Attention-Driven Hybrid Ensemble Approach With Bayesian Optimization for Accurate Energy Forecasting in Jeju Island’s Renewable Energy System

Abstract

The rapid integration of renewable energy sources into power grids has created an urgent need for accurate energy demand and supply forecasting models capable of managing the inherent variability of renewable energy generation. The combination of fluctuating consumer demand patterns and high variability across different energy sources presents significant challenges in maintaining a reliable balance between supply and demand. To address these challenges, we propose a Attention-driven Bayesian-Optimized Hybrid Ensemble Framework (ABHEF), evaluated on Jeju Island’s energy mix data. ABHEF integrates state-of-the-art models—ConvBiLSTM (Convolutional Bidirectional Long Short-Term Memory), ETCN (Enhanced Temporal Convolutional Network), TFT (Temporal Fusion Transformer), and DAT (Dual Attention Transformer)—to capture both short-term fluctuations and long-term trends in energy data. The proposed framework is evaluated on actual energy demand and supply data from Jeju Island, along with key weather attributes, thereby enhancing the model’s real-world applicability and accuracy. Bayesian optimization was applied to each model to determine optimal hyperparameters, to ensure the peak predictive performance. The output of the base models was stacked, and four meta-models (Gradient Boosting, LGBM, Ridge, and CatBoost) were applied. Among these, CatBoost demonstrated the best performance and was selected as the final meta-model. For hourly supply prediction, our selected meta-model achieved a 52% reduction in MAE and a 50% reduction in RMSE compared to BiLSTM, the best-performing standalone time-series model, validated through a consistent evaluation of accuracy metrics across all models on the same dataset. For hourly demand predictions, it achieved a 43% reduction in MAE and a 34% reduction in RMSE. For daily supply predictions, it demonstrated a 76% reduction in MAE and a 77% reduction in RMSE, while for daily demand predictions, the reductions were 70% in MAE and 69% in RMSE. These results highlight the superior accuracy of the proposed framework, offering significant benefits for energy management and resource planning in renewable energy systems.

Keywords

• Hybrid deep learning
• renewable energy forecasting
• Bayesian optimization
• time-series forecasting
• meta-model stacking