Methodology

Every number on the map is either computed from archived public IEMOP market files or a labeled constant with a primary source. This page lists the sources, the access mechanics, the units, and the caveats. Nothing here is an estimate of ours dressed as a fact.

Datasets archived from IEMOP

IEMOP (the WESM market operator) publishes these on public pages under iemop.ph/market-data/ with no login. The public window is rolling (about 90 days per dataset), so this repository archives the files daily; the git history becomes the durable public archive. Access mechanic: each page lists files via wp-admin/admin-ajax.php (action display_filtered_market_data_files) and serves them via a ?md_file=<base64 path> parameter.

Courtesy: fetches are sequential with a 250 ms sleep and abort after 5 consecutive errors. IEMOP data is republished here as-is for public research with attribution; if IEMOP objects to any part of this mirror we will take it down.

Named choke points (primary sources)

Corridor lines on the map follow the real routed geometry of the lines between the named endpoints above, taken from the OpenStreetMap grid layer described in the next section. A corridor that cannot be resolved onto that geometry would fall back to a straight schematic link and say so on hover; on the current bake all five corridors resolve.

The grid itself: real geometry from OpenStreetMap

NGCP publishes no public GIS. Its Transmission Development Plan maps are raster PDFs, and the network model behind WESM (impedances, ratings) is distributed to members only. The only public source of real routed line geometry is OpenStreetMap, so the map draws it and labels it what it is: community-mapped data under ODbL, not an NGCP document.

• Pull. pipeline/fetch_grid_geometry.py pulls three raw Overpass extracts into data/raw/OSMGRID/ (committed, so the bake reproduces without hitting Overpass): power lines tagged 500/230/138 kV inside the PH boundary, 350 kV HVDC lines plus all power cables, and every mapped substation. 138 kV is in scope because the Mindanao grid and the island interconnections run on it. The pull is a manual/monthly refresh, not a nightly cron; the grid changes slowly.
• Bake. pipeline/grid_geometry.py filters to transmission-level features and writes grid_lines.geojson and grid_nodes.geojson. A substation with no voltage tag still counts as transmission when the HV network physically lands on it (Tabango, the Leyte end of the Cebu-Leyte crossing, is mapped with no tags at all).
• Binding equipment pinned on. The RTDCV/DAPCV league names equipment in station code form (5DAAN_4TAB2, 1EHVNGS_TR2). Codes parse into station tokens, tokens resolve against normalized OSM substation names (plus a short verified alias table), and line constraints must find a plausible shortest path between their two stations in the snapped network graph; ambiguous tokens are settled by voltage level (transformers) or path geography (lines), never by guessing. Whatever fails stays in the unmatched list in grid.json, shown as unmatched. Matched segments draw red with their receipts on hover.
• Caveats. Geometry only, no ratings; OSM completeness varies (several named binding stations, including the Bauang Power Plant switchyard, are not mapped at all); coverage share is reported in grid.json rather than implied.

Nodal prices: what the market publishes, and the model built on the geometry

DIPCEF (nodal LMP results, final) prices about 1,200 resource nodes per 5-minute interval and decomposes each LMP into SMP + loss + congestion. The raw hourly zips are too heavy to commit and IEMOP's window rolls, so pipeline/nodal_prices.py derives one compact JSON per day into data/derived/nodal_daily/ (the nightly cron tops it up): the regional SMP series, each node's hourly mean deviation from its regional SMP, each node's hourly scheduled MW, and the day's pricing-flag tally (OK / PSM / SEC).

The derived record is surfaced in the product: the map's Prices mode draws each node's persistent deviation from its regional price (clean market days only), each dot sitting on the faint transmission network it belongs to, so the price field reads as the grid coloured node by node rather than dots on a blank basemap. The studio's Analysis section carries a Nodal prices view with the full per-node table. Both label the statistic as an observed locational deviation, not a congestion premium, for the reason stated next.

Placement resolves through a priority of public evidence, every dot tagged with its source: a named OSM substation on the HV network, a named OSM power-plant site, the repo's named-generator pin, a locality centroid for load and delivery codes that name their locality, and a DOE plant-list row whose location column names the municipality (the last two placed at the OSM place centroid, city-precision). Region must agree in every path and ambiguity is a miss, never a guess. The resolved share of scheduled MW per grid ships inside nodal_obs.json as a scoreboard, so the resolution number is public and its drift visible.

What the derived record shows, stated plainly: the published LMP_CONGESTION column is zero through the market suspension window and small and intermittent after real-time pricing resumed on 2026-05-01, nonzero on 28 of the 70 sampled days (median 0.56 PhP/kWh where it fires, up to 19 PhP/kWh on 2026-05-26) and nonzero on 1.18 percent of clean-day node-hours. The SMP is region-constant per interval, so most within-region locational separation rides the loss column, inter-island congestion appears as the regional SMPs splitting, and the small nodal congestion that is priced is sparse rather than a persistent per-node charge. Administered intervals also carry a bimodal SMP (a base value and the same value about 2.2% higher); the regional reference series takes the per-interval mode and reports how many intervals were multimodal.

pipeline/nodal_dcopf.py is the honest tier of nodal modeling this data supports: a reduced backbone (class-typical reactances scaled by real routed length, class-default ratings, both labeled estimates) solved as B-theta linear programs on HiGHS. A replay mode places the day's observed per-node scheduled MW on the network and asks where the flows load it; the defensible validation is whether the equipment RTDCV actually recorded at its limit ranks among the model's most-loaded branches. An OPF mode re-dispatches within each unit's observed-day capability at grid-fuel proxy costs and reads nodal prices off the bus-balance duals. Resolution of DIPCEF resource codes onto buses is partial and reported (share of MW resolved); nothing from this model is presented as an observed number. Branch ratings carry provenance tags: the operator's own observed operating limits (RTDCV/DAPCV BINDING_LIMIT) where the equipment matched, replay-calibrated floors where the observed dispatch visibly carried more than the class default, and labeled class defaults for the rest. The public NGCP TDP attachment is the summary report and carries only grid-level facility aggregates, so per-line TDP ratings cannot be keyed from it; that is the exact boundary, not a hand-wave. The OPF's own measured finding, recorded in the artifact: at the current resolution the re-dispatch concentrates each grid's unresolved generation onto its few resolved plant buses, so modeled price LEVELS are not usable; the probe reports binding geography and the honest gap, and the zonal engine remains the price model.

