RenewablesPublished Oct 06, 2026·Updated Oct 06, 2026·12 min read

Why Most Solar Underperformance Goes Unnoticed for Weeks

Q: How do you separate soiling from cloud cover?

Build a weather-stratified baseline using clearness index. Compute separate PR medians for clear sky (> 0.75) and cloudy (< 0.5) days, then compare today's output against the matching category. This eliminates false alarms from natural cloud variation.

Q: What if there's no irradiance sensor?

Three fallbacks in order of reliability: transpose GHI to POA using a geometrical model like Hay-Davies (±8–12% accuracy), inverse derivation from production assuming a healthy plant (risky as legacy losses become normal), or install a POA sensor for +25–30% accuracy gain with ROI in 6–12 months.

Q: How do you validate a new baseline?

Test on a known-good period such as 30 days post-maintenance. Run alerts and check for zero false triggers. If alerts fire, the baseline is biased. Calibrate iteratively by comparing against known incidents and adjusting thresholds.

Q: Does this work for wind?

Yes, but the approach is simpler. Wind power follows a cubic curve, so detection compares actual output against the turbine's power curve lookup table. No irradiance needed, just an anemometer. Hidden wind losses include blade pitch issues (~8%), gearbox wear (~3–5%), and wind sensor fouling.

Q: How often should you update the baseline?

Daily calculation for acute detection (string trips, stuck trackers), weekly reset to avoid long-term drift, and quarterly audit comparing baseline versus actual. Soiling detection requires at least 5–7 days of data, while chronic degradation needs 90+ days.

A well-maintained 5 MW solar plant averaging 2.1% performance loss loses $52,000–78,000 per year undetected. How daily detection with an Expected Production Baseline recovers 97% of lost revenue.

Nadine DabouzCEO & Co-Founder

LinkedIn &nearr;

Key takeaways

→SCADA alarms only trigger above 5–8% loss, letting 2–3% silent losses accumulate unnoticed
→A 5 MW plant loses an average of $38,000/year from undetected issues (strings, soiling, trackers)
→An Expected Production Baseline (EPB) detects these anomalies in 24 hours instead of 30 days
→ROI of daily detection reaches 6.78× from year two onward

The Problem: The Gap Between SCADA and Reality

Traditional SCADA systems monitor operational state, not energy performance. ### Why SCADA Isn't Enough

Traditional SCADA systems (Supervisory Control and Data Acquisition) monitor **operational state**, not **energy performance**.

They trigger alarms for: - Inverter outages - Isolation faults (> 1 MΩ) - Network overload - Voltage swings (± 10%)

But they **miss**: - Performance loss of 1–3% (plant runs normally) - Gradual degradation (reset monthly) - Partial failures (one tripped string = 5–7% loss only)

**Result:** A plant showing 100% availability but producing at 96–98% capacity.

### The Perverse Cycle of Monthly Reporting

Day 1: String trips (-6% output) — Not visible in SCADA (plant is stable)

Days 2–30: Loss accumulates — No one investigates ("SCADA says it's fine")

Day 31: Monthly report arrives — Loss detected, but it's "gone now" — Corrective action on day 35+

Month's total loss: -6% × 30 days ≈ $4,200 USD (5 MW)

**Timing is the critical issue.**

Anatomy of Hidden Losses: Three Mechanisms

Three failure modes account for 85% of invisible production losses: tripped strings, soiling, and stuck trackers. ### 1. Tripped Strings (35% of invisible losses)

**Physics:** A string (series-connected panels) disconnects when: - DC breaker trips (surge, transient short circuit) - MC4 connector corrodes (salt spray, poor crimp) - String fuse blows (internal cell defect)

**Production Impact** — For a plant with **100 parallel strings**: - 1 tripped string = **1/100 = 1% loss** - 3 strings = **3% loss** - 5 strings = **5% loss**

**Critically:** A tripped string doesn't drop central voltage (other 99 strings in parallel compensate). Inverter sees only a small voltage swing (< 5%) and **does not alarm.**

