Forecast of Daily Reference Crop Evapotranspiration Using Universal Deep Learning Models in China

Evapotranspiration is a key process in ecosystems, and accurate prediction of short-term daily reference crop evapotranspiration (ETo) is crucial for real-time irrigation decision-making and regional water resource allocation. Due to the complex nonlinear relationship between meteorological factors and ETo, deep learning models are often trained using meteorological observation data from several to dozens of stations to estimate ETo at the station scale. These ETo estimation models are then driven by weather forecast data to achieve ETo prediction.
we first evaluated the accuracy of daily temperature forecasts for the next 15 days based on four years of weather forecast data collected from 2,381 stations. Subsequently, based on different input variables from this stations, we developed three ETo estimation cases and five ETo forecast cases. For each case, five time series DL models derived from improvements to RNNs were employed (i.e., LSTM, BiLSTM, GRU, CNN-BiLSTM, CNN-BiLSTM-Attention). The results can be concluded as below.
The results revealed that the differences in ETo estimating performance between the DL models were significantly smaller than the variations between different training strategies.With the average Root Mean Square Error (RMSE) of the five DL models decreasing from 0.55 mm d-1 to 0.48 mm d-1. Furthermore, when we directly use a larger volume of weather forecast data to train the models, the forecasting accuracy of ETo has been significantly improved, and among the five DL models, the GRU performed the best. Specifically, the RMSE values for the GRU model's future ETo forecasts on the 1st, 4th, 7th, and 15th days have decreased from 0.70, 0.87, 1.00, and 1.33 mm d-1 to 0.51, 0.56, 0.61, and 0.67 mm d-1, respectively.
We have provided a portion of weather forecast data and historical meteorological data for your reference and use. These data cover multiple key meteorological indicators such as temperature, humidity, wind speed, and precipitation, with high accuracy and reliability. Additionally, we offer a pre-trained model based on deep learning algorithms, specifically designed for predicting short-term daily reference crop evapotranspiration (ETo). In the "research" folder, there are three main sections: the Estimating Model, the Forecasting Model, and the Data. The Data section includes a portion of the data we use for model training, comprising a complete set of meteorological data necessary for calculating reference evapotranspiration. The Estimating Model section contains models designed for estimating reference evapotranspiration, while the Forecasting Model section includes models used for forecasting reference evapotranspiration.
You can utilize this model free of charge to enhance your real-time irrigation decision-making and regional water resource allocation efforts. By providing these data and the model, we aim to facilitate and assist your related research and practices.

蒸散发是生态系统中的关键过程，而短期逐日参考作物蒸散发（reference crop evapotranspiration, ETo）的精准预测，对于实时灌溉决策与区域水资源配置至关重要。由于气象因子与ETo之间存在复杂的非线性关系，现有研究通常利用数座至数十座气象站点的观测数据训练深度学习模型，以站点尺度的ETo估算为目标。随后将这些ETo估算模型接入天气预报数据，即可实现ETo预测。

本研究首先基于2381个气象站点收集的4年天气预报数据，对未来15天的逐日气温预报精度进行了评估。随后，基于各站点的不同输入变量，构建了3组ETo估算场景与5组ETo预测场景。针对每个场景，均采用了5种基于循环神经网络（Recurrent Neural Network, RNN）改进的时序深度学习模型，即长短期记忆网络（Long Short-Term Memory, LSTM）、双向长短期记忆网络（Bidirectional Long Short-Term Memory, BiLSTM）、门控循环单元（Gated Recurrent Unit, GRU）、卷积神经网络-双向长短期记忆网络（Convolutional Neural Network-Bidirectional Long Short-Term Memory, CNN-BiLSTM）以及卷积神经网络-双向长短期记忆网络-注意力机制（CNN-BiLSTM-Attention）。研究结果总结如下。

结果表明，不同深度学习模型间的ETo估算性能差异，显著小于不同训练策略带来的性能波动——5种深度学习模型的平均均方根误差（Root Mean Square Error, RMSE）从0.55 mm·d⁻¹降至0.48 mm·d⁻¹。进一步研究发现，当直接使用更大体量的天气预报数据训练模型时，ETo的预测精度得到了显著提升；在5种深度学习模型中，门控循环单元（GRU）的表现最优。具体而言，GRU模型在第1、4、7、15天的未来ETo预测的均方根误差分别从0.70、0.87、1.00、1.33 mm·d⁻¹降至0.51、0.56、0.61、0.67 mm·d⁻¹。

本数据集提供了部分天气预报数据与历史气象数据供使用者参考与调用。这些数据涵盖气温、湿度、风速、降水等多项关键气象指标，具备较高的精度与可靠性。此外，我们还提供了一款基于深度学习算法的预训练模型，专门用于短期逐日参考作物蒸散发（ETo）预测。在「research」文件夹中，共包含三个核心板块：估算模型、预测模型与数据集。其中「数据集」板块包含了模型训练所用的部分数据，涵盖了计算参考作物蒸散发所需的全套气象数据；「估算模型」板块收录了用于参考作物蒸散发估算的模型；「预测模型」板块则包含了用于参考作物蒸散发预测的模型。

您可免费使用本模型，以优化实时灌溉决策与区域水资源配置工作。通过提供上述数据与模型，我们旨在为您的相关研究与实践提供便利与支持。