Abstract: For construction schedule simulation, parameters updating is the key to ensuring its accuracy. However, it is difficult for its existing methods to learn and extract the local nonlinear fluctuation characteristics of the parameters, and the updating accuracy needs further improvement. This paper takes advantage of the deep learning model's capability of mining more hidden information from parameter sequences and adopts a new modeling idea of decomposition-predict-integration. We construct a new dynamic updating model of construction simulation parameters using the techniques of empirical mode decomposition, and improved long short-term memory (EMD-P-ILSTM). To improve modeling efficiency, we use an improved Beetle Antennae algorithm based on an adaptive step factor to automatically optimize the hyperparameters of the LSTM network model. The EMD method is used to decompose a parameter sequence into several stationary sub-sequences; a partial autocorrelation function is used to select a time window for each sub-sequence automatically. A case study shows that, compared with the unimproved LSTM, BPNN, SVM, or Bayesian update method, our new model can track effectively the complicated changes in construction parameters and achieve a higher prediction accuracy.

Key words: high arch dam, construction simulation parameter update, long-short-term memory network, improved Beetle Antennae algorithm, empirical mode decomposition, partial autocorrelation function