**Real Case Study: 20 MW Park, Provence (May 2025)** - **SCADA detection date:** May 17 (monthly report) - **Actual fault date:** ~May 2 (retroactive) - **Affected strings:** 8 of 2 inverters (~4% loss) - **Undetected days:** 15 days - **Production lost:** 15 days × 4% × 20 MW ≈ **12 MWh = $1,800 USD** - **Root cause:** Salt-water corrosion on connectors (coastal zone) - **Intervention time:** 3 hours once identified

---

### 2. Soiling — 28% of invisible losses

**Physics:** Soiling reduces irradiance reaching cells. The soiling ratio (surface fraction covered) ranges from 5–15% in agricultural/coastal zones to 2–5% in urban areas.

**Detection Problem:** Aggregated plant data **cannot distinguish** soiling (irradiance reduction) from cloud cover (irradiance reduction). **Both produce identical signals.** Without a co-located irradiance sensor, diagnosis is impossible.

**Real Case: 8 MW Park, Nouvelle-Aquitaine (March–April 2025)** - **Detected loss:** 8–12% over 28 days - **Type:** Spring pollen + dust - **Cause:** Panel orientation toward NW (pollen exposure max) - **SCADA:** No alarm (plant produced every day) - **Financial impact:** 28 days × 10% × 8 MW = **22.4 MWh = $3,360 USD** - **Solution:** Mechanical cleaning (1 day) = $800 USD; ROI = 4 days

**Key point:** Without baseline, you confuse real performance with weather.

---

### 3. Stuck Trackers — 22% of invisible losses

**Mechanism:** A single-axis tracker should follow the sun (< 2° error) from 05:00 to 19:00. If stuck (motor failure, friction): - **Morning (06:00–09:00):** Panel facing wrong way — **+8% loss** - **Noon (11:00–14:00):** Panel perpendicular — **-3% loss only** - **Evening (16:00–19:00):** Panel facing away — **+12% loss**

**Signature: Deceptively Hard to Spot** — The daily total shows only -4.2% loss buried in noise. The deviation varies by hour, making it look like normal cloud variation at midday but clearly anomalous at dawn and dusk — detectable only with a precise baseline.

**Case Study: 12 MW Park, Provence (May 2026)** - **Discovered:** May 14 (via monthly report) - **Actual loss:** ~5% over 10 days - **Cause:** Bearing wear on tracking motor - **Production lost:** 10 days × 5% × 12 MW = **6 MWh = $900 USD** - **Intervention:** Replace bearing (3 hours; part $150 USD) - **Detection if daily dashboard:** 1–2 days instead of 10

Global Quantification: Annual Cost of Invisible Losses

For a utility-scale 5 MW plant, undetected losses total approximately $38,000 per year. For a **utility-scale 5 MW plant** in Mediterranean climate, invisible losses break down as follows: tripped strings (8–12 events/yr, 2.5% avg loss, 8 days duration) cost $9,600; soiling in agricultural zones (4–6 periods, 6–10% loss, 14 days) costs $18,900; stuck trackers (2–3 events, 4.5% loss, 7 days) cost $6,300; and other minor causes add $3,200. **Total annual cost: $38,000** at an average 2.1% loss.

**Baseline assumption:** Standard detection = 30-day lag.

**With daily detection** = 1–2 day lag — **97% recovery = +$36,900/yr**

Cause	Frequency/yr	Avg Loss	Avg Duration	Annual Cost
Tripped strings (1–5)	8–12 events	2.5%	8 days	$9,600
Soiling (agricultural zones)	4–6 periods	6–10%	14 days	$18,900
Stuck trackers	2–3 events	4.5%	7 days	$6,300
Other (minor)	various	~1%	ongoing	$3,200
TOTAL ANNUAL	—	2.1% avg	—	$38,000

Solution: Expected Production Baseline (EPB)

An EPB forecasts the theoretical output of a healthy plant daily, based on irradiance, cell temperature, and system specs. ### Fundamental Principles

An EPB estimates the **theoretical output of a healthy plant** each day based on irradiance, cell temperature, equipment specs, and system efficiency (cable, transformer, inverter losses of ~2–4%).

