Total organic carbon（TOC）is one of the indicators for evaluating the organic matter abundance and hydrocarbon generation potential of hydrocarbon source rock. In this paper，taking the cored well of shale oil reservoir in Niuzhuang Depression，Dongying Sag as an example，the TOC-related sensitive parameters were obtained through cross plots based on the experimental analysis of TOC from core and logging data. The conventional TOC calculation models for continental shale oil reservoir，namely the ΔlogR method and the multiple regression analysis method，were used to predict the TOC of well and lake facies shale oil reservoir in the study area，but the correlation and performance were not good. Therefore，this paper proposed combining machine learning models，i.e.，the principal component analysis（PCA）model and an improved convolutional neural network（CNN）model，to form the PCA-CNN model. In this model，the PCA model was employed to reduce the dimensions of data and remove redundant information and noise information，and then，the CNN model was used to predict the TOC of shale oil reservoir，which could improve sample data quality and prediction accuracy of TOC. The PCACNN model was applied to predict the TOC of six cored wells shale oil reservoir in Niuzhuang Depression，and the results reveal that for continental shale oil reservoir，the proposed model can accurately predict TOC，and the compliance rate is up to 96%.