The monitoring and fault prediction of power substations is one of the most important issues for critical nodes in future smart grid. An accurate prediction and real-time evaluation of the operational status for substation equipment can improve the stability and reliability of the power grid, reduce fault risks, and ensure the safe operation of the power system. In practice, however, the challenges such as the vast amount of data generated, the complexity of equipment interactions, and the need for timely and accurate fault detection pose significant hurdles. A novel method is proposed for state prediction and real-time evaluation of substation metering systems in this paper. It aims to predict state quantities and evaluate actual performance based on these predictions. By systematically sampling and preprocessing environmental and metering data, applying advanced clustering and imbalance learning techniques, and utilizing a decision tree-neural network hybrid algorithm, the proposed method addresses these challenges and enhances the accuracy and reliability of substation monitoring.