Evidence
How was each claim checked?
Forecast9
Day-ahead forecast error on the test period (115,719 daylight hours, 13 points): 31.2 W per kWp for the model retrained monthly, 39.1 for raw physics, 38.2 for the earlier run A, 43.6 for same-time-yesterday.
How it was checked
Read test_monthly.csv and test_static.csv (row ALL).
Limits
Scored versus a physics model, not versus measurement (archive weather; E08).
results/test_monthly.csv
Forecast accuracy on real rooftops.
How it was checked
No metered rooftop output or ground station is held, so no comparison can be run.
because no metered rooftop output is held
Limits
The target is a PV-model output on Open-Meteo archive weather: versus a physics model, not versus measurement. Data requests in order: metered rooftop output (20-50 consumers, 15-minute or hourly, 12 months), ground GHI (IMD or a state station), MOSDAC.
results/test_monthly.csv
Whole-area P10 to P90 band covers 77.2% of hours (monthly retrain), below the 80% target: the area band is a little narrow.
How it was checked
Read area_test.csv, row new_monthly (coverage 0.772, 8,991 hours).
Limits
Band method is an approximation; error correlation 0.341 measured on 2024 only.
results/area_test.csv
The change of bias before and after October 2025 is caused by the weather-feed switch.
How it was checked
Review notes in the model README: the flip is mostly the 2025 monsoon, not only the feed switch; the cause is not established.
because the cause (weather models, archive source or a wet 2025 monsoon) is not separated
Limits
Do not describe the flip as proven to be the feed switch.
../README.md (review notes)
The newest archive days (from about 6 Nov 2026) are provisional: the archive is back-filled from forecasts.
How it was checked
Stated in the model README: archive within 1-18 W/m2 of the forecast from about 6 Nov. Dropping the last 4 test days changes nothing material (31.30 vs 38.33 vs 39.20). Only the model README states this: no independent check was made here (review fix M16, 18 Nov 2026).
because only the model README states it and no independent check was made here
Limits
Re-download the archive in a few weeks before re-scoring recent days.
../README.md
The 0.954 factor aligns the satellite-based model to PVGIS-ERA5 and was fitted on 2019 to 2021. After the factor the Hyderabad annual model/PVGIS ratio rises from 0.981 (2019) to 1.031 (2023). On the six-point average the model reads +2.0% (2022) and +2.63% (2023) against PVGIS. SME's ratios: 0.986 and 1.043 for Hyderabad; this project found 0.981 and 1.031 (difference stated, not hidden).
How it was checked
Annual sums of pvgis_1kwp and a re-run of the model for 2019 to 2023 (a_items_1_5_8.py).
because the claim carries the SME label [AGENT]: here only its local-file parts are reproduced (the 0.9538 factor fitted on 2019-2021 and the +2.0% / +2.63% 2022-2023 biases: checked - correct, local files only, see E43); the SME's ratios 0.986 and 1.043 are checked - wrong as I read them (0.981 and 1.031 here); 'lowest-yield years of 2005-2023' and 2024-2025 cannot be checked, because only PVGIS 2019-2023 is held
Limits
PVGIS-ERA5 is itself a model; the SME's ratios for 2019 and 2023 (0.986, 1.043) are not reproduced exactly; 2024 and 2025 cannot be checked, because PVGIS data held ends in 2023. SME note on 18 Nov 2026: the fit file is not on disk and the direction is supported by an independent test; in this folder pitch/data/pvmodel_1kwp/calibration_factor.txt (pooled 0.9538, fitted 2019-2021) and bias_check_2019_2023.csv are on disk.
pvgis_1kwp/hyderabad.csv
PVGIS-ERA5, Hyderabad, 17 degree tilt, 14% losses: long-term 1,501 kWh per kWp a year, yearly range 2005-2023 1,418 to 1,548. Roof-mounted (building) setting 1,441.6, about 4% lower: the default free-standing setting is optimistic for rooftops. Global Solar Atlas (satellite, optimal tilt 21 degrees): 1,589.
How it was checked
SME reproduced the PVGIS values and read the Global Solar Atlas value; not re-run here (the 2005-2023 series is not on disk; pvgis_1kwp holds 2019-2023 only, mean 1,456.4, range 1,418.0 to 1,521.1).
(SME label [checked-correct]; not re-checked here; the 2019-2023 part checked here against the local file)
Limits
PVGIS-ERA5 is itself a model. A roof-mounted case is offered on the Forecast page.
data/pvgis_1kwp/hyderabad.csv
Our model's annual yield, Nov 2025 to Oct 2026: Hyderabad 1,579.7 kWh per kWp; mean of the 13 TGSPDCL points 1,604.0 (range 1,558.9 to 1,638.7). That is above the PVGIS band (the 13-point mean about 3% above 1,550) and close to the Global Solar Atlas value (Hyderabad 0.6% below 1,589).
