A Causality-Driven Dynamical Framework for Multi-Step-Ahead Daily Streamflow Forecasting in Data-Limited Scenarios

Achieving model interpretability and high-precision multi-step-ahead daily streamflow prediction remains a challenge in data-limited watersheds. First, causal learning is performed to quantify the causal strength of different factors, based on the causal-oriented representation learning predictor (CReP). Second, the multi-head self-attention mechanism integrated with causal information (CI-MHSA) is developed to enable efficient fusion of spatiotemporal features, and then a coupled framework called CI-MHSA-CReP is proposed. CI-MHSA-CReP is applied to 5–15-day-ahead streamflow prediction at four stations with a training dataset length of 1200. It is found that the predicted values generated by CI-MHSA-CReP exhibit excellent agreement with the observed values during low-flow, normal-flow, and high-flow periods, with minimal time-lag effects. Meanwhile, the model delivers highly consistent prediction performance across 5-step, 10-step, and 15-step forecasting. The average NSE values of CI-MHSA-CReP reach 0.830, 0.713, and 0.647 for 5-step, 10-step, and 15-step predictions respectively, representing an improvement of 4.3–75.8% compared with those of spatiotemporal information conversion machine (STICM), CReP, Transformer, and MHSA-LSTM. Moreover, CI-MHSA-CReP demonstrates superior peak flow prediction performance at all the stations, and the average ARE of CI-MHSA-CReP is 56.2–67.2% lower than that of the four comparative models. Under small-sample conditions, CI-MHSA-CReP still maintains satisfactory 5-step prediction performance, with the NSE ranging from 0.787 to 0.959. Therefore, CI-MHSA-CReP serves as a reliable tool for daily streamflow prediction in data-scarce watersheds.