Data-driven workforce forecasting for transportation infrastructure: A comparative analysis of deep learning and statistical time series models

Reliable workforce forecasting is critical for transportation agencies facing fluctuating labor demands and constrained budgets. This study compares statistical and deep learning models for predicting workforce demand using ten years of person-hour data from the Arkansas Department of Transportation (ARDOT). Three forecasting techniques were evaluated across short- (one-year) and long-term (two-and-a-half-year) horizons: autoregressive integrated moving average (ARIMA), long short-term memory (LSTM), and one-dimensional convolutional neural network (1-D CNN). Model performance was assessed using mean absolute error (MAE) and root mean squared error (RMSE), complemented by parametric tests to compare deep learning stability. Results indicate that the 1-D CNN achieved the best overall performance and stability, with average MAE and RMSE of 4.6% and 5.8% for short-term forecasts and 4.9% and 6.4% for long-term forecasts. The ARIMA model performed competitively in the short term (5.0% MAE, 6.5% RMSE) but exhibited reduced responsiveness over longer horizons, while the LSTM suffered from instability and overfitting despite capturing nonlinear dependencies. These findings highlight CNNs as robust and computationally efficient tools for modeling complex temporal workforce dynamics, particularly where interpretability is balanced with predictive power. The results demonstrate that deep learning can substantially enhance workforce planning in transportation infrastructure management, supporting more proactive and data-driven staffing strategies. Future research should expand to multi-agency datasets and hybrid model designs to integrate the interpretability of statistical models with the adaptability of deep learning frameworks.