How it was checked
Recomputed here by hosting/yield_check.py: sum of column P_W_per_kWp_calibrated in pitch/data/pvmodel_1kwp/hyderabad.csv (8,760 hours) and of target_w per city in results/test_predictions_monthly.parquet, 1 Nov 2025 to 31 Oct 2026. The audit's 1,604 is the 13-point mean.
(checked here, local files; result in results/annual_yield_check.csv)
Limits
Modelled, not metered. High yield means more export (the cautious side for hosting capacity) but it overstates energy and savings.
results/annual_yield_check.csv
The model uses a 0.954 factor fitted on 2019-2021 (the lowest-yield years); it reads +2.0% high in 2022 and +2.6% high in 2023 against PVGIS on the six-point average.
How it was checked
SME: fit file not on disk, direction supported by an independent test. Here: data/pvmodel_1kwp/calibration_factor.txt (pooled 0.9538) and bias_check_2019_2023.csv are on disk; six-point mean of cal_bias_pct is +2.0 (2022) and +2.63 (2023).
(checked here, local files; the SME's label on direction: supported)
Limits
SME ratios 0.986 and 1.043 are not reproduced (0.981 and 1.031 here); the basis of the SME's ratios is unknown.
data/pvmodel_1kwp/bias_check_2019_2023.csv
| Table | Source file | Rows | Loaded at (UTC) |
|---|---|---|---|
| annual_yield | results/annual_yield_check.csv | 10 | 2026-11-20T05:58:18+00:00 |
| area_band | results/area_test.csv | 4 | 2026-11-20T05:58:18+00:00 |
| cap_unit_sensitivity | results/hosting_cap_unit_sensitivity.csv | 24 | 2026-11-20T05:58:18+00:00 |
| circle_point | data/irradiance_openmeteo_tgspdcl/README.md | 20 | 2026-11-20T05:58:33+00:00 |
| evidence_item | scripts/evidence_items.json | 56 | 2026-11-20T05:58:31+00:00 |
| feeder_dtr | results/feeder_demo.json | 20 | 2026-11-20T05:58:18+00:00 |
| feeder_meta | results/feeder_demo.json | 3 | 2026-11-20T05:58:18+00:00 |
| forecast_accuracy[monthly] | results/test_monthly.csv | 12 | 2026-11-20T05:58:18+00:00 |
| forecast_accuracy[static] | results/test_static.csv | 12 | 2026-11-20T05:58:18+00:00 |
| forecast_by_point[monthly] | results/test_by_point_monthly.csv | 39 | 2026-11-20T05:58:18+00:00 |
| forecast_by_point[static] | results/test_by_point_static.csv | 39 | 2026-11-20T05:58:18+00:00 |
| forecast_example | results/live_area_example.txt | 24 | 2026-11-20T05:58:18+00:00 |
| forecast_hourly | results/test_predictions_monthly.parquet, results/test_predictions_static.parquet | 231438 | 2026-11-20T05:58:33+00:00 |
| growth_class | TGSPDCL_summary/net_meter_clean_22_months.csv (computed by hosting/growth_classes.py) | 18 | 2026-11-20T05:58:18+00:00 |
| hosting_bootstrap | results/hosting_p5_bootstrap_ci.csv | 9 | 2026-11-20T05:58:18+00:00 |
| hosting_runs[indian] | results/indian_runs.csv | 4200 | 2026-11-20T05:58:18+00:00 |
| hosting_runs[load025] | results/hosting_lv_runs.csv | 540 | 2026-11-20T05:58:18+00:00 |
| hosting_runs[load050] | results/hosting_lv_load050_runs.csv | 540 | 2026-11-20T05:58:18+00:00 |
| hosting_summary[indian] | results/indian_summary.csv | 21 | 2026-11-20T05:58:18+00:00 |
| hosting_summary[load025] | results/hosting_lv_summary.csv | 18 | 2026-11-20T05:58:18+00:00 |
| hosting_summary[load050] | results/hosting_lv_load050_summary.csv | 18 | 2026-11-20T05:58:18+00:00 |
| kusum_cap_sensitivity | results/hosting_kusum_cap_sensitivity.csv | 16 | 2026-11-20T05:58:18+00:00 |
| kusum_grid | results/hosting_kusum.csv | 16 | 2026-11-20T05:58:18+00:00 |
| net_meter_category_month | TGSPDCL_summary/net_meter_clean_22_months.csv | 163 | 2026-11-20T05:58:18+00:00 |
| net_meter_gap | TGSPDCL_summary/net_meter_clean_22_months.csv (derived) | 23 | 2026-11-20T05:58:19+00:00 |
| net_meter_section_month | TGSPDCL_summary/net_meter_clean_22_months.csv | 10099 | 2026-11-20T05:58:18+00:00 |
| net_meter_series_break | TGSPDCL_summary/net_meter_clean_22_months.csv (derived) | 7 | 2026-11-20T05:58:19+00:00 |
| nowcast_backtest | nowcast/results/backtest.csv | 6 | 2026-11-20T05:58:33+00:00 |
| real_dataset | computed from the files above | 8 | 2026-11-20T05:58:31+00:00 |
