Trend detection and forecasting is an important area of machine learning and a crucial task for researchers, news agencies, organizations, and more. In this paper, we propose an auto-encoder LSTM model with attention units, attend2trend, for the task of trend detection and forecasting. The model utilizes the attention units to assign different weights to the input values based on their importance to the predicted value(s). We used two large datasets from Twitter and Wikipedia to evaluate our model. Our preliminary results show that attend2trend predicts trending topics with high accuracy compared with other statistical and deep learning models.