An SBOA-optimized hybrid grey-seasonal model for nonlinear energy production forecasting

Purpose To address the challenges of low accuracy and poor generalization in energy time series forecasting, which stem from its nonlinear, non-stationary, and limited-sample-size nature, this study proposes a hybrid forecasting framework. The model was designed to be highly accurate and robust, thereby enhancing the reliability of the energy production forecasts. Design/methodology/approach A hybrid forecasting framework centred on an ICEEMDAN-FANGBM-Holt-Winters (I-F-H) structure is proposed. The original energy sequence was first decomposed by ICEEMDAN into a trend sequence and multiple IMFs. The trend and low-frequency components were then modelled using the FANGBM(1,1) model, while seasonal mid-frequency and high-frequency components were modelled by Holt-Winters. The secretary bird optimization algorithm was simultaneously applied to optimize all model parameters for minimal forecast error. Findings The proposed I-F-H model consistently achieved superior performance over all benchmark models (including I-F-S, I-F-A and traditional single models) in forecasting natural gas, solar and nuclear power production using quarterly Chinese data. This advantage was evidenced by a significantly lower mean absolute percentage error (MAPE), underscoring the model's exceptional precision and stability, particularly when capturing complex seasonal patterns. Originality/value A multi-scale hybrid forecasting framework centred on the I-F-H architecture was constructed, effectively leveraging the complementary strengths of the grey model in capturing trends and the Holt-Winters method in identifying seasonal patterns. The framework was constructed according to a component-characteristic matching strategy, in which trend-dominant components are modelled by FANGBM(1,1), while seasonal fluctuation components are handled by Holt-Winters. The SBOA was employed to synchronously optimize all model parameters globally, thereby enhancing the collaborative forecasting capability of the framework. Empirical verification demonstrated that the I-F-H structure achieved higher accuracy and stability in quarterly energy forecasting, thereby providing a novel approach for predicting complex seasonal time series.