MBA Capstone · Tata Power Ajmer Distribution Limited (TPADL) · 2025
Author: Divyansh Jain (084015)
111,792 consumers scored · ₹24.02 Cr net NPV · 2.91× BCR · 8.68-month payback
India's power sector loses $17B annually to electricity theft. National AT&C losses hit 16.12% in FY 2023-24. TPADL's AT&C ~18% — 3 pts above RDSS target.
This project builds a three-layer, rule-based, interpretable theft detection framework that:
- Scores all 111,792 consumers across 12 months of smart meter data
- Produces prioritised inspection dispatch list
- Quantifies financial impact at TPADL and Tata Power group level
Designed for field operability — every formula, weight, threshold documented. Billing engineers can read, audit, calibrate. No data science expertise required.
Raw Dataset (164,076 consumers)
│
▼
┌─────────────────────────────────────────────────┐
│ PREPROCESSING │
│ P1 → Null out months ≤10 kWh (noise) │
│ P2 → Exclude consumers with ≤5 valid months │
│ P3 → Compute Coverage Ratio per consumer │
└─────────────────────────────────────────────────┘
│ │
111,792 SCORED 52,284 EXCLUDED (Watchlist)
│
▼
┌─────────────────────────────────────────────────┐
│ LAYER 1 — CPI (40%) Consumption Pattern Index │
│ CTS · CVI · DSR · ZSB · SCAS · FCS · PDS │
└─────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────┐
│ LAYER 2 — LCI (35%) Load Compliance Index │
│ LFR · MOI · MCGI · ZCSS │
└─────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────┐
│ LAYER 3 — VHS (25%) Violation History Score │
│ Violation Count (step function) │
└─────────────────────────────────────────────────┘
│
▼
Final_Score = (0.40 × CPI) + (0.35 × LCI) + (0.25 × VHS)
│
OVERRIDE RULES
O1: Violation ≥2 AND CPI ≥55 → HIGH [0 fired]
O2: MDI_Ratio ≥ 2.0 → HIGH [49,859 fired]
O3: Load ≥25kW AND LFR <3% → HIGH [122 fired]
│
▼
HIGH (≥65) · MEDIUM (35–64) · LOW (<35)
| Parameter | Value |
|---|---|
| Source | Tata Power Ajmer Distribution Limited |
| Raw consumer base | 164,076 |
| Scored pool (after preprocessing) | 111,792 |
| Excluded (watchlist) | 52,284 (31.9%) |
| Data period | Jan–Dec (12 months) |
| Features per consumer | 34 raw → 39 engineered |
| Tariff source | AVVNL Tariff 2023 |
Rate category breakdown:
| Category | Consumers | Share |
|---|---|---|
| DS-LT1 (Domestic Single-Phase) | 94,970 | 84.95% |
| NDS-LT2 (Non-Domestic LT) | 14,574 | 13.04% |
| SP-LT5 (Special Purpose Industrial) | 1,339 | 1.20% |
| PSL-LT3 (Public Street Lighting) | 399 | 0.36% |
| AG-MS-LT4 (Agricultural) | 372 | 0.33% |
| Others (HT/Mixed) | 138 | 0.12% |
| Variable | Weight | Signal |
|---|---|---|
| CTS — Consumption Trend Slope | 20% | Sustained falling trend → bypass installed |
| CVI — Consumption Volatility Index | 20% | Erratic month-to-month → partial bypass |
| DSR — Drop Severity Ratio | 15% | Extreme peak-to-trough gap |
| ZSB — Z-Score Statistical Break | 10% | Single-month cliff → mid-year bypass |
| ⭐ SCAS — Summer Consumption Anomaly | 15% | Low summer vs off-season → Rajasthan bypass |
| ⭐ FCS — Flat Consumption Score | 10% | Near-zero CoV → tampered register |
| ⭐ PDS — Peer Deviation Score | 10% | Below peer median → under-recording |
⭐ New in v3.0 — fills gaps in original 4-variable CPI
