Accurate temperature forecasting plays a pivotal role in environmental management, agriculture, and energy planning, where reliable predictions underpin informed decision-making and resource allocation. However, effectively modeling the intricate seasonal patterns, long-term trends, and nonlinear fluctuations inherent in temperature data remains a persistent challenge. To address these complexities, we introduce STR-NBEATS, a hybrid framework that integrates Seasonal-Trend Decomposition using Regression (STR) with the Neural Basis Expansion Analysis (N-BEATS) deep learning architecture. STR decomposes temperature time series into interpretable trend and seasonal components, along with a remainder term. This enables a tailored forecasting strategy wherein predictable cyclical behavior is handled through a simple seasonal naive approach, and more complex trend and residual dynamics are captured by the highly flexible N-BEATS network. We rigorously benchmark our approach against robust and well-established forecasting methods capable of modeling seasonality, including an STL-based hybrid model combined with Exponential Smoothing (STL-ETS) and an automated seasonal ARIMA model. Empirical results on real-world datasets demonstrate that STR-NBEATS consistently outperforms these strong benchmarks, achieving lower error metrics and delivering forecasts that are both more accurate and interpretable. By enhancing the fidelity of temperature predictions and providing clearer insights into underlying climatic patterns, STR-NBEATS offers a valuable tool for stakeholders seeking to navigate the challenges of a changing environment with greater confidence and precision.