Session: 01-04: AI for Energy Sustainability IV

Paper Number: 131519

131519 - Towards Improving High Spatiotemporal Weather Forecast Accuracy With Data-Driven Modeling 

Abstract: 

Mesoscale weather forecasting with high spatial and temporal resolution can enhance the accuracy of accessing initial and boundary conditions for microscale simulation model development and analysis. The dominant approach to weather forecasting is pursued by the numerical weather prediction (NWP) method, utilizing discretized grids and solving partial differential equations to represent atmospheric states. However, as computational models become more complex and detailed, there is a notable increase in computational costs coupled with a challenge in maintaining accuracy – particularly when the computational domain approaches the urban boundary layer. One potential solution is to explore model reduction methods in conjunction with machine learning-based techniques to enhance the speed of weather forecasting while maintaining high levels of accuracy.

 

This paper first demonstrates that intricate patterns in weather data can be effectively identified using neural network models such as Recurrent Neural Networks (RNN). Subsequently, the reduced-order model (ROM) is introduced as a promising alternative or complementary tool to generate low-dimensional surrogate models renowned for their high accuracy and computational efficiency. While widely employed in various fields and not extensively explored in fluid mechanics, this approach is demonstrated by presenting a reliable RNN model and a ROM. These models are specifically designed for higher spatiotemporal resolution simulations, demonstrating their capability to maintain high accuracy compared to reference solutions.

Mesoscale weather forecasting with high spatial and temporal resolution can enhance the accuracy of accessing initial and boundary conditions for microscale simulation model development and analysis. The dominant approach to weather forecasting is pursued by the numerical weather prediction (NWP) method, utilizing discretized grids and solving partial differential equations to represent atmospheric states. However, as computational models become more complex and detailed, there is a notable increase in computational costs coupled with a challenge in maintaining accuracy – particularly when the computational domain approaches the urban boundary layer. One potential solution is to explore model reduction methods in conjunction with machine learning-based techniques to enhance the speed of weather forecasting while maintaining high levels of accuracy.

 

This paper first demonstrates that intricate patterns in weather data can be effectively identified using neural network models such as Recurrent Neural Networks (RNN). Subsequently, the reduced-order model (ROM) is introduced as a promising alternative or complementary tool to generate low-dimensional surrogate models renowned for their high accuracy and computational efficiency. While widely employed in various fields and not extensively explored in fluid mechanics, this approach is demonstrated by presenting a reliable RNN model and an ROM. These models are specifically designed for higher spatiotemporal resolution simulations, demonstrating their capability to maintain high accuracy compared to reference solutions.

 

 

 

Presenting Author: Shivesh Sharma Cleveland State University

Presenting Author Biography: Shivesh is a Graduate Research Assistant and Ph.D. student in the M3TFluiD lab in the Mechanical Engineering Department at Cleveland State University.

Authors:

Shivesh Sharma Cleveland State University
Maede Najian Cleveland State University
Navid Goudarzi CSU Ohio