<p>Double-input-rule-modules (DIRMs) stacked fuzzy system has attracted extensive attention due to its good prediction performance and interpretability. However, with the increase of input variables, the model structure will be very miscellaneous, and the predicted results still have time delay phenomenon. To further compress the structure of the model and improve its performance, an ensemble model, which combines the subtractive clustering-based double-input-rule-modules stacked deep fuzzy model and the long-term trend (SCMDIRM-DFM+LT), is proposed for time series prediction. In the proposed ensemble model, the original time series are firstly decomposed by the seasonal and trend decomposition using Loess (STL) method to extract stable components—the long-term trend features. The remaining random components of the original time series after removing the stable components are used to construct the subtractive clustering-based double-input-rule-modules stacked deep fuzzy model (SCMDIRM-DFM). In the SCDIRM-DFM, the subtractive clustering method is adopted to construct the DIRMs and to compress its structure. This ensemble model is applied to predict the subway passenger flow, the traffic flow, and the power network load, respectively, and compared with some popular models, e.g., the traditional fuzzy model, the shallow modular fuzzy model, the deep fuzzy model, and the deep neural network. Experimental results show that the prediction models considering the long-term trend are better than the ordinary prediction models. The SCMDIRM-DFM+LT performs best among all the comparative models, and it has compressed structure and much less time delay.</p>