# GridLoss AI: full description

## What it is
GridLoss AI turns a DISCOM's existing data (smart meters, energy accounting, billing, asset sensors, supply) into one reconciled view of aggregate technical and commercial (AT&C) loss and a ranked list of actions. The same screens answer five questions in order: See (how much is lost), Trace (where and why), Predict (what happens next), Act (what to do and the rupee impact), Prove (can the result be trusted).

## Current figures (60-day window, synthetic data)
- Revenue loss: Rs 5,877,018
- AT&C loss: 12.94% of energy input
- Energy lost: 807,248 kWh of 6,240,598 kWh input
- Consumers: 2,000; feeders: 10

## Loss ledger (how the numbers tie)
Input energy minus billed energy plus the collection gap is split into five buckets: technical loss, theft, metering error, billing gap and collection gap. Each bucket is given in kWh and in rupees (kWh at the feeder's realised tariff; collection in exact rupees). Network totals equal the sum of feeders, and each feeder equals the sum of its line, transformer and HT-consumer nodes. The identity is exact, so a figure on the Overview equals the same figure in any drill-down.

## Network and data
One distribution network: 220 kV receiving substations, 33/22/11 kV feeders, distribution transformers, then consumers. Industrial consumers connect at 11, 22 or 33 kV, commercial at HT or LT, domestic at 230 V single phase or 415 V. Supply comes from 16 conventional and renewable sources (coal, nuclear, gas, solar, wind, hydro, small hydro, biomass). Data is synthetic and seeded; anomalies are injected with known labels (theft, reverse energy, meter failure, billing mismatch, technical loss) and used only to score models. Browse every table at /data/tables.

## Capabilities and the eight problems
1. Load forecasting: /forecast/overview, /forecast/series, /forecast/accuracy. Feeder, substation, transformer and category level, with intervals and baselines.
2. Supply constraint: /supply/load-shedding/scenario, /supply/load-shedding/critical-loads. Linear-programme plan against blind rotation.
3. AT&C loss attribution: /ledger/network, /ledger/feeders/{id}.
4. Theft and anomaly ranking: /investigations, /investigations/{consumer_id}. Probability with evidence text.
5. Grid congestion: /forecast/congestion, /forecast/growth.
6. Renewable variability and DSM: /supply/renewables/overview, /forecast, /dsm, /variability.
7. Asset performance: /assets/failure-risk, /assets/failure-risk/{dt_id}, /assets/failure-evaluation.
8. Billing integrity: /leakage/summary, /leakage/accounts, /leakage/evaluation.

## Method summary
Supervised gradient-boosted models with out-of-fold or held-out evaluation, rule signals kept transparent, a closed ledger for attribution, per-level gradient-boosted forecasters compared with seasonal-naive baselines, and a linear programme for load shedding. Each module publishes its metrics at /submission/testing, with baselines and limitations.

## Honest limits
Labels come from anomalies the generator injected, so scores show that methods recover a known pattern; they are not field performance. Some naive baselines match the model on this synthetic data and the pages say so. A pilot must backtest on real utility history and inspection outcomes. Deviation-settlement bands and tariffs are labelled assumptions.

## Data privacy
Synthetic data only. Consumers are pseudonymous identifiers. The application runs inside the utility environment and calls no external service at run time.

