A Transformer-Based Regression Scheme for Forecasting Significant Wave Heights in Oceans

Pokhrel, Pujan; Ioup, Elias; Simeonov, Julian; Hoque, Tamjidul; Abdelguerfi, Mahdi

doi:10.1109/joe.2022.3173454

Cited by 16 publications

(6 citation statements)

References 72 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Note that ML/regression algorithms are generally designed for the observation space, that is, the dependent and the independent variables are derived from the observations. This has been previously explored by (Pokhrel et al, 2022) for DA. However, the setup of our study, as described in Sections 2.2 and 2.4, allows the model to learn the residuals of the analysis values (which provide the closest "true" estimate) and model predictions.…”

Section: Deep Symbolic Regression (Dsr)mentioning

confidence: 99%

A machine‐learning and data assimilation forecasting framework for surface waves

Pokhrel,

Abdelguerfi,

Ioup

2024

Quart J Royal Meteoro Soc

Self Cite

View full text Add to dashboard Cite

In this paper, we combine Deep symbolic regression (DSR) and Ensemble Optimal Interpolation‐based Data Assimilation (DA) method to correct the error in the forecasts from the numerical model, WaveWatch III. In our experiments, the DA and DSR training is performed on the hindcasts and then the model is integrated forward in time with both the numerical model and the symbolic expressions generated from the DSR procedure to generate the forecasts. The DSR method is utilized in this paper to generate the symbolic equations that correct the model error in the WaveWatch III/ DA system. The proposed algorithm takes the zonal (u) and meridional (v) wind components from Global Forecast System (GFS) forecasts, wave heights from WaveWatch III, and geographical coordinates (latitude and longitude) to model physical relationships not included in the original numerical model. The DA is performed using JASON‐2 and SARAL altimeter measurements, and the independent testing uses the in situ buoys The RMSD of the proposed method is better than the numerical model with/without DA for up to 42 hours with only 12 days of assimilation spin‐up cycle. The symbolic equation generated from the proposed framework can be used to correct the predictions from WaveWatch III for weather prediction.This article is protected by copyright. All rights reserved.

show abstract

Section: Deep Symbolic Regression (Dsr)mentioning

confidence: 99%

A machine‐learning and data assimilation forecasting framework for surface waves

Pokhrel,

Abdelguerfi,

Ioup

2024

Quart J Royal Meteoro Soc

Self Cite

View full text Add to dashboard Cite

show abstract

“…This ratio of loss from multiple sources improves the training process when numerical data, observational data, or both are noisy. The proposed buoy forecasting task is inspired by [36], but we forecast multiple buoy parameters, test additional numerical models (ERA5 and HYCOM), and apply our physicsregularized training methodology, as main differences. So, we show, in an experimental approach, that we may use complex solutions calculated by numerical climatology and ocean flow models as a means of regularizing surrogate PINN models.…”

Section: Related Workmentioning

confidence: 99%

Forecasting Buoy Observations Using Physics-Informed Neural Networks

Schmidt,

Pokhrel,

Abdelguerfi

et al. 2024

Preprint

View full text Add to dashboard Cite

Methodologies inspired by physics-informed neural networks (PINNs) were used to forecast observations recorded by stationary ocean buoys. We combined buoy observations with numerical models to train surrogate deep learning networks that performed better than with either data alone. Numerical model outputs were collected from two sources for training and regularization: the hybrid circulation ocean model and the fifth ECMWF reanalysis experiment. A hyperparameter determines the ratio of observational and modeled data to be used in the training procedure, so we conducted a grid search to find the most performant ratio. Overall, the technique improved the general forecast performance compared with nonregularized models. Under specific circumstances, the regularization mechanism enabled the PINN models to be more accurate than the numerical models. This demonstrates the utility of combining various climate models and sensor observations to improve surrogate modeling.

show abstract

“…Xu et al (2023)tested the performance of the Transformer model in predicting lake water levels and discussed the influence of the Yangtze River basin on fluctuations in lake water levels. Pokhrel et al (2022) utilized the Transformer model to forecast significant wave heights in the ocean. Despite the Transformer model's strong performance, there are several challenges in the field of water level prediction.…”

Section: Introductionmentioning

confidence: 99%

Short term prediction of water level based on deep learning during the flood season, in the downstream area of The Three Gorges Reservoir

Mao,

Xiong,

et al. 2024

Preprint

View full text Add to dashboard Cite

Accurately predicting river water levels is crucial for managing water resources and controlling floods. In this study, we propose a water level prediction model based on a deep learning method (Transformer model) to improve the accuracy and efficiency of predicting inland river water levels. Water level data from seven hydrological stations were collected from the downstream area of the Three Gorges Reservoir, which confirmed the effectiveness of the model. the proposed model was improved by three main algorithms: the wavelet thresholding denoising (WTD) algorithm, the maximum information coefficient (MIC) algorithm, and the linear exponential (LIMEX) loss function. We are dedicated to improving the accuracy of predicting peak water levels during flood seasons. The results show that the proposed MIC-TF-LINEX model has achieved superior performance in predicting water levels compared to other models, such as traditional Transformer, Back Propagation (BP) Neural Network, and Bi-directional Long Short Term Memory (BILSTM). Furthermore, extending the forecast period will also affect the accuracy of the water level forecasting model. When the prediction duration is 8 hours, the R2 value is 0.9989, the MAE is 0.1020, the MSE is 0.0166, and the MAPE is 0.0060. When the prediction timeframe is within 56 hours, the MSE of the prediction result is still less than 0.1 meters. This study provides a highly accurate and well-suited method for predicting floods.

show abstract

A Transformer-Based Regression Scheme for Forecasting Significant Wave Heights in Oceans

Cited by 16 publications

References 72 publications

A machine‐learning and data assimilation forecasting framework for surface waves

A machine‐learning and data assimilation forecasting framework for surface waves

Forecasting Buoy Observations Using Physics-Informed Neural Networks

Short term prediction of water level based on deep learning during the flood season, in the downstream area of The Three Gorges Reservoir

Contact Info

Product

Resources

About