In the context of badminton, identifying and analyzing technical movement variations is essential for coaches to assess and improve the quality of player actions. However, due to limited viewing angles and movement speed, coaches often have difficulty in observing the details of players' movement variations, especially those related to intravariation classes, i.e. variations within the same movement from different viewing angles. This research highlights the shortcomings of previous studies that only focused on general movement recognition without considering pose intravariation in badminton. To address this issue, this study proposes the use of the Human Pose Estimation MoveNet model and the ConvLSTM deep learning model equipped with the Squeeze and Excitation Block attention mechanism. The dataset used in this study is a badminton pose dataset with intravariation classes, consisting of 6 types of movements captured from 3 different camera angles. Each frame in the dataset contains 26 keypoints extracted using the MoveNet model to represent human body positions. This dataset effectively captures variations in movement due to different viewing angles, enabling the proposed model to address pose intravariation challenges. This study achieved notable results, with the proposed model yielding 99\% accuracy on training data and 98\% on testing data, outperforming baseline models in pose classification for badminton. The combination of the MoveNet Human Pose Estimation model and the ConvLSTM deep learning model, enhanced by the Squeeze and Excitation Block attention mechanism, proved effective in addressing pose intravariation challenges. Compared to previous approaches, the proposed model not only improves accuracy but also offers better clarity in classifying variations within similar movements from different angles, making it a valuable tool for coaches to analyze and refine players' techniques.