New statistical engine simulates France’s temperatures, heatwaves included

Imagine being able to replay French weather thousands of times, generating plausible alternate histories in which the summer of 2003 unfolds differently, or in which a heatwave strikes a region that has never yet experienced one. A team of French researchers has now built exactly that kind of machine, not with supercomputers running physical climate models, but with a carefully constructed statistical model that can churn out realistic daily temperature fields across an entire country in a fraction of the time. The work, published in the journal Advances in Statistical Climatology, Meteorology and Oceanography, describes a spatio-temporal weather generator calibrated on a few dozen weather stations yet capable of producing coherent temperature simulations at any point over metropolitan France.

Weather generators occupy a crucial niche in climate science. Physical climate models are the gold standard for projecting future conditions, but they are computationally expensive, which limits how many simulations researchers can run, and their coarse spatial resolution makes them poorly suited for local studies. Historical records, meanwhile, offer only a single realization of the climate, which can lead to a serious underestimation of variability. Stochastic weather generators sidestep both problems by treating observed weather as one of infinitely many possible trajectories of a random process. The goal is to simulate numerous plausible sequences that share key statistical properties with the observations, allowing risk analysts to sample the climate far more comprehensively than the historical record alone would permit.

The new generator, developed by Caroline Cognot of EDF R&D and Université Paris-Saclay, together with Liliane Bel, David Métivier and Sylvie Parey, was motivated by a practical problem in the energy sector. Decision-makers need to understand the risks of events that affect large areas simultaneously, such as a heatwave that drives air-conditioning demand across the whole country while simultaneously reducing the output of thermal power plants that rely on river cooling. A useful model must therefore capture spatial dependence in temperature, generate simulations for sites where no measurements exist, and cleanly separate the contribution of long-term climate change from stationary seasonal effects, so that a model calibrated on the present climate can be adapted to future conditions.

The mathematical heart of the approach is a decomposition. At every location, the daily average temperature is split into deterministic terms, namely a long-term trend and an annual seasonal cycle in both the mean and the variance, and a stochastic residual representing the underlying climate variability. Rather than fitting a simple linear trend, the team used LOESS regression, a non-parametric smoothing technique that fits local linear models with tri-cubic weights. This choice matters: at Toulouse-Blagnac station, for example, the fitted trend reveals an average warming of about two degrees over the observation period, but with occasional low peaks and a varying pace that a straight line could never capture. Seasonality is modeled with trigonometric polynomials, and the same structure is applied to the variance, which the data show is systematically larger in winter than in summer.

Once the deterministic components are stripped away, the residuals are assumed to form a stationary, isotropic Gaussian random field in space and time, fully characterized by its covariance function, which quantifies how strongly temperatures at two points separated by a given distance and time lag co-vary. The crucial design decision here was to reject a separable covariance model, in which the spatial and temporal structures simply multiply together. Such models are computationally convenient but fail to capture the complex interplay between space and time that meteorological variables exhibit. Instead, the researchers adopted a non-separable Gneiting-Matérn covariance model, in which an effective spatial correlation scale stretches with time lag, governed by a separability parameter estimated at 0.663 in summer and 0.883 in winter, values far from zero that confirm a strongly non-separable behavior in the data.

Fitting such a model to 41 French stations from the European Climate Assessment and Dataset, spanning 31 years and over eleven thousand days, poses a computational challenge, because the full covariance matrix of the residuals is far too large to invert. The team turned to weighted composite likelihood, an estimation strategy that replaces the full likelihood with a product of pairwise likelihoods, weighting to zero pairs of stations farther apart than about 650 kilometers. Parameter uncertainty was quantified with the Godambe information matrix, and the estimation was implemented in the Julia programming language with automatic differentiation. To extend the deterministic components from the sparse station network to any point in the country, the researchers used ordinary kriging, correcting for altitude with the International Standard Atmosphere model so that mountain sites do not distort the interpolated trends.

Validation proceeded in two stages. First, one hundred simulations were generated at the 41 fitting stations and compared with observations using a battery of indicators: pairwise correlations in space and time, conditional probabilities of joint threshold exceedance, and a novel diagnostic called the quantile exceedance ratio, which measures the proportion of the domain exceeding a given temperature quantile on any given day. The non-separable model reproduced observed correlations almost perfectly along the one-to-one line, while a purely temporal model, lacking any spatial component, produced discontinuous, patchy temperature maps that bore little resemblance to reality. A separable model performed nearly as well but systematically overestimated low correlations and misjudged covariances at non-zero time lags, underscoring the value of the non-separable formulation.

The second stage pushed the generator onto a 1,071-point grid covering France, comparing the simulations against the high-resolution E-OBS gridded observational dataset. Here the model faced its sternest test: could it produce heatwaves of realistic spatial extent and duration? The answer was largely yes. Simulated summers included short, localized hot spells, brief but geographically vast events, and multi-week episodes resembling the reference 2003 European heatwave in both intensity and duration. For heat episodes covering at least 40 to 60 percent of the country at moderate quantiles, observed values fell comfortably within the range of the simulations, which even generated events longer than any observed. The model did tend to overestimate the frequency of heat days over small areas while underestimating those covering the largest extents, a limitation traceable to the Gaussian distribution’s known lack of asymptotic dependence in the extremes. Cold events were reproduced for short durations, though persistent, large-scale cold spells remained harder to generate.

Perhaps the most striking demonstration of the generator’s flexibility is its handling of climate change. Because the warming trend is isolated as a deterministic component, the stochastic machinery calibrated on 1985 to 2015 data can be recombined with updated trends to project forward. When the team applied this procedure to the period 2016 to 2022, the 95 percent simulation intervals covered 92.21 percent of the observed mean temperatures, close to the nominal coverage. Crucially, when the heatwave-rich year 2022 was examined, most observed spatial heat episodes fell within the simulated range, and the model generated episodes longer than any recorded, meaning it properly accounts for the possibility of heatwaves exceeding anything in the historical archive.

The authors are candid about the limits of their approach and the road ahead. Improving the representation of extremes may require heavier-tailed marginal distributions, such as skew exponential power or extended Pareto distributions, or explicit extreme-value modeling, and the Gaussian dependence structure, while adequate for temperature alone, may not suit other variables or cross-correlations. The team plans to extend the generator to the whole European continent, where varying climates will introduce new non-stationarities, and to incorporate precipitation, wind speed and solar radiation, likely by weaving the temperature model into a weather-type framework driven by Markov chains. For now, though, the study delivers something genuinely practical: a fast, statistically rigorous engine that turns a handful of thermometers into a country-wide simulation of thermal climate, heatwaves and all, ready to feed the risk assessments on which grid operators, farmers and public health planners increasingly depend.

Subject of Research: A spatio-temporal stochastic weather generator for daily temperature over France

Article Title: A spatio-temporal weather generator for the temperature over France

Article References: Cognot, C., Bel, L., Métivier, D., & Parey, S. (2025). A spatio-temporal weather generator for the temperature over France. Advances in Statistical Climatology, Meteorology and Oceanography, 11(2), 203-228. https://doi.org/10.5194/ascmo-11-203-2025

Image Credits: AI Generated

DOI: 10.5194/ascmo-11-203-2025

Keywords: weather generator, stochastic simulation, temperature, France, heatwaves, spatio-temporal statistics, Gaussian process, covariance function, climate change, ECA&D, E-OBS, climate risk