MCP — Measurement-Correlate-Predict: converting short-term campaigns into long-term estimates

A 12-month LiDAR campaign produces, by definition, 12 months of data. The financing contract for a wind farm covers 20 years of operation. The question the lender needs to answer is: does the wind measured in 2025 represent the wind that will blow from 2026 to 2045? The answer, almost always, is no — and MCP is the method used to correct this.

The wind resource varies every year. A period of La Niña can raise the average speed in the Northeast by 8–12% compared to the climate of 20 years; a period of atmospheric blockade can suppress the wind for whole months. A campaign measured in an exceptionally windy year will generate an overestimated P50 — and the park will deliver less energy than the previous model, pressing the DSCR exactly when the bank expects debt service coverage.

The MCP (Measure-Correlate-Predict) solves this by correlating the short campaign with a long-term reference and projecting the site's speed to complete historical climatology.

The problem that MCP solves

Be V̄_c the average speed measured in the campaign and V̄_LT the average long-term speed of the site. If the campaign captured a period above the historical average, V̄_c > V̄_LT — and use V̄_c directly as a resource estimate overestimate the P50 in production. The adjustment factor quantifies this difference:

adjustment_factor = V̄_LT_site / V̄_campaign_measured

adjustment_factor > 1  →  campaign underestimated the historical resource
adjustment_factor < 1  →  campaign overestimated the historical resource
adjustment_factor = 1  →  campaign representative of the climatology (rare)

An adjustment factor of 0.89, for example, means that the measured period was 11% above the historical average. The campaign-based P50 would be 11% optimistic — which, for a park of 100 MW, represents a difference of AEP that can make the difference in debt service coverage (see NT.06).

The three inputs of MCP

The MCP uses three distinct series. Measured campaign It is the hourly speed (or in 10 min) of the LiDAR for the height of analysis in the measurement window — months to a maximum of 2 years. - simultaneous reference It is the same variable extracted from the source of reanalysis (ERA5, MERRA-2) or from a long-term weather station, for the same period of the campaign. long-term reference It is the complete series of the same source, covering the available history — typically 20+ years for ERA5 and MERRA-2.

The quality of the simultaneous reference defines the quality ceiling of the MCP. If the reference does not capture the local variability of the wind (complex terrain, coastal effects, inversions), the regression will have low R² and the adjustment will be of low confidence — regardless of how good the campaign is. The choice of reference and verification of its local representativeness is the first technical decision of the MCP, and is discussed in NT.02.

Linear regression: mechanics and formulas

The linear adjustment minimizes quadratic residues between measured V_site and model prediction in the overlap period:

V_site = a + b × V_ref       (regression equation)

b = slope     →  change in V_site per unit of V_ref
a = intercept →  estimated V_site when V_ref = 0

slope, intercept = np.polyfit(V_ref_overlap, V_site_overlap, degree=1)

The adjustment metrics quantify the quality of the regression in the simultaneous period:

R²   = 1 − SS_res / SS_tot       (fraction of variance explained)
RMSE = √( mean(residuals²) )       (error in m/s)
MAE  = mean(|residuals|)           (mean absolute error in m/s)
bias = mean(predicted − measured)     (systematic bias in m/s)

The long-term prediction applies the same equation to the long-term reference series:

V̂_site_LT = a + b × V_ref_LT    (clamped at ≥ 0 m/s)
P50 = mean(V̂_site_LT)           (mean of the predicted LT series)
adjustment_factor = P50 / V̄_campaign

What each metric says to the engineer

R² It is the fraction of the wind variance of the site that the reference can explain in the simultaneous period. R²=0.84 means that 84% of the hourly variability of the campaign is captured by the reference. The alert threshold is R² < 0.70 — below that, the reference does not have sufficient local representativeness to make a bankable adjustment. The most common cause is the horizontal distance between the site and the grid point of the reanalysis, or the presence of local effects (channeling by valley, wake of another park, variation of coastal regime) that the regional model does not capture.

