{
  "id": "ensemble-forecasting",
  "title": "Ensemble Forecasting",
  "lang": "en",
  "source_lang": "en",
  "summary": "Ensemble forecasting uses multiple model simulations to estimate future conditions and the uncertainty surrounding a prediction, particularly in weather, climate, and hydrology.",
  "category": "geography",
  "subcategory": "",
  "type": "concept",
  "aliases": [
    "Ensemble prediction",
    "Ensemble prediction system"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/ensemble-forecasting/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/ensemble-forecasting/",
      "json": "https://aiwiki.page/zh/ensemble-forecasting.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Ensemble forecasting is a prediction method that uses multiple simulations, called ensemble members, to represent a range of possible future conditions. Rather than producing only one forecast, an ensemble varies initial conditions, model formulations, or other uncertain inputs and examines how the resulting predictions differ. Widely used in meteorology, it extends numerical weather prediction from a single estimated outcome toward an estimate of the probability distribution of future weather. Related methods are used in climate prediction, ocean forecasting, and hydrology. An ensemble samples uncertainty; it does not automatically capture every possible outcome or provide perfectly calibrated probabilities."
    },
    {
      "heading": "Scientific basis",
      "text": "Uncertainty and predictability\n\nWeather forecasts begin with an estimated state of the atmosphere, constructed from observations and a previous forecast through data assimilation. Observations have errors, coverage is incomplete, and the atmospheric state cannot be specified exactly at every location and scale.\n\nAtmospheric evolution is also nonlinear. Through chaotic dynamics, initially small differences can grow and eventually produce substantially different weather patterns. This sensitivity limits detailed prediction even if the governing physical equations are known. Predictability varies with atmospheric conditions, the variable being forecast, and the spatial and temporal scale of interest.\n\nForecast models introduce additional uncertainty. Their grids cannot explicitly resolve every cloud, turbulent eddy, or surface feature, so some processes require approximate representations. Uncertainty also arises from numerical methods, imperfect physical knowledge, and boundary conditions such as sea-surface temperature or soil moisture.\n\nA single deterministic forecast follows one modeled trajectory through these uncertainties. An ensemble follows multiple trajectories to investigate which outcomes are relatively consistent and which depend strongly on uncertain inputs.\n\nA sampled predictive distribution\n\nFor a forecast quantity \\(X\\), let the predictions from \\(N\\) ensemble members be \\(x_1,\\ldots,x_N\\). The collection can be interpreted as a sample from an estimated predictive distribution.\n\nFor an event \\(A\\), an elementary ensemble probability estimate is\n\n\\[\\widehat{P}(A)=\\frac{1}{N}\\sum_{i=1}^{N}\\mathbf{1}(x_i\\in A),\\]\n\nwhere \\(\\mathbf{1}\\) equals 1 when the member predicts the event and 0 otherwise.\n\nFor example, if 15 of 50 members predict daily precipitation above a specified threshold at a location, the raw ensemble frequency is 30%. This is not necessarily the final forecast probability: statistical processing may adjust it for systematic errors, insufficient spread, or differences between model-grid values and local observations.\n\nMembers are generally not independent, equally plausible samples of all possible futures. They share observations, model components, and design assumptions. Their frequencies therefore require empirical evaluation before being treated as reliable probabilities."
    },
    {
      "heading": "How ensembles are constructed",
      "text": "Operational systems often combine several methods to represent different sources of uncertainty.\n\nInitial-condition ensembles\n\nAn initial-condition ensemble starts forecasts from slightly different estimates of the current atmospheric state. Perturbations are designed to represent analysis uncertainty and, in some systems, emphasize structures likely to grow rapidly.\n\nMethods include:\n\n- Bred vectors, obtained by repeatedly evolving, rescaling, and reintroducing differences between model states.\n- Singular vectors, which identify perturbations with strong growth over a selected interval under a specified mathematical measure.\n- Ensemble data assimilation, in which parallel analyses use perturbed observations, model states, or other components of the assimilation process.\n\nSimply adding arbitrary noise is generally inadequate. Perturbations must represent plausible uncertainty while preserving important relationships among atmospheric variables.\n\nModel-uncertainty ensembles\n\nModel uncertainty can be sampled by changing physical parameterizations, varying uncertain parameters, or introducing stochastic perturbations to selected model tendencies. For example, members may differ in how they represent cloud formation or unresolved convection.\n\nA multimodel ensemble combines predictions from distinct forecasting models. This can expose differences that a single model cannot sample, although the models may still share structural errors. The choice of models and their weighting affects the resulting distribution.\n\nBoundary-condition and coupled ensembles\n\nSome applications perturb ocean, land, sea-ice, or atmospheric-composition conditions as well as the atmosphere itself. Coupled prediction systems allow interactions among these components to evolve during a forecast.\n\nFor seasonal prediction, uncertainty in ocean conditions can be particularly important because ocean anomalies evolve more slowly than individual weather systems. Predictable influences associated with El Niño–Southern Oscillation, for example, can alter the probabilities of seasonal temperature and rainfall patterns without determining the weather on a particular day.\n\nLagged ensembles\n\nA lagged ensemble combines forecasts initialized at different times but valid for the same future period. It can expand the available sample without launching every member simultaneously. However, older forecasts have longer lead times and may have different error characteristics, so they are not automatically interchangeable with newer members."
