Shakti
Use शक्ति, So that you don't waste yours
The problem Shakti solves
The Problem SHAKTI Solves
Electricity theft remains one of the largest contributors to non-technical losses in power distribution systems, causing severe financial strain on utilities and governments. In India alone, state-owned DISCOMs lose tens of thousands of crores annually due to theft, inefficient inspections, and delayed detection—losses that are ultimately borne by taxpayers.
Current detection mechanisms are largely reactive, rule-based, and fragmented. They rely on static thresholds, manual inspections, and isolated meter analysis, which:
fail to detect subtle or intermittent theft,
generate high false positives,
lack spatial and grid-level context, and
do not scale effectively across large distribution networks.
As a result, inspection resources are often misallocated, genuine theft cases go undetected for long periods, and infrastructure issues such as faulty transformers remain hidden within consumer-level data.
SHAKTI addresses this gap by transforming raw meter and grid data into actionable, explainable intelligence—enabling utilities to detect electricity theft early, prioritize inspections effectively, and strengthen overall power distribution governance using advanced machine learning.
Our Approach:
- Data Ingestion
Smart Meter Data: Consumption patterns at consumer level
Transformer / Feeder Data: Aggregate load, losses, imbalance
Contextual Data: Weather, season, tariff category, location
This gives both micro (user) and macro (grid) visibility.
- Parallel Analysis Layer (Independent Checks)
Multiple analyses run simultaneously to avoid single-point bias:
Anomaly Detection (Usage Patterns)
Detects sudden drops, spikes, flat-lining, or abnormal load curves using ML models.
Peer Comparison (Local Neighbors)
Compares a consumer with similar users (same area, tariff, load class).
Transformer Loss Localization
Identifies abnormal technical vs non-technical losses at feeder/DT level.
Voltage / Power Quality (PQ) Analysis
Checks voltage dips, fluctuations, power factor anomalies linked to theft.
Seasonal / Agricultural Pattern Checks
Ensures changes align with crop cycles, irrigation periods, or seasonal demand.
- Risk Signal Generation
Each parallel module outputs risk indicators:
A → Anomaly score
P → Peer deviation score
T → Transformer loss contribution
S → Seasonal inconsistency score
These are independent evidences, not raw guesses.
- Normalized Risk Scoring
All risk signals are normalized to a common scale.
A weighted risk score is computed:
Final Risk=𝑤1*(𝐴𝐴)+𝑤2*(𝑃𝑃)+𝑤3*(𝑇𝑇)+𝑤4*(𝑆𝑆)
Weights reflect real-world importance, not just statistics.
- Temporal Suspicion Tracking
Risk scores are tracked over time, not judged on a single day.
Ensures consistency of suspicion (reduces false positives).
One-time anomalies are filtered out.
- Human Inspection Priority List
Consumers/feeders are ranked by persistent risk.
Output is:
Explainable (why flagged)
Evidence-backed (multiple signals agree)
Field teams inspect only high-confidence cases, saving cost and effort.
- Feedback Loop (Implicit)
Inspection results improve:
Model weights
Thresholds
Pattern recognition accuracy over time
Challenges we ran into
Key Challenges :
- Real-time data accumulation :
Our project requires simple but essential datasets. Initial aim was to scrape real time government data, but we were soon halted as the data we were able to find had no correlation to the other datasets, rendering the process in vain, combining them into one would result in a clutter of data which won't fit in with the rest.
Solution -> Use synthesized datasets and merge them into one.
Training the model from scratch required multiple iterations to fine-tune the weights of different risk signals. This iterative process was time-consuming but necessary to ensure balanced, accurate, and explainable predictions.
Although it may appear to be a temporary solution, this approach effectively demonstrates the full scope and scalability of the system. Designed for B2B and government use cases, SHAKTI is built to continuously improve as it ingests live, curated data from official sources. Increased data availability directly enhances model accuracy, risk prioritization, and decision reliability over time.
Tracks Applied (1)
The Requestly "Creative Use" Prize
Requestly