Swarm Intelligence-Enhanced TimesNet for Adaptive Residual Energy Prediction in Power Communication Networks

Predicting residual energy in communication nodes of private power networks is crucial for maintaining stable power grids, yet current methods often fail to capture the complex relationships between periodic features and sudden energy fluctuations in power time-series data. This article proposes a novel prediction model integrating TimesNet and a swarm intelligence-based adaptive attention reweighting (SI-AAR) mechanism. The model employs channel-independent slicing to encode heterogeneous data, extracts dynamic patterns through period matrix reconstruction and Inception convolution, and dynamically allocates attention weights using neighboring node information via the SI-AAR module to enhance spatial anomaly detection. Experimental results on real datasets demonstrate that the proposed method reduces Mean Squared Error (MSE) and Mean Absolute Error (MAE) by 8.18% and 11.19%, respectively, compared to TimesNet. Ablation studies highlight the SI-AAR module's significant contribution, achieving a 7.66% reduction in MSE.