AI-Free Weather Model Learns France’s Rain Regimes From Ten Stations Alone

What if a handful of rain gauges scattered across France could teach a computer everything it needs to know about the country’s large-scale weather patterns? That is the striking promise of a new stochastic weather generator developed by Emmanuel Gobet and David Métivier of École Polytechnique and INRAE, together with Sylvie Parey of EDF R&D, published in the journal Advances in Statistical Climatology, Meteorology and Oceanography. Their model, called a seasonal hierarchical hidden Markov model, or SHHMM, learns the daily rhythm of French rainfall from just ten weather stations with complete records spanning 1956 to 2019 – and it does so without feeding in a single satellite image, pressure map, or circulation index. The hidden states it discovers turn out to correspond closely to the well-known North Atlantic weather regimes that climatologists usually need heavy machinery to identify, yet the whole thing is light enough to simulate a 64-year sequence of French rain in about a hundredth of a second.

The motivation comes from a practical and increasingly urgent problem: industries that operate many installations spread across a territory need to know how often those installations will be hit by the same extreme weather event at the same time. Prolonged, spatially extensive droughts are a particular worry for the French electricity system, because nuclear plants depend on river water for cooling, and a nationwide dry spell can throttle several reactors simultaneously. Parliamentary missions in France have explicitly called for quantifying such hydro-stress on nuclear generation. Agriculture faces analogous risks when entire growing regions share the same drought. Yet historical observations offer only one realization of climate variability, and climate models, though valuable, are too computationally expensive to run in the thousands of variants needed for robust statistics of rare events. Stochastic weather generators fill this gap by churning out thousands of statistically plausible weather histories at negligible cost.

Mathematically, the new model builds on hidden Markov models, a statistical framework dating back to precipitation applications in the early 1990s. The central idea is elegant: imagine an invisible variable, one per day, that represents the overall weather regime over France. This hidden state evolves as a Markov chain, meaning that tomorrow’s regime depends only on today’s. Crucially, the model is seasonal – its transition probabilities are 366-periodic functions of the calendar day, parameterized by trigonometric polynomials so that they evolve smoothly through the year rather than jumping between months. Given the hidden regime, rain occurrence at each station follows a simple Bernoulli probability, and the model enforces conditional independence between stations. That constraint is deliberate: it forces every scrap of spatial correlation to be absorbed by the hidden states, which is exactly what makes them interpretable.

The authors added two refinements that prove decisive for performance. First, they introduced a one-day local memory: today’s wet or dry status at a station depends not only on the hidden regime but also on yesterday’s status at the same station. Without this autoregressive touch, the model systematically produced dry and wet spells that were too short, overestimating brief episodes and underestimating the long droughts that matter most for risk assessment. Second, after generating rain occurrences, the model attaches rainfall amounts using mixtures of two exponential distributions whose parameters also vary seasonally and with the weather regime, linked across stations by a Gaussian copula whose correlation strength itself depends on the regime. Model selection using the integrated complete-data likelihood landed on four hidden states, one day of memory, and first-degree seasonal polynomials – a parsimonious configuration of roughly 276 parameters instead of the more than 33,000 a day-by-day approach would demand.

The interpretability results are the paper’s showpiece. The four learned states form recognizable regional patterns: one in which it rains nearly everywhere in France, one that is wet in the north and drier in the south, its mirror image, and a fourth in which rain is scarce nationwide. When the authors averaged winter sea-level pressure anomalies from the ERA5 reanalysis conditional on each hidden state, coherent large-scale structures emerged over the whole North Atlantic – from a model trained exclusively on ten French rain gauges. State four resembles a blocking pattern, and the set of four regimes lines up closely with the classic Euro-Atlantic regimes – the positive and negative phases of the North Atlantic Oscillation, the Atlantic Ridge, and blocking – and with weather types previously derived from French rainfall data, though with anomalies centered more tightly on France, as expected from the training data.

The temporal dynamics of the regimes also ring true. The dry regime is the stickiest, with the highest probability of persisting from one day to the next, which explains the long nationwide droughts it can generate. It almost never flips directly into the wettest regime; an intermediate state intervenes, mirroring the intuition that a countrywide dry day is rarely followed by a countrywide soaking. Decoding the hidden states across history with the Viterbi algorithm, the model even detects famous events: a 27-day stretch of the dry regime beginning in early June 1976, matching the notorious drought of that summer, and a 15-day dry sequence during the August 2003 heatwave.

Validation tests show the generator reproducing observed statistics remarkably well. At nearly all stations, the historical distributions of dry and wet spell lengths fall within the envelope of five thousand simulated 64-year runs. Spatial correlations between stations, which range from near zero to about 0.5, are well captured, confirming the conditional independence assumption a posteriori. Most impressively, the model nails the distribution of large-scale droughts – episodes during which at most twenty percent of stations are wet – across all four seasons, and it can even produce dry spells longer than any yet observed. A comparison against the widely used WGEN-type generator, which links stations through pairwise Gaussian copulas and fourth-order local Markov chains, was revealing: the WGEN approach overestimated short dry episodes and underestimated long ones, because pairwise correlations simply cannot encode temporally persistent, territory-wide dry states. The hidden-state model scales as the square of the number of regimes rather than the square of the number of stations, giving it a decisive computational edge.

The authors then turned the generator loose on climate change. Training it on downscaled projections from the French DRIAS service, they compared climate models against the statistical envelope of natural variability during a reference period, a more meaningful benchmark than comparing against a single historical trajectory. Under the high-emission RCP8.5 scenario with the IPSL projection, the fitted transition matrices shift: the rainy regimes become more self-persistent in summer, projecting longer stretches of heavy rain – a pattern consistent with known biases of the European regional model used, which the generator makes easy to expose. Training the generator on future scenarios allows researchers to resample thousands of synthetic climates and extract return periods of compound extremes that no ensemble of physical models could afford to simulate directly.

Limitations remain candidly acknowledged. The conditional independence assumption requires carefully chosen, well-separated stations – a pair like central Paris and Orly, thirteen kilometers apart, would break it – and the model generates weather only at training sites, not high-resolution maps. Coastal stations such as La Hague and Chassiron, with their idiosyncratic maritime climates, fit less well, and the Gaussian copula underestimates joint extremes at the closest pairs, pointing toward heavier-tailed alternatives like the Student copula. Still, the package, released as an open-source Julia library with a fully reproducible tutorial, opens a practical path toward stress-testing entire energy systems and agricultural regions against synthetic weathers – including future ones – and toward estimating, at last, the true odds of the drought that half of France dreads sharing at once.

Subject of Research: An interpretable seasonal hidden Markov model for stochastic multisite precipitation generation and climate risk assessment in France

Article Title: Interpretable seasonal multisite hidden Markov model for stochastic rain generation in France

Article References: Gobet, E., Métivier, D., & Parey, S. (2025). Interpretable seasonal multisite hidden Markov model for stochastic rain generation in France. Advances in Statistical Climatology, Meteorology and Oceanography, 11(2), 159-201. https://doi.org/10.5194/ascmo-11-159-2025

Image Credits: AI Generated

DOI: 10.5194/ascmo-11-159-2025

Keywords: stochastic weather generator, hidden Markov model, precipitation modeling, France, drought, extreme events, weather regimes, climate change scenarios, energy systems, North Atlantic Oscillation, rainfall statistics, climate risk