A Combined PSO-LSTM Prediction Model for Dam Deformation

[Objective] Dam deformation results from the nonlinear effects of multiple complex environmental factors. Traditional mathematical models for dam deformation monitoring have difficulty reflecting the complex nonlinear relationships between effect variables and environmental variables, often leading to unsatisfactory prediction results. By leveraging the long-short-term memory (LSTM) model and particle swarm optimization (PSO) algorithm from artificial intelligence technology, a combined PSO-LSTM dam deformation prediction model is established, offering a novel approach for enhancing the accuracy of dam deformation prediction. [Methods] By applying PSO for global optimization of LSTM hyperparameters, a combined PSO-LSTM dam deformation prediction model was established. This method both addressed the deficiencies of traditional prediction models in describing nonlinearity between variables and enhanced the appropriateness of LSTM hyperparameter values. The specific methods included: constructing environmental variable factors based on the interaction mechanism between dam deformation and environmental variables; inputting deformation training sets to determine the range of hyperparameters to be optimized and training the network hyperparameters using the LSTM model; setting the particle position information as the hyperparameters to be optimized and using the PSO algorithm to optimize the LSTM hyperparameters; and outputting dam deformation predicted values at different prediction time points using the parameters obtained from training. [Results] Utilizing deformation monitoring data from concrete gravity dams and concrete arch dams, this study established a traditional monitoring statistical model, a standalone LSTM prediction model, and a combined PSO-LSTM model. The results showed that: (1) the combined PSO-LSTM model achieved the smallest RMSE and MAE values and the largest R2 value, indicating excellent prediction accuracy. Compared to statistical models for monitoring and standalone LSTM models, it demonstrated significantly improved prediction performance. (2) Due to its strong nonlinear learning capabilities, the combined PSO-LSTM model could effectively extract nonlinear characteristics from complex datasets, thereby achieving good prediction performance even with poor-quality deformation monitoring data. [Conclusion] (1) The combined prediction model established based on LSTM and PSO algorithms effectively extracts nonlinear characteristics between environmental variables and effect variables, leading to improved prediction performance. (2) The PSO-LSTM prediction model demonstrates good versatility. Its fundamental principles apply not only to concrete dams but also to earth-rock dams and other hydraulic engineering projects. However, when applying the model, the configuration of neurons in the LSTM model’s input layer must be tailored to the structural characteristics, operational conditions, and influencing factors of different dam types.