The loss-surface validation. Because the market's within-region nodal structure is loss-dominated (the congestion column is small and sparse, 1.18 percent of clean-day node-hours), network physics can be tested against it at scale: replay the observed injections on the reduced backbone, take marginal loss factors per node (class-typical resistances scaled by real routed length, labeled estimates), and compare the modeled deviation against each node's observed deviation, grid by grid. The comparison is Spearman rank correlation across nodes plus the error after a per-grid affine fit (the loss-reference convention is an affine choice, so slope and intercept are fitted and reported). The result recomputes nightly as clean days accumulate and ships in loss_surface.json with a self-computing verdict: grids where physics ranks the market's deviations are listed as validated, and grids that fail the test are listed as failing, not hidden. Both surfaces draw it: a three-panel scatter figure in the README (scripts/loss_surface_fig.py) and the studio's Analysis, Loss validation view, which renders the same per-node points, fitted lines, and per-grid verdict from the baked artifact.

Two honest limits on that verdict. The comparison runs on the 15 clean market days (about 21 percent of the 70 sampled) that clear a 90-percent OK-flag filter, so administered PSM and SEC days are excluded; whether the excluded days line up with the choke-point episodes the rest of the site tracks is not yet checked. And because DIPCEF resources that share a substation resolve to the same model bus, the 314, 96, and 118 node counts collapse to 72, 25, and 37 independent buses; the reported spearman_ci95 uses those bus counts, and the 0.4 validated cutoff is a moderate-correlation reporting convention, not a power calculation.

Data-center sites

Only facilities with a citable public source are pinned; each pin carries its source URL and a mw value only where a public figure exists (hollow pins have none). Pins are placed at the named city or municipality, never at a street address. This is NOT a complete inventory: Cushman & Wakefield counts 24 operational facilities (73 MW) with 22 MW under development and 89 MW in planning (APAC DC Update); DataCenterMap lists 44 Philippine facilities.

Demand-side anchors (all labeled)

• DICT: Philippine data-center capacity could reach 1.5 GW by 2028 (BusinessWorld, Oct 2025). A forecast, not a measurement.
• DOE: 300 to 1,500 MW of additional peak demand once incoming data centers are factored in; DOE has no official published projection of data-center capacity requirements (PCIJ, Jan 2026).
• Meralco: set to deliver 1,000 MW for 10 data centers (PCIJ, Jan 2026).
• Data Center Philippines alliance: 473 MW pipeline (thePhilBizNews, Feb 2026).
• Operational capacity today is contested: about 200 MW (DICT) vs about 630 MW (Mordor Intelligence), per PCIJ. Shown only as a labeled range.

Market and bill anchors

• WESM was suspended 26 March to 1 May 2026 under the fuel-shock emergency (EO 110); market-driven pricing resumed 1 May (Tribune).
• May 2026, meaning the WESM billing period of 26 April to 25 May (IEMOP reports on billing periods, not calendar months; all four figures below reproduce on that period from our own archive to within 0.1% and none reproduce on calendar May). Note that its first five days, 26 to 30 April, were still under administered pricing, so this is not a fully post-resumption month: dropping them lifts the system figure about 5%. System average P7.79/kWh, up 38.5% from April; Luzon P7.02, Visayas P10.20, Mindanao P9.28. The same report gives a system supply margin of 3,629 MW, down from 4,427 MW in April, a distinct margin series IEMOP reports alongside monthly-average supply (21,374 MW) and demand (15,755 MW); those averages are a different basis and do not subtract to the margin, so this page uses only the 3,629 MW supply-margin figure for the share arithmetic (IEMOP May 2026 report via Power Philippines; BusinessWorld).
• Meralco June 2026: overall rate P14.4833/kWh (+P0.1488); generation charge P9.0704/kWh, of which the spot market supplied 10% of the energy at a WESM price of P7.0281/kWh, contributing about P0.70/kWh to that charge (the price is not itself a slice of the bill; the blended charge averages across all sources) (BusinessWorld; Meralco advisory).
• Sual coal plant: two 647 MW units at Sual, Pangasinan (Wikipedia). Among the largest units on the Luzon grid and the worked example here because its trips recur, but NOT the largest: GNPower Dinginin in Mariveles, Bataan runs two 668 MW supercritical units (NS Energy). The market corroborates it without being asked: the WESM Luzon contingency reserve requirement (commodity Fr in the RTD regional summaries) sits at exactly 668 MW, because contingency reserve is sized to the largest single unit it must cover.

Units and conversions

• LWAP and LMP components arrive in PhP/MWh and are divided by 1,000 for display in PhP/kWh.
• LOAD_CURTAILED arrives as MW per 5-minute interval; summed to MWh by multiplying by 5/60.
• Daily LWAP is the unweighted mean of that day's 5-minute values (a display series, not a settlement quantity).

