Dynamic Association Modeling of Industrial Power Intelligence-Demand Based on Multimodal Feature Alignment

Abstract

Modern industrial power systems as well as integrated energy systems face the crucial and demanding challenge of power demand forecasting. Because of this, improvements in the accuracy of power demand forecasts are severely impeded. An industrial power demand model that uses machine learning and incorporates several kinds of data (such as weather, production, and economic indicators) produces more accurate and thorough power demand predictions compared to systems that rely on a single model. To better capture the complex elements driving industrial power consumption, this strategy integrates diverse data sources and analytical approaches. The result is enhanced forecast accuracy and stability, leading to a more dependable power system. To accurately forecast power use in the near future, this research suggests a hybrid deep learning model that combines Gated Recurrent Units (GRU) and Long Short-Term Memory (LSTM) units. The model takes into account generation and consumption data from the past to represent both the short-term variations and the long-term relationships in power use using SHapley Additive exPlanations (SHAP) model. We ran many trials and used industry-standard measures like R2, MAE, MSE, and RMSE to assess the model's performance. Having an R2 score of 0.9902, a MAE of 0.0124, or RMSE of 0.0187, the suggested SHAP-GRU-LSTM model outperformed solo GRU and LSTM as well as many benchmark models found in the literature.