In this article, precipitation forecasts from the South China regional hourly updated cycle mesoscale model (GTRAMS-3 km-RUC) are verified based on observations of Fujian Province during May to September in 2017-2018. A correction approach on categorical hourly precipitation forecast using convolutional neural network (CNN) is built and trained, taking into account of issues such as sample imbalance, selection of feature and over-fitting of the system. Comparisons of test between CNN and frequency matching (FM) method on the 2017-2018 test set and the 2019 practical data set are also made for their correction effects. The results show that all these correction methods can improve the original forecast at different degrees, but underperform with precipitation over 15 mm·h-1. CNN performs better than FM, and CNN with correlation coefficient discrimination (CCD) is the most effective for heavy rainfall category. In addition, we apply two different solutions to extract input features of the neural network system. The CNN model converges faster when principal component analysis (PCA) solution is used, but it leads to over-fitting earlier and severely， which implies that the system’s generalization ability needs to be improved further. In contrast, CCD solution displays its longer promotion period and more potential state. It seems that CNN correction method improves forecast ability mainly by reducing the missing ratio for categorical precipitaition forecasts and the false alarm ratio for the rainfall and light rain forecasts.