Physics-Informed Wind Forecasting via Neural Network Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for long-term wind resource forecasting are inefficient and inaccurate, struggling to balance accuracy and execution time, and are prone to errors due to the complexity of weather and climate dynamics.
Innovation Solution
A method utilizing a physics-informed machine learning model to predict weather or climate parameters, combined with a deep learning spatiotemporal neural network to forecast wind parameters, leveraging past spatiotemporal data and the outputs of the first model to improve forecasting accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Numerical Weather Prediction (NWP) is used for long-term forecast, then the model can capture physical dynamics, but the trade-off between accuracy and execution time deteriorates and the model becomes exposed to nature chaos
Solution Approach 1:
The forecast system is segmented into two distinct components: a physics-informed machine learning model for long-term weather parameter prediction and a deep learning spatiotemporal neural network for wind parameter forecast. This segmentation allows each model to specialize in its strength, with the physics-informed model handling physical dynamics over long timescales and the neural network processing temporal-spatial patterns efficiently.
Solution Approach 2:
The physics-informed machine learning model acts as an intermediary that provides weather parameter predictions (temperature, pressure, humidity) as input features to the deep learning spatiotemporal neural network. This intermediary layer bridges the gap between physical physics-based modeling and data-driven pattern recognition, combining their advantages for improved forecast accuracy.
2Productivity
If pure Machine Learning models are used to learn patterns from past data, then the model can project predictions efficiently for short-term scales, but the model cannot project predictions in an efficient way for long-term scales
Solution Approach 1:
The system segments the forecasting task by timescale: the physics-informed machine learning model handles long-term weather parameter prediction where physical laws dominate, while the deep learning spatiotemporal neural network handles short-term wind parameter forecasting where pattern recognition excels. This segmentation optimizes both productivity and reliability across different timescales.
Solution Approach 2:
The model transitions from using only historical data patterns to incorporating physics-informed parameters such as temperature, pressure, and humidity fields. This parameter change enables the system to capture both the statistical patterns learned by machine learning and the physical constraints imposed by atmospheric dynamics, improving long-term forecast reliability.
3Adaptability or versatility
If hybrid methods combining NWP and ML in a cascade manner are used, then the model can leverage both approaches, but the accumulation of errors from both sides occurs
Solution Approach 1:
The physics-informed machine learning model serves as a mediator that processes physical weather parameters and transforms them into features suitable for the deep learning neural network. This intermediary approach allows the system to leverage both physical modeling and machine learning without direct cascade connection, reducing error accumulation while maintaining adaptability.
Solution Approach 2:
The system replaces traditional NWP mechanical solving of partial differential equations with a physics-informed machine learning model that learns physical relationships from data. This substitution maintains the physical consistency of NWP while avoiding its computational burden and error propagation, improving overall forecast reliability.
Data Source
AI summary
The invention relates to a method for forecasting wind parameters in an area of interest, the method comprising obtaining at least a first model configured to output a prediction over time of a first parameter for the area of interest, obtaining past spatiotemporal data for the area of interest to form a training database, training a second model for forecasting wind parameters for the area of interest when past spatiotemporal data of the area of interest are inputted in the second model, the training of the second model depending on the training database and on predictions of the first parameter obtained with each first model for the area of interest, and operating the trained second model for forecasting wind parameters for an area of interest.