Site Calibration

Global NWP models operate at coarse spatial resolutions (e.g., 9km or 3km grid cells) and cannot resolve local terrain features, surface roughness, or thermal effects unique to a specific wind farm or solar plant. Site calibration establishes a statistical mapping between raw model output and on-site meteorological mast or SCADA measurements to correct for these systematic biases.

• Corrects for orographic speed-up over ridgelines not captured in mesoscale models
• Accounts for thermal circulations like sea breezes or valley flows
• Removes systematic phase errors in diurnal timing of wind ramps or cloud passages

Model Output Statistics is the foundational statistical post-processing technique used in site calibration. It builds a regression relationship between historical NWP predictor variables and local observations to correct raw forecast output.

• Predictors include model wind speed at multiple pressure levels, temperature, and geopotential height
• Regression coefficients are trained on a multi-year historical archive of paired forecasts and observations
• MOS equations are typically re-derived seasonally to capture changing atmospheric regimes
• The technique corrects both conditional bias (error dependent on forecast magnitude) and unconditional bias (systematic offset)

A Kalman Filter provides a recursive, self-correcting framework for site calibration that updates bias estimates in real-time as each new observation arrives. Unlike static MOS, it continuously adapts to changing conditions without requiring a full retraining period.

• Maintains a state vector representing the current bias and trend of the forecast error
• Updates estimates using a Kalman gain that optimally weights the prediction against the latest measurement noise
• Particularly effective for intra-day and hour-ahead forecast horizons where conditions evolve rapidly
• Automatically tracks seasonal transitions and sensor drift without manual intervention

Newly constructed renewable assets lack the multi-year operational history required for robust site calibration. Transfer learning addresses this data scarcity by leveraging models trained on data-rich neighboring sites and fine-tuning them on limited local observations.

• A base model is pre-trained on a source farm with extensive SCADA and met mast records
• The model's feature extraction layers capture generalizable atmospheric relationships
• Only the final regression layers are fine-tuned on the target site's short observational record
• Reduces the cold-start period from 12+ months to as little as 3-6 months of local data

In wind farms, turbines operating downwind of others experience reduced wind speeds and increased turbulence. Generic power curve models ignore these wake effects, leading to systematic over-prediction. Site calibration must incorporate wake-aware corrections.

• Uses wind direction sectors to identify which turbines are waked for a given inflow angle
• Applies velocity deficit models (e.g., Jensen, Gaussian) parameterized with local turbulence intensity measurements
• Calibrates the effective roughness length and wake decay constant against SCADA data from multiple turbines
• Corrects for deep array effects where cumulative wake losses saturate in large wind farms

For solar photovoltaic sites, site calibration must account for time-varying losses that are not meteorological in origin. Soiling loss from dust accumulation and module degradation over time introduce a slow drift between expected and actual power output.

• Incorporates a soiling ratio derived from periodic cleaning events or co-located reference cell measurements
• Models annual degradation rate (typically 0.5-0.8% per year for crystalline silicon) as a linear trend component
• Uses STL decomposition to separate the degradation trend from seasonal weather-driven variability
• Calibrated models automatically flag anomalous underperformance events requiring field inspection

A statistical post-processing technique that corrects systematic biases in raw Numerical Weather Prediction (NWP) output. MOS establishes a regression relationship between historical model forecasts and local meteorological mast observations.

• Corrects for consistent over- or under-prediction at a specific site
• Accounts for unresolved terrain features and local roughness
• Forms the statistical backbone of many site calibration workflows
• Requires a historical archive of paired forecast-observation data

A recursive Bayesian algorithm that optimally estimates a dynamic system's state from noisy sensor measurements. In site calibration, it adaptively corrects systematic forecast biases in real-time as new SCADA or met mast data arrives.

• Updates bias estimates sequentially without reprocessing the full historical archive
• Adapts to seasonal changes in vegetation, sensor drift, or turbine degradation
• Provides a running estimate of uncertainty alongside the corrected forecast
• Computationally lightweight, ideal for on-site edge deployment

A machine learning paradigm where a model pre-trained on a data-rich source wind or solar farm is fine-tuned on limited data from a target site. This accelerates deployment for newly constructed assets with short operational histories.

• The pre-trained model captures general atmospheric physics and turbine response
• Fine-tuning adapts the model to local microclimatic effects and terrain
• Reduces the cold-start problem for greenfield projects
• Often combined with physical constraints to prevent unphysical extrapolation

The computational simulation of reduced wind speed and increased turbulence downstream of a wind turbine rotor. Site calibration must account for wake effects because they cause significant energy losses in densely packed wind farms.

• A turbine in a wake may experience 10-40% lower wind speeds than the free-stream flow
• Wake deficits are a function of atmospheric stability, wind direction, and thrust coefficient
• Calibrating a farm-level model requires isolating wake effects from NWP biases
• SCADA data from upwind turbines provides ground truth for wake model tuning

The reduction in photovoltaic panel conversion efficiency caused by the accumulation of dust, pollen, and debris on the glass surface. This degradation factor must be explicitly modeled during site calibration to avoid misattributing soiling-driven power loss to forecast error.

• Soiling rates vary by location, season, and panel tilt angle
• Rain events provide natural cleaning, creating a sawtooth performance pattern
• Calibration models that ignore soiling will develop a slow, spurious negative bias
• On-site soiling measurement stations provide the ground truth for correction

A metric quantifying the relative improvement of a forecasting model over a reference baseline, typically persistence. It is defined as one minus the ratio of the model error to the reference error.

• A positive skill score indicates the model outperforms the naive baseline
• Site calibration should demonstrably improve the skill score versus the raw global model
• Skill is often stratified by forecast horizon, season, and weather regime
• Provides a standardized way to communicate calibration value to asset operators