An Machine Learning-Based Adaptive Prediction Model for Floods in the Headwater Region of the Tarim River Basin

The Tarim River is the longest inland river in China. Its headwater regions suffer from weak monitoring of meteorological conditions, snow cover, and floods, as well as relatively insufficient research on the formation mechanisms of snowmelt floods, posing significant challenges for high-precision flood forecasting and early warning. Based on the runoff generation processes in the headwater regions from 2000 to 2023, this study proposed a set of flood-influencing factors from three aspects: hydrometeorology, solar radiation characteristics, and underlying surface conditions. Principal component analysis was employed for dimensionality reduction to extract key input variables for runoff prediction models for six tributaries, namely the Kumarak River, Toshkan River, Taxkorgan River, Yarkant River, Karakash River, and Yurungkash River. The cumulative variance contributions of the first four principal components were 88.83%, 88.24%, 87.07%, 87.61%, 87.93%, and 86.48%, respectively, all exceeding 85%, thereby retaining most of the information from the original data. Four-layer neural network prediction models based on the LSTM algorithm were developed for the six tributaries. The Nash-Sutcliffe efficiency (NSE) values during the prediction period were 0.9751, 0.9573, 0.9648, 0.9929, 0.9477, and 0.9785, respectively, indicating overall satisfactory simulation performance, particularly for accurate predictions of low to medium flows below 600 m³/s. The error rates for peak flood flow predictions ranged from 5.55% to 16.72%, while the error rates for three-day flood volume predictions ranged from 2.37% to 15.76%. The errors for peak occurrence time were generally within one day. This research provides a technical reference for flood prediction and regulation in the Tarim River Basin.