The chronological engine and the studio

The studio's engines solve the dispatch as linear programs through HiGHS (the July 2026 solver pass): the snapshot views as a single-hour LP, Chronology and the Backcast as one LP over the 24 coupled hours. Prices are the balance-row duals, real locational marginal prices: an importing grid can price at its neighbour's marginal block plus the small wheeling cost, an exporter at the importer's minus it, and the price-setter label says so (import, export, storage, shortage, or the named fuel). Unserved load prices at the WESM offer cap of P32/kWh (Tripartite Committee Joint Resolution No. 2 s.2013, permanent since December 2015): the market's own ceiling, a published rule rather than a fitted value, applied identically in every engine, the map's browser clear included. One stated exception: if a what-if edit pushes a fuel cost above the cap, the shedding penalty rises just above that cost instead (shedding must stay strictly dearer than serving), and the hour still labels 'shortage'. The secondary price cap (P7.423/kWh when the 72-hour rolling GWAP breaches P12.413/kWh, ERC Resolution 26 s.2025) is stated here as the other published cap, and its trigger arithmetic is now computed rather than cited: the 72-hour rolling mean of the published 5-minute GWAP series (generator-weighted-average-price-final; archived here since July 2026 after a round-7 audit found the earlier "not published" claim false) runs per region in market_ops.json, and for the combined Luzon-Visayas region the operator also publishes. The computation is a finding in both directions. Read the offer-cap-held series, not the raw one: the raw file carries intervals priced up to P165.05/kWh, five times the market's own P32/kWh offer cap, which are violation and scarcity coefficients rather than clears (the operator's own price-substitution record caps at exactly the offer cap, with no exceptions). Held at the cap, LUZON breaches zero windows and its peak falls below the threshold, so its entire breach count came from those intervals. That correction does not erase the finding elsewhere: the System row still breaches held at the cap, as does the combined Luzon-Visayas row, and Visayas and Mindanao run hot either way. Two things that were listed here as unknowns are in fact published: the rule monitors the EX-ANTE rolling GWAP while GWAPF is the final series, and under ERC Res. 26 s.2025 the regional or island cap applies only while a grid interconnection is on outage, which makes the System row the default trigger and the Visayas and Mindanao rows conditional. The observed price record shows no day pinned at either the current P7.423 or the prior P6.245 cap level, so the gap between the computed trigger and the operational one narrows but does not close; what remains unreproduced is the weighting and the imposition and lifting mechanics. The widest-swing as-bid scenario day lands just UNDER the stated threshold (P12.23 against P12.413); an earlier version of this page said it crossed, which was true of the day the old hourly binning selected. Impositions announce themselves in the archived advisory stream when they happen. Tiny deterministic epsilons (at most fractions of a centavo, largest on the last enumerated block) make the optimum unique on the model's flat cost plateaus so both solver builds land on the same answer. The inputs and the checks:

