LSTM-GAT networks based on ResNet structure for prediction of complex multivariable systems

Based on Residual Neural Network (ResNet) structure, a long short term memory and graph attention networks (LSTM-GAT) model is proposed to predict complex multivariable systems. In order to enhance the correlation of adjacent time series data, a sliding window is designed to divide the data for improving the robustness of the model to abnormal disturbance. The convolution layer is used to extract local features, and the pooling layer is adopted to realize the aggregation of data features. A residual block composed of LSTM network is designed to capture the long-term dependency of the extracted local features. In order to capture the non-Euclidean features of the time series data, the data in the window is transformed into the topology structure. GAT is used to model the topology data, and the attention mechanism is adopted to achieve effective data feature updating. The prediction results are obtained by mapping the extracted deep features through the fully connected layer. Finally, by compared with the existing network models, the superiority and applicability of the proposed network model is verified.