Construction and Application of VS-DBN Anti-Theft Diagnosis Model Based on Neural Architecture Search

Abstract

Anti-theft work is important to the normal operation and safety guarantee of power system. But currently, many anti-theft diagnostic methods are no longer able to meet the demand for accurate detection. Therefore, the study introduced deep belief networks and improved them to address their shortcomings, designing a volume sparse deep belief network based on neural architecture search. On this basis, the construction of an anti-theft diagnosis model was studied and implemented. In the results, the hidden layers number obtained by the neural architecture search was all 3, and the fluctuation range of the number of neurons searched was [80100]. The true occurrence rate of the receiver operation characteristic curve after model processing has been effectively improved, and it is infinitely close to 1. At the same time, the region value of this curve is as high as 0.98, which is 0.52 higher than the untreated region value. In addition, the feature recognition accuracy of this model is stable at around 95%, and the highest accuracy reaches 98.26%. And the recognition loss of this model is around 0.2, with a minimum of only 0.11. This indicates that the anti-theft diagnosis model has excellent performance and can be well applied in practical anti-theft diagnosis. This provides a reliable safety guarantee for the operation of the power system.

Keywords

• Deep belief network
• neural architecture search
• anti-theft of electricity
• sparse volume
• particle swarm optimization