A multi-target prediction model for dam seepage field

Abstract

Prediction of dam behavior based on monitoring data is important for dam safety and emergency management. It is crucial to analyze and predict the seepage field. Different from the mechanism-based physical models, machine learning models predict directly from data with high accuracy. However, current prediction models are generally based on environmental variables and single measurement point time series. Sometimes point-by-point modeling is used to obtain multi-point prediction values. In order to improve the prediction accuracy and efficiency of the seepage field, a novel multi-target prediction model (MPM) is proposed in which two deep learning methods are integrated into one frame. The MPM model can capture causal temporal features between environmental variables and target values, as well as latent correlation features between different measurement points at each moment. The features of these two parts are put into fully connected layers to establish the mapping relationship between the comprehensive feature vector and the multi-target outputs. Finally, the model is trained for prediction in the framework of a feed-forward neural network using standard back propagation. The MPM model can not only describe the variation pattern of measurement values with the change of load and time, but also reflect the spatial distribution relationship of measurement values. The effectiveness and accuracy of the MPM model are verified by two cases. The proposed MPM model is commonly applicable in prediction of other types of physical fields in dam safety besides the seepage field.