    },
    {
      "heading": "Interpreting ensemble output",
      "text": "Ensemble mean and spread\n\nThe ensemble mean is\n\n\\[\\bar{x}=\\frac{1}{N}\\sum_{i=1}^{N}x_i.\\]\n\nA common measure of spread is the ensemble standard deviation,\n\n\\[s=\\sqrt{\\frac{1}{N-1}\\sum_{i=1}^{N}(x_i-\\bar{x})^2}.\\]\n\nThe mean often has lower average squared error than an individual member because some member-specific errors cancel. Nevertheless, it is not necessarily a physically representative outcome. If half the members place a narrow rainband in one location and half place it elsewhere, the mean may depict a broad region of moderate rain that no member actually predicts.\n\nSpread measures disagreement among members, not the forecast error itself. Larger spread often indicates greater uncertainty, but this relationship must be verified statistically. A tightly clustered ensemble can still be wrong if its members share the same bias or omit an important source of uncertainty.\n\nProbabilities, quantiles, and scenarios\n\nCommon products include:\n\n- Threshold probabilities, such as the probability that wind speed exceeds a specified value.\n- Quantiles and prediction intervals, describing portions of the forecast distribution.\n- Plume diagrams, showing member trajectories through time at a location.\n- Spaghetti plots, displaying selected contours from multiple members to reveal differences in large-scale patterns.\n- Clusters, grouping members with similar circulation patterns or event evolutions.\n- Neighborhood probabilities, estimating the chance of an event occurring within a specified area rather than at one exact grid point.\n\nThese products answer different questions. The probability of heavy rain somewhere within a region is not the same as its probability at a particular site. Likewise, a probability for a 24-hour accumulation cannot be interpreted as a probability for each individual hour.\n\nClusters and individual members can illustrate coherent scenarios that a mean conceals. Their relative frequencies, however, depend on the ensemble design and the clustering method."
    },
    {
      "heading": "Forecast calibration and verification",
      "text": "Ensemble quality is assessed using many past forecasts matched with observations. Evaluation must distinguish several properties.\n\nReliability, also called probabilistic calibration, describes agreement between forecast probabilities and observed frequencies. Among comparable occasions assigned a 30% probability, the event should occur approximately 30% of the time. Reliability is a property of a collection of forecasts, not something established by one outcome.\n\nResolution is the ability to distinguish situations with different observed event frequencies. Sharpness describes how concentrated forecast distributions are, independently of the observations. A useful system seeks sharp forecasts subject to adequate calibration; narrow intervals are not beneficial if they repeatedly exclude the observed outcome.\n\nVerification tools\n\nA rank histogram records where observations fall among sorted ensemble members. A roughly flat histogram is consistent with appropriate ensemble dispersion under suitable assumptions, but it does not prove overall forecast quality. Excess observations outside the ensemble range can indicate insufficient spread, while asymmetry may indicate bias. Observation errors and differences between observation and model scales also affect the result.\n\nThe Brier score evaluates probabilities for a binary event:\n\n\\[\\mathrm{BS}=\\frac{1}{M}\\sum_{j=1}^{M}(p_j-o_j)^2,\\]\n\nwhere \\(p_j\\) is the forecast probability, \\(o_j\\) is 1 if the event occurred and 0 otherwise, and \\(M\\) is the number of forecast cases. Lower values indicate better performance.\n\nThe continuous ranked probability score evaluates a complete predictive distribution for a scalar quantity. Reliability diagrams, interval coverage, and comparisons with climatology or other reference forecasts provide complementary information. For spatial weather fields, additional methods assess displacement, scale, and event structure.\n\nStatistical post-processing\n\nRaw ensembles frequently exhibit systematic bias or insufficient dispersion. Statistical post-processing uses archived forecasts and observations to adjust the predictive distribution. Methods include regression-based distributional models, weighted mixtures, and machine learning.\n\nReforecasts, also called hindcasts, are retrospective forecasts generated for past dates using a consistent forecasting system. They help estimate model errors and provide a training record for calibration, particularly when the current operational model has a short archive.\n\nCorrecting individual locations, variables, or lead times separately can disrupt relationships among them. Applications requiring coherent spatial fields or time sequences therefore also need methods that preserve or reconstruct dependence."
