time or space) could render more extensive changes necessary for at least one model
• [Strategic long-term applications] The extent to which the spatial and temporal dimensions of the models to be linked can be harmonised should be considered, and the model(s) modified accordingly.
• [Strategic long-term applications] As computational power may be the core constraint, a preliminary simplified test application could offer important insights into the sensitivity of model results to the attained increase of detail, and thus to the added value of engaging in the model linking process to begin with.

• Common assumptions, key variable definitions, and underlying scenario narratives must be harmonized to the extent possible. When interpreting results, the lack of full harmonisation across all model assumptions must be transparently reflected in the analysis. This includes paying close attention to how variables have been defined in the models to be linked, to ensure the definitions do not differ.
• Detailed model documentation should be developed and then used during the model linking, to better understand the model assumptions made for model structure, variables and parameterisation. Adequate time within the linking process should be devoted to discussing such assumptions and definitions across modelling teams.
• [Strategic long-term applications] A full mapping of assumptions (explicit and otherwise) and their drivers for the models to be linked can facilitate the process of harmonising assumptions and variable definitions, ensuring consistency across both explicit and background assumptions across the linked model system, and legitimising the produced model framework and its results.

• Standardisation (e.g., of data templates and automating processes) can be decisive. The use of dedicated tools developed by the energy-/climate-economy modelling community can help with data exchange, as well as add transparency for variable definitions, model structure, and assumptions.
• [Strategic long-term applications] Using standardised interfaces or APIs may be more resource-intensive but is highly beneficial, as it can reduce errors and provide a distinct part of the linked model system.

• Before linking, the epistemic foundations of the models must be mapped to assess where the rationale of the models may be in conflict; if the areas appear critical, and serious inconsistencies unavoidable, the usefulness of the model linking activity may need to be reconsidered.
• Documenting the linking process itself can contribute to understanding how the model results should be understood, as well as to legitimising the exercise, in terms of both scientific rigour and policy credibility.
• Practitioners may need to assess how critical data exchange points (e.g., inconsistencies between variable definitions, or compromises made to facilitate data exchange) may introduce errors as well as how these errors may propagate, especially when two-way feedbacks are considered between the models (as the latter increase non-linearity and variability); this will help them explicitly discuss and highlight potential issues when interpreting model results.
• A separate interpretation stage should be included to explicitly consider and discuss how the potentially different model rationales and various compromises made during the model linking should be interpreted for the results. Special attention should be paid to the emergent properties from the linked system, in terms of whether they can be validated against empirical data.
• [Strategic long-term applications] In the light of the epistemological points discussed, for strategic model linking the inconsistencies should be minimized a priori, even at the cost of having to alter one or more of the models – or changing the models to be linked. It is inadvisable to create “permanent” models for which the interpretation of results has obvious ambiguities.