Power Quality State Estimation for Distribution Grids based on Physics-Aware Neural Networks - Harmonic State Estimation

Data set for the paper "Power Quality State Estimation for Distribution Grids based on Physics-Aware Neural Networks - Harmonic State Estimation"

This upload contains

• Training set
• Validation set
• Test set
• Admittance matrices per frequency

used for the paper as pickle files and weights of trained models as zip files.

Weights represent the model with the best validation loss recorded within the first 3000 Epochs of training.

Code for reading in the data sets, preprocessing and state estimation is available in the linked repository.

To replicate the results of the paper follow these steps:

1. clone the linked repository
2. save the provided pickle files in the data folder of the linked repository
3. optional: unzip weights and save them in the data folder, otherwise train a model yourself instead

Version 1.1:

Added data and model weights for the IEEE33 grid to improve comparability.

For the IEEE33 grid, all data (train, test, validation) is saved in one pickle file; see the release tag 1.1.0 in the accompanying GitHub repository for details on the data format. Moreover, the training set size of the new grid was increased from 35040 to 131400 samples to incorporate simulation results that capture a broader range of system states. 

The code was slightly updated to account for inclusion of the IEEE33 grid. Therefore, model weights and input data are now expected in either `cigrelv` or `ieee33` subfolder.

Added Transformer and CNN model weights for IEEE33 and CNN weights for the CIGRE grid. The Transformer model is trained with a smaller batch size since the model did not fit into GPU memory using the same batch size as in other models. This change results in more gradient updates and significantly longer training times, thus the amount of epochs was reduced to achieve a fairer comparison (batch sized reduced from 16384 to 1024, epochs reduced from 3000 to 375, total amount of gradient updates increased from 27000 to 48375). The training of the PANN model over 3000 Epochs is significantly faster than that of the Transformer model trained over 375 epochs (approximately 2.5 hours vs 12.5 hours).