Predicting construction safety accidents using VMD-BiGRU architecture with improved snake optimization algorithm

Purpose:
The construction industry has persistently high occupational accident rates due to diverse project types and complex operations. Conventional time series models struggle with the accident data's nonlinear and multi scale features, limiting predictive accuracy. This study develops a robust forecasting framework to support proactive construction safety management.

Design/methodology/approach:
We propose a hybrid VMD-ISO-BiGRU framework. This model uses Variational Mode Decomposition (VMD) for multiscale feature extraction, Improved Snake Optimization (ISO) for adaptive hyperparameter tuning, and Bidirectional Gated Recurrent Units (BiGRU) to capture bidirectional temporal dependencies. Using weekly OSHA data from January 2015 to October 2024, VMD decomposes non stationary sequences into intrinsic mode functions (IMFs); ISO optimizes BiGRU parameters; and the model is trained and validated for short and long term predictions.

Findings:
The framework substantially outperforms baselines, reducing short term Root Mean Square Error (RMSE) by 94% (0.5478 vs. 8.8289) for BiGRU and long term RMSE by 90% (1.1726 vs 11.2458) for BiGRU. It effectively captures multiscale patterns, nonlinearities, and bidirectional temporal dependencies in accident data.

Practical implications:
The framework enables early identification of high risk periods, supporting proactive interventions, resource allocation, and policy development to reduce accidents and improve industry sustainability.

Originality/value:
By integrating signal decomposition, optimization, and deep learning, the VMD ISO BiGRU model offers a novel, interpretable, and high precision tool for forecasting construction safety accidents.