What are you looking for?
51 Résultats pour : « Portes ouvertes »

L'ÉTS vous donne rendez-vous à sa journée portes ouvertes qui aura lieu sur son campus à l'automne et à l'hiver : Samedi 18 novembre 2023 Samedi 17 février 2024 Le dépôt de votre demande d'admission à un programme de baccalauréat ou au cheminement universitaire en technologie sera gratuit si vous étudiez ou détenez un diplôme collégial d'un établissement québécois.

Construction Engineering Research and Innovation Sustainable Development, the Circular Economy and Environmental Issues Infrastructure and the Built Environment

AI-Enabled Rapid Assessment of Storm Surge and Waves

Waves hitting a lighthouse during a storm

A Promising Approach: Machine Learning

Hurricanes are among the most destructive natural hazards, posing significant threat to coastal regions in Canada and North America, essentially through storm surges and waves. Increasing coastal populations, rising sea levels and climate change exacerbate the risk of economic and human loss in hurricane-prone areas. Accurate and efficient storm surge and wave modeling is essential for effective risk mitigation and decision-making. While traditional statistical and empirical models are effective, their regional specificity and limitations in handling nonlinear dynamics hinder their broader applicability. High-fidelity numerical models, although highly accurate, are computationally demanding, complicating real-time predictions and large-scale risk assessments. Data-driven approaches, particularly machine learning (ML), provide a promising alternative. By learning complex relationships between hurricane parameters (speed, direction, intensity) and resulting storm surge and wave heights, ML models can deliver faster and more accurate predictions.

Proposed Hybrid Model

This study leverages deep neural networks (DNNs) to predict storm surges and significant wave heights for both landfalling and bypassing hurricanes. Specifically, we developed a novel hybrid model to simultaneously identify a low-dimensional representation of high-dimensional systems, while mapping input parameters to the resulting low-dimensional latent space. This model determines the low-dimensional latent space required for the development of a suitable regression model to achieve the best accuracy. By simultaneously optimizing the dimensionality reduction algorithm and the regression model, a robust, generalizable and meaningful model can be obtained by promoting a balance between the capabilities of dimensionality reduction and regression techniques.

The proposed model consists of a deep autoencoder (DAE) and a deep neural network (DNN), dubbed DAE-DNN. The DAE model enables the discovery of low-dimensional representation from a high-dimensional space—peak storm surge or significant wave height over a large area. The DNN model identifies the nonlinear relationship between the input parameters (six storm parameters) and the identified latent space (latent output). To train the hybrid model, we designed a weighted loss function to encourage a balance between DAE and DNN training in achieving the best accuracy, while accounting for system constraints. The architecture of the proposed hybrid model is illustrated in Fig. 1, where DAE and DNN are trained simultaneously.

Hybrid architecture of the combined DAE/DNN model
Fig. 1. Architecture of the proposed hybrid model

As shown in Fig. 2, once trained, the predictive model can efficiently predict the high-dimensional vector (peak storm surge/significant wave height) based on any given input scenario (storm parameters). Therefore, only the DNN (which predicts latent output vectors) coupled with the decoder (which predicts the high-dimensional output vector over the entire region) are required.

Fig. 2. Architecture of the predictive model
Fig. 2. Architecture of the predictive model

Case Study

The performance of the hybrid model is evaluated through a case study using the synthetic data from the North Atlantic Comprehensive Coastal Study (NACCS) covering critical regions in New York and New Jersey. Storm surge and significant wave height responses are extracted from the simulation results of the ADvanced CIRCulation (ADCIRC) and the Steady State Spectral WAVE (STWAVE) models, respectively. A total of 289 coastal locations were selected as input for the DAE model. In addition, corresponding storm parameters related to peak storm surge and wave height (344 storms in the selected coastal region) were also downloaded from the NACCS database. These served as input for the DNN model.

hybrid model prediction
Fig. 3. Performance of the hybrid model for peak storm surge and significant wave height prediction for training (left) and validation (right)

Both storm surge and significant wave height are predicted using the proposed hybrid model, which consists of a coupled DAE and DNN model, trained simultaneously. Additionally, we compared the proposed framework with two decoupled models consisting of a dimensionality reduction technique (PCA and DAE) and a DNN-based regression model, which are trained separately. As shown in Fig. 3, both models for peak storm surge and significant wave height were well trained, as training results indicate that loss functions decrease with increasing numbers of epochs, both for training and validation.

Application

A case study corresponding to the prediction of storm surge and significant wave height highlights the prediction capacities of the proposed hybrid model. With the storm parameter values listed in Table 1 and the trained DAE-DNN hybrid models, a peak storm surge and significant wave height can be obtained.

Table 1. Storm parameters for storm surge and significant wave height prediction

Storm parameters
Comparison of simulated values ​​with real data
Fig. 4. ADCIRC-based and simulated-based peak storm surge (left column), and STWAVE-based and simulated significant wave height (right column) of the selected storm scenario

Fig. 4 shows excellent agreement between the simulated peak storm surge and significant wave height, and their numerical-based simulation results.

Conclusion

In this study, a novel hybrid machine learning model has been proposed for rapid prediction of peak storm surge and waves over an extended coastal region for both landfalling and bypassing storms. The proposed technique was further compared with two decoupled models consisting of a dimensionality reduction technique (PCA and DAE) and a DNN-based regression model, which are trained separately. The hybrid model outperformed the decoupled PCA-DNN and DAE-DNN models. For example, the R2 (MSE) used to train the storm surge-based model was 0.97 (0.02 m2), 0.89 (0.07 m2), 0.92 (0.06 m2) for the hybrid model, decoupled DAE-DNN and decoupled PCA-DNN, respectively. Similarly, the R2 (MSE) used to train the significant wave height-based model was 0.90 (0.02 m2), 0.84 (0.06 m2), 0.84 (0.08 m2) for the hybrid model, decoupled DAE-DNN and decoupled PCA-DNN, respectively. Consequently, the hybrid model demonstrated high accuracy and computational efficiency, and could be readily integrated into an early warning system or used for probabilistic risk assessment, and early wave and storm surge predictions.

Additional Information

For more information on this research, please read the following paper:

Naeini, S. S., & Snaiki, R. (2024). A novel hybrid machine learning model for rapid assessment of wave and storm surge responses over an extended coastal region. Coastal Engineering, 190, 104503. https://doi.org/10.1016/j.coastaleng.2024.104503