An Improved LSTM Prediction Method for Monthly Precipitation Based on Boundary Correction and Double Decomposition.

Based on deep learning, precipitation prediction has been a hot topic in hydrological research in recent years. Monthly precipitation data is a typical small sample data that cannot meet the demands of deep learning for large datasets. Therefore, this paper proposes an improved method for predicting monthly precipitation using a Long Short-term Memory Network (LSTM) by integrating the ideas of signal decomposition and boundary correction. First, to address the " end effect" of sequence decomposition, this paper uses the waveform feature matching extension method to expand the boundaries of the original sequence. Then, the original sequence is decomposed twice by using Extreme-point Symmetric Mode Decomposition (ESMD) and Variational Mode Decomposition (VMD). ESMD is used to extract different sorts of scale information of the monthly precipitation sequence, obtaining several model components with decreasing frequencies and one residual component. After cutting off the most severe extension parts that "pollute" the internal data of each sub-sequence, VMD further smooths the high-frequency components. Finally, LSTM is applied to each sub-sequence for prediction, and the predicted results are reconstructed to obtain the final prediction. The monthly precipitation in Badong County, Hubei Province is selected as an example for verification. Through model comparison analysis, the results show that compared to traditional single SVM and LSTM models, the prediction model combining signal decomposition algorithm has more advantages in monthly precipitation prediction; The method of boundary correction is incorporated into the combined model by using the ESMD algorithm improves the overall prediction accuracy of the model. The prediction effect of high-frequency components is a key factor determining the accuracy of combined model prediction. The method proposed in this paper not only performs the best on the selected evaluation indicators but also significantly improves the fitting effect of extreme data points. In particular, even when facing a small sample dataset, the Nash-Sutcliffe Efficiency coefficient (NSE) of monthly precipitation prediction can reach 0.960 0 by using this model. [ABSTRACT FROM AUTHOR]