Reservoir classification and evaluation are critical during the oil and gas exploration and production, and traditional production forecasting often require complex methods such as reservoir modeling, which is time-consuming and will delay the E&P process. With the rapid development of artificial intelligence technology, the production forecasting method based on machine learning has become a new hot spot for research. Since the characteristics of various reservoirs are different, it is difficult for conventional machine learning models to meet the required accuracy due to a large number of unknowns. In this paper, we proposed an improved hierarchical XGBoost (h-XGBoost) modeling method, which can effectively extract the features of different reservoirs and build a hierarchical data model. The study is conducted on tight conglomerates reservoir at the Mahu area in Xinjiang, and the results verify the effectiveness of the method.