Epilepsy seizures can cause great physical and mental trauma to patients, and accurate prediction of epileptic seizures can assist doctors in treating patients in time. In order to accurately predict seizures, this paper proposes a seizure prediction method that integrates electroencephalogram(EEG) features and multichannel EEG interaction features. Firstly, the multi-scale symbolic alignment transfer entropy is proposed to analyze the multi-channel EEG signal interaction information, generate the synchronization matrix, and screen the important EEG channels related to seizures by significance analysis to reduce the interference of unnecessary features to the classification. Then, the power spectral density energy diagram (PSDED), which characterizes the EEG signal, and the synchronization matrix diagram (SMD), which describes the interaction characteristics of brain channels, are generated for the EEG signals after screening the channels. The two feature maps are fused, and the deep convolutional neural network (DCNN) is used to classify and identify the EEG signals of epilepsy patients, which improves the learning ability and generalization ability, and the classification accuracy can reach 96.825%. Finally, a prediction evaluation system is used on the basis of classification to evaluate the seizure prediction performance, and the prediction sensitivity reaches 96.66% and the false detection rate can reach 0.03/h when the seizure prediction range (SPH) is 10 minutes and the seizure onset period (SOP) is 10 minutes. When the SPH is 30 minutes and the SOP is 10 minutes, the prediction sensitivity reaches 93.17% and the false detection rate can reach 0.05/h. Compared with the results of existing studies, the proposed method has better prediction sensitivity and lower false detection rate.