Abstract: For digital twin hydraulic monitoring perception networks, traditional single-point time series prediction models fail to capture spatial relationships among the devices, and cause missing correlation features; Uncertainty issues arising from strong subjectivity in model structure and parameter design. To address these issues, this paper presents a multi-parameter time series prediction model for monitoring perception networks based on the Bayesian optimization and Hyperband (BOHB), self-learning graph structures, and Bidirectional Long Short-Term Memory (BiLSTM) networks. First, a self-learning graph structure is generated to extract spatial features of the perception network using graph neural networks. Then, the bidirectional Long Short-Term Memory networks are used to extract temporal features, and the BOHB method is used to optimize hyperparameters and improve prediction accuracy. Finally, the model is applied to proactive predictions of future states of the monitoring perception network. We have verified that our new model has achieved optimization rates higher more than 4.35%, 33.14%, 20.47%, 9.09% and 15.03% in R2, RMSE, MAE, MAPE and RMSRE respectively, enjoys higher accuracy and stronger generalization ability compared with a variety of previous prediction models, and has significant performance advantages.

Key words: digital twin water conservancy, monitoring network, self-learning dynamic graph structure, graph neural network, bidirectional long short-term memory network, Bayesian optimization