A Multitask Framework for Optimizing Smart Grid Energy Consumption Using RegClassXNet and Dynamic Cluster Adjustment

With urbanization accelerating electricity demand, advanced energy management is vital for building sustainable cities. This paper presents a novel framework using real-world data from commercial, residential, and industrial buildings over six years to optimize electricity consumption. The proposed RegClassXNet model, integrating EfficientNet, Xception, and Swin-Transformer, performs multitask predictions for both classification and regression objectives. Proportional Dynamic Cluster Adjustment (PDCA) is introduced to address data imbalance, and a hybrid attribute refinement process synthesizes relevant features to enhance predictive accuracy. Our model achieves 95.0% R-squared, 2.1 MAE, 1.8 RMSE, and 3.2 MSE, significantly outperforming existing methods such as CNN, LSTM, and RF. New stability metrics, including Label Variability Consistency Index (LVCI), Temporal Prediction Stability Measure (TPSM), and Output Correlation Coefficient (OCC), ensure robust and consistent predictions. The framework was evaluated on a dataset comprising 500,000 energy consumption records, utilizing a distributed training approach on a high-performance GPU cluster. Simulations illustrate the framework’s capability to optimize energy usage across building types, adjust for environmental impacts, and support effective energy-saving strategies. This work offers a transformative approach to sustainable energy management, paving the way for adaptive, data-driven smart grid systems.