Improving daily reference evapotranspiration forecasts: Designing AI-enabled recurrent neural networks based long short-term memory

Improving daily reference evapotranspiration forecasts: Designing AI-enabled recurrent neural networks based long short-term memory

Predicting daily reference evapotranspiration (ETo) plays a significant role in numerous environmental and agricultural applications. It aids in optimizing agricultural practices, enhancing drought resilience, supporting environmental conservation efforts, and providing critical data for research. By leveraging advanced technologies and accurate modeling techniques, stakeholders can make informed decisions that promote sustainability and resilience in the face of changing climatic conditions. The main purpose of this investigation was to forecast the daily ETo trends at Melbourne and Sydney stations in Australia, where several cutting-edge machine learning methodologies were employed. The modeling approach encompassed the implementation of Neural Network (NN), Deep Learning (DL), Recurrent Neural Networks (RNN), RNN based Long Short-Term Memory (RNN-LSTM), and Convolutional Neural Network based LSTM (CNN-LSTM) to forecast daily ETo using historical meteorology data. During the model development stage, the optimal variables were determined successfully via heatmaps for precise assessment of ETo in both stations. The predictive models were built by incorporating both the training subset (80 %, covering the years 2009 to 2020) and the testing subset (20 %, ranging from 2021 to 2024) independently to forecast ETo. The results confirmed that the RNN-LSTM attained higher prediction accuracy as compared to NN, DL, RNN, and CNN-LSTM models. Conversely, based on the visual representations and assessments, one can grasp the significant resemblance between the forecasts of the RNN-LSTM model and the actual data. By combining RNNs with LSTM units, models can leverage the strengths of both approaches to improve their ability to process sequential data effectively. This integration allows for better capturing of both short-term and long-term dependencies in the input sequences. Upon careful evaluation, it became clear that the error values associated with the RNN-LSTM models were negligible at the designated stations during the testing phase, with an RMSE of 0.0011 mm for Melbourne, and 0.022 mm for Sydney, followed by RNN, DL, and NN respectively. The proposed modeling approach can be beneficial in monitoring and managing water and crop planning which relies on precise ETo predictions.