Emotions of learners are fundamental and significant in e-learning as they encourage learning. Machine learning models are presented in the literature to look at how emotions may affect e-learning results that are improved and optimized. Nevertheless, the models that have been suggested so far are appropriate for offline mode, whereby data for emotional categorization is saved accessible at any time. However, the model can only observe the data once, thus necessitating a real-time response for real-time emotion categorization. A new MERS in EP model using DC-SSA (Dwarf Combined Squirrel Search Algorithm) is proposed in this work that comprises of following steps: In the first phase, the multimodal inputs (text, audio, and video) are pre-processed. Then, the features are extracted from all the input modalities. From the text, improved BOW and thematic features are extracted. From the video, AAM, improved LBP feature and SLBT are extracted. From the audio, improved MFCC, chroma and spectral features are extracted. The features are fused under IFLF to determine the final features. The emotion recognition process takes place via EC combines RNN, Bi-GRU and CNN, respectively. To make the recognition process more accurate, the weights of CNN are tuned by the algorithm introduced, DC-SSA. Based on the recognition, the performance of the learners is validated.