To improve the accuracy of watershed runoff volume prediction， and considering the lack of model transparency and physical interpretability of data-driven hydrological model， a new runoff forecast model named Graph Attention neTwork and Dual-stage Attention mechanism-based Long Short-Term Memory network （GAT-DALSTM） was proposed. Firstly， based on the hydrological data of watershed stations， graph neural network was introduced to extract the topology of watershed stations and generate the feature vectors. Secondly， according to the characteristics of hydrological time series data， a runoff forecast model based on dual-stage attention mechanism was established to predict the watershed runoff volume， and the reliability and transparency of the proposed model were verified by the model evaluation method based on attention coefficient heat map. On the Tunxi watershed dataset， the proposed model was compared with Graph Convolution Neural network （GCN） and Long Short-Term Memory network （LSTM） under each prediction step. Experimental results show that， the Nash-Sutcliffe efficiency coefficient of the proposed model is increased by 3.7% and 4.9% on average respectively， which verifies the accuracy of GAT-DALSTM runoff forecast model. By analyzing the heat map of attention coefficient from the perspectives of hydrology and application， the reliability and practicability of the proposed model were verified. The proposed model can provide technical support for improving the prediction accuracy and model transparency of watershed runoff volume.

Deep learning model with convolution structure has poor generalization performance in few-shot learning scenarios. Therefore， with AlexNet and ResNet as examples， a derivative-free few-shot learning based performance optimization method of convolution structured pre-trained models was proposed. Firstly， the sample data were modulated to generate the series data from the non-series data based on causal intervention， and the pre-trained model was pruned directly based on the co-integration test from the perspective of data distribution stability. Then， based on Capital Asset Pricing Model （CAPM） and optimal transmission theory， in the intermediate output process of the pre-trained model， the forward learning without gradient propagation was carried out， and a new structure was constructed， thereby generating the representation vectors with clear inter-class distinguishability in the distribution space. Finally， the generated effective features were adaptively weighted based on the self-attention mechanism， and the features were aggregated in the fully connected layer to generate the embedding vectors with weak correlation. Experimental results indicate that the proposed method can increase the Top-1 accuracies of the AlexNet and ResNet convolution structured pre-trained models on 100 classes of images in ImageNet 2012 dataset from 58.82%， 78.51% to 68.50%， 85.72%， respectively. Therefore， the proposed method can effectively improve the performance of convolution structured pre-trained models based on few-shot training data.