Research on power consumption prediction of hot-rolled strip based on transfer ensemble learning

Energy, has a huge impact on both industry and daily life, making its prediction and consumption planning crucial. Given the challenges posed by limited data samples, high dimensionality, strong correlation, and low long-term model prediction accuracy in current power consumption forecasting for hot-rolled strips, this paper proposes a power consumption prediction model based on online multi-transfer learning optimization. In this work, data samples were effectively reused through sample migration. Moreover, sub-data sets were trained in parallel, forming GRU weak learners, and models among multiple weak learners were reused through the model migration method. The weak learning buffer pool strategy screened out several weak learners with the best prediction results. A continuously updated learning was then employed to establish strong learner models, which enables real-time prediction of power consumption data. The model also introduces a threshold early warning method for fault detection, enabling real-time prediction and providing a basis for intelligent energy management. Result shows that this work addresses the limitations of power consumption prediction due to limited samples and high coupling.