Forecasting reference crop evapotranspiration using deep learning model and online training

【Objective】 Reference crop evapotranspiration (ET0) is a critical parameter for irrigation and water management. This paper proposes a method for real-time forecasting ET0 using weather forecast data and a deep learning approach. 【Method】 The study was conducted in Xiaoshan District, Hangzhou City, Zhejiang Province. Hourly measured weather data and 1-7 day forecasted weather data from April 24, 2021 to December 31, 2023 were used as the dataset. The forecasting accuracy of the weather data was analyzed. A deep learning model based on the backpropagation (BP) neural network algorithm was developed and deployed for online training using Alibaba Cloud servers. 【Result】 The accuracy of the input parameters was generally reliable, with minimum temperature forecasts being more accurate than maximum temperature forecasts. Forecasting accuracy decreased as the lead time increased. Errors were observed in forecasting weather types and wind scales. The ET0 predicted by the model closely matched those calculated using real-time data, demonstrating high forecasting accuracy. During the training period, the model achieved a maximum accuracy of 91.56%, with an average root mean square error (RMSE) of 0.828 mm/day and a mean absolute error (MAE) of 0.667 mm/day. During the testing period, the model achieved an accuracy of 84.75%, with the average RMSE and MAE being 1.049 mm/day and 0.829 mm/day, respectively. 【Conclusion】 By using publicly accessible weather forecast data and an online-trained BP neural network model, real-time ET0 forecast can be achieved with high accuracy. This approach offers valuable support for farmers, enabling informed and timely irrigation decisions.