Physics-Informed Neural Network (PINN) for Thermal Modeling of Power Transformers

📌 Project Overview

This repository implements a Physics-Informed Neural Network (PINN) using PyTorch to predict the Hot-Spot Temperature (HST) in power transformers. Unlike standard purely data-driven models, this approach embeds the governing 1D Heat Diffusion Partial Differential Equation (PDE) directly into the neural network's loss function.

By forcing the network to respect physical laws, the model achieves robust, highly accurate predictions even when physical sensor measurements are noisy or sparse.

---

📐 The Physics Engine & Governing Equations

The model solves the 1D heat equation with a localized heat source term generated by electrical load losses ($I^2 \cdot R$):

$$\frac{\partial T}{\partial t} = \alpha \frac{\partial^2 T}{\partial x^2} + \gamma I^2(t)$$

Where:

• $T(x,t)$ is the winding temperature across space and time.
• $\alpha$ is the thermal diffusivity coefficient of the copper/insulation medium.
• $I(t)$ is the normalized transformer load current factor.
• $\gamma$ represents the physical heat generation constant.

The optimization space minimizes a dual-component objective function:

$$ \mathcal{L}_{\text{total}} = \mathcal{L}_{\text{data}} + \lambda \mathcal{L}_{\text{physics}} $$

---

🛠️ Repository Tree Configuration

transformer-pinn/
│
├── data/
│   └── synthetic_thermal_data.csv    # Tabular physics-baseline dataset
│
├── src/
│   ├── __init__.py
│   ├── physics.py                    # PyTorch autograd PDE residual loss
│   ├── model.py                      # Multi-layer Perceptron with Tanh activations
│   ├── data_utils.py                 # StandardScaler pipeline and tensor utilities
│   └── pinn_transformer_model.pth    # Saved trained neural network weights
│
├── .gitignore                        # Excludes untracked background clutter
└── README.md                         # Project documentation