Carbon Price Prediction Based on Comprehensive Feature Selection and WOA-Driven MultiHead-Attention-BiGRU Model

Accurate forecasting of carbon prices is essential for improving the allocation of carbon resources and promoting the growth of carbon markets. To address prediction accuracy challenges, this research proposes an integrated forecasting framework combining comprehensive feature engineering with enhanced deep learning techniques. The methodology systematically implements three innovative components. First, technical indicators are derived from raw market data to construct an expanded feature set that captures complex market patterns. Second, a hybrid feature selection approach integrating extreme gradient boosting (XGBoost), and LASSO algorithms is developed, with shapley additive explanations (SHAP) value analysis incorporated to ensure model transparency and interpretability. Third, an optimized gated recurrent unit (GRU) architecture enhanced by multi-head attention mechanisms (MAM) is established, where whale optimization algorithms (WOA) automatically tune hyperparameters to maximize prediction performance. Experimental results across three carbon markets demonstrate the framework’s superior accuracy compared with seven benchmark models. The proposed model achieves mean absolute percentage error (MAPE) values of 0.515, 0.444, and 1.016 in respective markets, showing significant improvements over alternative approaches. These findings not only provide market participants with reliable decision-making references but also contribute methodologically to time-series forecasting research.