Hybrid-Order Representation Learning for Electricity Theft Detection

Electricity theft is the primary cause of electrical losses in power systems, existing methods usually detect anomalies in electricity consumption data on the first-order information and ignore the second-order representation learning that can efficiently model global temporal dependency and facilitates discriminative representation learning of electricity consumption data.

We propose a novel and lightweight end-to-end Hybrid-Order Representation Learning Network (HORLN) to identify electricity thieves. To the best of our knowledge, this is the first attempt to incorporate second-order information with regular first-order based deep architectures for electricity theft detection.

For more details, please refer to our paper Hybrid-Order Representation Learning for Electricity Theft Detection.

Requirements

• python3.7
• numpy
• pytorch
• scikit_learn (for implementation of the comparing method RF)
• xgboost (for implementation of the comparing method XGB)
• tsai (for implementation of the comparing method MiniRocket)

Dataset

We conducted our experiments on a public real-world dataset (link). We preprocessed the dataset and randomly split it into three sets for training, validation, and testing. The preprocessed data has been saved in the datasets/electricity.zip file, please unzip the file and put it to the datasets/electricity folder.

Train

The command to train our HORLN has been written in Ours/train.sh. Please run the bash file and the HORLN will be trained for 200 epochs. The parameters of the models will be saved in the checkpoints/electricity_elec_horln folder. Please add a parameter --gpu_ids -1 when running with CPU.

cd Ours
bash train_horln.sh

Validate and Test

After training, the saved models of each epoch could be validated and tested by the following commands, respectively. The prediction results will be saved in the results/electricity_elec_horln directory, and two summary tables named electricity_elec_horln_validation.csv and electricity_elec_horln_evaluation.csv will also be generated. Please add a parameter --gpu_ids -1 when running with CPU.

cd Ours
bash val_horln.sh

cd Ours
bash test_horln.sh

Evaluate

By validating the trained model in the validation set, the optimal value of the threshold for F1 score calculation will be automatically calculated for each epoch. Please find the threshold of the epoch you would like to evaluate (e.g., the epoch with the largest F1 score in the validation set) in the electricity_elec_horln_validation.csv file and revise the value of the epoch argument and the best_threshold argument in eval.sh accordingly. With the following command, the performance of the trained model on the testing set with the threshold determined on the validation set could be evaluated.

cd Ours
bash eval_horln.sh

Implementation of State-of-the-art Methods

For RF, XGB and MiniRocket, we adopt implementations from open-source packages and select the parameter settings based on our dataset. Please use cd command and change the working directory to the corresponding folders, and then run the bash file to realize training and evaluation in one step. For CNN and PFSC, we implement the structures in previous works and revise some parameters for fair comparison; for Wide&Deep, and HybridAttention, we reimplement the official code. Please change the working directory to the corresponding folders, and then run the corresponding bash files in the order of train->val->test->eval, which is the same as running our HORLN.

Citation

If you use this code for your research, please cite our paper.

@ARTICLE{HORLN,  
        author={Zhu, Yuying and Zhang, Yang and Liu, Lingbo and Liu, Yang and Li, Guan bin and Mao, Mingzhi and Lin, Liang},  
        journal={IEEE Transactions on Industrial Informatics},   
        title={Hybrid-Order Representation Learning for Electricity Theft Detection},   
        year={2022},  volume={},  number={},  pages={1-1},  
        doi={10.1109/TII.2022.3179243}}