Slope (b) It indicates the sensitivity of the wind of the site to the reference. Slope close to 1.0 indicates that the relationship is almost proportional; slope < 1 indicates that the site has less variability than the reference (depreciated microclimate); slope > 1 indicates local amplification. Negative or zero Slope is physically implausible and indicates or data corrupted or completely inappropriate reference — MCP should not be accepted in these cases.

RMSE It is the basis for calculating the uncertainty of MCP: lower values indicate that the reference follows the time-to-time site with lower residual dispersion. For 12 months in the Northeast with ERA5, typical RMSE is between 0.8 and 1.5 m/s. Values above 2.0 m/s indicate that the reference is not adequately representing the local regime.

Bias It is the mean systematic difference between predicted and measured in the overlap. Bias near zero indicates that the reference does not have a systematic offset in relation to the site — which is expected if the regression was adjusted correctly. Bias high may indicate the simultaneous period captured an unusual event (a cold front, a week of blockage) that distorted the mean overlap.

Uncertainty of the MCP and the P90 speed

The uncertainty of MCP measures how long-term adjustment may err in the estimate of V_LT_site. The v1 formulation, conservatively simple, uses the RMSE relative to the adjusted mean:

σ_MCP (%) = RMSE / V̄_LT_site × 100

P90 = P50 × (1 − 1,2816 × σ_MCP / 100)
# Where 1.2816 = Φ−1(0.90): z of the 90th percentile of the normal distribution

A critical point regarding interpretation: the MCP P50 and P90 are wind speed estimates, not AEP estimates. Conversion requires the wind turbine power curve — a nonlinear relationship that amplifies uncertainty near rated wind speed. For the AEP P90 required by the lender, MCP uncertainty is only one component; it must be combined by RSS with the other sources (measurement uncertainty, flow modelling and interannual variability), as discussed in NT.06.

The warnings of MCP and what they mean

Alerts are warnings, not locks: BORA allows to calculate MCP even when triggered, so that the engineer can test alternative settings. What is not allowed is to delete the report alert — if MCP was calculated with overlap of 4 months or R²=0.65, this limitation needs to appear explicitly in the MCP section of the report, with the technical justification for why the result is still usable (or with the recommendation not to use it for final decision).

Availability and source of reference

The current version of BORA accepts two types of source: manually imported CSV and local NetCDF file from ERA5. For the Brazilian Northeast, the most used data are:

• . ERA5 time (ECMWF): coverage since 1940, resolution ~31 km, available via Copernicus CDS. Best option for most sites in the Northeast — documented performance in NT.02.
• MERRA-2 (NASA): coverage since 1980, resolution ~50 km, access via GESDISC. Alternative to ERA5 especially in regions where ERA5 has higher biases.
• Local weather station) : reference closer to the site, but with typically shorter history and no guarantee of continuity. Useful when the terrain makes the correlation with gross grid reanalysis inadequate.

The source traceability is stored in the file - .boraThis allows the auditor, in due diligence, to verify which version of the ERA5 was used and whether the long-term period covers the minimum acceptable — part of the audit trail discussed in NT.05.

Limitations of the linear method and what comes next

The linear MCP assumes that the relationship between site speed and reference speed is stationary, isotropic and without directional influence. These hypotheses are reasonable in many places in the Northeast — where the alisis regime dominates and the sector variability is lower — but fail in:

• Sites in complex terrain where the acceleration of flow is different by sector (channeling by valley, escarpment effect)
• Sites close to the coast where the sea breeze regime has a distinct directional correlation with the alisis regime
• Sites with high IT where speed variance is expanded in a nonlinear way

For these cases, the future methods — MCP sectorial, matrix MCP, variance ratio — solve the linear limitations by applying different regressions per wind sector or by preserving the speed distribution rather than just adjusting the mean. The comparison between methods will be the subject of a future note as implementation progresses.