Solar plant performance ratio: separate weather from equipment losses

What the asset manager is really asking

A monthly generation report says March was 6% below budget. The asset manager for a portfolio of rooftop and ground-mount plants has to answer one thing before anything else: was it the weather, or was it us? Weather shortfalls are nobody's fault and need no action. Equipment, soiling and outage losses are recoverable, and each has a different owner: the O&M contractor, the cleaning crew, the DISCOM or the inverter supplier. Energy alone cannot separate these. PR can, if the inputs are right.

Inverter, meter and irradiance files

Aggregate irradiance to kWh/m² by summing interval readings × interval length in hours, and only after dropping night-time negative offsets that some sensors report.

Definitions

• Final yield Y_f = E_AC ÷ P_DC (kWh ÷ kWp, in hours)
• Reference yield Y_r = H_POA ÷ G_ref, with G_ref = 1 kW/m² (also in hours)
• PR = Y_f ÷ Y_r
• Period PR = Σ E_AC ÷ (P_DC × Σ Y_r), a ratio of sums, not the mean of daily PRs
• Temperature-adjusted PR (optional) = E_AC ÷ (P_DC × Y_r × (1 + γ × (T_cell − 25))), where γ is the power temperature coefficient from your module datasheet

The temperature adjustment removes the expected summer dip so that a hot month is not read as degradation. Contracts and standards define their own correction methods; if your PR guarantee specifies one, use that method and not this simplified version.

Ten days on a synthetic 500 kWp plant

The data is synthetic. 4 March is heavily overcast, 6 March has a 3.5-hour grid outage, the modules gather dust through the first eight days, and they are cleaned on the morning of 9 March.

import pandas as pd

# Synthetic daily data for a hypothetical 500 kWp (DC) rooftop plant
P_DC_KWP = 500.0
G_STC = 1.0          # kW/m², the irradiance at standard test conditions
GAMMA = -0.0035      # per °C power temperature coefficient: take this from your module datasheet

d = pd.DataFrame({
    "date": pd.date_range("2026-03-01", periods=10, freq="D"),
    "e_ac_kwh": [2380, 2352, 2310, 1190, 2275, 1500, 2236, 2200, 2440, 2425],
    "h_poa_kwh_m2": [5.90, 5.88, 5.85, 3.00, 5.82, 5.80, 5.78, 5.75, 5.95, 5.92],
    "h_ghi_kwh_m2": [5.30, 5.28, 5.26, 2.80, 5.24, 5.22, 5.20, 5.18, 5.35, 5.33],
    "grid_down_hours": [0, 0, 0, 0, 0, 3.5, 0, 0, 0, 0],
    "t_cell_avg_c": [42, 43, 43, 33, 44, 44, 45, 45, 43, 44],
    "cleaned": [False] * 8 + [True, False],
})

d["ref_yield_h"] = d["h_poa_kwh_m2"] / G_STC
d["final_yield_h"] = d["e_ac_kwh"] / P_DC_KWP
d["pr"] = d["final_yield_h"] / d["ref_yield_h"]
d["pr_ghi_wrong"] = d["final_yield_h"] / (d["h_ghi_kwh_m2"] / G_STC)
d["pr_temp_corr"] = d["e_ac_kwh"] / (P_DC_KWP * d["ref_yield_h"] * (1 + GAMMA * (d["t_cell_avg_c"] - 25)))

cols = ["e_ac_kwh", "h_poa_kwh_m2", "grid_down_hours", "pr", "pr_ghi_wrong", "pr_temp_corr"]
out = d.set_index(d["date"].dt.strftime("%d-%b"))[cols]
print(out.round(3).to_string())

clean_days = d[d["grid_down_hours"] == 0]
print("Period PR, all days:", round(d["e_ac_kwh"].sum() / (P_DC_KWP * d["ref_yield_h"].sum()), 3))
print("Period PR, excluding grid-outage day:",
      round(clean_days["e_ac_kwh"].sum() / (P_DC_KWP * clean_days["ref_yield_h"].sum()), 3))

Output:

        e_ac_kwh  h_poa_kwh_m2  grid_down_hours     pr  pr_ghi_wrong  pr_temp_corr
date
01-Mar      2380          5.90              0.0  0.807         0.898         0.858
02-Mar      2352          5.88              0.0  0.800         0.891         0.854
03-Mar      2310          5.85              0.0  0.790         0.878         0.843
04-Mar      1190          3.00              0.0  0.793         0.850         0.816
05-Mar      2275          5.82              0.0  0.782         0.868         0.837
06-Mar      1500          5.80              3.5  0.517         0.575         0.554
07-Mar      2236          5.78              0.0  0.774         0.860         0.832
08-Mar      2200          5.75              0.0  0.765         0.849         0.823
09-Mar      2440          5.95              0.0  0.820         0.912         0.875
10-Mar      2425          5.92              0.0  0.819         0.910         0.878
Period PR, all days: 0.766
Period PR, excluding grid-outage day: 0.795

Three stories in one column

The cloudy day is not a problem. Energy on 4 March halved to 1,190 kWh, yet PR was 0.793, in line with its neighbours. A report that ranks days by energy would flag it; PR correctly does not.

The outage day is. On 6 March irradiance was normal and PR fell to 0.517. The event log explains it: 3.5 hours without the grid. That loss belongs to the grid, not the plant, and for a contractor's PR guarantee it is usually excluded under the contract's own rules. Here, excluding it moves the period PR from 0.766 to 0.795. Report both, and say which one the contract uses.

Soiling hides in a slow slide. From 1 to 8 March PR drifts from 0.807 to 0.765 with no event at all. After cleaning on 9 March it jumps to 0.820. That step size, divided by the days since the last clean, is your soiling rate, and it is the input to a cleaning-frequency decision that weighs water and labour cost against recovered energy.

GHI, AC capacity and other PR inflators

• GHI in place of POA. Horizontal irradiance is lower than tilted irradiance for a south-facing tilted array, so dividing by GHI inflates PR. In this synthetic data it adds about 0.09 on every day, turning an ordinary plant into an apparently excellent one.
• AC capacity in the denominator. Plants are often built with more DC than AC capacity. Using kW AC in place of kWp DC inflates PR. Always state which capacity you used.
• A dirty or drifting pyranometer. If the sensor gathers dust like the modules, both numerator and denominator fall together and PR hides the soiling. A sensor cleaning log, or a comparison with a nearby reference or satellite estimate, catches this.
• Curtailment and export limits. When the DISCOM or a zero-export controller limits output, the energy is lost by instruction, not by fault. Flag those intervals from the event log.
• Averaging daily PRs. The mean of daily ratios weights a cloudy day the same as a sunny one. Use the ratio of sums for any period figure.
• Temperature effects read as degradation. Summer PR is normally lower. Compare the same month year on year, or use the temperature-adjusted series with the datasheet coefficient.

Presenting to the asset manager

Lead with a loss waterfall, not a table: budget energy, then the weather effect (from irradiance versus the budget's irradiance), then grid outage and curtailment, then soiling, then residual equipment loss, ending at actual energy. Put the daily PR chart underneath with event markers for outages and cleanings. The one sentence that matters is who owns each loss. A shortfall that is mostly weather needs no action; a residual equipment loss that grows month after month needs an inverter or string-level investigation, and that is where string current data comes in.

Building the skill

Normalising an outcome by the input that nobody controls, and then attributing the remainder, is a modelling habit that the Data Science programme builds on real time-series data. The same idea of a physical measurement checked against a book figure appears in fuel station wet-stock reconciliation. If you are coming from an energy O&M role, the free skill check will map what your current work already shows.