Optimized variational mode decomposition–bidirectional long–short-term memory framework for short-term energy load forecasting

This study proposes an optimized variational mode decomposition (VMD)–bidirectional long–short-term memory (BiLSTM) framework for short-term energy load forecasting under non-stationary and noisy conditions. VMD is used to decompose complex energy time series into stable intrinsic mode functions, while a BiLSTM network is used to capture bidirectional temporal dependencies. Bayesian optimization is adopted to systematically tune key hyperparameters, and permutation-based feature importance analysis is incorporated to refine the input feature space. Experimental results indicate that the proposed framework consistently outperforms benchmark BiLSTM and empirical mode decomposition–BiLSTM models, achieving substantially lower forecasting errors and an R2 value close to unity. These findings demonstrate that the integration of signal decomposition, feature selection and optimized deep learning architectures provides an effective and robust solution for short-term energy consumption forecasting in smart grid and industrial energy management applications.