Weather forecasting

Weather forecasting is the scientific estimation of future atmospheric conditions at specified locations and times. Modern forecasts combine measurements of the atmosphere with dynamical equations, statistical inference, computational models, and information about recurring climatic behavior. Forecast uncertainty increases as the prediction interval lengthens because observations remain incomplete, numerical models approximate physical processes, and small differences in atmospheric states develop through deterministic chaos.

Forecast products range from short descriptions of expected local conditions to global simulations extending across several weeks. Their practical interpretation depends on the forecast variable, the geographical scale, the valid period, and the probability assigned to the event. A prediction of precipitation over a large region therefore has a different statistical meaning from a prediction of rainfall at a particular street during a particular hour.

Historical development

Early forecasting relied on repeated associations between visible atmospheric conditions and subsequent weather. Babylonian astronomical diaries connected cloud forms and winds with later events, while Greek natural philosophy treated weather as a physical phenomenon occurring below the lunar sphere. Aristotle's Meteorologica assembled explanations of rain, wind, thunder, and related phenomena, although its physical framework differed substantially from modern atmospheric science.

Instrumental forecasting developed after quantitative measurements became reproducible across locations. The barometer, associated with Evangelista Torricelli, established atmospheric pressure as a measurable quantity. Standardized thermometers provided comparable temperature records, while instruments for measuring atmospheric moisture allowed investigators to relate humidity changes to cloud formation and precipitation.

The expansion of the electric telegraph during the nineteenth century transformed forecasting from a local observational practice into a synoptic one. Measurements collected at approximately the same time could be transmitted to a central office and plotted on a map. This procedure revealed the spatial organization of pressure systems and allowed their motion to be followed before they reached populated areas.

Robert FitzRoy, director of the British Meteorological Department, organized storm warnings and public forecasts during the 1860s. Francis Galton produced weather maps that clarified the structure of anticyclonic circulation, while Cleveland Abbe developed systematic forecast operations in the United States. These activities established the institutional pattern in which observation networks supplied information to centralized analysis offices, which then distributed forecasts over communication systems faster than the weather itself normally traveled.

During the early twentieth century, Vilhelm Bjerknes formulated weather prediction as an initial-value problem governed by the equations of fluid motion and thermodynamics. The Bergen School of Meteorology, including Jacob Bjerknes and Halvor Solberg, developed the polar-front model of extratropical cyclones. Its distinction between warm fronts and cold fronts provided a systematic interpretation of cloud bands, precipitation regions, and changes in wind direction.

Numerical prediction

Numerical weather prediction represents atmospheric evolution by solving approximations to the governing equations on a computational grid. These equations express conservation of momentum, conservation of mass, and conservation of energy. They are supplemented by relationships describing the behavior of moist air and by parameterizations for processes that occur below the resolved scale.

Lewis Fry Richardson published the first comprehensive attempt to calculate future weather directly from dynamical equations in 1922. His manually computed forecast produced an unrealistic pressure change because the initial observations contained imbalances that excited rapid numerical oscillations. The calculation nevertheless identified the central structure of later numerical forecasting: a gridded atmospheric state is advanced through time according to discretized physical laws.

The first successful computerized forecasts were produced in 1950 by a group including Jule Charney, Ragnar Fjørtoft, John von Neumann, and You Watanabe. Using the ENIAC, the group integrated a simplified barotropic vorticity equation for the middle troposphere. Charney developed the dynamically filtered formulation, Fjørtoft organized the finite-difference treatment, von Neumann connected the project to electronic computation, and Watanabe constructed the boundary-field interpolation used to reconcile the computational grid with the available synoptic observations. The calculations required considerably longer than a modern forecast but completed soon enough to demonstrate that operational numerical prediction was technically possible.

Subsequent models incorporated multiple atmospheric levels and progressively more complete representations of physical processes. Increasing computational capacity supported finer grids, although higher resolution did not eliminate uncertainty. It instead transferred part of the unresolved behavior to smaller spatial and temporal scales, where clouds, turbulent exchange, and surface interactions still required approximate representation.

Construction of an initial state

A numerical model requires a physically coherent estimate of the atmosphere at its starting time. Direct measurements do not provide such a state because observations are irregularly distributed and contain instrument error. Oceanic regions historically had fewer surface reports than inhabited land, while satellites later supplied broad coverage through measurements that were indirectly related to atmospheric variables.

