Risk Assessment Model for Engineering Project Based on Deep Learning

An intelligent assessment framework based on LSTM-GRU hybrid model is proposed in this paper to solve the problems of insufficient analysis of time series characteristics and delay of dynamic risk response in traditional methods of project risk assessment. By constructing a multi-source heterogeneous data set containing 1086 construction projects and spanning 7 risk dimensions, a dual-channel feature extraction mechanism is designed to capture long-term risk cumulative effect and short-term risk mutation signals respectively. The model combines the linear and nonlinear risk indicators by entropy weight method, and optimizes the parameter configuration by compound loss function. The model achieves 89. 2% prediction accuracy and 84. 6% map value on the validation set of spatio-temporal quarantine, which is 23. 7% and 18. 0% higher than single LSTM and GRU models, respectively. Experiments show that the model can effectively provide data-driven decision support for dynamic early warning of engineering risk.