• Observed day profiles (profiles.json): per-grid hourly demand is NATIVE LOAD: dispatched generation plus net market imports plus recorded curtailment, all from the same RTD regional summary rows. Generation alone would self-balance every grid by construction and erase the observed inter-island flows (the Visayas net-imports roughly a quarter of its own generation across the window; the flow identity between the three grids' import and export columns closes exactly), so a generation-based replay would never need the corridors this product is about. The observed hourly price is the mean of that hour's 5-minute LWAP, and, where the archive carries it, the hourly regional clearing price (MCP) rides alongside as the second target; the observed corridor flows ride as the third. Days without full 24-hour demand coverage on all three grids are dropped, not filled.
• Daily scheduled outages: each replayed day subtracts the DEVIATION of the operator's matched scheduled-out MW from the MARKET-window mean, per grid and fuel (the PASA layer's OUTRTD mapping). Deviation, not the raw MW: the static availability derates already carry the average outage state, so subtracting the raw schedule would double-count it; and the baseline is the market days the backcast replays, so the adjustment washes out over the scored window. Hydro is excluded here because its daily variation is already the observed water budget; unmatched codes carry no MW, so the adjustment is a floor. Snapshot views keep the static fleet (one reference hour has no day to deviate from).
• Observed corridor availability: the Leyte-Luzon HVDC's per-hour limit scales by the fraction of the hour the link was actually unblocked, INFERRED from the operator's own advisory stream (every blocking and de-blocking is announced with a timestamp; the archive carries 95 paired block windows in the window, median 10 minutes, touching about 7 percent of replay hours). The dedicated HVDC-limits dataset (HVDCRTD) is header-only in every archived file, but the per-interval HVDC schedule (RTDHS: flow, congestion flag, overload MW per corridor per 5 minutes; archived here since July 2026) is the operator's own corridor record, and both backcast modes now score their corridor flows against it directly (profiles.json flows_rtdhs), independent of the demand-identity construction; the two observed records themselves agree to within about half a MW on hourly means (market_ops.json flow_record), and the operator's per-interval CONGESTION_FLAG supplies a binding-share target (the corridor bound in 50 percent of VISLUZ1 and 42 percent of MINVIS1 intervals across the archived window) that the model's at-cap share is scored against. The advisory-window inference still sets the replay caps; it is labeled here and the levers compose on the base limit. MVIP has no observed block events in the window and stays at its nameplate cap. We measured the alternative of feeding the operator's own RTDHS binding-schedule caps into the LP for both corridors (pipeline/corridor_cap_probe.py): it makes the modeled at-cap share track the observed binding share by construction, but on the independent judge, the price backcast, it lowers Luzon price MAE only slightly while worsening price correlation on every grid, badly on Visayas, because the Leyte de-rate to near-zero on flagged hours decouples Visayas from its observed import pattern. A model that tracks the observed price shape worse is not an improvement, so the caps stay advisory-based and the corridor under-binding versus RTDHS stays a reported boundary, not a constructed match (the measured deltas live in market_ops.json corridor_cap_probe, regenerated from the current archive by pipeline/corridor_cap_probe.py).
• Solar shape: the 24-hour availability profile is the model's labeled clear-sky-ish assumption (fleet_ph.SOLAR_PROFILE), not measured irradiance. Other fuels hold their derated availability across the day. We measured the alternative of replaying each day's OWN observed solar energy (the DIPCEF dailies, market_ops.json solar_wind_observed) by scaling that shape to the observed daily total (pipeline/vre_probe.py): it collapses the Luzon price correlation while barely moving MAE, because Luzon's observed solar runs well above the clear-sky credit and the flat shape dumps that extra energy into midday, crushing modeled daytime prices below what actually cleared. The missing piece is a per-resource hourly solar series, which the daily energy cannot supply, so the backcast keeps the clear-sky credit and the per-day observed solar stays a reported observation, not a replay input (the measured deltas live in market_ops.json vre_probe, regenerated from the current archive by pipeline/vre_probe.py).
• Unit commitment: a full production-cost model commits each thermal unit as an integer decision, so a committed unit holds a minimum-stable floor instead of idling to zero. We built that as a mixed-integer program on the same solver (HiGHS solves MIP on both engine builds): binary commitment plus a generic minimum-stable level on the thermal blocks (coal, gas, geothermal, oil, biomass, at NREL ATB / typical fractions, labeled generic), priced by solving the MIP, fixing the commitments, and reading the balance duals (pipeline/uc_probe.py). Measured on the same backcast, commitment lowers the price correlation everywhere, because a minimum-stable floor applied at the aggregate fuel-block level forces must-run chunks far larger than any single unit and collapses off-peak prices in the wrong hours. The honest fix is per-PH-unit heat rates and minimum-stable levels, which no public dataset carries (RTDSL is archived and gives per-resource MIN/MAX operating limits, but its resources are coded and most floors are VRE self-schedule pins, so it does not de-fabricate thermal min-stable without a unit registry). So the LP block model stays the default engine and commitment is the reported finding, not a swap (the measured deltas and verdict live in market_ops.json uc_probe, regenerated by pipeline/uc_probe.py).
• Price-model levers: the same measure-first gate ran on the levers that could lift the cost-mode price backcast (pipeline/price_model_probes.py, measured 2026-07-19). Reserve withholding at the day's scheduled MW moved no metric on any grid (the flat committed-coal tranche absorbs it without changing the marginal block), so it stays off and the number is the finding. The water budget's opportunity-cost channel priced hydro-marginal hours in under five percent of the window, so hydro alone cannot rescue the Mindanao shape. A STYLIZED offer book, the leave-one-out median of the operator's own offer curves per grid and hour of day (estimated from bids, never from prices; a day is never priced by a curve that saw its own book), closes 88 percent of the Luzon correlation gap between the cost stack and the same-day replay (pooled correlation 0.38 to 0.69, median within-day correlation 0.19 to 0.83, evening-peak MAE P7.89 to P5.58) and on the Visayas slightly beats the same-day book, so it earns the third-engine slot: cost floor, typical bidding, actual day. Two levers are named rather than dropped: a monthly fuel-price index has no sourced series (the observed books already carry each day's fuel-cost level, so the lever rides the stylized book), and an observed unplanned-unavailability layer is blocked on a per-day unavailable-MW parser for the NSO stream (the dated 935 MW July 1 case remains its one-day proof). All tables live in data/derived/price_model_probes.json.
• Per-grid hydro capacity: recalibrated July 2026 from the DOE plant lists after the observed schedules contradicted the old allocation (the model gave the Visayas 10 MW of hydro; DIPCEF showed its plants clearing up to 377 MWh a day). The split now follows the fleet's installed shares (Luzon 66.7 / Visayas 1.4 / Mindanao 31.8 percent), prorated onto the sourced 3,836 MW national hydro total. Calibrated against capacity and observed dispatch, never against prices. Mindanao's wettest observed days still exceed the derated modeled capacity by about 3 percent; the bake reports that residual instead of hiding it.
• Storage: the day LP optimises the sourced storage fleet (DOE 634 MW BESS; Kalayaan 685 MW pumped storage) across the hours: it cycles only when the price spread beats the round-trip loss and idles on a flat day, which the retired heuristic never admitted. Energy durations are stated assumptions because the sources publish MW, not MWh. The snapshot views keep storage out of the energy stack, as before.
• Parity: both engines build the SAME linear program as the same text, byte for byte (coefficients serialized from integer micro-units), and the golden fixtures pin the text's sha256 plus the solved outputs to P0.02/kWh and 1 MW: the Python reference is highspy, the browser runs the HiGHS wasm build, and a drift in either the model construction or the solve fails the suite. The retired coordinate-descent clear stays in the pipeline tests as a cost cross-oracle (the LP may never cost more).
• Backcast: every full-coverage market day is replayed with the base model against the observed hourly LWAP and the error is stated per grid (MAE, bias, correlation). The high-hour hit rate reports n/a when the flat cost model cannot rank hours, instead of a fake 100%. Nothing is tuned to fit; the evening residual is the scarcity and offer premium the cost model cannot see, and it stays the finding.
• Per-plant fleet (fleet.json): the DOE List of Existing Power Plants (grid-connected; Luzon and Mindanao as of April 2025, Visayas March 2025), parsed per unit from the DOE's own PDFs (Internet Archive captures, since doe.gov.ph refuses non-PH requests). Every fuel section must reconcile to the PDF's own subtotal before the artifact is written. Dependable capacity is the DOE figure, not a model derate.
• Hydro water budgets: on days covered by the derived DIPCEF dailies, hydro is energy-limited to the day's observed water: the LP may not dispatch more hydro MWh than the operator's own per-resource schedules show for that day, scaled proportionally with hydro capacity edits and the hydrology lever so what-ifs stay coherent. Classification is grid-connected WESM hydro matched to the DOE fleet (pumped storage and batteries excluded; anything that hints hydro but stays unclassified is listed and NOT counted, so the budget is a verified floor). Embedded hydro never appears in the nodal schedules, which matches the model's grid-connected scope.
• Gas fuel budget (Malampaya): the Leyte-Luzon gas fleet burns Malampaya gas, which the DOE rates at 429 MMscfd at full field (doe.gov.ph/natgas/malampaya-gas-field) and expects to deplete commercially around 2027 (BusinessWorld). The Malampaya-supply lever models that cliff as a day-level gas ENERGY budget, built into the LP the same way as the hydro water budget (lp_model.build_day_lp and lpText.buildDayLp, a gas_ constraint pinned by a golden case in both engines): the sum of natural-gas dispatch across the day may not exceed the budget, set as a percent of the gas fleet's flat-out day. Off at 100 percent, a what-if below it, so a user can price the supply cliff without inventing a figure the DOE has not stated.
• Reserve: an optional toggle withholds the mean scheduled requirement per grid (RTD reserve rows) from the reserve-capable stack (coal, gas, oil, geothermal, hydro, biomass, storage discharge), so tight hours price the withheld capacity. Closer to the co-optimised market than the old demand-inflation approximation, but still not a reserve PRICE: the reserve products stay a market layer this model does not clear.

