Multivariate rolling decomposition hybrid learning paradigm for power load forecasting

Very-short-term power load forecasting (VSTLF) is essential for supporting government planning in the transformation and expansion of power grids, as well as in the formulation of sustainable green development strategies. The traditional hybrid VSTLF method utilises future information using the one-off data decomposition method. This causes information leakage and significantly reduced the practical applicability of model engineering. Although rolling decomposition solves these issues, the prediction model shows poor performance and inefficiency in its utilisation of annual power load data. This study proposes a flexible VSTLF framework for a multivariable rolling decomposition hybrid learning paradigm to overcome these deficiencies. Specifically, a real-time rolling decomposition module based on annual power load features is utilised to adaptively denoise and strengthen the learnability of the sequences; this approach avoids using future information and improves data utilisation efficiency. Then, the main forecasting module, which integrates the long short-term memory network, attention mechanism, and ensemble learning optimised by the multiobjective dung beetle optimiser, performs point and interval predictions. Experimental results from two real-world power load datasets from China and Spain demonstrate that the proposed system achieves a mean absolute percentage error of approximately 2 % for point predictions and high-quality average interval scores for interval predictions. Performance evaluations demonstrated that the proposed optimiser can effectively obtain solutions along the Pareto frontier, outperforming traditional optimisation algorithms in terms of both convergence and diversity. This accurate and stable VSTLF system is expected to improve the efficiency of power grid planning and support effective energy management policies.