A deep learning architecture for energy service demand estimation in transport sector for Shared Socioeconomic Pathways

Abstract

Abstract Meeting current global passenger and freight transport energy service demands accounts for 20% of annual anthropogenic CO2 emissions, and mitigating these emissions remains a considerable challenge for climate policy. Pursuant to this, energy service demands play a critical role in the energy systems and integrated assessment models but fail to get the attention they warrant. This study introduces a novel custom deep learning neural network architecture (called TrebuNet) that mimics the physical process of firing a trebuchet to model the nuanced dynamics inherent in energy service demand estimation. Here we show, how TrebuNet is designed, trained, and used to estimate transport energy service demand. We find that the TrebuNet architecture shows superior performance compared with traditional multivariate linear regression and state of the art methods like densely connected neural network, Recurrent Neural Network, and Gradient Boosted machine learning algorithms when evaluated for regional demand projection for all modes of transport demands at short, decadal, and medium-term time horizons. Finally, TrebuNet introduces a framework to project energy service demand for regions having multiple countries spanning different socio-economic development pathways which can be replicated for wider regression-based task for timeseries having non-uniform variance.