Examination for predicting ground settlement based on measurement records by using a neural network model

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. The long term settlement of those structures is often measured. In this paper, we examined the applicability of a neural network model for settlement prediction using measurements in the early stage after construction. Simulations using a basic network model showed that when the measurement data used for teaching the neural network accumulated, the prediction was in good agreement with the measurement data, and the variance of predicted values was low. However, the basic model could not predict the settlement behaviour precisely, when the amount of teach data was limited as would be in the early stage after construction. Some improvement was necessary for this model to conduct early settlement prediction. To achieve a higher accuracy in long term settlement prediction from early stage measurements, several improvements to the model are proposed, which generate additional data points and improve the prediction accuracy. Firstly, a cubic spline interpolation technique is used to generate additional data between measurements and regulate the input to constant time intervals. Secondly, statistical techniques are used to weed out predicted data points that are outside preset parameters. Short-term predicted values that satisfy with clear statistical criteria (low co-variance or low standard deviation) are added to the network teach data. The improved network model simulations showed that the accuracy of settlement prediction based on early stage measurements improved significantly.