High-accuracy and unsupervised feature extraction for smart grid transformer loss prediction based on deep belief networks

Abstract

Addressing the limitations of existing transformer loss prediction techniques for smart grid, which rely heavily on labeled data and exhibit weak feature extraction capabilities, we propose a DBN-BP integrated algorithm combining deep belief networks and Back propagation neural networks for transformer loss prediction. Theoretically, three sub networks with tanh, LeakyReLU, and sigmoid activation functions perform unsupervised feature extraction, subsequently integrating the outputs through a back propagation neural network. During experimentation, historical transformer loss data and corresponding meteorological conditions served as training inputs to generate predictions. Comparative analysis is conducted using six algorithms. Experiment results indicate DBN-BP demonstrates superior real-time performance with 0.85 ms, outperforming TFT with 1.48 ms and Informer with 1.18 ms. When the training sample size is reduced to 50 instances, DBN-BP maintained the lowest absolute 0.56% error, representing a 26.1% reduction compared to 0.92% error for LSTM. When sample size increased to 500 instances, DBN-BP absolute error further decreased to 0.28%, consistently outperforming other algorithms. It demonstrates that the proposed algorithm requires minimal labeled data, autonomously extracts high-dimensional robust features, and shows significant potential for engineering applications in transformer loss analysis of smart grid.