A robust hybrid framework for load forecasting in smart grid

• Propose a robust hybrid framework for smart grid load forecasting. • Propose an improved method for selecting key features and resisting anomalies. • Design an improved decomposition method with adaptive noise thresholding. • Use an improved optimization algorithm based on chaotic opposition learning. In the context of the rapid development of smart grid, accurate load forecasting is imperative for ensuring reliable performance and efficient management of power systems. However, existing studies frequently concentrate on the complexity and non-linearity of load sequences while overlooking data integrity risks in temporal patterns. In order to address the aforementioned security challenges, this paper proposes an integrated deep learning framework DITBAG, which combines Temporal Convolutional Network, Bidirectional Long Short-Term Memory Network and attention mechanisms for load forecasting. Firstly, our method employs Dynamic Robust Pearson’s Coefficient to select key features and filter out potential anomalies. Then, it employs Improved Singular Spectrum Analysis with adaptive noise thresholding to decompose data for effectively separating perturbations from legitimate load patterns. Additionally, Golden Black-winged Kite Algorithm integrates with Chaotic Opposition Learning and Levy Flight Strategy for tuning hyperparameters, maximizing prediction accuracy and preventing model tampering. Comparative experiments conducted on latest benchmarks demonstrate that our model has the lowest MAE, MAPE and RMSE, and the highest R2, with average improvements of 48.85%, 51.18%, 47.51% and 0.28%, respectively. The findings demonstrate that the proposed model attains enhanced prediction accuracy and stability. Our approach significantly advances load forecasting by offering a reliable and accurate method that enhances efficient operation and management of smart grid.