Seizure poses life-threatening risks to individuals with epilepsy. Predicting seizures is challenging due to individual variations of EEG signal patterns. Channel selection is a crucial task in the effectiveness of seizure prediction devices. This work suggests utilizing a single EEG channel for designing small-sized, power-efficient, wearable prediction devices. The performances of all channels are not consistent when employing deep learning models. It leads to variations in the ranking of the channel’s performance in each instance. In this work, each channel’s performance is evaluated in five rounds, ultimately deriving a ranking of each channel based on the average performance across all rounds. The CNN-LSTM1D model is employed to calculate the performance of each channel. The channel securing the first rank is considered for efficient seizure prediction. Finally, an ensemble learning technique is employed that integrates CNN1D, DenseNet1D, and CNN-LSTM1D models to utilize a majority voting approach. Ensemble learning delivers superior performance compared to individual models using the CHB-MIT database. This research introduces effective strategies for both patient-specific and patient-independent seizure prediction. The patient-specific and patient-independent models provide a sensitivity of 0.9768 and 0.9425, and a specificity of 0.9681 and 0.9294, respectively. It outperforms existing state-of-the-art works.