The planning layers (July 2026 pass)

Five layers added in the market-operations pass. Each is either computed from the archive or a sourced list; none runs an optimizer.

• Load sweep: the snapshot solve repeated over steps of added flat 24/7 load on one grid (the data-center shape), on top of the ran scenario. Each step is the same coupled clear the Run button executes; the view marks the step where the importing corridor saturates, where the marginal fuel flips, and where unserved load begins. Reference-hour arithmetic, not a forecast; a single hour carries no daily water budget, so the sweep prices hydro at capacity (Chronology is where the budget binds).
• Window band: the ran scenario replayed over every full-coverage market day in the archive, reported as per-hour price percentiles (10th to 90th) and the distribution of daily means. The sample is the archive's own observed days; nothing synthetic is drawn.
• What set the price (Chronology): each hour is classified as unserved load, a saturated corridor on the grid's importing side, or the marginal fuel block. This is a reading of the solve's own outputs, named per hour.
• LT Plan (projects.json): the DOE's committed and indicative private-sector power project lists ("As of 31 December 2025", Internet Archive captures of the DOE's own PDFs), parsed with the same reconciliation gate as the fleet: every fuel section must sum to its printed subtotal, and every grid to the DOE's LVM summary (committed 13,839.48 MW; indicative 119,226.39 MW). Rows carry grid, fuel, MW, and the proponent's target date, and deliberately no project name (names wrap unpredictably in the PDF layout; a wrong name on a right number is still a defect). ESS is tracked separately, as the DOE's own summaries do. Transmission candidates come from the NGCP TDP 2025-2050 (plus the September 2025 revision for MVIP Stage 2): only figures the TDP itself states as transfer capacity carry MW (Luzon-Visayas HVDC bipolar +440 MW, MVIP Stage 2 +450 MW); conductor thermal ratings are not transfer limits and are not imputed. Apply writes ordinary scenario edits.
• PASA (pasa.json): the operator's own outage schedules used in real-time dispatch (OUTRTD), one row per resource out per day. The files carry resource codes and no MW, so codes map to DOE-fleet plants through a hand-maintained alias table verified against fleet.json at bake time; unmatched codes are listed with no MW, making the matched MW a floor. Grid per code comes from the WESM numeric area prefix (01-03 Luzon, 04-08 Visayas, 09-14 Mindanao), an inferred mapping spot-checked against named plants. The view re-runs the live reliability Monte Carlo with the day's matched MW off the stack first; the out MW stops drawing its forced-outage rate while the same plant's in-service units keep theirs.
• Emissions (emissions.json): operational CO2 factors per technology. Coal 0.874 and gas 0.337 tCO2/MWh are the IPCC 2006 fuel-combustion defaults converted at the EMB's published Philippine average heat efficiencies (39% coal, 60% CCGT); oil 0.533 is the EMB's own diesel figure; geothermal carries the midpoint of the IPCC 6-79 gCO2/kWh range; hydro, solar, wind, and storage discharge are zero operational (their lifecycle emissions are real and out of scope); biomass is excluded rather than assigned a contested factor. The DOE National Grid Emission Factor (0.7181 Luzon-Visayas, 0.8173 Mindanao, 2019-2021) is baked as a cross-check anchor, never an input.
• Bill mix history: Meralco's published supply mix moved WESM 6%, 7%, 10% across April, May, June 2026 (PSA 74/73/69, First Gas and IPP 20/20/21) while the generation charge climbed P8.3864 to P8.7942 to P9.0704/kWh; each month cites the Meralco advisory and an independent news report. Only the June per-source movement is public in clean form and only that is shown.
• Run report: a saved run exports as one self-contained HTML file (scenario edits, daily summaries, the per-hour binding tally, CO2, and this page's provenance statement), so a shared result carries its own caveats.

