CPMA: Spatio-Temporal Network Prediction Model Based on Convolutional Parallel Multi-head Self-attention

Long distance pipeline plays a vital role in long distance water supply facilities, and the prediction of pipeline leakage is always a difficult research problem. Data such as water flow and pressure of pipelines have obvious spatial and temporal characteristics. To solve the problem that traditional temporal prediction network LSTM lacks spatial characteristics in pipeline leakage prediction, a convolutional parallel multi-head self-attention module (CPMA) is proposed to capture temporal and spatial characteristics of time series data through the CPMA module. The parallel structure in the CPMA module can give full play to the complementary characteristics of shared weights of convolutional neural network and flexible weight calculation of self-attention mechanism, and according to the similarity of the calculation of convolutional neural network and self-attention mechanism in the feature mapping part, the input features are unified for feature mapping in the CPMA module, effectively reducing the calculation amount. Combined with the memory ability of LSTM for time series features, the leakage prediction of pipeline is realized. The experimental results show that the model proposed in this paper has the best performance in terms of prediction accuracy, convergence rate and robustness compared with the existing common models in the field of time series prediction. The pipeline leakage model established by using this prediction method can effectively improve the detection efficiency of pipeline leakage, so as to provide better decision support for water management.