🎓 Lesson 2 D2

Sky Camera Optics and Radiometric Calibration

A sky camera is a specialized upward-facing digital camera that captures images of the sky to detect clouds and estimate solar irradiance, and radiometric calibration ensures its brightness measurements match real-world light levels.

🎯 Learning Objectives

  • Explain how lens distortion and spectral sensitivity affect sky image interpretation
  • Calculate pixel-to-radiance conversion coefficients using calibration target measurements
  • Analyze calibrated sky images to classify cloud cover and estimate diffuse-to-global irradiance ratios
  • Design a field calibration protocol compliant with ISO 9060:2018 Class A requirements

📖 Why This Matters

In AI-augmented renewable forecasting, sky cameras serve as the 'eyes' of short-term solar irradiance prediction systems—providing real-time cloud motion and opacity data critical for grid stability and energy trading. Without accurate radiometric calibration, AI models ingest biased inputs, leading to systematic over/under-prediction of PV generation. A 5% radiometric error can propagate into >10% forecast error at 15-minute horizons—directly impacting ancillary service dispatch and curtailment decisions.

📘 Core Principles

Sky camera optics rely on equiangular (fisheye) projection to map the full 180° hemispherical sky onto a circular image. Key optical challenges include radial distortion, chromatic aberration, and cosine response error (deviation from Lambert’s cosine law). Radiometric calibration bridges the gap between digital counts and physics: it comprises dark current subtraction, flat-field correction (to remove vignetting and pixel non-uniformity), spectral responsivity characterization (via monochromator scans), and absolute calibration against a NIST-traceable reference source (e.g., integrating sphere or calibrated pyranometer under clear-sky conditions). The final output is a per-pixel radiance lookup table (LUT) or polynomial coefficient set enabling DN → W·sr⁻¹·m⁻²·nm⁻¹ conversion.

📐 Radiometric Conversion Model

The fundamental radiometric conversion relates raw pixel values to spectral radiance using a linearized model corrected for dark signal and optical non-uniformity. For narrowband sky cameras (e.g., 450–900 nm), this is often expressed as a gain-offset model per channel.

Pixel Radiance Conversion

L_λ(x,y) = g_λ(x,y) × [DN(x,y) − DN_dark(x,y)]

Converts raw digital number to spectral radiance at wavelength λ for pixel location (x,y)

Variables:
SymbolNameUnitDescription
L_λ(x,y) Spectral radiance W·sr⁻¹·m⁻²·nm⁻¹ Radiance at pixel (x,y) and wavelength λ
g_λ(x,y) Gain coefficient W·sr⁻¹·m⁻²·nm⁻¹ per DN Calibrated radiometric sensitivity per pixel and band
DN(x,y) Raw digital number DN Uncorrected pixel value from ADC
DN_dark(x,y) Dark signal offset DN Temperature-dependent bias measured in zero-light condition
Typical Ranges:
Clear-sky zenith (650 nm): 180 – 280 W·sr⁻¹·m⁻²·nm⁻¹
Overcast sky (550 nm): 20 – 60 W·sr⁻¹·m⁻²·nm⁻¹

💡 Worked Example

Problem: A calibrated sky camera records a pixel value DN = 3,247 in the 650 nm channel. Measured dark current = 124 DN (averaged over 100 dark frames). Flat-field normalized gain coefficient = 0.082 W·sr⁻¹·m⁻²·nm⁻¹ per DN. What is the spectral radiance at that pixel?
1. Step 1: Subtract dark current: DN_corrected = 3247 − 124 = 3123 DN
2. Step 2: Apply gain coefficient: L_λ = 3123 × 0.082 = 256.086 W·sr⁻¹·m⁻²·nm⁻¹
3. Step 3: Verify typical range: Clear-sky zenith radiance at 650 nm ranges from ~180–280 W·sr⁻¹·m⁻²·nm⁻¹ — result falls within expected bounds.
Answer: The spectral radiance is 256.1 W·sr⁻¹·m⁻²·nm⁻¹, consistent with midday clear-sky conditions.

🏗️ Real-World Application

At the National Renewable Energy Laboratory’s (NREL) Solar Radiation Research Laboratory (SRRL) in Golden, CO, the SkyCam-3 system—a dual-fisheye, 12-bit CMOS imager with integrated temperature-stabilized filter wheels—undergoes biannual radiometric calibration using a 1.5-m diameter integrating sphere (Labsphere Spectralon® source) traceable to NIST SRM 2252. Calibration coefficients are embedded in metadata and applied in real time by the AI forecasting engine SolCast, improving 5-minute GHI forecast accuracy by 14% RMSE reduction versus uncalibrated input.

📋 Case Connection

📋 CAISO Zone 12 Solar Ramp Event Mitigation

Unpredicted cloud-edge ramp rates exceeding 150 MW/min causing reserve shortfalls

📋 ERCOT South Texas Wind Farm Forecast Recalibration

Persistent under-forecasting during cold-air advection events due to mesoscale NWP boundary layer errors

📋 Hawaii Island Microgrid Solar Forecast Hardening

Volcanic haze and trade-wind cloud variability caused 30–50% forecast error spikes during monsoon season

📚 References