Landslides are a challenging problem in India and the world. Different weather conditions and soil properties could trigger landslides. The Machine learning (ML) models could predict landslides’ movements. The ML model may overfit the high-dimensional feature of weather and soil data. Dimension reduction techniques could reduce the dimension of the features. The autoencoder model was developed to reduce the data dimension in this experiment. The four months, April to August 2022, time series data of the landslide monitoring station (LMS) from the five landslides in Himachal Pradesh, India, were considered to train the ML models. The dimension reduction autoencoder long-short-term memory (DR-A-LSTM), an autoencoder and LSTM model ensemble, was developed. Furthermore, the ensemble of principal component analysis (PCA) and LSTM (PCA-LSTM) model was developed to reduce the dimension by PCA. The DR-A-LSTM model was compared with the simple LSTM and PCA-LSTM models to predict landslide movements. The data was split in the 80:20 ratio to train and test the ML models. The simple LSTM model produced 82.3% accuracy in the training data and 71.8% in the testing data. The simple LSTM model showed overfitting in the training data. The PCA-LSTM model produced 76.5% accuracy in training and 88.2% in testing. Next, the DR-A-LSTM model produced 97.8% accuracy in training and 100% in testing. The findings of this experiment suggest that the DR-A-LSTM model performed better than the simple LSTM model and PCA-LSTM. As a result, the DR-A-LSTM model could be developed for real-time landslide movement predictions.