关键词: 高拱坝, 施工仿真参数, 时序概率预测, 差分整合移动平均自回归, 长短时记忆网络

Abstract: Most previous studies for updating the simulation parameters of high arch dam construction are based on probability prediction alone or point-wise prediction considering their time series characteristics; such methods are usually faced with difficulty in quantitative description of their randomness while considering their time series characteristics. To couple both factors, this study uses an Autoregressive Integrated Moving Average model (ARIMA) to predict the probability along with consideration of parameter time series characteristics, and develops a new intelligent updating model of the simulation parameters of high arch dam construction based on ARIMA and a Long Short-Term Memory model (LSTM) that can learn the advantages of complex nonlinear characteristics of parameter time series. This new model uses ARIMA to predict the linear part of the parameter time series, and uses LSTM to train and predict the residuals output by the ARIMA model. We fuse the predicted linear part and the nonlinear part of the predicted residuals, and then make probability predictions at the 95 % confidence interval to obtain the final result. Thus, we can calculate the construction simulation parameters that describe both randomness and temporal characteristics, and have achieved a new method of higher accuracy (with MSE of 0.518, MAE of 0.519 and RMSE of 0.720) than the model of ARIMA, ARIMA-BP neural network, or random forest (RF). Compared with traditional simulations, the simulation results of a high arch dam construction system are greatly improved using the simulation parameters predicted by our new method.

Key words: high arch dam, construction simulation parameter, timing probability prediction, ARIMA, LSTM