Abstract: Temperature is one of the key influencing factors on the growth of wine grapes. Disasters, such as high temperatures, low temperatures, and frost, can severely impact the wine grape cultivation. Temperature prediction and correction technologies are widely used in agricultural production, providing accurate temperature forecast services. This study focused on the issue of bias in agricultural meteorology temperature forecasting. A systematic investigation was also made to correct the temperature forecast errors and application using recurrent neural networks and ensemble learning. The forecast and actual datasets were collected in the daily forecasts of 15 regions in Ningxia in 2018. The root mean square error and mean absolute error were used to verify the systematic deviation in the forecast data of the European Centre for Medium-Range Weather Forecasts (ECMWF) model. The systematic deviations were experimentally verified in the ECMWF model's 2m maximum and minimum temperature forecasts. The results show that the ECMWF forecasts exhibited systematic deviations: The forecasts that initialized at 00:00Z were provided for better predictions of the minimum temperature, but there were generally larger errors in the maximum temperature. While the forecasts that were initialized at 12:00Z performed better for the maximum temperature, there were significant biases in the minimum temperature. The correction of the maximum and minimum temperature forecasts was conducted using three models: a recurrent neural network (RNN), a long short-term memory (LSTM), and a gated recurrent unit (GRU). An ensemble learning model was constructed to integrate these three models in order to further improve the accuracy of the model. The correction was evaluated by the mean absolute error and the forecast correction improvement rate. A comparison was made between the conventional correction method (Random Forest, RF) and the proposed approach. Experiments showed that all four models effectively reduced the systematic deviation of the ECMWF model's maximum and minimum temperature forecasts at 2 m above the ground in the study areas. Their correction performance was generally superior to that of RF. But there were some differences in the performance among the four models. The RF exhibited the weakest performance of the correction, where an average forecast correction was improved by 25.86%. The RNN showed the relatively limited correction in some scenarios, with an average improvement of 32.24%. LSTM and GRU achieved a similar correction performance, both reaching 37.2%. The ensemble learning model achieved the best correction performance, reaching 38.61%. The effectiveness of ensemble learning was highlighted in the temperature forecast correction. The strong performances of the four models significantly outperformed the RF. The correction results confirmed the effectiveness of recurrent neural networks in the meteorological data. The ensemble learning model (Ensemble) exhibited superior robustness and stability over the various scenarios, indicating the relatively high correction performance. The ensemble model with the highest average forecast correction enhanced both predictive accuracy and robustness, compared with the individual models. In addition, an application system was designed and implemented to correct 2m maximum and minimum temperature forecasts. A rapid and accurate correction of regional forecast data was realized using the ensemble learning model. The finding can provide a reliable reference for the wine grape cultivation, particularly for disaster prevention and decision-making against extreme weather events.