### Essential Components

**1. Irradiance: GHI vs POA**

**GHI (Global Horizontal Irradiance)** = total radiation on horizontal surface. Public data available (Météo-France, ERA5), but requires transposition to array tilt.

**POA (Plane of Array)** = irradiance at panel surface. Exact measurement via direct sensor, but costly ($3,000–5,000) with maintenance required.

**Recommendation:** - Small parks (< 1 MW): GHI transposition - Medium parks (1–20 MW): POA sensor + hybrid - Large parks (> 20 MW): 2–3 redundant POA sensors

**2. Cell Temperature (Tc)**

Cell temperature depends on ambient temperature, irradiance, and wind speed. With a typical temperature coefficient of **α ≈ -0.4%/°C**, a 5 MW park at 65°C (40°C ambient + 25°C thermal gain) loses **-16% power vs STC**.

**3. Progressive Degradation Model**

Normal degradation runs at 0.5–0.8%/yr. Accelerated degradation above 1.5%/yr flags investigation. The baseline must account for both aging and seasonal effects.

---

### Practical Implementation: 4 Steps

**Step 1: Build Historical Baseline (3–6 months)** — Collect 90+ days of clean data: actual production (inverter), irradiance (POA preferred), temperature (ambient + wind), and plant status (all strings active, no maintenance). Convert to daily Performance Ratio and compute the median, ignoring outliers.

**Step 2: Model Production–Irradiance Relationship** — Use linear or polynomial regression on historical data (POA, cell temperature, cloudiness vs production), or a physics-based direct model.

**Step 3: Set Alert Thresholds** — Compare actual to expected daily. Ignore ±2% (instrumental noise), exclude scheduled maintenance days, and apply weather filters.

**Step 4: Alert & Investigation Workflow** — Daily calculation at 16:00, automatic alert classification, O&M team action by next morning. Example: a -5% deviation triggers a moderate alert, strings are checked, 2 tripped strings reset by 10:00 — lost revenue limited to 1 day × 5% ≈ $280 vs 30 days if delayed.

Deviation	Urgency	Action
> -3%	Below normal; possibly noise	Monitor 2 days
-3% to -7%	Moderate alert	Visual inspection
-7% to -12%	High alert	Full diagnosis today
< -12%	Critical	Immediate intervention

Complete Case Study: 20 MW Park, Provence

Over a full year, daily detection recovered 109 MWh worth $16,350 with a 6.78× ROI from year two. ### Context - **Location:** Drôme, elevation 400m, Mediterranean climate - **Type:** Single-axis trackers + centralized inverters - **Spec:** 20 MW AC, 22 MW DC, ~45 GWh/yr expected - **Period:** January–December 2025

### Event 1: Spring Soiling (March 15–April 10)

Average loss detected at 8.2%. SCADA took 27 days to detect; EPB flagged it in 2 days. Production lost dropped from 54 MWh (27-day lag) to just 4 MWh (2-day lag). **Gain from daily detection: $7,500.** Cleaning cost was $1,200 — **ROI of 6.25× in 3 days.**

### Event 2: Stuck Trackers (May 23)

Estimated loss of 4.5% average (7–10% morning/evening). Standard detection came on May 31 (8 days after); EPB+alert detected it on May 24 (1 day after). Additional production saved: 6.3 MWh. **Revenue saved: $945** at $300 intervention cost.

