Overcoming data sparsity in industrial energy-storage systems: A dynamic graph attention approach

Accurate short-term energy consumption forecasting is critical for optimizing industrial production scheduling and achieving low-carbon objectives. However, real-world industrial systems often face challenges such as sparse data and complex spatiotemporal dependencies. These factors severely limit the performance of traditional prediction models. To address these issues, a deep learning framework named RG-BiAtt is proposed. This model integrates a residual graph convolutional network with a bidirectional attention mechanism. A parallel spatiotemporal architecture is employed to decouple spatial and temporal feature extraction. Additionally, a dynamic graph construction strategy combines Pearson Correlation Coefficient and Transfer Entropy to capture linear and nonlinear dependencies without predefined topologies. Experimental results on real-world industrial datasets demonstrate that RG-BiAtt outperforms ten state-of-the-art baselines. The proposed model achieves a Root Mean Square Error reduction of up to 44.9%. Stability analysis further confirms that the model maintains high accuracy even under coarse-grained forecasting conditions. In terms of practical deployment, the framework is estimated to reduce annual energy costs by over 788600 USD. Furthermore, it facilitates a reduction in carbon dioxide emissions of approximately 3254 tonnes per year. These findings confirm the value of RG-BiAtt for sustainable smart manufacturing. • Proposes RG-BiAtt framework for sparse industrial energy forecasting. • Dynamic PCC-TE graph captures dependencies without predefined topology. • Reduces RMSE by 44.9% against SOTA models in coarse-grained forecasting. • Saves $788600 and mitigates 3254 tonnes of CO 2 emissions annually.