Estimating method for power marketing data based on multi-scale convolutional variational autoencoder

Anomalous power marketing data can significantly impact the interests of various stakeholders in the development of power enterprises. To improve the detection of anomalies in such data, we propose a method based on the Multi-Scale Convolutional Variational Autoencoder (MSCVAE). First, a multiscale attribute matrix is constructed to represent the system's state at different time steps within the multivariate time series of power marketing data. Next, a convolutional variational autoencoder is employed to generate a reconstruction matrix from this attribute matrix, while an attention-based ConvLSTM is used to capture temporal patterns. To address the class imbalance issue inherent in the power marketing dataset, a novel threshold-setting strategy based on the confusion matrix is introduced. The proposed model is evaluated on two real-world power marketing datasets and compared with several benchmark models. Experimental results demonstrate that the MSCVAE outperforms the comparison models in terms of recall rate, F1 score, and G-mean, and exhibits superior performance and robustness in detecting anomalies in power marketing data.