Our study proposes a Dual-Path Time Series Decomposition and Fusion Multi-Level Ensemble Model (DP-TDFM) architecture divided into two paths. The first path handles non-stationary time series using STL decomposition to separate the sequence data into trend, seasonality, and residual components. These components are then processed by our Multi-Level Ensemble Model (MLEM), which incorporates algorithms such as Random Forest, Support Vector Regression (SVR), and Decision Tree, with a neural network in the hidden layer serving as the final prediction model. The second path employs the GatedTabTransformer, integrating trend and seasonality features alongside external environmental factors as augmented features (AF). Experimental results indicate that, even when individual models exhibit overfitting, our DP-TDFM architecture maintains stable overall performance and achieves the highest prediction accuracy among all models across five-time points, demonstrating more stable and smoother prediction results. This model effectively addresses several challenges in prediction tasks, including overfitting, sparse data, and long-distance dependencies.