SSA-Optimized CNN-BiLSTM-Attention Hybrid Model for Electricity Load Forecasting

Abstract-Accurate short-term power load forecasting is pivotal for optimizing energy dispatch, ensuring grid stability, and achieving carbon neutrality. Traditional models struggle with spatiotemporal feature extraction and hyperparameter optimization, while deep learning models often face suboptimal generalization due to manual tuning and sensitivity to renewable fluctuations. This study proposes a hybrid CNN-BiLSTM-Attention model optimized by the Sparrow Search Algorithm (SSA)to address these challenges. The model integrates multi-scale feature learning:CNN captures spatial correlations in meteorological and load data, BiLSTM extracts bidirectional temporal dependencies, and an attention mechanism dynamically weights critical time steps.SSA enables dual optimization of both model architecture and hyperparameters, avoiding local optima and enhancing convergence. Attention weight visualization improves interpretability by quantifying the impact of factors like temperature and historical load on predictions. Experimental validation using Australian smart grid data demonstrates superior performance:the SSA-optimized model achieves a coefficient of determination ($R^{2}$)of 0.9827, a 1.29% increase over the unoptimized hybrid model, and a mean absolute percentage error (MAPE)of 1.8313%, a 24.6% reduction in prediction errors. It also shows robust performance under extreme weather and renewable energy fluctuations, highlighting its utility in high-RES penetration environments. This research provides a robust framework for intelligent load forecasting, facilitating data-driven grid scheduling and carbon-aware energy management. The SSA-driven optimization strategy offers a novel approach for enhancing hybrid deep learning models, paving the way for more efficient and sustainable power systems.