A Novel Prediction Method for Smart Meter Error Using Multiview Convolutional Neural Network
Xuan Tong, Jun Ma, Li Ma, Sen Yan, Qiu Hua Tang, Zhaosheng Teng, Cheng Da · IEEE Sensors Journal · 2024
The accurate prediction analysis of smart meter (SM) error is highly important for optimizing electricity energy management and promoting grid development. However, most studies ignore the effects of environmental stresses, leading to insufficient feature information and limited prediction performance. To address this problem, we propose a novel fusion framework for the prediction analysis of SM error based on a MVCNN. First, a feature fusion structure is proposed where the time, temperature, and humidity information are effectively extracted and integrated. Then, the 1-D composite convolution is further presented to enhance the feature diversity using the depthwise separable convolution (DSC) and dilated convolution (DIC). In addition, batch normalization (BN) is deployed for the adjustment of feature distribution, which can speed up the training and avoid overfitting. Finally, various experiments are conducted based on the measurement error dataset of SMs from high dry heat areas in China to validate the proposed MVCNN. The root mean square error (mse) of prediction is 0.0311, and the mean absolute error of prediction is 0.0163. Compared with some advanced prediction models, the MVCNN can obtain superior multisource feature fusion ability and excellent prediction performance, which can contribute to device selection and improvement under extreme environments.