Accurate electricity consumption forecasting for industrial energy management

• A novel three-step methodology combining signal decomposition, feature selection, and DL optimization for TSF. • The integration of BO-VMD for adaptive signal decomposition, which automatically optimizes VMD parameters to improve predictive quality. • The application of RFR-SHAP for objective and data-driven feature selection, effectively reducing dimensionality and enhancing model focus and interpretability. • Design a hybrid Bi-NOA-LSTM model to capture temporal patterns with improved representation power, applying SOMA to optimize hyperparameters for better forecasting performance. • Empirical validation of the proposed integrated approach on real-world industrial datasets, showcasing superior forecasting performance and robustness. Accurate electricity consumption forecasting is crucial for effective Energy Management Systems (EMS) in industrial environments. This study introduces a robust, three-stage methodology for Time Series Forecasting (TSF) of industrial electrical energy consumption, designed to enhance predictive accuracy and interpretability. First, a novel hybrid approach, Bayesian Optimization-Variational Mode Decomposition (BO-VMD), is employed to adaptively decompose the raw signal into optimal components by automatically tuning VMD's parameters. Second, a Random Forest Regressor (RFR) combined with SHapley Additive exPlanations (SHAP) is utilized for effective and justifiable feature selection, identifying the most relevant variables and improving model efficiency. Finally, a sophisticated forecasting model, SOMA-Bi-NOA-LSTM, is proposed. This model leverages a Multi-Input/Multi-Output (MIMO) architecture, a Bidirectional (Bi) structure to capture long-range dependencies, and Non-Output Activation Long Short-Term Memory (NOA-LSTM) layers for better information preservation. The SOMA optimization algorithm is applied to adjust the model's hyperparameters. Tested on real industrial electrical consumption data with a 1-hour forecasting horizon, the proposed BO-VMD-RFR-SHAP-SOMA-Bi-NOA-LSTM model demonstrates superior predictive accuracy, achieving a Mean Squared Error (MSE) of 0.429 kW², a Root Mean Squared Error (RMSE) of 0.655 kW, a Mean Absolute Error (MAE) of 0.505 kW, and an R² of 0.99977. These results underscore the model's significant potential for enhancing EMS efficiency and optimizing industrial energy management.