A diffusion-model-based approach for forecasting energy demand in New Zealand’s transport sector

Accurate forecasting of energy demand in the transportation sector is critical for effective policy-making and resource management. However, predicting energy consumption under the influence of abnormal events that can cause significant changes in travel behaviours, such as natural disasters, pandemics, and wars, poses significant challenges. This is because existing approaches predominantly rely on historical tendencies without incorporating context-specific information. To address this gap and enhance forecasting accuracy, this study introduces a model that integrates a dependency-based mechanism with context-specific information, utilizing key variables such as crude oil prices, GDP, population statistics, overseas tourist arrivals, and vehicle kilometres, to predict end-use energy consumption in transportation, including periods affected by abnormal events. The model optimizes dependencies among these variables, leveraging a Graph Recurrent Unit (GRU) and a Denoising Diffusion Probabilistic Model (DDPM) to effectively capture temporal dynamics and provide robust probabilistic predictions. While this model can be applied to any country, in this paper, a case study focuses specifically on New Zealand’s transportation sector to demonstrate the model’s ability to improve yearly energy demand forecasts. The model’s performance is evaluated against various benchmarks, including both classic and AI-based forecasting methods, showing superior results, particularly in scenarios affected by the COVID-19 pandemic. Finally, the study offers insights into future energy demand in transportation in New Zealand under both persistent and recovering after the pandemic condition.