Recently， Approximate Query Processing （AQP） of aggregate functions is a research hotspot in the database field. Existing approximate query techniques have problems such as high query response time cost， high storage overhead， and no support for multi-predicate queries. Thus， a deep autoregressive model-based AQP approach DeepAQP （Deep Approximate Query Processing） was proposed. DeepAQP leveraged deep autoregressive model to learn the joint probability distribution of multi-column data in the table in order to estimate the selectivity and the target column’s probability distribution of the given query， enhancing the ability to handle the approximate query requests of aggregation functions with multiple predicates in a single table. Experiments were conducted on TPC-H and TPC-DS datasets. The results show that compared with VerdictDB， which is a sample-based method， DeepAQP has the query response time reduced by 2 to 3 orders of magnitude， and the storage space reduced by 3 orders of magnitude； compared with DBEst++， which is a machine learning-based method， DeepAQP has the query response time reduced by 1 order of magnitude and the model training time reduced significantly. Besides， DeepAQP can handle with multi-predicate query requests， for which DBEst++ does not support. It can be seen that DeepAQP achieves good accuracy and speed at the same time and reduces the training and storage overhead of algorithm significantly.