<p>Accurate and interpretable time-series forecasting is essential for supporting operational decision-making in industrial systems, where models must capture multi-scale temporal patterns while remaining transparent and reliable. However, industrial time-series data pose several persistent challenges: The availability of temporal patterns simultaneously at multiple scales, ranging from short-term operational fluctuations to long-term structural trends; the way forecasting models are often optimized purely for statistical accuracy, limiting their practical relevance for operational decision support; and the lack of interpretability in many high-performing deep learning models. This paper proposed FTE, a hybrid deep learning architecture that jointly models multi-scale temporal patterns and context-dependent dynamics. FTE combines a Trend Extraction Branch using multi-scale convolutions with an attention-based Temporal Dependency Branch to capture both long-term and short-term dependencies, while a gated fusion mechanism adaptively balances their contributions. Experiments on the ENTSO-E European electricity load dataset showed that FTE consistently outperformed all other baselines, achieving lowest MAE of 132.28 MW and MAPE of 1.88%, while maintaining stability across both single-step and multi-step day-ahead prediction horizons. SHAP-based analysis further confirmed that FTE produced interpretable, temporally coherent attributions aligned with physical demand behavior. Overall, FTE provides a robust and explainable forecasting framework that enhances the practical usefulness of predictions in industrial settings by improving accuracy and interpretability. The model is intended to support downstream decision-making processes, with future work examining its impact on real operational outcomes.</p>