Predicting coarse particle displacements due to turbulent flows at near-threshold conditions via LSTM models

Bed particle motion as bedload transport in riverine flows is a topic of interest in scientific and engineering fields as it is responsible for erosion and sedimentation, which are essential for hydraulic structures design and maintenance [1] but also for river and basin management. The physics of particle motion as the bedload is governed by stochastic processes which interrelated various parameters and conditions (i.e., particle-particle and particle-flow interrelations). Therefore, applying physically-based or hydrodynamic modeling is not always intuitive because of the complex dynamics. In these situations, in which physics is complex, data-driven modeling approaches may yield an efficient alternative approach since it solely considers the relations among the data. Artificial intelligence models (as for data-driven approach) have offered robust predictive performance in various fields of study. In addition, for time-series and sequential forecasting, a beneficial approach is to choose a model that relates previous states to predict future events.

This study contributes to developing a Long Short-Time Memory (LSTM) neural network modeling to predict the particle displacements near-threshold conditions. In order to prepare the data needed for the study, experimental tests were conducted in a hydraulic laboratory on a tilting recirculating flume with a 2000 (length) cm × 60 (width) cm dimension. Laser Doppler Velocimetry (LDV) was applied to record the flow velocity time-series upstream of the particle with 350-hertz frequency. Also, a He-Ne laser with a photomultiplier tube was used to track the particle motion [2]. Data were pre-processed with some statistical approaches for outlier detections and normalization purposes [3]. Therefore, different training and validation datasets ratios were considered, and the results were analyzed with some statistical measures (i.e., MAPE and RMSE).

The proposed input-output architecture (based on the hydrodynamic forces acting on the bed particle) was a function of the future particle displacement and local instantaneous streamwise flow velocity (about 1 diameter upstream of it). Accordingly, the proposed LSTM model achieved high particle displacement prediction accuracy even for lower percent data conditions for model training.

 

References

[1] Michalis, P., Saafi, M. and Judd, M. (2012). Wireless sensor networks for surveillance and monitoring of bridge scour. Proceedings of the XI International Conference Protection and Restoration of the Environment - PRE XI. Thessaloniki, Greece, pp. 1345–1354.

[2] Diplas, P., Celik, A.O., Dancey, C.L., Valyrakis, M. (2010). Non-intrusive method for Detecting Particle Movement Characteristics near Threshold Flow Conditions, Journal of Irrigation and Drainage Engineering, 136(11), pp.774-780, DOI:10.1061/(ASCE)IR.1943-4774.0000252.

[3] Valyrakis, M., Diplas, P., Dancey, C.L. (2011). Prediction of coarse particle movement with adaptive neuro-fuzzy inference systems, Hydrological Processes, 25(22). pp.3513-3524, DOI:10.1002/hyp.8228.