Using and Improving Neural Network Models for Ground Settlement Prediction

Abstract

Earth-fill structures such as embankments, which are constructed for the preservation of land and infrastructure, show significant amount of settlement during and after construction in lowland areas with soft grounds. Settlements are often still predicted with large uncertainty and frequently observational methods are applied using settlement monitoring results in the early stage after construction to predict the long term settlement. Most of these methods require a significant amount of measurements to enable accurate predictions. In this paper, an artificial neural network model for settlement prediction is evaluated and improved using measurement records from a test embankment in The Netherlands. Based on a learning pattern that focuses on convergence of the settlement rate, the basic model predicted settlements which were in good agreement with the measurements, when the amount of measured data used as teach data for the model exceeded a degree of consolidation of 69 %. For lower amounts of teach data the accuracy of settlement prediction was limited. To improve the accuracy of settlement prediction, it is proposed to add short-term predicted values that satisfy predefined statistical criteria of low coefficient of variance or low standard deviation to the teach data, after which the model is allowed to relearn and repredict the settlement. This procedure is repeated until all predicted values satisfy the criterion. Using the improved network model resulted in significantly better predictions. Predicted settlements were in good agreement with the measurements, even when only the measurements up to a consolidation stage of 35 % were used as initial teach data.