Adaptive hybrid potential evapotranspiration (PET) prediction method based on automatic machine learning

Abstract In arid areas, estimation of crop water demand through potential evapotranspiration (PET) forecast has a guiding effect on water-saving irrigation, to cope with the crisis of water shortage. Neural network-based PET prediction methods is considered to have huge application potential because of its small prediction error. However, the physical conditions and data quality in different regions make the choice of neural network different, making it difficult to provide a general PET prediction method. So an adaptive hybrid model based on automatic machine learning for short-term PET prediction is proposed coupling neural network and PET formula. Process is divided into two stages: learning and forecasting. Learning stage includes three modules: meteorological data reconstructing, adaptive data set generation and adaptive hybrid model (PET calculation formula + neural network) selecting. Forecast stage includes two modules: adaptive data set generation and rolling prediction. 105 standard weather stations in Xinjiang were used as data sets (43 of them had missing data) to test model. According to modules, networks and PET formulas used in the prediction process, corresponding labels were generated in each dataset forecast result. Ratio of training set and test set for each data set was 8:2. Grid search was used to optimize the best hyperparameter combination. In test set, the average absolute error (MAE) and average squared error (MSE) of the model prediction were 0.338mm and 0.270, achieving high prediction accuracy. The mean prediction error is smaller to any single mixed model. We demonstrate that the neural network applicability varies among the used data sources, and Gate Recurrent Unit (GRU) and 1 Dimension convolutional neural network (1DCNN) are more suitable for the selected datasets, while Long Short Term Memory network (LSTM) and Multilayer Perceptron (MLP) are not applicable. Combined with the analysis of the labels, We find evidences that applicability of neural networks and PET formulas is independent of geographic region and degree of drought. In 2023, method of rolling prediction for 1-15 days is verified, and the verification results show that PET prediction error based on neural network is significantly smaller than useing weather forecast data to calculate PET. In addition, by comparison,we determined that adaptive input length can effectively reduce the prediction error, MAE was 27.52% smaller than fixed input length, and MSE was 45.76% smaller than fixed input length. The proposed method realized the automatic machine learning of PET forecast, can predict PET more accurately, and can be further expanded by adding neural networks and PET formulas to improve its generalization ability.