This is the codebase for SCVAE.
This section of the README walks through how to train and evaluate the SCVAE model.
The training code reads time series from a directory of .csv files. In the data folder, we have provided synthesized data including normal data and 6 types of anoamly data.
For creating your own data, simply prepare the time series data in the same format. The .csv files should at least contain colums of "instantaneous_output_power","instantaneous_global_radiation","clearsky_Tamb" for power yeild, irradiation intensity and environment temperature data. You should also scale the data at the time dimension using Z-Score standardization.
To train your model, you should first decide some hyperparameters.Here are some reasonable defaults for a baseline:
"--h_dim 512 --z_dim 128 --mode 2 --learning_rate 1e-5 --batch_size 256 --seq_shape (-1, 12, 6)"
Once you have setup your hyper-parameters, you can run an experiment like so:
python scripts/SCVAE_train.py --h_dim 512 --z_dim 128
The above training script saves checkpoints to .pth files in the "saved_model" directory.
Once you have a path to your model, you can evaluate your model's reconstruction results using synthesized data like so:
python scripts/SCVAE_eval.py --model_path ./saved_model/SCVAE_savedmodel.pth --anomaly_type all
Once you have a path to your model, you can also get the latent variable s of synthesized data as its representations like so. The latent variables are saved as a set of .npy files in the "re_z" folder.
python scripts/get_latent_var.py --model_path ./saved_model/SCVAE_savedmodel.pth --is_trainsites True
Once you have a set of latent variables files, you can do some clustering and classification experiments to analysis these latent variables like so:
python scripts/cluster_and_classify.py --latent_var_path ./re_z --task clustering --latent_mode diff --anomaly_type whole_anomaly
python scripts/cluster_and_classify.py --latent_var_path ./re_z --task classification --latent_mode diff --anomaly_type whole_anomaly --classifier XGB