The price of carbon emission rights play a crucial role in carbon trading markets. Therefore, accurate prediction of the price is critical. Taking the Shanghai pilot market as an example, this paper attempted to design a carbon emission purchasing strategy for enterprises, and establish a carbon emission price prediction model to help them reduce the purchasing cost. To make predictions more precise, we built a hybrid deep learning model by embedding Generalized Autoregressive Conditional Heteroskedastic (GARCH) into the Gate Recurrent Unit (GRU) model, and compared the performance with those of other models. Then, based on the Iceberg Order Theory and the predicted price, we proposed the purchasing strategy of carbon emission rights. As a result, the prediction errors of the GARCH-GRU model with a 5-day sliding time window were the minimum values of all six models. And in the simulation, the purchasing strategy based on the GARCH-GRU model was executed with the least cost as well. The carbon emission purchasing strategy constructed by the hybrid deep learning method can accurately send out timing signals, and help enterprises reduce the purchasing cost of carbon emission permits.