Performance comparison of various time-series forecasting models for bridge sufficiency rating prediction

The rapid increase in the number of bridges worldwide has intensified the need for effective maintenance strategies to ensure structural safety and economic efficiency. Accurate predictions of future bridge performance are essential for preventing unexpected failures and optimizing road network maintenance planning. However, existing prediction models frequently overlook the time-series characteristics inherent in bridge inspection data, thereby limiting their accuracy. This study aims to develop improved prediction models by integrating sequential data patterns using advanced deep-learning techniques. Data from the National Bridge Inventory were utilized. As most NBI data lacked explicit sequential structures, preprocessing techniques were applied to generate meaningful time-series patterns. Deep-learning models, including deep neural networks (DNNs), convolutional neural networks, long short-term memory (LSTM), and Transformers, were developed and evaluated using cross-validation to optimize their performance. Results showed that the LSTM model improved prediction accuracy by approximately 46% compared to the baseline DNN model. The Transformer model further improved accuracy by approximately 7% over the LSTM, highlighting its superior ability to capture long-term dependencies. These findings highlight the potential of the Transformer model as a powerful tool for predicting bridge performance, thereby supporting effective maintenance planning and reducing the risk of structural failures.