Runoff prediction is an important and challenging nonlinear time series problem in hydrology. It is difficult to capture complex local dynamic changes and uncertainties due to nonlinear and uncertain hydrological data. Therefore, a multi-time window runoff prediction method based on Self-Attention-LSTM model is proposed. By introducing Self-Attention mechanism to dynamically learn and highlight important information at different time scales, convolutional neural network is used to extract meteorological and hydrological characteristics at different time points, and key information and LSTM output are integrated. The features are mapped to future multi-step runoff sequences to improve the model's ability to understand and predict runoff data. In this experiment, meteorological and hydrological characteristics and runoff data of Qingshuijiang River Basin in Guizhou were selected from a dataset driven by regional meteorological elements in China, and the runoff forecast in the next 3 h (one step), 1 day (8 steps) and 3 days (24 steps) was taken as the target. The experimental results showed that, compared with the classical LSTM and Conv-LSTM models, the efficiency coefficient (NSE) of Self-Attention-LSTM in runoff prediction is significantly improved. In 3 h prediction, the NSE increases by 8.2% and 7.6% respectively. In the 1-day forecast, increases of 4.1% and 9.8%; In the 3-day forecast, the increase was 6.8% and 5.0%, and the effectiveness and superiority of the Self-Attention-LSTM model in the multi-time window runoff prediction were proved through comparative analysis experiments, which improved the utilization rate of the internal information of meteorological data, and provided beneficial exploration and new ideas for the field of runoff prediction.