This study proposes a three-stage framework for improving fuzzy time series forecasting by balancing interpretability and predictive accuracy. First, the optimal number of fuzzy sets is selected using a hybrid evaluation metric that integrates dynamic time warping distance, spectral reconstruction error from Fourier transforms, an interpretability index, and a robustness measure. This partitioning strategy ensures stability and enhances model reliability. Next, forecasts are generated using multiple fuzzy models that leverage the selected partition. Finally, a novel ensemble model WISE-Blend (Weighted Integrated Smoothing Ensemble) is introduced to combine the base forecasts based on confidence-driven weighting. Empirical results on seven diverse time series datasets show that WISE-Blend consistently outperforms recent fuzzy forecasting methods, achieving lower forecasting errors across both general and financial domains. The findings affirm the effectiveness of combining optimal partitioning with robust ensemble integration in advancing fuzzy time series modeling.