A short-term energy consumption forecasting model for charging stations based on an enhanced DBO-optimized hybrid CNN-transformer architecture
Abstract
Accurate energy consumption prediction is crucial for the operational stability of smart grids and the integration of sustainable energy. To address the limitations of existing models in capturing the complex temporal dynamics of charging station loads, this paper proposes a novel hybrid model — Combined-DBO-Transformer. We first confirm that the Transformer architecture inherently outperforms traditional LSTM and GRU models in modeling long-term dependencies and complex temporal patterns. To fully unleash the potential of the Transformer, we introduce a significantly enhanced Dung Beetle Optimizer (DBO), which strategically integrates three advanced mechanisms: a Chebyshev chaotic map for enriching population diversity, a Golden Sine Algorithm for improving local search capability, and an adaptive t-distribution mutation for dynamically balancing exploration and exploitation. Extensive experimental evaluation demonstrates that our model achieves state-of-the-art performance, with MAE, RMSE, and MAPE as low as 0.0491, 0.0628, and 6.74%, respectively, significantly outperforming other optimizer-enhanced Transformer variants (e.g., BO, PSO, SSA). Ablation studies further indicate that the adaptive t-distribution mutation is the core driver of performance improvement, and the synergistic combination of all components is essential for achieving superior prediction accuracy, accelerated convergence, and enhanced robustness. This framework not only provides a powerful tool for energy prediction but also offers a paradigm blueprint for developing hybrid optimization-deep learning models for complex time series tasks.