## Dashboard pages
- http://poc.nidhitek.com/overview (Overview, See): How much energy and revenue is lost, split by cause, with the feeder watchlist.
- http://poc.nidhitek.com/energy-flow (Energy flow, See): Sankey diagram from generation sources through the 220 kV substations and feeders to consumers, with conservation checks.
- http://poc.nidhitek.com/feeder-analytics (Feeder analytics, Trace): Ranked feeders and a substation to feeder to transformer to meter drill-down where parts add to the whole.
- http://poc.nidhitek.com/transformers (Transformers, Trace): All 88 distribution transformers with loss breakup, loading and meters.
- http://poc.nidhitek.com/consumers (Consumers, Trace): Consumer investigation with evidence, peer comparison and case workflow.
- http://poc.nidhitek.com/forecast (Load forecast, Predict): Demand forecast at feeder, substation, transformer and category level, with congestion and growth outlook.
- http://poc.nidhitek.com/renewables (Renewables and DSM, Predict): Solar, wind, hydro and biomass forecasts, deviation-settlement exposure in rupees.
- http://poc.nidhitek.com/asset-health (Asset health, Predict): Transformer failure risk with lead time and drivers.
- http://poc.nidhitek.com/queue (Action queue, Act): Prioritised cases with reason flagged, rupee value, action, owner and outcome tracking.
- http://poc.nidhitek.com/revenue-leakage (Revenue leakage, Act): Meter to bill to cash reconciliation and flagged accounts.
- http://poc.nidhitek.com/load-shedding (Load shedding, Act): Optimised load-shedding plan against blind rotation, protecting critical loads.
- http://poc.nidhitek.com/field-programs (Field programs, Act): Loss-reduction levers and expected recovery.
- http://poc.nidhitek.com/collections (Collections and subsidy, Act): Collection shortfall and agricultural subsidy tracking.
- http://poc.nidhitek.com/copilot (Copilot, Act): Question-answering over the ledger and cases, with guardrails.
- http://poc.nidhitek.com/model-evidence (Model evidence, Prove): Accuracy against injected ground truth, with limitations.
- http://poc.nidhitek.com/raw-data (Raw data, Prove): Every dataset the proof of concept runs on, paged, searchable and downloadable.
- http://poc.nidhitek.com/submission (Challenge submission, Prove): Dataset description, testing report and documentation for the Rodic challenge.
- http://poc.nidhitek.com/deployment (Deployment, Prove): Dataset contracts, ingest validator, footprint and scale path.