| Variable | Weight | Signal |
|---|---|---|
| LFR — Load Factor Ratio | 30% | kWh << sanctioned load × hours |
| MOI — MDI Overdrawal Index | 25% | Peak demand > 2× sanctioned capacity |
| MCGI — MDI-Consumption Gap Index | 30% | kWh inconsistent with MDI-implied draw |
| ZCSS — Zero Coverage Suspicion Score | 15% | Missing months on large connections |
| Violation Count | Score |
|---|---|
| 0 | 0 |
| 1 | 40 |
| 2 | 65 |
| 3 | 82 |
| ≥4 | 100 |
| Category | Count | % of Scored Pool |
|---|---|---|
| 🔴 HIGH Risk | 49,981 | 44.71% |
| 🟡 MEDIUM Risk | 550 | 0.49% |
| 🟢 LOW Risk | 61,261 | 54.80% |
| ⚪ EXCLUDED (watchlist) | 52,284 | — |
Override breakdown:
- O1 (Violation + CPI): 0 fires — 99.5% consumers have zero violation history
- O2 (MDI ≥ 2×): 49,859 fires — primary detection mechanism
- O3 (Large load + near-zero LFR): 122 fires
Max computed Final Score: 59.77 — no consumer reached 65 via pure scoring. Phase 2 data required to activate pure-score HIGH.
Top 10,000 consumers: 7.15 MU estimated stolen · ₹5.25 Crore gross annual loss
| Metric | Value |
|---|---|
| Gross Annual Revenue Loss Detected | ₹3.56 Crore |
| Recoverable (60% factor) | ₹2.13 Crore |
| Detection Capture Rate | 15.26% of commercial loss pool |
| Total Commercial Loss Pool | ₹23.29 Crore |
| Blended Tariff (AVVNL 2023) | ₹7.14/kWh |
| Scenario | AT&C Target | Annual Saving | 10-Yr NPV | PAT Impact | TPADL EBITDA |
|---|---|---|---|---|---|
| Conservative | 16.5% | ₹3.49 Cr | ₹21.47 Cr | ₹2.62 Cr | +16.3% |
| Base (RDSS Target) | 15.0% | ₹6.99 Cr | ₹42.94 Cr | ₹5.24 Cr | +32.6% |
| Optimistic | 13.0% | ₹11.65 Cr | ₹71.57 Cr | ₹8.74 Cr | +54.3% |
| Best Case | 12.0% | ₹13.98 Cr | ₹85.88 Cr | ₹10.48 Cr | +65.2% |
NPV at 10% WACC · PAT at 25% corporate tax · Distribution volume: 326.19 MU/year
| Metric | Value |
|---|---|
| Net Programme NPV (10-year) | ₹24.02 Crore |
| Benefit-Cost Ratio | 2.91× |
| Payback Period | 8.68 months |
| Full Programme Cost | ₹15.16 Crore (50,531 visits × ₹3,000) |
| Year-1 Break-even Threshold | 58.35 kWh/consumer/month |
| Perpetuity Break-even Threshold | 3.50 kWh/consumer/month |
TPADL-Theft-Detection-Engine/
│
├── 📁 python/
│ ├── tpadl_theft.py # Layer scoring engine (CPI + LCI + VHS)
│ └── tpadl_financials.py # Financial analysis module
│
├── 📁 data/
│ ├── combined_anonymized_unique_id_summary.csv # Raw 164,076-consumer dataset
│ ├── TPADL_Clean_Data.csv # Scored pool (111,792)
│ ├── TPADL_Layer1_CPI_v2.csv # Layer 1 outputs
│ └── TPADL_Layer2_LCI_v1_up.csv # Layer 2 outputs
│
├── 📁 outputs/
│ ├── TPADL_Complete_Theft_Analysis_Report.csv
│ ├── TPADL_Complete_Theft_Analysis_Report_Top10000.csv
│ ├── TPADL_Complete_Theft_Analysis_Report_Dashboard.csv
│ ├── TPADL_Financial_Analysis_Data.csv
│ ├── TPADL_Financial_Analysis_Report_*.csv
│ └── TPADL_Framework_v3.docx
│
├── 📁 website/
│ ├── index.html # Main project dashboard
│ └── data.html # Source data sheets (embedded Google Sheets)
│
├── 📁 report/
│ └── Project_Report.docx
│
└── README.md
| Tool | Use |
|---|---|
| Python 3.x | Detection engine, financial module |
| Google Colab | Primary development environment |
| pandas / numpy | Data processing and scoring |
| openpyxl | Multi-sheet Excel workbook generation |
