Neural network-based forecasting of modal frequency variations in Deriner Arch Dam based on reservoir and thermal fluctuations

Abstract Understanding the environmental sensitivity of dam dynamics is critical for effective structural health monitoring of large-scale infrastructure. This study explores the seasonal frequency behavior of the Deriner Arch Dam, one of Türkiye’s tallest arch dams, under operational conditions, using artificial neural networks (ANNs) to learn the underlying patterns governing modal response. A continuous 14-month dataset comprising daily records of ambient frequencies, reservoir water levels, and air temperatures was utilized. Across fourteen months of monitoring, Mode 1–4 frequencies changed 1.638–1.850, 2.204–2.829, 3.103–3.364, and 3.500–4.213 Hz, respectively. Simultaneously, reservoir elevation spanned 168.09–193.28 m and ambient temperature remained within − 1.67 °C and 35.9 °C. The analysis captures the dam’s dynamic adaptation to naturally fluctuating environmental loads, offering insight beyond static assessments or short-term experiments. A feedforward ANN regression model, trained using the Neural Network Fitting method, was able to accurately approximate modal frequency behavior as a function of temperature and hydrostatic conditions. The ANN models achieved correlation coefficient (R) values of 0.8541 (Mode 1), 0.7637 (Mode 2), 0.8299 (Mode 3), and 0.7969 (Mode 4), with corresponding Mean Squared Error (MSE) values below 0.005, demonstrating satisfactory agreement between predicted and measured frequencies. The resulting frequency prediction maps provide a data-driven reference for assessing future deviations and support a proactive approach to dam safety management. By focusing on an actively operating structure over an extended time frame, this study demonstrates the practical integration of AI tools in real-world structural health monitoring systems.