| real_fact | telangana_reports/mnre_2026_02.pdf; docs/Challenge16_DER_Final.md | 3 | 2026-11-20T05:58:31+00:00 |
| real_not_held | scripts/load_data.py (NOT_HELD) | 5 | 2026-11-20T05:58:31+00:00 |
| saturation | results/indian_months_to_full.csv (safe limit: indian_summary.csv P5 at 1.00 pu) | 21 | 2026-11-20T05:58:18+00:00 |
| weather_coverage | irradiance_openmeteo_tgspdcl/*.csv, prev_day1_tgspdcl/*.csv | 13 | 2026-11-20T05:58:31+00:00 |
| weather_daily | same files, every day, daily totals | 37310 | 2026-11-20T05:58:31+00:00 |
| weather_hourly | same files, last 14 days, hourly | 4368 | 2026-11-20T05:58:31+00:00 |
| weather_point | data/irradiance_openmeteo_tgspdcl/README.md | 13 | 2026-11-20T05:58:33+00:00 |
REAL = measured or published data held in this project. Weather actuals are satellite/reanalysis-based sunlight, not metered solar output. Our model, or a published model reference (PVGIS), run on real inputs. Not measured. Assumed typical values (street lengths, cables, houses, 3 kWp roofs, load per home). The pilot replaces them with the discom's data.
| Dataset | Tag | Period | Rows | Coverage | Last date | Source file | Can show | Cannot show |
|---|---|---|---|---|---|---|---|---|
| TGSPDCL rooftop net-meter records, monthly | REAL | 2025-01 to 2026-10 | 146,665 | 22 monthly files; 20 circles and 473 section offices (circle + section name) in 2026-10; 59,550 services in 2026-10; July 2025 de-duplicated; Aug 2025 has no Gadwal circle (kept missing) | Oct 2026 | data/TGSPDCL/ (22 monthly files, Open Data Telangana) -> data/TGSPDCL_summary/net_meter_clean_22_months.csv | Real billed services, billed kW (load column) and units, by month, circle and section office. | Per-transformer counts, hourly values, what 'units' means (import, export or net). |
| Open-Meteo weather actuals, 13 TGSPDCL points (hourly) | REAL | 2019-01-01 to 2026-11-09 | 895,440 | 13 points, 68,880 hourly rows each; satellite/reanalysis-based archive | 9 Nov 2026 | data/irradiance_openmeteo_tgspdcl/*.csv (Open-Meteo, CC BY 4.0) | Real sunlight (GHI, direct, diffuse), temperature, cloud, wind by hour for the area. | Ground-station measurements, metered solar output, the newest days are provisional (back-filled). |
| Open-Meteo day-before forecasts, 13 TGSPDCL points (hourly) | REAL | 2024-01-19 to 2026-11-09 | 319,891 | 13 points, 24,607 non-empty hourly rows each of 25,056 (the empty hours are before the first real forecast) | 9 Nov 2026 | data/prev_day1_tgspdcl/*.csv (Open-Meteo previous-runs API, CC BY 4.0) | What the weather model forecast one day before, to compare with the actuals above. | Metered solar output; it is a weather forecast, not a solar-power forecast. |
| MNRE state-wise renewable capacity table | REAL | 31.01.2026 | 1 | one Telangana row: rooftop 695.9 MW, solar total 5,065.10 MW | 31.01.2026 | data/telangana_reports/mnre_2026_02.pdf | State-level rooftop and total solar capacity on one date. | A time series (the other dates in the SME note are not held), anything per district, circle or transformer. |
| PM Surya Ghar installations, Telangana (quoted) | REAL | 27 Jul 2026 | 1 | 54,668 installations, 70,144 households (quoted from Lok Sabha Q 1698) | 27 Jul 2026 | docs/Challenge16_DER_Final.md (source file not held) | A state-level count of subsidised rooftop installations. | MW per state, per circle or per transformer. The source file is not held, so the figure cannot be checked yet. |
| PVGIS 1 kWp hourly output (published model reference) | MODEL | not a dated file | n/a | source files not found on this machine | n/a | data/pvgis_1kwp/*.csv | A published model reference for 1 kWp output. | Measured output. |
| Open-Meteo live forecast (called when you ask, LIVE badge) | REAL | not a dated file | n/a | 13 points, next day and the day after; falls back to the stored example when the call fails and says so | n/a | Open-Meteo API (CC BY 4.0) via backend/app/integrations/open_meteo/ | A real weather forecast fed into our shipped model. | The power estimate is modelled: it is not scored against metered output. |
