نوع مقاله : مقاله پژوهشی
نویسندگان
1 دانشجوی دکتری آبوهواشناسی، دانشکده برنامهریزی و علوم محیطی، دانشگاه تبریز، تبریز، ایران
2 دانشجوی دکتری آب و هواشناسی، دانشکده علوم انسانی، دانشگاه زنجان، زنجان، ایران
3 گروه جغرافیا، دانشکده علوم انسانی، دانشگاه زنجان، زنجان
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Air temperature forecasting in semi-arid regions is relevant to water resource management, agriculture, and environmental planning. This study evaluated a multivariate long short-term memory (LSTM) model for forecasting monthly mean air temperature at the Urmia synoptic station and compared it with statistical and machine-learning models. Daily observations of 55 meteorological variables from January 1951 to August 2025 were quality-controlled and aggregated to monthly resolution. Meteorological and calendar features were used to construct 120-month input sequences for simultaneous forecasts of the following 12 months. Data were split chronologically, and hyperparameters were tuned using validation data. On the independent test set, LSTM achieved a root mean square error (RMSE) of 1.683 °C and an R² of 0.963. Across 10 matched test origins, LSTM and SARIMA yielded RMSEs of 1.620 and 1.642 °C, respectively, whereas SARIMA had a lower mean absolute error. Neither model was superior across all metrics. A separate 64-month LSTM hindcast yielded RMSEs of approximately 2.06–2.21 °C. Forecasts for 2026–2030 are station-specific projections based on historical observations; they are not climate-change scenarios. These findings demonstrate LSTM’s forecasting ability at the Urmia station, while evaluation across multiple stations is needed to assess spatial generalizability.
کلیدواژهها [English]