Model Output Statistics (MOS)

MOS corrects systematic errors inherent in NWP models caused by unresolved topography, parameterized physics, and grid-scale averaging. It establishes a statistical transfer function between model-predicted variables and observed local weather.

• Uses multiple linear regression to relate NWP predictors (e.g., 850mb temperature, relative humidity) to observed predictands
• Accounts for model climatology drift by retraining on rolling historical windows
• Corrects conditional biases that vary with weather regime (e.g., model over-predicts wind speed during stable conditions)

The skill of a MOS equation depends critically on selecting physically meaningful predictors from the NWP output that correlate with the target observation.

• Direct model outputs: Temperature, humidity, wind components at standard pressure levels
• Derived quantities: Thickness fields, vorticity advection, frontogenesis functions
• Temporal tendencies: 12-hour changes in model variables capturing frontal passages
• Geographic variables: Station elevation, distance to coast, land-use classification

MOS equations are developed individually for each observation site, capturing local microclimatic effects that the coarse NWP grid cannot resolve.

• A coastal station's temperature equation includes sea surface temperature and wind direction terms to model sea breeze cooling
• Mountain valley stations incorporate valley geometry and drainage flow predictors
• Urban heat island stations include boundary layer depth and anthropogenic heat flux proxies
• Each equation is valid only for its specific forecast projection hour and season

MOS traditionally employs stepwise multiple linear regression to build parsimonious equations, though modern implementations use advanced techniques.

• Forward selection: Iteratively adds the predictor that most reduces residual variance
• Backward elimination: Removes predictors that fail statistical significance thresholds (typically p < 0.05)
• Regularization methods: Ridge regression and LASSO prevent overfitting when predictor pools are large
• Machine learning extensions: Random forests and gradient boosting capture non-linear predictor interactions that linear regression misses

Beyond deterministic corrections, probabilistic MOS outputs full predictive distributions for risk-based decision-making in energy trading and grid operations.

• Logistic regression for binary events (e.g., probability of cloud ceiling below 1000 ft)
• Quantile regression directly estimates specific percentiles of the forecast distribution without assuming normality
• Ensemble MOS ingests multiple NWP ensemble members as predictors, producing spread-skill relationships that quantify forecast uncertainty
• Bayesian model averaging weights competing MOS equations by their posterior predictive performance

MOS equations are deployed in real-time operational forecasting pipelines at national meteorological centers and private energy forecasting firms.

• Equations are frozen for a season or year to maintain operational stability, then redeveloped as model upgrades occur
• Update frequency matches the parent NWP cycle (typically 1-4 times daily)
• Decay of skill: MOS accuracy degrades with forecast horizon as NWP errors grow non-linearly
• Requires archived model output spanning multiple years for robust equation development—a minimum of 2 years of homogeneous data is standard practice

The foundational physics-based input to the MOS pipeline. NWP solves discretized equations of atmospheric dynamics to simulate future states. However, raw NWP output contains systematic biases due to unresolved topography and parameterization approximations. MOS treats NWP variables as predictors in a regression framework, mapping them to observed local weather. Without NWP, MOS has no dynamic input signal.

A recursive Bayesian algorithm that serves as an adaptive alternative to static MOS equations. While classical MOS relies on a fixed historical training period, a Kalman filter updates its correction coefficients sequentially as each new observation arrives. This makes it highly effective for removing time-varying biases that drift seasonally or diurnally. It is often deployed for online, real-time forecast correction at wind farm control centers.

A flow-dependent post-processing method that bypasses rigid regression equations. AnEn searches a historical archive for past NWP forecasts that are similar to the current target forecast. The corresponding historical observations form the predictive distribution. Unlike MOS, AnEn naturally preserves the joint covariance between multiple weather variables and can generate sharp probabilistic forecasts without assuming a parametric error distribution.

The strictly proper scoring rule used to evaluate the probabilistic output of a MOS system. CRPS measures the integrated squared distance between the forecast's cumulative distribution function (CDF) and the observation. It simultaneously assesses calibration (statistical consistency) and sharpness (concentration). Minimizing CRPS is the standard objective for training modern probabilistic MOS implementations.

The training data backbone for modern MOS development. ERA5 provides a globally complete, hourly gridded record of atmospheric states spanning decades. Because MOS requires a long, stable archive of model-forecast-to-observation pairs, ERA5 is often used as a proxy for the forecast model when direct NWP archives are unavailable. It enables site screening and MOS equation derivation for locations without dense sensor networks.

A statistical technique that extends classical linear MOS to produce non-parametric prediction intervals. Instead of predicting only the conditional mean, quantile regression estimates specific percentiles (e.g., 10th, 50th, 90th) of the target variable. When trained with pinball loss, it directly outputs the forecast distribution without assuming Gaussian errors, making it ideal for renewable energy applications where uncertainty is asymmetric.