Wind Speed Time Series Forecasting with LSTM and TCN Models Using Physics-Informed Vectors

This repository implements a state-of-the-art machine learning framework that integrates Physics-Informed Vectors (PIV) with Long Short-Term Memory (LSTM) and Temporal Convolutional Networks (TCN) for improved wind speed forecasting. The project enhances predictive accuracy by incorporating domain-specific physical knowledge and custom hybrid loss functions as performed in physics informed machine learning.

Overview of the Proposed Methodology

Below is a block diagram that outlines the framework, including data preprocessing, Physics-Informed Vector computation, and model training using LSTM and TCN architectures:

Table of Contents

• Installation
• Project Structure
• Running the Project
• Custom Hybrid Loss Function
• Results
• Publication
• License

Installation

To get started with the project, follow the steps below:

1. Clone the repository:

   git clone <repository-url>
   cd <repository-name>
   

Set up a virtual environment

python -m venv venv source venv/bin/activate # On Windows use venv\Scripts\activate

pip install -r requirements.txt

Project Structure

The following is an overview of the repository's folder structure. The dataset folders (20_Percent, 50_Percent, 75_Percent, and All) contain data from different geographic location. The data is divided by percentage to see that the proposed scheme outperform the exixiting methods even if the data is limited.

.
├── 20_Percent
│   ├── canada_data
│   │   ├── ...
│   ├── chilli_data
│   │   ├── ...
│   ├── kazakistan
│   │   ├── ...
│   └── Mongolia
│       ├── ...
├── 50_Percent
│   ├── canada_data
│   │   ├── ...
│   ├── chilli_data
│   │   ├── ...
│   ├── kazakistan
│   │   ├── ...
│   └── Mongolia
│       ├── ...
├── 75_Percent
│   ├── canada_data
│   │   ├── ...
│   ├── chilli_data
│   │   ├── ...
│   ├── kazakistan
│   │   ├── ...
│   └── Mongolia
│       ├── ...
└── All
    ├── canada_data
    │   ├── ...
    ├── chilli_data
    │   ├── ...
    ├── kazakistan
    │   ├── ...
    └── Mongolia
        ├── ...

Running the Project

1. Navigate to the Desired Dataset Folder: Change directory to the specific dataset folder (e.g., 20_Percent, 50_Percent, or 75_Percent).
2. Modify Shell Script Parameters (if necessary): Open the corresponding .sh file (e.g., run.sh) and adjust parameters such as NUM_EPOCHS, BATCH_SIZE, and INITIAL_LR as needed.
3. Execute the Shell Script: Run the script in your terminal: bash run.sh

Custom Hybrid Loss Function

Our model employs a custom hybrid loss function that incorporates Physics-Informed Vectors (PIV) and traditional Mean Squared Error (MSE). The PIV term initially has a strong influence on predictions, guiding the model towards physical accuracy, but this influence gradually decays during training.

The loss function is defined as follows:

This hybrid loss function improves model convergence and predictive accuracy by initially aligning the model closely with physics-based predictions and later tuning it to focus on data-driven corrections.

Results

The integration of Physics-Informed Vectors (PIV) significantly improves predictive performance across multiple datasets. Below are the performance metrics across different subsets of data:

Key Performance Metrics

• Mean Absolute Error (MAE):
  
• Mean Squared Error (MSE):
  
• R-Squared (R²):
  

The graphs above illustrate the improvements achieved by incorporating PIV into standard LSTM and TCN models. The percentage improvements are noted in terms of accuracy and convergence speed.

Citation and Publication

This work has been presented and published at the 2024 14th Asian Control Conference (ASCC).

Title: Integrating Physics-Informed Vectors for Improved Wind Speed Forecasting with Neural Networks
Authors: Laeeq Aslam, Runmin Zou, Ebrahim Awan, Sharjeel Abid Butt
Conference: 2024 14th Asian Control Conference (ASCC)
Pages: 1902-1907
Year: 2024
Organization: IEEE
Link to Paper

Citation:

@inproceedings{aslam2024integrating,
  title={Integrating Physics-Informed Vectors for Improved Wind Speed Forecasting with Neural Networks},
  author={Aslam, Laeeq and Zou, Runmin and Awan, Ebrahim and Butt, Sharjeel Abid},
  booktitle={2024 14th Asian Control Conference (ASCC)},
  pages={1902--1907},
  year={2024},
  organization={IEEE}


}