The balance of human energy intake and energy consumption is one of the standards for maintaining health. Unbalanced intake may cause consequences such as cell damage and obesity. The estimation of energy intake is of great significance to human health management. The current method of assessing energy intake is mainly through dietary review, but it is time-consuming because of increasing the burden of the person to be evaluated. Therefore, developing a simple and fast way to estimate energy intake is urgent. After energy intake, metabolites generated by digestion and metabolism are excreted as waste. Wastes such as urine, etc., contain many chemical species, which can systematically reflect the dietary status and disease processes. This research aims to establish a classification model based on SERS techniques, which is highly sensitive, non-destructive, and identifiable molecular fingerpring. Peak statistics, and unsupervised and supervised clustering algorithms are utilized to analyze SERS data collected from volunteer groups of energy intake with 1 500, 2 030, 2 700 kcal·day-1. Since there is a certain amount of overlapping of Raman peaks in many organic molecules, it is difficult to analyze and assign SERS peaks. This study adopts a comparative analysis of an unsupervised PCA and a supervised OPLS-DA algorithm for classification and prediction. It was found that the scattering distribution of different categories in PCA has a large extent, so the model shows poor categorization. After correcting the background by first-order derivative difference, the scatter map presents the classified trend. The OPLS-DA algorithm can decompose the X matrix information into the Y-related and unrelated two components by presetting the Y’s label to achieve good classification after orthogonal signal correction processing. The results show that the OPLS-DA algorithm can be well-classified for three or each two different energy intake levels. Both the specificity and accuracy of the ROC analysis have reached 100%. The permutation test of 200 times also illustrates the model with good accuracy and predictability. The results indicate that the levels of human energy intake can be directly estimated by analyzing the SERS signal of the urine. This method can rapidly analyse urine in 2 minutes with simple manipulation and accurate discriminant results, which shows great potential in medical health applications.