Energy-limited hydro, first scoped out for want of a citable monthly figure, later shipped from a better source: the archive's own DIPCEF per-resource schedules give OBSERVED daily hydro energy per grid, day resolution instead of monthly, no interpolation (see the hydro water budgets bullet above). Still not carried, and why: security-constrained unit commitment and minimum up/down times would require per-unit commitment data the Philippine sources do not publish, so modeling them would be fabricated fidelity. RAMP RATES are a different case and this page used to get it wrong: the operator DOES publish them. Every RTDOE offer row carries a piecewise ramp-rate curve (RR_BREAK_QUANTITY1-5 with RR_UP1-5 and RR_DOWN1-5, in MW per minute by MW band), populated on essentially every resource in the interval, in the same hourly file this project already downloads for the offer books. So the question is not whether ramp limits could be modelled but whether they would bind, and that is now measured rather than asserted (market_ops.json ramp_probe, pipeline/ramp_probe.py). They would not, at this engine's resolution. Aggregate each unit's one-hour ramp capability, capped at its own offered range because a machine cannot exceed it, and compare against the largest hour-to-hour demand RISE anywhere in the archived observed profiles (1,223 MW on Luzon, 308 on the Visayas, 307 on Mindanao). The margin depends on how the fleet is counted, so the figure reported is the conservative one: resources that were actually GENERATING that hour, not everything that offered; each unit's SLOWEST published ramp band, not its fastest; and the WORST of six sampled hours spanning weekday and weekend, morning pickup and evening peak. On that strict basis the fleet still out-ramps the worst demand rise by 3.5 times on Luzon, 2.4 on the Visayas and 4.3 on Mindanao. (Counting every offered resource at its fastest band lifts those to 6.1, 3.4 and 6.0, which is why the strict read is the one quoted.) A per-fuel hourly ramp constraint is therefore inert by construction, and building one would be fidelity theater: engine complexity for no accuracy. It is measured out, not skipped. Two honest qualifications travel with that: the result is a grid and fuel-aggregate one at HOURLY resolution, which is this engine's resolution, and per RESOURCE an hourly ramp limit could still bind for about a third of the book (the tightest are oil and coal machines moving under a tenth of their range per hour, so a per-UNIT model would need them); and the 5-minute case is a different question this measurement does not answer.

