Scholarlink

Accurate temperature forecasting is essential for climate monitoring, renewable energy management, and disaster preparedness. This study evaluates classical statistical time-series models for medium-range temperature prediction on the multivariate Jena Climate dataset (2009 to 2016). It develops a comprehensive preprocessing pipeline involving stationarity diagnostics, seasonal decomposition, and feature selection to identify the key drivers of temperature dynamics. Baseline forecasters (mean, naïve, drift, and simple exponential smoothing) are compared against ARMA, ARIMA, and seasonal ARIMA (SARIMA) models estimated using the Levenberg–Marquardt algorithm. Additionally, a Box–Jenkins transfer-function model incorporates exogenous humidity and wind variables to enhance forecast accuracy. The best SARIMA configuration achieves an RMSE of 5.02  \(^{\circ }\) C, representing a 32% improvement over the strongest baseline, while the Box–Jenkins model lowers the 1-step ahead MAE to 0.28  \(^{\circ }\) C. Discuss residual autocorrelation, seasonality inversion, and computational challenges in scaling seasonal models to 10-minute data resolution. All code is publicly available to ensure reproducibility.

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

Temperature Forecasting on the Jena Climate Dataset Using SARIMA and Box–Jenkins Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dinesh Chandra Gaddam,
  • Reza Jafari

摘要

Accurate temperature forecasting is essential for climate monitoring, renewable energy management, and disaster preparedness. This study evaluates classical statistical time-series models for medium-range temperature prediction on the multivariate Jena Climate dataset (2009 to 2016). It develops a comprehensive preprocessing pipeline involving stationarity diagnostics, seasonal decomposition, and feature selection to identify the key drivers of temperature dynamics. Baseline forecasters (mean, naïve, drift, and simple exponential smoothing) are compared against ARMA, ARIMA, and seasonal ARIMA (SARIMA) models estimated using the Levenberg–Marquardt algorithm. Additionally, a Box–Jenkins transfer-function model incorporates exogenous humidity and wind variables to enhance forecast accuracy. The best SARIMA configuration achieves an RMSE of 5.02  \(^{\circ }\) C, representing a 32% improvement over the strongest baseline, while the Box–Jenkins model lowers the 1-step ahead MAE to 0.28  \(^{\circ }\) C. Discuss residual autocorrelation, seasonality inversion, and computational challenges in scaling seasonal models to 10-minute data resolution. All code is publicly available to ensure reproducibility.