The ash control of the traditional remedia sorting process is limited by the lack of process models, and it is difficult to make breakthroughs. Although the mechanism modeling made by some studies can explain the physical parameters of the sorting system in the model, for the re-intermediate sorting control process, the mechanism modeling covers the process insufficiently comprehensively and the range of process parameters is difficult to define. After the introduction of high-precision ash analyzers in industrial sites in recent years, this problem has mainly focused on reducing ash measurement delays. Based on the random forest algorithm (Random Forest) and empirical mode decomposition, the density, magnetic content and ash data measured in the industrial site are denoised, and the mathematical model of the heavy medium sorting system is established. In order to solve the problem of high ash delay in the re-intermediate sorting process, a pre-correspondatory method for ash separation is proposed: the input at t moment (density value m magnetic content value M) corresponds to the output (ash value h) at the moment of t + T (T is the delay time) for model training. After comparing and optimizing BP neural network, random forest algorithm and algorithm based on least squares principle, it is finally concluded that the random forest algorithm has the best modeling effect. Then, the random forest estimate can be used as a guide value for coal sorting industrial sites, which is helpful to improve the efficiency of heavy medium sorting and improve the quality of refined coal.