Boundaries and queued builds, each with its named input

• Reserve co-optimisation: the replay is scored; the joint LP keeps its named input: WESM co-optimises energy and reserves, and this page long claimed the per-participant reserve offers were unpublished. A round-7 audit of the full IEMOP market-data surface found that claim false: RTDOR publishes every resource's reserve offer curve per commodity (Fr, Dr, Ru, Rd) in the same breakpoint schema as the generation offers. The archive derives daily reserve books from it (data/derived/reserve_daily/, gated so each grid-hour-commodity book covers the scheduled reserve or the day is refused), and the reserve replay now scores them: each book cleared at the operator's scheduled MW for the book's exact interval, against the official RSVPR price at that same interval, 96 days by 12 grid-commodity pools (market_ops.json reserve_validation). Every pool's mean residual is negative, and the hours where the marginal offer sits above the official price are noise-level (8.5 percent of scored hours, by at most P0.033/kWh); that one-signed pool residual is the co-optimisation opportunity-cost wedge itself, measured per pool, biggest on regulation products and the tight islands and near zero on dispatchable reserve (Luzon Dr matches the official price within half a centavo in 80.0 percent of hours). The post-convergence build queue tightened the target: the operator's own final per-resource cleared reserve (DIPC reserve results final) is now archived, and the RTDOR book replay scored against it under-prices on every pool as it did against the RTD price, the same one-signed wedge confirmed against the authoritative final clearing (market_ops.json reserve_results). The per-resource joint energy+reserve LP that would close the wedge was then PROTOTYPED from the raw hourly per-resource offer curves (RTDOE energy, RTDOR reserve) on a sample of grid-hours, and the prototype located exactly why it cannot reproduce the official RSVPR (market_ops.json joint_lp_probe). The reserve requirement CLEARS SHORT on the scarce hours: the operator's own RTDSUM shows the scheduled reserve at 11 to 36 percent of the market requirement on those hours. Where the requirement is met, the marginal Dr offer cleared to the scheduled quantity reproduces RSVPR to within about a centavo, so the offer stack already explains those hours (the pooled replay scores them). Where it clears short, the official price sits above the ENTIRE offer stack, an administered reserve-scarcity value (up to about P25/kWh over the marginal offer on the sample) that the public offers cannot explain, and a joint LP that forces the full requirement goes infeasible on exactly those hours, as the real market could not meet it either. So the wedge is administered reserve scarcity on the requirement-short hours, not a co-optimisation internal recoverable from offers; fixing the energy dispatch to the observed schedule (DIPCEF) and registered capacity (CAPEG) tracks RSVPR only where the requirement is met and cannot recover the scarcity uplift where it is not. The wedge therefore stays measured, not closed, and the dispatch model keeps its stated approximation: withhold the requirement, show the official price beside it.
• Sub-zonal topology (a Leyte-Cebu split): WESM expresses inter-island congestion as one price per island grid, so a split region has no published price series to validate against; only sample-day DIPCEF nodal LMPs could check it, which makes the layer diagnostic, not scorable.
• The secondary-cap trigger is computed; its operational imposition is the remaining boundary: the 72-hour rolling mean of the published per-region GWAP (GWAPF) now runs in market_ops.json. Held at the market's own P32/kWh offer cap, which excludes the violation-priced intervals, it still crosses the P12.413/kWh threshold on the System, Luzon-Visayas, Visayas and Mindanao series, though no longer on Luzon, while the observed price record shows no day pinned at either the current or the prior cap level (the clamp scan is published beside the breach counts). The named missing input is the resolution's operational text: ERC Resolution 26 s.2025 as carried in the trade press gives the two numbers, not the algorithm (series designation, weighting, imposition and lifting mechanics, effectivity), so the computed series is the trigger's input, not a reconstruction of the cap itself.
• Must-run instructions are displayed, not dispatched, and the inertness finding is scoped to the must-run subset: the processed must-run list peaks at a median of 9 MW across its 176 instructed grid-hours in 14 weekly files (the per-interval instruction median is 5.7 MW), two orders of magnitude below the offer books it would constrain, so a must-take layer built from it would be inert and is not built. The round-9 audit showed that finding does not extend to the family: the sibling MOT-raise re-dispatch record carries 104 thousand instructions across the archived weeks at a pooled 55 MW median and a 668 MW maximum, larger than one Sual unit, so the full out-of-merit record is not inert. It is archived and measured (market_ops.json so_instructions), the daily instruction log names the binding element in the operator's own remarks, and the engine layer the record sizes, an administered-dispatch overlay on the replay, was then MEASURED against the model and the measurement decided against building it (market_ops.json admin_dispatch). The raises are material in MW (7 to 11 percent of dispatched generation on hours with a raise, highest on Mindanao) but price-inert in the model's per-fuel block engines: the raised MW is mostly coal (about two thirds) with roughly a fifth on gas, and coal is already the modeled marginal fuel on about 93 percent of the raise hours. Every raise hour already prices at or just above coal's administered floor (the modeled price tops out at P6.04/kWh across all of them), so a per-fuel administered floor (the form a raise takes once the named units are aggregated to their fuel) cannot lift a price coal already sets; the only move it could make is downward, coal displacing water, which does not add the administered premium the overlay was meant to model. This is the minimum-stable-coal-floor outcome below, one layer up. The out-of-merit effect is a per-resource fact (which named unit runs for security), which the aggregated LP drops; the engine layer it truly sizes is the per-resource joint energy+reserve clear, not a per-fuel overlay.
• Minimum-stable coal floors were built, measured, and removed: with the offer book fuel-classified (84.6 percent of offered MW through the alias table, hand-verified abbreviations, and a fuel-tolerant fleet match), carving the sourced 40 percent technical minimum of online coal down to the price floor left the books byte-identical, because the coal fleet ALREADY offers its committed tranche at the floor. Commitment behavior is in the bids; a floor layer would model what the market already models.
• The negative-price hours are a resolution limit, named and sized: the negative clearing prices are an ex-ante dispatch outcome, solar flooding, not a settlement artifact. The 5-minute market crosses zero intermittently; an hourly replay on hour-mean demand cannot, it averages the intra-hour dip away. A sub-hourly replay was then prototyped and measured (market_ops.json subhourly_probe), and the picture is sharper than a clean resolution refinement: the crossing is a KNIFE-EDGE. On the crossing days the offer book's floor-priced supply comes within 0.66 to 2.53 percent of native load at the midday solar peak, and the observed regional load-weighted price does cross negative on four of the six closest days (down to about -P10/kWh). But it splits into two boundaries, neither of which a per-grid 5-minute replay closes. The near-zero band is a true knife-edge: the offer book clears at exactly zero over a wide MW range, and its own 48-block compaction moves about 4 percent of the floor-priced supply across the zero line, wider than the 0.66 percent tightest margin, so whether a shallow interval reads -0.001 or +0.001 turns on supply accounting finer than the offers give. The DEEP negatives are not an aggregate crossing at all: at the day's most negative interval the observed price is about -P10/kWh while physical native load (9,870 MW) sits ABOVE the entire aggregate floor-priced supply (9,834 MW), so the per-grid clear must price positive there. That -10 is a load-weighted average of nodal prices pulled to the floor by curtailed-solar pockets, which the per-grid model cannot produce at any time resolution. A clean 5-minute clear on the model's own demand basis reproduces none of the day's 50 observed negative intervals. Sub-hourly granularity is necessary but not sufficient; the sign flips stay a named boundary, a near-zero resolution knife-edge over a curtailed-solar nodal floor, not a clean missing input.
• The rest of the public surface is triaged and named, not ignored: the IEMOP market-data sitemap carries 58 dataset pages; this archive keeps 22 of them as raw daily CSVs and fetches several more transiently to commit as compact derived books (the RTDOE and RTDOR offer books, the DIPCRF reserve-results finals, and now the Regional Merit Order Table). The exact split between raw and derived is not a number this page should assert precisely, because it is exactly the kind of hand-count that drifts. Of the remainder, the projection series (day-ahead, hour-ahead and week-ahead price and schedule projections) project rather than record and are out of scope for a replay scored on the real-time record; the Original/Raw and interim variants duplicate final series already archived. One page slipped both nets until a round-10 audit: the Regional Merit Order Table (the rtd_smerit files), a real-time dispatch record, neither a projection nor a duplicate, which this triage had counted in no class at all. It is now consumed. The lesson is that a hand-maintained disposition drifts, so this count is one to re-scrape and re-check against the live sitemap rather than trust standing. What it buys is in market_ops.json mot_dispatch_cut: the operator publishes, per region per 5-minute interval, its whole offer stack split into what cleared and what did not, so the not-dispatched total is the market's OWN published economic headroom. Over 15 days spread weekly across the window (2026-04-21 to 07-13) that averages 2,452 MW on Luzon (17.7 percent of the stack), 527 MW on Mindanao and 249 MW on the Visayas, where it falls to zero: on four separate days the Visayas cleared its entire offer stack, with no economically offered MW left to call at the tightest interval. That is a tighter read of the supply question than registered capacity, which cannot tell offered-and-unused from unavailable. Two checks travel with it, one of them only partly passing. MOT's dispatched MW reconciles to the RTDSUM generation for the same interval, which is what a cleared-MW reading predicts and an as-bid one does not, but only the Visayas and Mindanao land inside one percent; Luzon runs about two percent high and that residual is open, not explained. The obvious candidate is already ruled out, since pumped storage and batteries charging would show as dispatched offers without being generation and they come to 0.2 MW on a sampled Luzon interval against a gap of roughly 160 MW. The RTDSUM import, export and loss columns, which the earlier note flagged as the untested lead, do not close it either, and that is now measured rather than guessed: import averages 40 MW on Luzon, about a quarter of the gap; and the dispatch-minus-generation gap tracks RTDSUM's OWN energy-balance residual (generation plus imports minus requirement, losses and exports, which fails to close by about 1.5 percent of Luzon generation) at a correlation near 0.9 (159 against 167 MW across the 15 days, close and moving together). So the gap is essentially RTDSUM's own unexplained balance term, the part the import, export and loss columns leave over rather than anything they explain. What remains is most likely a self-scheduled or must-run generation split that sits in both dispatched MW and generation but not in the economic requirement the balance nets, and IEMOP does not publish that split, so the residual stays a genuine boundary with its missing input named. The second check: the partially-cleared resource this module reads off the cut is the one IEMOP separately names as price setter on 88 to 96 percent of intervals per grid against a 22 to 35 percent random-draw baseline. MCP and MOT come out of the same dispatch solve, so that agreement checks this parse rather than confirming the setter independently. MOT is not finer-grained than the offer books already archived, its Block column being the same tranche index as RTDOE's breakpoints, so the sub-hourly accounting boundary above stands untouched. The settlement-side family the round-8 disposition left queued is now measured: the per-resource reserve results (DIPC reserve results final) drive a tighter reserve validation and archive the per-resource cleared schedules for the queued joint clear (market_ops.json reserve_results); the ancillary-services registered capacities (CAPER) size the reserve book against its registration base (reserve_registration); the indicative administered prices and the prices used in settlement are sampled as a measured record (settlement_side), where the administered price is the operator's cost-based counterfactual, not the as-bid market price, carrying the island premium on the corridors (Visayas and Mindanao administered prices run well above Luzon's), while the settlement congestion component is empty at the one-price-per-island granularity WESM settles at, so neither adds a receipt the flows table does not already carry; and the day-ahead prices are measured as a diagnostic spread against the real-time settlement, both-signed, staying out of the replay's dispatch scope as a projection. What stays named with its landing spot: the RTD per-resource prices and schedules zips (5-minute per-resource validation, the sub-hourly refinement below) and the reserve market clearing price; the indicative modified administered prices ride with the administered-prices family. The System Operator's dispatch-instruction family (the full daily instruction log, the MOT-raise re-dispatch list, and its valid-discrepancy list) was MISSING from this triage until the round-9 audit named it; all three are now archived and measured (market_ops.json so_instructions). Beyond IEMOP, the long-term demand outlook is now consumed: the DOE Power Development Plan 2023-2050's peak-demand forecast (its Table 28, per grid, 2021-2050) is parsed into a demand path in the LT Plan view (demand_path.json, pipeline/pdp_demand.py), each year's three grid values reconciled to the plan's own Philippines total within 2 MW or the build refuses. It is a LABELED DOE forecast, not this site's projection: national peak grows about 5.3 percent a year from 19,003 MW in 2025 to 68,483 MW in 2050, and the demand-side data-center anchors (DICT, DOE, Meralco) sit on top of that baseline growth. The DOE PDF 403s non-PH requests, so the source is the Internet Archive's capture of the DOE's own URL (the pdftotext extraction committed for a reproducible parse, like the plant lists). None of the remaining named features contradicts a written claim; each is a deepening of a table that already exists.
• Offer-based dispatch is NOT a boundary, and it shipped: IEMOP publishes the real-time generation offers and the regional merit-order tables in the same public window (verified July 2026), and the replay now dispatches against the observed books: the offer-mode backcast and the studio's Observed-offers engine toggle (rounds 3 and 4 of the parity loop), with the derived daily books committed under data/derived/offer_daily/. The flows table stays as the quantified evidence of what the COST proxy misses: the observed corridors run on offer differentials, about 2 percent of market-window hourly clearing prices are NEGATIVE (below any cost floor; the share reaches half of all hours inside the suspension's administered weeks, which the backcast excludes), and the cost stack reproduces neither. The offer mode reproduces the corridor directions; the negative hours remain the named sub-hourly refinement above.