| TGSPDCL guideline and agreement texts, CSTEP feeder report, MNRE table (PDFs) | REAL | not a dated file | n/a | source folder not found on this machine | n/a | data/telangana_reports/*.pdf | The published rules: 50% DTR cap wording, 56 kWp self-certification threshold, the CSTEP feeder method. | The TGERC regulation text and the 2016 gazette are not held; the unit of the 50% cap is not stated in either text. |
- Real transformer data (kVA, consumers, phase, 15-minute DTR or feeder load). Not public; the DTR list has to come from the department. Every network value here is assumed.
- Metered rooftop solar output. Not public and not downloadable (PVGIS stops at 2023; no metered source was found). Forecast accuracy is versus a physics model, not versus measurement.
- Ground-measured sunlight (GHI). No ground station file is held. The Open-Meteo actuals are satellite/reanalysis-based; NIWE or IMD stations would be the check.
- Hourly or 15-minute net-meter / smart-meter data. Only monthly counts, billed kW and units are public. The meaning of the 'units' column (import, export or net) is not documented.
- TGNPDCL (northern discom) net-meter data. No TGNPDCL file was found; the net-meter data covers TGSPDCL only.
- REAL Telangana rooftop solar, MNRE state table: 695.9 MW (as on 31.01.2026; telangana_reports/mnre_2026_02.pdf)Table headed 'State-wise (Location based) installed capacity of Renewable Power as on 31.01.2026', Telangana row. State total, not per transformer.
- REAL Telangana solar power total, MNRE state table: 5,065.1 MW (as on 31.01.2026; telangana_reports/mnre_2026_02.pdf)Same table, 'Solar Power Total' column (ground mounted 4,360.49 + rooftop 695.9 + the third solar column 8.71).
- REAL PM Surya Ghar installations, Telangana: 54,668 installations (as on 27 Jul 2026; docs/Challenge16_DER_Final.md (quotes Lok Sabha Q 1698))Quoted from Lok Sabha Q 1698; the source file is not held, so it cannot be checked yet (the expert review found it correct). 70,144 households in the same answer; no MW per state found.
A traffic light per transformer: reason, room left, early warning. Network values are assumed examples until the discom supplies its own data. Screening only, never approved capacity.
Every claim in this app, how it was checked, and what it cannot tell you. Labels: checked - correct, checked - wrong, cannot be checked yet, because ...
Limit on the street results (E01 to E03 generic sensitivity; E55 Indian headline). P5 is a screening value and can be optimistic. Checks run on the generic 400 V model (not repeated on the Indian model): in 91 of 540 roof-by-roof trials "more roofs is never better" fails, and OpenDSS gives 69.0 vs 73.35 kWp (urban) and 16.35 vs 18.0 kWp (rural), both unbalanced. Treat it as a range, not a guarantee.
P5 as defined here (Passes screening tests). P5 = 5th percentile over 200 random trials (Indian street model, the headline; 30 random trials on the generic 400 V sensitivity) of the largest rooftop solar (3 kWp roofs, panel kWp) that passes the screening tests at the sunniest hour (voltage rise +6%, transformer 100 per cent of rating, cable 100 per cent, unbalance 3%); each trial redraws which homes adopt and (except the balanced and three-phase cases) each home's phase. Each trial value is the largest number of 3 kWp roofs that passes, so it is a multiple of 3 kWp; the 5th percentile interpolates linearly between the sorted values (numpy), so the printed P5 can fall between 3 kWp steps (for example 34.35). P50 is the median of the same trials. Screening value, not a guarantee.
Label notes: where an item says SME label, the label is the SME's as given: [checked-correct] = official text read by the SME, [secondary] = news or think-tank, [cannot be checked] = no source. The project could not re-read any gazette or live site, so none of these is upgraded to checked here.