Data Drift Detection and Assessment for AI-hybrid Models Applied on Electrical Energy Consumption

Abstract

Data drift evaluation is crucial in the operational step in the industry. In the real world, several drift types are usually contributing to drift detection, which may come from input data and output data distributions. In addition, the application context and the interpretation of these drift types add complexity to drift analysis. In this work, we apply drift detection in the specific domain of electrical transmission network systems. Three drift types, covariate, label, and concept drift, are considered and implemented on systems based on Physics-Informed Neural Networks (PINNs). The experimental results show the impact of each drift type and the evolution of their contributions when drift occurs in the industrial system. A contextual interpretation of the obtained results is also developed in this specific application domain for the three drift types.