洪水预报能够提供有用的未来洪水信息，对于洪灾防治的相关决策起重要的支持作用。未来气候变化导致极端天气事件的增加，形成的洪水灾害将给全人类带来更加严峻的考验。因此，探索更为有效的方法提升洪水预报的精度和可靠性，对于洪灾防治有着重要意义。本论文在深度学习洪水预报建模、洪水预报输入变量选择和输入序列长度确定、无测站流域和少数据样本流域洪水预报三个方面开展系统研究，以期提升深度学习洪水预报模型的可靠性和预报精度，拓宽基于深度学习方法的洪水预报模型的适用性。论文取得的主要研究成果如下：（1）基于研究区历史观测降雨和径流数据，探讨了不同输入序列长度对长短期记忆网络（LSTM）、门控循环单元（GRU）和人工神经网络（ANN）模型洪水预报精度的影响规律。研究结果显示输入序列长度对建立的深度学习洪水预报模型的预报效果有显著影响且存在最优输入序列长度。预见期较小时，LSTM、GRU和ANN的洪水预报精度随着输入序列长度的增加呈现出先提升后略微下降的变化特征；预见期较大时在输入序列大于最优输入序列长度后，ANN模型洪水预报精度显著下降且预报结果出现较大波动，而LSTM和GRU模型的洪水预报精度下降不明显。在最优输入序列长度下，3个模型在捕捉洪水过程和洪峰流量方面均能取得不错效果。（2）利用线性层处理前期径流，以反映前期降雨和前期径流与预报径流之间关系的差异，建立了改进的序列到序列洪水预报模型（seq2seq-lrp），将新模型与注意力机制耦合建立了seq2seq-lrp-a模型。研究结果显示在预报短预见期洪水时，seq2seq-lrp和seq2seq-lrp-a模型显著优于对比模型，随着预见期的延长，seq2seq-lrp-a模型相较seq2seq-lrp模型具有更好的预报精度。以前期径流和降雨作为输入时，短预见期洪水预报结果更多地受前期径流影响，随着预见期延长前期降雨数据对预报结果的影响程度逐渐加大。（3）基于流域产汇流过程，提出了全新的降雨驱动的深度学习产汇流框架，基于此框架，建立了具有物理机理的卷积权重洪水预报模型。在渔潭、陈大和新桥子流域的洪水预报结果表明建立的卷积权重模型具有很好的洪水预报性能且其参数具有实际的物理意义，模型确定的产流系数和汇流系数能够很好地反映流域产流和汇流特征。建立的卷积权重模型在无测站流域和少数据样本流域进行洪水预报时具有很强的适用性。

Flood forecasting can provide useful information on future flood, which plays an important role in supporting flood mitigation decisions. In the future, climate change will lead to increase in extreme weather events and flood disasters will bring more challenges to all mankind. Therefore, it is of great significance to explore more effective methods to improve the accuracy and reliability of flood prediction. In this research, we made progress in three aspects of flood prediction, including model development, selection of input variable and the length of the input sequence, and flood prediction in ungauged basins and basins with a small number of data samples, improved the reliability and accuracy of deep learning models for flood forecasting, and extended the applicability of the deep learning models for flood forecasting. The main results are as follows:(1) Based on rainfall and runoff data observed in the study area, the effects of the input sequence length on the accuracy of flood prediction for long short-term memory (LSTM), gated recurrent unit (GRU) and artificial neural network (ANN) models were investigated. The results show that the length of the input sequence has a significant impact on the accuracy of flood prediction with above models above and there is an optimal length of input sequence. When the lead time is small, the flood prediction accuracies of LSTM, GRU and ANN models increase first and then decrease slightly with the increase of the input sequence length. When the lead time is large and the length of the input sequence is larger than the optimal value, the flood prediction accuracy of ANN model decreases significantly with the increase of the input sequence length and the predicted runoff fluctuates greatly. However, LSTM and GRU models show no obvious decrease in the accuracy of flood prediction under this circumstance. Under the condition that the length of the input sequence is equal to the optimal length, the three models all perform well in capturing flood process and peak discharge.(2) Based on the different effects of previous rainfall and runoff on predicted runoff, a seq2seq model using a linear layer to process previous runoff (seq2seq-lrp) was established, and the seq2seq-lrp-a model was established by coupling the seq2seq-lrp model with the attention mechanism. The results show that the seq2seq-lrp and seq2seq-lrp-a models are significantly superior to the benchmark models in predicting flood with a short lead time. With increasing lead time, the seq2seq-lrp-a model has higher prediction accuracy than the seq2seq-lrp model. The above models almost completely rely on previous runoff data to predict runoff in the short lead time when the previous rainfall and runoff are used as the only inputs. With the increase of the lead time, the dependence on the previous rainfall data gradually increases.(3) Based on the process of runoff generation and flow convergence in the watershed, a new rainfall-driven deep learning model of runoff generation and flow convergence (DL-RGFC) is proposed. Based on DL-RGFC, a convolutional weighted model (Conv-Weight) with physical mechanisms is proposed. The results of flood prediction in Yutan, Chenda and Xinqiao watersheds show that the Conv-Weight model has good performance for flood prediction with parameters that have physical meanings. The runoff generation coefficient and the flow convergence coefficient in the Conv-Weight model reflect the characteristics of runoff generation and flow convergence in the watershed, respectively. The results also show that the Conv-Weight model can be used to predict runoff in ungauged basins and basins with a small number of data samples.