What this map does NOT claim

• It does not claim data centers raised Philippine electricity prices. Current data-center load is small against a roughly 15 GW Luzon peak; the price history in the window is driven by fuel, outages, weather, and the market restart.
• It does not predict brownouts. It shows observed curtailment, observed reserve shortfalls, and arithmetic on published margins.
• It does not accuse any company. Data centers are also investment; the map shows where the grid binds and where the announced load lands.
• Corridor and grid lines follow real routes as mapped in OpenStreetMap (community data, ODbL; not NGCP documents); data-center pins are city-precision; forecasts belong to their named owners.

Known gaps: what a siting or trading decision still needs

This page states its limits section by section; here is the short list a decision-maker would want up front. None of these three is built in, and none is a promise of a future feature: the underlying queue and curtailment data are not public, so this is a stated scope boundary, not a gap we plan to close.

• Interconnection-queue position. Nothing here says where a candidate site sits in NGCP's connection queue, or how long that queue runs ahead of it.
• Curtailment risk once connected. The map and studio show curtailment already recorded in the archive; neither forecasts the curtailment a new interconnection would face once the generation queued ahead of it clears.
• PPA-versus-merchant framing. The nodal deviation and the capture-price work show how thin the market's own margin already runs; neither is a bid or a contract price, and this tool does not weigh a power-purchase agreement against a merchant position for any site.

Code MIT; baked data products CC-BY-4.0. Attribution when redistributing: Power Dispatch Studio (2026), IEMOP public market data archive, https://github.com/xmpuspus/power-dispatch-studio.