@Article{ee.2026.085666,
AUTHOR = {Saud Aljubairi, Mahbub Hassan, Hridoy Deb Mahin, Md Ashequl Islam, Md Ehtesamul Haque, M M Hafizur Rahman},
TITLE = {An Integrated Framework of Explainable Artificial Intelligence and Large Language Models for Shared Mobility Energy Efficiency Prediction and Policy Synthesis},
JOURNAL = {Energy Engineering},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/energy/online/detail/27991},
ISSN = {1546-0118},
ABSTRACT = {Urban transport in rapidly growing Asian cities represents a growing share of energy consumption and greenhouse gas emissions, yet energy efficiency determinants of coexisting shared mobility modes remain poorly quantified in tropical developing-country contexts. This study develops an integrated Explainable AI and Large Language Model framework for energy efficiency prediction and evidence-based policy synthesis across four urban mobility modes (Solo Car, Shared Ride-Hailing, Motorcycle Taxi, Electric Minibus), using a 6000-record physics-grounded synthetic dataset calibrated to Bangkok and Dhaka. A key methodological contribution, a route-group holdout protocol eliminating intra-route leakage largely unaddressed in prior trip-level transport energy studies, shows XGBoost achieving the best held-out <math id="mml-ieqn-1"><msup><mi>R</mi><mrow><mn>2</mn></mrow></msup></math> of <math id="mml-ieqn-2"><mn>0.731</mn></math>, while naive random splitting inflated this figure by <math id="mml-ieqn-3"><mn>0.08</mn></math> to <math id="mml-ieqn-4"><mn>0.12</mn></math> points, confirming leakage as a substantive performance threat. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) jointly identified occupant count (mean <math id="mml-ieqn-5"><mrow><mo stretchy="false">|</mo></mrow><mi>ϕ</mi><mrow><mo stretchy="false">|</mo></mrow><mo>=</mo><mn>598.98</mn></math> kJ/pax-km), road gradient (<math id="mml-ieqn-6"><mn>304.69</mn></math>), and load factor (<math id="mml-ieqn-7"><mn>266.38</mn></math>) as the dominant energy drivers, with Spearman rank-agreement of <math id="mml-ieqn-8"><mi>ρ</mi><mo>=</mo><mn>0.29</mn></math> to <math id="mml-ieqn-9"><mn>0.80</mn></math> across modes. Electric Minibus recorded a median energy intensity of <math id="mml-ieqn-10"><mn>228</mn></math> kJ/pax-km, <math id="mml-ieqn-11"><mn>8.2</mn></math> times more efficient than Solo Car, while ride-hailing pooling simulations projected a mean energy reduction of <math id="mml-ieqn-12"><mn>38</mn><mi mathvariant="normal">%</mi></math> on affected trips. A structured three-stage LLM synthesis chain, with a boundary-risk audit, translated these findings into mode-specific policy recommendations with explicit confidence assessments. Occupancy optimisation offers greater near-term energy reduction potential than vehicle technology substitution alone, and the framework offers a replicable methodology for transport energy analysis in data-scarce settings, with extension to further cities identified as a priority for future work.},
DOI = {10.32604/ee.2026.085666}
}