| Chart.js 4 | Interactive dashboard visualisations |
| HTML / CSS / JS | Project website (no framework) |
| Google Sheets | Live data embedding |
| Google Fonts | Rajdhani + Inter + JetBrains Mono |
# Clone repo
git clone https://github.com/[username]/TPADL-Theft-Detection-Engine.git
# Install dependencies
pip install pandas numpy openpyxl
# Run detection engine
python python/tpadl_theft.py
# Run financial analysis
python python/tpadl_financials.pyOpen website/index.html directly in browser. No server required.
website/data.html requires internet access (embedded Google Sheets).
-
Override O1 = 0 fires — 99.5% consumers have zero violation history. Enforcement-history-dependent systems fail in first-cycle Indian DISCOM contexts. Behavioral scoring only viable approach.
-
Override O2 dominates — 49,859 of 49,981 HIGH consumers elevated because MDI exceeded 2× sanctioned load. MDI = most powerful indicator in monthly data without electrical parameter readings.
-
Agriculture flag rate 30.9% — AG-MS-LT4 highest proportional flagging despite 0.33% of scored pool. SCAS correctly identifies summer irrigation bypass patterns.
-
Max score 59.77 — Phase 2 data (billing + DT mapping + V/I readings) required to activate pure-score HIGH classifications.
-
₹24.02 Cr NPV · 2.91× BCR · 8.68-month payback — Investment case robust across all 16 parameter combinations. Even most conservative scenario = positive NPV.
| # | Contribution |
|---|---|
| 1 | Interpretable rule-based architecture — No black-box ML. Every rule auditable by billing engineers, satisfying legal enforceability for Indian revenue tribunals |
| 2 | Network validation via override rules — Override O3 formalizes zone-level dormant connection detection not addressed in prior frameworks |
| 3 | Field-linked financial quantification — First study linking consumer-level detection to ACS-ARR gap, EBITDA, PAT addback, EPS accretion, BCR, and consumer-level break-even thresholds |
| Limitation | Impact | Phase 2 Fix |
|---|---|---|
| No monthly billing data | CDI and BMR uncomputable | P1 priority data request |
| No V/I/PF readings | CT tampering undetectable | P2 priority |
| No DT mapping | Zone-level balancing impossible | P1 priority |
| Monthly granularity only | Temporal evasion undetectable | Hourly data (P2) |
| Single year of data | YoY comparison unavailable | Accumulates automatically |
| Estimated leakage only | Not for demand notices | Confirmed field evidence required |
- Power Finance Corporation (2024). Report on Performance of State Power Utilities 2023-24
- Savian et al. (2021). Non-technical losses: systematic review. Renewable & Sustainable Energy Reviews, 147, 111205
- Kawoosa et al. (2023). ML ensemble for energy theft detection. IET Generation, Transmission & Distribution, 17(16)
- AVVNL. Tariff for Supply of Electricity 2023
Divyansh Jain — 084015
MBA Capstone Project · Tata Power Ajmer Distribution Limited · 2025
⚠️ Disclaimer: This framework is a targeting tool, not a verdict. Every HIGH-flagged consumer requires physical field verification before enforcement. All leakage figures are peer-benchmark estimates — must not be used in demand notices without confirmed field evidence.