ارزیابی مقایسه‌ای مدل‌های یادگیری ماشین و یادگیری عمیق در پیش‌بینی پارامترهای اقلیمی (مطالعه موردی: ایستگاه سینوپتیک کاشان)

Climate change has emerged as one of the most significant environmental challenges of the twenty-first century, with profound impacts on water resources, agriculture, ecosystem stability, and socio-economic development. In arid and semi-arid regions such as central Iran, rising temperatures, increasing climate variability, and the growing frequency of drought events have intensified the need for accurate prediction of climatic variables. Reliable forecasting of temperature and precipitation plays a crucial role in supporting water resources management, agricultural planning, drought mitigation, and climate adaptation strategies. Although conventional statistical approaches have been widely employed for climate forecasting, their ability to capture the complex nonlinear relationships and temporal dependencies inherent in meteorological data remains limited. Recent advances in artificial intelligence, particularly machine learning (ML) and deep learning (DL), have provided powerful alternatives capable of modeling highly nonlinear processes and extracting hidden temporal patterns from large climate datasets.



The present study developed and evaluated an integrated machine learning and deep learning framework for simulating and forecasting four key climatic variables in Kashan, a representative arid city located in central Iran. The investigated variables included daily maximum temperature, minimum temperature, mean temperature, and precipitation. The analysis was based on 16,437 daily observations collected from the Kashan Synoptic Station during the period 1980–2024. Prior to model development, the dataset underwent a comprehensive preprocessing procedure including quality control, detection and imputation of missing values, Min–Max normalization, and feature engineering. Additional predictor variables were generated using calendar-based information together with temporal features including Lag-1, Lag-7, and a 7-day moving average, allowing the models to better capture short-term persistence and seasonal variability in climatic conditions.



To preserve the chronological nature of the time series and prevent information leakage, the dataset was divided into three consecutive subsets consisting of a training period (1980–2016), a validation period (2017–2020), and an independent testing period (2021–2024). Three predictive algorithms representing different modeling philosophies were implemented and compared: Random Forest (RF), Long Short-Term Memory (LSTM), and Transformer. Model performance was quantitatively assessed using four widely accepted evaluation metrics, namely Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²). These complementary performance measures provided a comprehensive assessment of prediction accuracy and model robustness across different climatic variables.



The comparative analysis revealed substantial differences among the investigated algorithms. The Transformer model consistently demonstrated the highest predictive capability for temperature-related variables, indicating its superior ability to capture long-range temporal dependencies and complex sequential patterns within climatic time series. For daily maximum temperature, the Transformer achieved an excellent correlation coefficient of R = 0.967, together with MAE = 1.490°C and RMSE = 1.913°C, indicating high predictive accuracy and low estimation error. Similar superiority was observed for minimum and mean temperature predictions, confirming the effectiveness of attention-based architectures in modeling gradually evolving climatic processes. These findings suggest that the self-attention mechanism enables the Transformer to identify both short-term fluctuations and long-term temperature trends more effectively than conventional recurrent or tree-based algorithms.



The performance of precipitation forecasting differed considerably from that of temperature prediction. Daily precipitation is characterized by high variability, intermittent occurrence, and strong stochastic behavior, making it substantially more difficult to model than temperature. Under these conditions, the LSTM model exhibited relatively better performance than the Transformer and Random Forest, achieving R² = 0.178 and RMSE = 13.371 mm. Although the overall predictive accuracy remained moderate, the results indicate that recurrent neural networks retain an advantage for modeling short-term temporal dependencies associated with rainfall occurrence. Conversely, the Random Forest model consistently showed weaker performance, particularly for long-term forecasting. Owing to its tree-based structure, RF was unable to effectively extrapolate persistent warming trends beyond the range represented in the historical training data, limiting its applicability for climate projection.



The comparative results highlight that no single algorithm is universally optimal for all climatic variables. Instead, model performance depends strongly on the statistical characteristics and temporal dynamics of the target variable. Transformer models appear particularly suitable for variables exhibiting relatively smooth temporal evolution, such as air temperature, whereas LSTM networks remain competitive for highly irregular variables such as precipitation. These findings emphasize the importance of selecting forecasting models according to the physical and statistical behavior of individual climatic parameters rather than relying on a single modeling framework.



Overall, this study demonstrates the considerable potential of modern deep learning techniques for climate forecasting in arid and semi-arid environments. The proposed framework provides reliable predictions of key meteorological variables while offering valuable insights into the comparative strengths and limitations of different artificial intelligence models. The results can support evidence-based decision-making in water resources management, agricultural planning, climate change adaptation, drought risk reduction, and sustainable natural resource management. Furthermore, the developed methodology provides a transferable framework that can be extended to other regions with similar climatic conditions and can be integrated with future climate scenarios to improve long-term environmental planning and resilience under changing climatic conditions.