Query cost models are the key parts of database workload management and performance tuning. Firstly, it is difficult, even impossible, to precisely estimate the costs of different relational operators due to the complexity of database systems and competition of computer resources. Secondly, most existing research work uses general query information without taking advantage of actual operators because of the complexity of query plans. Thirdly, most previous research work does not address the problem of predicting actual execution time of a query but rather predicts the query performance by the cost the like query optimizers generate. To reduce the complexity of workload management, his paper proposes an elaborate cost prediction model based on recurrent neural network through learning from operator behavior and detailed runtime information. In particular, the model uses a special kind of recurrent neural network, called long-short term memory (LSTM). Given an ad-hoc query, the model is able to predict its running time before it starts to run. It is more meaningful than the state-of-the-art query optimizers of existing database systems which only estimate costs in arbitrary units. It is also better than query progress indicators which cannot predict cost before the query runs. This research provides a novel approach to solve the key problem in database workload management. Verified by the experiments, the accuracy of the model is over 71% which shows the method is feasible to some degree.