### Annual Summary 2025

11 minor-to-moderate incidents over the year. Production lost dropped from 127 MWh (standard detection) to 18 MWh (EPB). **Gross recovery: 109 MWh = $16,350/yr.** Infrastructure cost: $8,000 capex + $2,000/yr opex. **Year 1 ROI: 2.05× (11 months). Ongoing ROI: 6.78×.**

Metric	Value
Number of events	11 minor–moderate incidents
Output lost (standard detection)	127 MWh
Output lost (EPB)	18 MWh
Gross recovery	109 MWh = $16,350/yr
Infrastructure cost	$8,000 capex + $2,000/yr opex
Year 1 ROI	2.05× (11 months)
Ongoing ROI	6.78×

Implementation: Technical Checklist

Deploying daily detection requires POA sensors, a calculation pipeline, and 3–6 months of historical data for calibration. ### Sensor Infrastructure - POA irradiance sensor — **$4,000** (add 3× for redundancy) - Temperature + anemometer — **$800** - Data acquisition unit — **$1,500** - Connectivity (4G or fiber) — **$200/month**

### Software & Compute - Calculation pipeline (Python/Node) — **$5k–15k dev** or **$500–2k/mo SaaS** - Real-time dashboard — included - Alerts (email, SMS, Slack) — included

### Data & Sources - External weather (ERA5, Copernicus) — free - Plant history (min 3–6 months) — existing data - Equipment specs — datasheet

### Validation & Tuning - Compare baseline vs 2–3 months independent data - Test alerts (target < 5% false positives) - Document seasonal anomalies (winter, rain, etc.)

### Deployment Checklist - Install POA + temperature sensors - Gather 3–6 months plant history (output + weather) - Build/deploy baseline model (custom or SaaS) - Tune alert thresholds (test on known incidents) - Validate baseline (test period, zero false alarms) - Train O&M team (alert interpretation, response) - Deploy to production with 24/7 monitoring - Quarterly audit (baseline drift check)

Industry Standards

IEC 61724-1 and ISO 9060 define the reference framework for performance monitoring and solar radiation measurement. ### Standards to Follow

**IEC 61724-1** (Performance monitoring) — High relevance, defines PR and losses.

**IEC 62446-1** (Grid safety) — Moderate relevance, for data validation.

**ISO 9060** (Solar radiation) — High relevance, sensor validity.

**IEA PV-SYSTEM** — Very high relevance, international benchmark.

### Key Metrics

**Performance Ratio (PR)** = Actual Output / STC Theoretical. Healthy plant: 80–85% (system losses included). PR < 75%: investigate.

**Specific Yield** (kWh/kWc/day). Sunny zone: 4–5 kWh/kWc/day. Seasonal range: 3–6 winter, 5–7 summer.

**Capacity Factor** = Annual Output / (Capacity × 8,760h). France average: 13–15% (utility). Optimal southern zones: 16–18%.

Standard	Relevance
IEC 61724-1 (Performance monitoring)	High — defines PR and losses
IEC 62446-1 (Grid safety)	Moderate — data validation
ISO 9060 (Solar radiation)	High — sensor validity
IEA PV-SYSTEM	Very high — international benchmark

LumenAI Integration

LumenAI embeds this EPB methodology with automated baseline calculation, smart alerts, and assisted diagnosis. LumenAI embeds this EPB methodology via:

1. **Sensor ingestion** — direct POA, T, wind links 2. **Auto-baseline** — simple setup (capacity, location) 3. **Smart alerts** — adaptive thresholds per plant type 4. **Assisted diagnosis** — suggests root cause (string, soiling, tracker) 5. **Benchmarking** — compare your plant vs regional peers

Frequently asked questions

How do you separate soiling from cloud cover?

Build a weather-stratified baseline using clearness index. Compute separate PR medians for clear sky (> 0.75) and cloudy (< 0.5) days, then compare today's output against the matching category. This eliminates false alarms from natural cloud variation.

What if there's no irradiance sensor?

Three fallbacks in order of reliability: transpose GHI to POA using a geometrical model like Hay-Davies (±8–12% accuracy), inverse derivation from production assuming a healthy plant (risky as legacy losses become normal), or install a POA sensor for +25–30% accuracy gain with ROI in 6–12 months.

How do you validate a new baseline?