izikpi/electricity-theft-detection

Hybrid CNN-LSTM deep learning model for detecting electricity theft in power distribution systems, using real consumption data from the Electricity Company of Ghana. # AI-Based Electricity Theft Detection in Distribution Systems A hybrid CNN-LSTM deep learning model that detects electricity theft from customer consumption profiles in power distribution networks. Developed as a final-year undergraduate project at the Kwame Nkrumah University of Science and Technology (KNUST), using real consumption data from the Electricity Company of Ghana (ECG). ## Problem Non-technical losses from electricity theft (meter tampering, illegal connections, bypassing) are a significant source of revenue loss for utilities. Manual inspection is slow and costly. This project frames theft detection as a binary classification problem on monthly consumption data, where abnormal consumption patterns are flagged automatically for inspection. ## Approach The pipeline has three stages: 1. **Preprocessing** (`src/preprocessing.py`) — missing and non-numeric meter readings are mean-imputed, features are standardised, and four statistical descriptors (interquartile range, mean, standard deviation, skewness) are engineered per customer to summarise the shape of the yearly profile. The combined feature set is then projected to a 2-D embedding with t-SNE, which exposes the cluster structure separating fraudulent from normal customers. 2. **Hybrid model** (`src/hybrid_cnn_lstm.py`) — a 1-D CNN feature extractor is wrapped in a `TimeDistributed` layer and followed by an LSTM. The CNN learns local consumption patterns while the LSTM models how those patterns evolve, and a sigmoid output produces the theft/normal decision. 3. **Baselines** (`src/baselines.py`) — standalone CNN, LSTM and RNN models trained on the same features and split, used as comparison points. ## Results On a stratified 70/30 train-test split of the ECG dataset, the hybrid CNN-LSTM model outperformed all three baselines across every metric: | Model | Accuracy | F1-score | Precision | Recall | | ---------- | -------- | -------- | --------- | ------- | | CNN-LSTM | 94.44% | 76. …