Deep learning surrogate model to explain decision variable synergies in energy system modelling

Abstract

Energy system models are powerful tools for planning the low-carbon transition. They identify optimal technology portfolios but rarely explain why certain combinations outperform others. Understanding decision variable synergies across multiple sectors remains a key gap in the literature. This study addresses that gap with a two-stage methodology applied to the Piemonte region (Italy) for 2050, using 17 decision variables and two objective functions (total annual costs and CO₂ emissions). In the first stage, an expert-based analysis of six electrification technologies is conducted using the EPLANopt model. It reveals a fundamental behavioral dichotomy: heat pumps reduce CO₂ independently of grid decarbonization, while Power-to-X technologies require an effective renewable share above approximately 50% to yield a net environmental benefit. In the second stage, a deep learning surrogate model trained on thousands EnergyPLAN simulations enables systematic sensitivity and interaction analysis across all 17 variables. Explainability artificial intelligence methods identify solar capacity and synthetic gas as the dominant variables. Pairwise analysis across all 136 variable combinations reveals a dominant PV surplus–Power-to-Gas valorisation chain. Third-order analysis shows that battery storage acts as a critical mediator between PV generation and Power-to-Gas conversion, producing emergent synergies that cannot be decomposed into pairwise effects. These findings provide a physically interpretable, data-driven basis for sequencing technology deployments in regional energy planning.