Accurate temperature forecasting plays a crucial role in a wide range of applications, including agriculture, water resource management, energy planning, transportation, and environmental management. Numerical Weather Prediction (NWP) models are among the primary tools for forecasting atmospheric variables. Although the Weather Research and Forecasting (WRF) model is widely used for regional weather prediction, its raw forecasts systematic errors (biases) arising from uncertainties in the initial and boundary conditions, physical parameterization schemes, model resolution, and the complex interactions between atmospheric processes and local environmental conditions. These errors reduce forecast accuracy and highlight the need for effective post-processing techniques. In recent years, statistical, machine learning, and deep learning approaches have been increasingly adopted to improve NWP forecasts by correcting systematic errors.

This study evaluated and compared four post-processing methods, namely Moving Average (MA), Ridge Regression, Random Forest (RF), and Long Short-Term Memory (LSTM), for improving 2-m air temperature forecasts from the WRF model at synoptic stations in Isfahan Province, Iran. Hourly WRF outputs were combined with observational data from synoptic stations. Data from 2021–2024 were used for model training, while data from 2025 were reserved for independent evaluation. Model performance was assessed using Mean Error (ME), Mean Absolute Error (MAE), and the Standard Deviation of Error (STD), together with analyses of the temporal and spatial characteristics of forecast errors.

The raw WRF forecasts exhibited a systematic cold bias (ME ≈ −1.07 °C; MAE = 2.03 °C), with error magnitude varying throughout the diurnal cycle and across seasons, indicating that forecast errors depend on prevailing atmospheric conditions. Although most stations showed a cold bias, no clear relationship was found between forecast bias and station latitude or longitude, suggesting that local atmospheric and geographic factors play a more important role than geographic location alone.

All four post-processing methods improved the WRF forecasts, although their effectiveness differed. The MA method, which relies primarily on the historical temperature error, was effective in correcting systematic bias, reducing MAE by approximately 21%. Ridge Regression, Random Forest, and LSTM all utilized the same set of meteorological predictors describing the atmospheric state. However, their performance differed because of their modeling capabilities. Ridge Regression models only linear relationships between meteorological features and forecast errors, resulting in a modest MAE reduction of approximately 24%. In contrast, RF captures nonlinear relationships among the predictors, reducing MAE by approximately 46%. LSTM further exploits both nonlinear relationships and long-term temporal dependencies, providing the best overall performance with an approximately 54% reduction in MAE. Likewise, STD decreased by approximately 8%, 11%, 34%, and 42% for MA, Ridge, RF, and LSTM, respectively.

These results demonstrate that the superiority of RF and especially LSTM is not simply due to the use of meteorological predictors, since Ridge employed the same input features. Rather, it arises from their greater ability to extract complex nonlinear and temporal relationships between atmospheric conditions and forecast errors, enabling more effective correction of flow-dependent errors. Consequently, RF and particularly LSTM produced more stable forecasts by substantially reducing both temporal and spatial variations in forecast errors. These findings highlight the importance of combining informative meteorological predictors with advanced learning algorithms and support the operational application of machine learning and deep learning techniques for post-processing WRF temperature forecasts.