The characteristics of underwater wireless communication (UWC) systems show the considerable challenges presented by the highly variable and unpredictable nature of underwater acoustic channels which includes multipath propagation, Doppler shifts and environmental changes. Reliable and efficient communication systems depend on accurate predictions of the channel state. This study proposes a deep learning (DL) architecture that incorporates Convolutional Neural Networks (CNN) together with Bidirectional Long Short-Term Memory (BiLSTM) with self-attention. The architecture was designed to be trained on multiple tasks. The resulting architecture produced an MSE of 0.022 and prediction accuracy of 92.5%, and achieved a model generalization score of 0.89. Compared to the traditional methods stand-alone baselines, the proposed architecture was superior during evaluation and provides a computationally efficient model with diminished training time and energy consumption. This paper presented an architecture developed as a practical and efficient solution to predicting underwater channels for real-time conditions to provide adaptive communication in extreme underwater conditions.