The study deals with deep learning-based handcrafted affect detection using EEG techniques to enhance the field of human-computer interaction and mental health assessment. The EEG signals that were recorded during positive, negative and neutral emotional states were recorded with a headband called MUSE under the influence of film clips. The features used in the frequency domain included delta, theta, alpha, beta and gamma bands, time domain features included the amplitude, power and entropy, spatial features accounted their regions reflecting activity within the brain, non-linear features included Lyapunov exponent and fractal dimension, and functional connectivity features included the phase synchronization and coherence. The data were inspected and classified by three models of a deep learning Gated Recur-rent Unit (GRU), hybrid GRU-LSTM, and Transformer networks. Transformer model outperformed the others with 97 percent accuracy in comparison to 91 percent by GRU and 93 percent by hybrid GRU-LSTM. The presence of these out-comes confirms that deep learning, particularly Trans-formers, can be very effective to improve the accuracy of EEG-based emotion classification and opens the door to build affective computing systems.