Scholarlink

<p>Accurately forecasting daily temperature data is fundamental for a wide range of socioeconomic and environmental applications, including agriculture, ecological conservation, and energy management. However, temperature series often exhibit multiple seasonalities, evolving trends, and nonlinear dependencies that pose significant challenges to traditional modeling approaches. Recent advances in deep learning have shown promise in capturing such complexities, yet many solutions lack interpretability and struggle with highly intricate seasonal cycles. In this paper, we propose a novel hybrid framework that integrates <i>Seasonal-Trend Decomposition using Regression</i> (STR) with the powerful N-BEATS architecture, allowing the model to isolate multiple seasonal patterns before focusing on residual and trend variations. Empirical results using real-world daily temperature data demonstrate that the proposed STR–NBEATS method significantly outperforms benchmark models, reducing forecast errors across multiple metrics. A horizon-based error analysis reveals that while STR–NBEATS forecasts may appear smoother at shorter leads, this trait does not compromise overall accuracy. Moreover, the decomposition-driven methodology provides greater interpretability by distinguishing deterministic seasonal effects from stochastic fluctuations. These findings underscore the potential of hybrid, decomposition-based neural solutions for improving both accuracy and insight in complex environmental time series.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

STR-NBEATS: a novel hybrid framework integrating time series decomposition with deep learning for enhanced temperature forecasting

  • 期刊论文

  • 发表日期：

    2025年5月22日

查看全文
  • Mohammed Elseidi

摘要

Accurately forecasting daily temperature data is fundamental for a wide range of socioeconomic and environmental applications, including agriculture, ecological conservation, and energy management. However, temperature series often exhibit multiple seasonalities, evolving trends, and nonlinear dependencies that pose significant challenges to traditional modeling approaches. Recent advances in deep learning have shown promise in capturing such complexities, yet many solutions lack interpretability and struggle with highly intricate seasonal cycles. In this paper, we propose a novel hybrid framework that integrates Seasonal-Trend Decomposition using Regression (STR) with the powerful N-BEATS architecture, allowing the model to isolate multiple seasonal patterns before focusing on residual and trend variations. Empirical results using real-world daily temperature data demonstrate that the proposed STR–NBEATS method significantly outperforms benchmark models, reducing forecast errors across multiple metrics. A horizon-based error analysis reveals that while STR–NBEATS forecasts may appear smoother at shorter leads, this trait does not compromise overall accuracy. Moreover, the decomposition-driven methodology provides greater interpretability by distinguishing deterministic seasonal effects from stochastic fluctuations. These findings underscore the potential of hybrid, decomposition-based neural solutions for improving both accuracy and insight in complex environmental time series.