Abstract: Based on natural runoff data of Hubei Province, combined with the data of meteorology, satellite remote sensing, water conservancy projects, land use, socio-economy, etc, neural network models such as Multi-Layer Perceptrons （MLP）, Long and Short Term Memory Networks （LSTM） and Convolutional Neural Networks （CNN） were constructed to predict and analyze natural runoff in the Yangtze River Basin, Han River Basin and Qing River Basin. The results indicated that the constructed neural network models had a good prediction effect and could effectively capture the patterns and characteristics of natural runoff changes. The results of sensitivity analysis and importance ranking revealed that there were significant differences in the impact weights of climate change and human activities on runoff variations, and the differences between regions were obvious. Climate change was identified as the primary influencing factor, among its components, precipitation had the highest influence, followed by temperature, while evaporation had the least. Human activities were identified as secondary influencing factors, among its components, water conservancy projects had the highest weight, followed by land use. There were some differences in the weights of climate change and human activities in different basins, among which the weight of Hanjiang River Basin was the highest and the weight of Qing River Basin was the lowest. Predictions for natural runoff trends from 2023 to 2042 indicated a declining trend, with corresponding changes in the weights of influencing factors.

Key words: natural runoff, neural network, watershed difference, climate change, human activities, Hubei Province