Abstract: To address the limitations of conventional models in fully capturing the complex nonlinear characteristics of runoff sequences, a monthly runoff prediction model is proposed by integrating the Seasonal-Trend decomposition procedure based on Loess (STL) with convolutional neural networks (CNN) and long short-term memory neural networks (LSTM). In this model, the runoff sequence is first decomposed into trend components, seasonal components, and residual terms of random fluctuations using STL. The decomposed component sequences are then input to the CNN for convolutional operations and subsampling, and the CNN outputs feature sequences that capture temporal relationships. These sequences are further processed by LSTM and the predicted runoff values are obtained through fully connected layers. With the monthly runoff data from the Taolai River gauge station in the Heihe River Basin as an example, the prediction performance of three models, LSTM, STL-CNN, and STL-CNN-LSTM, is compared and analyzed. The validation results demonstrate that the model integrating STL and CNN-LSTM achieves the lowest prediction error and the highest accuracy. Compared to conventional models that directly analyze the original runoff sequence, this model significantly improves the ability to predict monthly runoff.

Key words: runoff forecast, Seasonal-Trend decomposition procedure based on Loess(STL), non-linear characteristics, convolutional neural networks(CNN), CNN-LSTM