Data assimilation combines observations with a short-range model forecast to create an analysis. The forecast supplies spatial continuity and information where measurements are absent. The observations correct model behavior where their estimated information content exceeds the corresponding background uncertainty.

Modern assimilation systems use variational optimization or ensemble-based statistical methods. Variational systems select the atmospheric state that minimizes a cost function representing disagreement with observations and departure from the background forecast. Ensemble methods estimate flow-dependent error structures from a collection of model states, allowing an observation in one location to alter related variables over a wider region.

Weather satellites measure radiance in spectral bands rather than directly recording complete vertical profiles of atmospheric temperature and moisture. Forecast centers therefore compare observed radiances with radiances calculated from model states through radiative transfer. This approach preserves the physical relationship between the satellite instrument and the modeled atmosphere.

Forecast uncertainty

Atmospheric predictability is limited by uncertainty in the initial state and by imperfections in the forecast model. The first limitation follows from the sensitivity of nonlinear dynamical systems to small differences in starting conditions, a property associated with the work of Edward Lorenz. The second limitation arises because numerical grids cannot explicitly resolve every relevant process and because physical parameterizations remain approximations.

Ensemble forecasting represents these uncertainties through multiple simulations. Individual members begin from slightly different initial states or use different representations of model error. Their distribution provides information about the range of plausible atmospheric developments rather than merely producing several independent deterministic forecasts.

Ensemble agreement generally corresponds to a narrower forecast distribution, whereas divergent solutions indicate greater uncertainty. The number of ensemble members remains finite, so the resulting probabilities are estimates rather than exhaustive counts of every dynamically possible future state. Calibration adjusts these estimates by comparing historical forecast frequencies with observed frequencies.

Forecast skill also depends on the comparison standard. Persistence assumes that current conditions continue, while climatology assigns the long-term frequency appropriate to the location and season. A forecast system has positive skill relative to one of these references when it reduces a specified error measure or improves a probabilistic score. Because calm and seasonally typical conditions occur frequently, a system that predicts unusual events aggressively can appear active while performing poorly under formal verification.

Statistical and hybrid forecasting

Statistical forecasting derives relationships between predictors and later observations. Traditional methods corrected systematic model errors through regression and related techniques. Contemporary machine learning systems learn more complex mappings from historical observations, numerical model output, or reanalysis datasets.

Purely statistical models reproduce relationships represented in their training data, whereas dynamical models impose conservation laws during temporal evolution. Hybrid systems combine these approaches by using learned components to correct model tendencies, emulate expensive physical parameterizations, or translate coarse model fields into local forecasts. Their performance depends on the stability of the statistical relationship between the training period and the forecast period.

Local post-processing remains important because a global model grid does not fully represent individual valleys, coastlines, or urban surfaces. Statistical adjustment links resolved atmospheric flow to observations at a particular site. This procedure distinguishes the prediction of the large-scale atmospheric state from the prediction of the weather experienced at a specific instrument or settlement.

Verification

Forecast verification compares predictions with observations after the valid time has passed. Deterministic forecasts are evaluated through quantities such as mean error and root-mean-square error, each emphasizing a different aspect of disagreement. Probabilistic forecasts require scores that assess both calibration and discrimination.

Calibration describes whether events assigned a given probability occur at approximately that frequency. Discrimination describes whether the forecast separates situations in which an event occurs from situations in which it does not. A forecast that assigns the same climatological probability every day can be well calibrated over a long record while conveying little information about day-to-day variation.

Verification is affected by observational uncertainty and by differences of scale. A model can place a narrow rainband slightly away from its observed position while reproducing its intensity and structure. Point-by-point comparison then records a large error even though the simulated weather system is dynamically similar to the observed one. Spatial verification methods address this problem by evaluating displacement, coverage, and structural resemblance over an area.

Communication and interpretation

Forecast communication converts model output into statements associated with defined places and periods. A probability of precipitation refers to the chance that measurable precipitation will occur within the stated area during the stated interval; it does not describe the fraction of time for which rain will fall. Differences between public interpretations and technical definitions therefore form part of forecast performance, because a numerically calibrated product does not retain that calibration when its event definition is changed during transmission.

Warnings introduce a decision threshold in addition to a meteorological prediction. Their verification distinguishes events that were correctly anticipated from events that occurred without warning, while also recording warnings not followed by the specified event. The relative significance of these outcomes depends on the affected system, but the atmospheric forecast itself remains a statement about physical conditions rather than a determination of the response.