Predictions of Sloshing Height Induced by Seismic Excitations Based on LSTM

Abstract

Present work aims to predict the sloshing height induced by seismic excitations by the Long-Short Term Memory (LSTM), which is a temporal recurrent neural network and has the time connection ability to reconstruct the main features of nonstationary records. Lushan and California earthquakes are selected as the external excitations, and thus yield two groups of time-series free surface elevations as the training samples. The internal structural parameters of LSTM, for example, the epochs, batch size and input shape, are systematically adjusted, and the errors and correlations between the predicted and actual results are analyzed. The input step is set as 50, and the proportion of the training set is 80%. The results show that the errors are respectively lower than 0.3% and 2.0%, and the correlations reach 0.998 for both cases. Overall, LSTM can well predict the sloshing heights because of the advantage of predicting the long sequential data.