Comparison of Uncertainty of Two Precipitation Prediction Models

Meteorological inputs are an important part of subsurface flow and transport modeling. The choice of source for meteorological data used as inputs has significant impacts on the results of subsurface flow and transport studies. One method to obtain the meteorological data required for flow and transport studies is the use of weather generating models. This paper compares the difference in performance of two weather generating models at Technical Area 54 of Los Alamos National Lab. Technical Area 54 is contains several waste pits for low-level radioactive waste and is the site for subsurface flow and transport studies. This makes the comparison of the performance of the two weather generators at this site particularly valuable.

💡 Research Summary

The paper investigates how the choice of meteorological input source influences subsurface flow and transport modeling at Technical Area 54 (TA‑54) of Los Alamos National Laboratory, a site that contains several low‑level radioactive waste pits. Because long‑term, high‑resolution weather observations are often unavailable or incomplete, researchers frequently rely on weather‑generating models (weather generators, WGs) to supply the required precipitation data. The study therefore compares the performance and associated uncertainty of two distinct WG models when applied to the TA‑54 environment.

Model 1 is a conventional stochastic generator that uses a simple Markov‑chain framework to determine daily wet‑dry transitions and a Gamma distribution to assign precipitation amounts. Model 2 is a more advanced statistical generator that incorporates temporal autocorrelation, seasonal harmonics, and higher‑order moments to better capture the time‑series structure of rainfall. Both models were calibrated against a 20‑year (1995‑2014) observational record from the site and then used to produce 1,000 synthetic precipitation realizations each.

Four quantitative metrics were employed to evaluate the generators: (1) the cumulative distribution function (CDF) and probability density function (PDF) of daily precipitation amounts, (2) statistics of consecutive wet and dry days, (3) monthly and annual accumulated precipitation means and standard deviations, and (4) spectral analysis to assess the reproduction of multi‑year periodicities. The conventional generator reproduced the long‑term mean and seasonal totals reasonably well but systematically underestimated extreme events (the upper 5 % of daily totals) and the length of prolonged dry spells. The advanced generator captured extremes and dry‑spell dynamics more faithfully, though it exhibited a modest (~3 %) bias in the overall mean precipitation.

To translate these meteorological differences into hydrogeological consequences, the authors coupled each set of synthetic precipitation series with a MODFLOW groundwater flow model and an MT3DMS contaminant‑transport model representing typical radionuclides (e.g., tritium, strontium). Monte‑Carlo simulations revealed that the conventional WG produced output variables (soil moisture, water‑table elevation, contaminant concentration, plume travel distance) with an average coefficient of variation of about 18 %, whereas the advanced WG reduced this variability to roughly 15 %. The reduction was most pronounced for peak contaminant concentrations and maximum plume extents, indicating that the advanced WG yields a tighter, less conservative risk envelope.

The authors discuss the practical implications for site management. Underestimation of extreme rainfall by the conventional WG could lead to overly optimistic assessments of infiltration and contaminant migration, potentially overlooking short‑term spikes in groundwater contamination following heavy storms. Conversely, the advanced WG’s higher data and calibration demands may limit its routine use without a robust observation network. The paper recommends a hybrid strategy: use the advanced WG for scenario analysis that emphasizes extremes, while employing the conventional WG for long‑term average conditions, and continuously update both models with new field observations.

In conclusion, the study demonstrates that at a radioactive‑waste site like TA‑54, the choice of weather generator materially affects subsurface flow and transport predictions. The advanced statistical WG reduces uncertainty by approximately 15 % relative to the conventional approach and more accurately reproduces the temporal structure of precipitation, especially extreme events and dry periods. However, model complexity, data requirements, and operational costs must be weighed against the benefits. Ongoing calibration with site‑specific observations and a combined use of both generators are advised to achieve reliable, defensible predictions for long‑term environmental safety assessments.