The invention provides a self-adaptive congestion control method based on deep learning in an information center network. The method comprises the following steps: firstly performing adaptive training, and then performing the congestion avoidance; inputting time series data into a deep network structure to train at the self-adaptive stage, wherein the time series data firstly learns the low-dimension features; and then training the time series data by taking the low-dimension features as the input of a GCRBM model; accumulatively learning the features of the time series data by a deep belief network DBN, and updating network parameters; at the congestion avoidance stage, collecting the prediction information of increment Rt+1 through the self-adaptive training stage, computing the weightedaverage interesting queue length Qavg of the queue at the t+1 moment through the prediction information, comparing the Qavg with the current queue so as to determine the congestion level of the network; and then encapsulating the congestion level into a NACK packet to feed back to a receiving end, wherein the receiving end adjusts a sending rate of the interesting packet according to the information. By using the self-adaptive congestion control method provided by the invention, the network status can be understood in advance, and a forwarding path can be dynamically selected, thereby improving the efficacy of the network transmission performance.