Optimising for the long game: methodological challenges in energy system optimisation pathways

TL;DR

This paper surveys national-scale energy system optimisation pathway studies to evaluate how methodological choices—especially foresight, horizon, aggregation, and investment dynamics—shape long-term results. Through a PRISMA-like systematic review of 330 peer-reviewed ESOM studies, it reveals a strong prevalence of perfect foresight or fully myopic designs, limited use of endogenous learning, and substantial end- and inter-milestone distortions, despite increasing short-term resolution and spatial detail. The authors document considerable transparency gaps and horizon-related biases, with horizon extensions rarely reported and many studies truncating analyses at policy years like 2050. They argue for a middle-ground approach with limited foresight, more systematic horizon design, improved interpolation/interpolation between milestones, and greater reporting clarity, to make long-term pathway insights more robust and policy-relevant. Publicly available data and code support reproducibility and future updates.

Abstract

Pathways that describe the optimal evolution of energy systems across multiple decades are important in energy system research and policy literature, with net-zero and similar climate policies being common drivers behind them. While there are many studies on aspects such as spatial and operational resolution, model features, and model transparency, there has been little attention on the methodological considerations of formulating pathway studies in mathematical optimisation terms, and how these methods have evolved over time. To address this, we conduct a systematic review of optimal pathway literature at or above national level focusing on the following: i) the implications of model foresight choices, ii) end effects and related issues that may bias model outcomes, iii) trade-offs in model resolution, and iv) investment dynamics. We showcase how modellers have dealt with these aspects in a large sample of studies spanning multiple decades, and provide recommendations to both modellers and model users on identifying issues that can bias model results and how to improve upon them. In particular, we identify opportunities to better balance long-term anticipatory planning with high operational and spatial detail in models, and to improve the communication and systematic treatment of those mathematical design choices that potentially distort model decisions across time.

Figures (17)

• Figure 1: The different horizons considered in a typical pathway exercise.
• Figure 1: Example of discount rates influencing the length of a model's decision horizon, adapted from bean_conditions_1984. a Cost trends of two mutually exclusive choices. b Cumulative discounted cost of both choices over time. c Difference between cumulative costs over time.
• Figure 2: Classification of typical foresight approaches seen in pathways. Dots represent model milestones, with white dots representing reconsidered investments. Adapted from keppo_short_2010kotzur_modelers_2021.
• Figure 2: Screening phases: relevant records (yellow) versus random relevant (blue). Figures produced using ASReview LAB version 1.6.6 developers_asreview_2025. i using Naive Bayes as classifier and a feature extractor. This phase ended after screening 66% of the sample. Screening phase ii using a two-layer neural network as classifier and a Doc2Vec feature extractor. This phase ended after screening 75% of the sample. Screening phase iii using a random sample of 50 leftover studies. No relevant records found.
• Figure 3: Simplified example of common model distortions caused by interactions between the lifetime of investments ($\mathbf{Life}_i$), foresight horizon (40 years), and milestone length ($\mathbf{ML}_m$) in a model with perfect foresight.