    },
    {
      "heading": "Development and operational use",
      "text": "The intellectual basis of ensemble forecasting grew from work on atmospheric predictability, nonlinear dynamics, and statistical prediction during the twentieth century. Operational adoption expanded as computing resources made repeated model integrations practical.\n\nA major milestone came in 1992, when both the European Centre for Medium-Range Weather Forecasts and the United States National Meteorological Center, later incorporated into the National Centers for Environmental Prediction, introduced operational ensemble prediction systems. Subsequent development expanded ensemble size, resolution, model-error representation, and coupling with other Earth-system components.\n\nModern systems operate across several ranges:\n\n- Short-range, convection-permitting ensembles address localized storms, intense rainfall, and other small-scale hazards.\n- Medium-range ensembles estimate the evolution of weather patterns over subsequent days and into the second week.\n- Subseasonal and seasonal ensembles assess shifts in probabilities over weeks to months, drawing partly on slower components of the climate system.\n\nThe forecast’s meaning changes with its range. Detailed local weather generally becomes less predictable with increasing lead time, while some large-scale or time-averaged anomalies may retain useful predictability."
    },
    {
      "heading": "Applications",
      "text": "Weather hazards\n\nEnsembles support early-warning systems by estimating the probability, timing, location, and intensity of hazardous conditions. For tropical cyclones, members can sample alternative tracks and intensities. Coupled coastal models may then translate atmospheric scenarios into possible storm surges.\n\nIn heavy-rainfall forecasting, ensembles help distinguish uncertainty in storm occurrence from uncertainty in placement or accumulation. They can identify plausible low-probability outcomes that a single forecast would omit, although their ability to estimate rare extremes depends on model quality and sampling.\n\nHydrological forecasting\n\nMeteorological ensemble members can drive rainfall–runoff and river-routing models to produce streamflow distributions and alternative hydrographs. Hydrological systems may also vary soil moisture, snow conditions, model parameters, and other uncertain inputs.\n\nThe resulting forecasts support assessment of flood thresholds, reservoir inflows, and drought conditions. Their uncertainty depends on both meteorological forcing and the representation of the drainage basin.\n\nEnergy and other weather-sensitive activities\n\nEnsembles are used to estimate uncertainty in wind and solar generation, electricity demand, transport conditions, and agricultural weather. Decision models can combine forecast probabilities with the different consequences of taking or not taking an action.\n\nA probability forecast does not by itself determine a decision. Different users can rationally respond differently to the same forecast because their exposure, costs, and tolerances for missed events or false alarms differ.\n\nClimate prediction and projection\n\nEnsemble methods also appear in climate-model ensembles. Initialized climate predictions use estimates of the current Earth-system state to forecast subsequent evolution. Long-term climate change projections instead examine responses under specified external forcing pathways.\n\nSpread across long-term projections may combine internal variability, model differences, and forcing-scenario differences. A collection of such projections should not automatically be interpreted as a calibrated probability distribution over future outcomes, especially when scenarios have not been assigned probabilities."
    },
    {
      "heading": "Limitations and interpretation",
      "text": "Finite ensemble size limits the sampling of rare events. In an equally weighted ensemble of \\(N\\) members, raw event frequencies change in increments of \\(1/N\\); an event predicted by no members is not necessarily impossible.\n\nShared model errors can leave the entire ensemble displaced from reality. Increasing member count improves sampling of the distribution represented by the system, but does not by itself correct a defective distribution. Computing resources must therefore be divided among ensemble size, spatial resolution, model complexity, and forecast-update frequency.\n\nUncertainty also depends on the quantity and scale being examined. An ensemble may predict a regional circulation pattern well while poorly locating individual thunderstorms. Errors in timing or position can substantially affect point forecasts even when the broader event is represented.\n\nFinally, verification observations are imperfect, forecasting systems change, and past performance may not fully describe unfamiliar or extreme conditions. Ensemble probabilities are most informative when their event definitions, calibration record, spatial scale, and known model limitations are made explicit."
    }
  ],
  "links": [
    "meteorology",
    "numerical-weather-prediction",
    "probability-distribution",
    "hydrology",
    "chaos-theory",
    "precipitation",
    "el-nino-southern-oscillation",
    "brier-score",
    "continuous-ranked-probability-score",
    "machine-learning",
    "storm-surge",
    "hydrograph",
    "drainage-basin",
    "climate-model-ensemble",
    "climate-change"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}