Multi-layer and multi-source features stacking ensemble learning for user profile

Abstract

Electricity bill-sensitive user profiling in the power industry is gradually being recognized as a research hotspot. A multi-layer feature construction method has been proposed, separating the mining of textual and numerical information, addressing the insufficient exploration of textual data in existing user profile processing methods. The complexity of electricity user profiles is addressed through the introduction of a two-stage predictive model based on Stacking ensemble learning. In the first stage, user sensitivity is predicted by utilizing the advantages of MLP (Multi-Layer Perceptron), CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory), and XGB (Extreme Gradient Boosting) in global, local, and missing value handling, respectively. In the second stage, electricity-sensitive users are identified by employing RF (Random Forest). The experimental results show that the MLS-SEL user profile model is higher than the models MVEM, SG and SMUPM in terms of both F1 value and accuracy rate. It is implied that users who could be more sensitive to fluctuations in electricity costs have been identified more accurately.