## API endpoints (from the OpenAPI schema)
- GET /assets/failure-evaluation: Honest evaluation of the failure model: pooled out-of-fold and strict time-split ROC-AUC, PR-AUC, precision and recall at
- GET /assets/failure-risk: Transformers ranked by expected cost avoided: calibrated probability of failure within 30 days,
- GET /assets/failure-risk/{dt_id}: One transformer: probability, band, drivers in words (top three, from tree contributions of the fold model that did not
- GET /assets/maintenance: Predictive maintenance: which transformer is degrading, and by when.
- GET /collections/risk: Propensity to pay, receivables segments and the action for each.
- GET /consumers/{consumer_id}/events: Consumer Events
- GET /consumers/{consumer_id}/peers: Consumer Peers
- GET /consumers/{consumer_id}/profile: Consumer Profile
- GET /contracts: The schema a DISCOM / IES-compatible feed has to satisfy.
- POST /copilot/ask: Ask a question about the current dataset.
- GET /copilot/selftest: One small live call to the model, to show whether key, model name and tool calling work, or the provider's error.
- GET /copilot/status: Whether the optional language-model pass is configured.
- GET /data-quality: Data Quality
- GET /data/tables: Data Tables
- GET /data/tables/{name}: Data Table
- GET /data/tables/{name}/download: Data Download
- GET /dts: Dts
- GET /dts/{dt_id}: Dt
- GET /dts/{dt_id}/trend: Dt Trend
- GET /energy/balance: The whole chain in one place: generation down to energy collected.
- GET /energy/flow: Generation to revenue as nodes and links, with the conservation checks that prove it.
- GET /evaluation: Evaluation
- GET /evaluation/curve: Evaluation Curve
- GET /feedback: Feedback
- GET /feeders: Feeders
- GET /feeders/{feeder_id}: Feeder
- GET /feeders/{feeder_id}/diagnosis: Why this feeder specifically is losing energy, and what to do about it.
- GET /feeders/{feeder_id}/trend: Feeder Trend
- GET /forecast/accuracy: Rolling-origin backtest (train before the final 168 h, test on it) per level and horizon bucket, model versus seasonal-naive baselines.
- GET /forecast/congestion: Ranked congestion watchlist for the next 7 days: overload probability, expected overload hours, first breach, peak loading with 80% band, action and owner.
- GET /forecast/growth: Augmentation priority list: robust (Theil-Sen) trend of daily peak demand projected 90 days ahead with an uncertainty band, flagged against rated capacity.
- GET /forecast/overview: Headline KPIs, accuracy summary, top congestion, top growth and cost of forecast error. Computed lazily once, then memoised.
- GET /forecast/series: Hourly history, 168 h forecast with 80% band, the capacity line and the backtest forecast for the last week, all in kW of metered demand.
- GET /generation/conventional: The conventional fleet, decomposed. Shares sum to 100% by construction.
- GET /generation/mix: Conventional and renewable circuits, by capacity and by actual output.
- GET /generation/renewable-profile: Renewable output shape over the day - the forecasting target.
- GET /health: Health
- POST /ingest/validate: Dry-run a candidate data source against the contracts.
- GET /inspection/triage: Which line to inspect next, and which defect to fix first.
- GET /investigations: The ranked queue - the actual product output for a field team.
- GET /investigations/{consumer_id}: Investigation
- POST /investigations/{consumer_id}/feedback: Capture what the field visit found. This is the loop the pilot needs.
- GET /leakage/accounts: Accounts ranked by expected leakage in Rs: leakage probability, estimated Rs for the 60-day window and annualised,
- GET /leakage/accounts/{consumer_id}: One account: leakage scores per type, every flagged reason, and the bill-by-bill reconciliation
- GET /leakage/evaluation: Out-of-fold evaluation against ground_truth BILLING_MISMATCH and METER_FAILURE: precision@K, recall, ROC-AUC, PR-AUC,
- GET /leakage/summary: Meter -> bill -> cash reconciliation for the network, by feeder and by consumer category (kWh and Rs, 60-day window),
- GET /ledger/feeders/{feeder_id}: One feeder, split into line, HT consumers and every transformer. Parts add to the feeder.
- GET /ledger/network: Network revenue loss and its breakup, plus every feeder and substation that sums to it.
- GET /ledger/nodes/{node_id}/consumers: The meters behind one transformer (or the HT consumers on a feeder), with the loss each owns.
- GET /loss/breakdown: Technical, commercial, billing and collection loss, kept distinct.
- GET /model/features: Which signals the supervised model actually leans on.
- GET /network/congestion: Where the network is running hot, by feeder and transformer.
- GET /network/load-transfer: Where a hot feeder could shed load, and the I2R that saves.
- GET /network/topology: The asset hierarchy: substations, feeds and the voltage ladder.
- GET /outcomes: Outcomes
- GET /overview: Command-centre KPIs: network, energy, risk and data-quality blocks.
- GET /submission/ai-engine: Inventory of every model: technique, inputs, outputs, how much is learned, human role; plus the Responsible AI framework.
- GET /submission/all: Dataset, testing report and documentation in one response.
- GET /submission/checklist: Challenge requirements and judging criteria mapped to evidence in the product, with gaps stated.
- GET /submission/dataset: Dataset description: source, size, key features, composition and limitations.
- GET /submission/documentation: Documentation pack: approach, technical implementation, tools and APIs, data-privacy compliance.
- GET /submission/testing: Testing report: accuracy per module with baselines, performance on the target host, and automated-test inventory.
- GET /subsidy/reconciliation: Authorised consumption against subsidy disbursed, with chain verification.
- GET /success-tests: Success Tests
- GET /supply/load-shedding/critical-loads: Synthetic critical-load register (hospitals, water works, traction, telecom) by feeder, sized against real feeder demand.
- GET /supply/load-shedding/scenario: Optimise load curtailment for a supply shortfall in a window of the next day and compare it with blind rotation.
- GET /supply/renewables/dsm: DSM deviation energy and penalty exposure (Rs) for the test week and annualised, by source type and fleet, for the
- GET /supply/renewables/forecast: Day-ahead forecast for one source type: actual versus model, persistence and same-time-yesterday on a test day,
- GET /supply/renewables/overview: Headline accuracy, DSM and variability figures for renewable generation (final-7-day backtest, 15-minute blocks).
- GET /supply/renewables/variability: Ramp rates (p95 MW per 15 minutes), coefficient of variation, capacity factor, renewable share of generation by hour

## Glossary
AT&C: aggregate technical and commercial loss. DISCOM: distribution company. DT: distribution transformer. DSM: deviation settlement mechanism for generation schedules. IES: India Energy Stack. WAPE: weighted absolute percentage error. ROC-AUC and PR-AUC: ranking-quality measures for classifiers.