Abstract: In order to solve the problem that the traditional TOC (Total Organic Carbon )content measurement method is costly and unable to obtain continuous distribution of TOC content, a statistical prediction method of TOC content is proposed.Due to the difference in lithology of the stratum, the difference in TOC content is very large.Therefore, the original log data was first clustered, and the strata of different lithologies were separated by clustering, and predictive models of TOC content were established for different strata.The correlation between the logging parameters and the TOC content was improved by the clustering method, which not only improved the accuracy of the model, but also made the model more convincing.Then the particle swarm optimization algorithm optimized the SVM model parameters, avoiding the manual parameter selection.The model was unstable, and then the SVM-RFE model with well logging parameters was established.The logging parameters were screened for each type, effectively avoiding the information redundancy and irrelevant parameters between the logging parameters.The performance of the model was reduced and the training time was increased.Finally, the SVR model was established for different formation lithology by using the optimized logging data and SOM classification results.Compared with other TOC content prediction models, the results show that the SOM-SVR model is more stable, more convincing, and the prediction error is small, the average relative error is about 6%, and the average absolute error is less than 0.2.It can be concluded that the SOM algorithm is used to cluster the different lithology strata and then establish the TOC content prediction model, which is more conducive to improve the accuracy of the model.