Dynamic hydrogen demand forecasting using hybrid time series models: Insights for renewable energy systems

Abstract

Hydrogen is gaining traction as a key energy carrier due to its clean combustion, high energy content, and versatility. As the world shifts towards sustainable energy, hydrogen demand is rapidly increasing. This paper introduces a novel hybrid time series modeling approach, designed and developed to accurately predict hydrogen demand by mixing linear and nonlinear models and accounting for the impact of non-recurring events and dynamic energy market changes over time. The model incorporates key economic variables like hydrogen price, oil price, natural gas price, and gross domestic product (GDP) per capita. To address these challenges, we propose a four-part framework comprising the Hodrick–Prescott (HP) filter, the autoregressive fractionally integrated moving average (ARFIMA) model, the enhanced empirical wavelet transform (EEWT), and high-order fuzzy cognitive maps (HFCM). The HP filter extracts recurring structural patterns around specific data points and resolves challenges in hybridizing linear and nonlinear models. The ARFIMA model, equipped with statistical memory, captures linear trends in the data. Meanwhile, the EEWT handles non-stationary time series by adaptively decomposing data. HFCM integrates the outputs from these components, with ridge regression fine-tuning the HFCM to handle complex time series dynamics. Validation using stochastic, non-Gaussian synthetic data demonstrates that this model significantly enhances prediction performance. The methodology offers notable improvements in prediction accuracy and stability compared to existing models, with implications for optimizing hydrogen production and storage systems. The proposed approach is also a valuable tool for policy formulation in renewable energy and smart energy transitions, offering a robust solution for forecasting hydrogen demand.