Reference crop evapotranspiration forecast using multi-model integrated output weather variables

Reference evapotranspiration (ET0) is one of the most important parameters to predict the drought degree in agricultural production and precision management. In this research, the multi-model ensemble method was applied to improve the forecasting precision for ET0. Genetic algorithm-support vector regression was utilized to forecast the weather variables output by European Centre for Medium-Range Weather Forecasts, National Centers for Environmental Prediction, Japan Meteorological Agency, and Korea Meteorological Administration models. The prediction accuracy and reliability of the model were evaluated by three indicators: root mean square error (RMSE), mean absolute percentage error (MAPE), and determination coefficient (R2). Penman-Monteith equation was used to forecast the ET0 from 1 to 7 days in the future, according to the optimal models and schemes. Finally, the applicability of the improved model was verified in agricultural test field, including the Yuncheng Station in Shanxi Province of China. The results show that the multi-mode integration improved the prediction performance of air temperature, actual water vapor pressure, and wind speed under a single mode, with the largest improvement in the daily maximum temperature, followed by the actual water vapor pressure, wind speed, and daily minimum temperature. The prediction performance showed a downward trend with the increase in the prediction period. There was no trend of multi-mode integration on the prediction performance of solar radiation. The forecast accuracy of a single model in the ET0 forecast decreased rapidly with the growth of the forecast period. By contrast, the multi-mode integration scheme improved the accuracy of ET0 forecasting and the stability in a long forecast period, and then balanced the tradeoff of weather forecasting accuracy and duration. In ET0 forecasts, the performance of the multi-model schemes was significantly better than that of the original single models. The best prediction performance was achieved in the forecasting precision of the recombination scheme with the optimal models. The smallest RMSE and MAPE were 0.65-0.81 mm/d and 19.43%-23.78%, respectively, and the highest R2 was 0.83-0.89. There was excellent forecast performance of the reorganization scheme in the seasonal ET0 forecast throughout the year, except for the long forecast period in summer. The forecast performance in spring and winter was the best of all the schemes, whereas, the forecast accuracy in autumn was also maintained at a high level. The reorganization scheme still showed better prediction performance in the ET0 prediction of the test field in the next 1-7 days, with the RMSE and the average absolute percentage error not exceeding 0.83 mm/d and 34.57%, respectively. The adaptability of numerical weather forecasts can be verified in the township areas under Yuncheng Station, providing accurate ET0 forecast information for local agricultural production. It was of great significance for agricultural water demand prediction and optimal management of water resources. However, the ET0 prediction performance and applicability of the multi-mode integration scheme need to be verified for the other agricultural production areas and seasons.