The increasing rate of dam failures in the United States, particularly since the 1970s, is a growing concern. This increase is attributed to a combination of aging infrastructure and more frequent extreme rainfall events due to climate change. The US National Inventory of Dams lists over 92,000 dams, many of which are aging and poorly maintained. Conventional dam design criteria focus on handling 'Probable Maximum Floods' (PMFs), which are based on extreme, single-event rainfall scenarios. However, recent dam failures, including Oroville (2017) and Michigan (2020), highlight that failures can occur under less extreme conditions. These failures often involve a sequence of rainfall events leading to high antecedent soil moisture and full reservoirs, exceeding capacity even with moderate final rainfall events. This study investigates whether sequences of high rainfall events, single extreme rainfall events, or a combination are the primary triggers of recent dam failures, how the likelihood of these factors has evolved, and what this implies for the changing risk of dam failure relative to historical design criteria.

Numerous studies have documented increases in the intensity and frequency of extreme precipitation events in the US. Climate projections suggest this trend will continue with global warming. However, few studies have examined the sequence of rainfall events preceding a large rainfall event at a continental scale. Persistent atmospheric circulation patterns can lead to recurrent storms, culminating in extreme flood peaks and volumes. Existing research also highlights the insufficiency of current hydrologic infrastructure design standards given the increase in extreme meteorological events, but lacks comprehensive analysis on a continental or global scale due to complexities in modeling hydroclimatic conditions and human behavior (reservoir operation).

The study analyzed data for 552 dam failures attributed to hydrologic factors (excluding snow-related failures) since 2000. The 20CRv3-ERA5 reanalysis rainfall dataset (1901–2021) was used. For each dam failure, the return period of the maximum 1-day rainfall (A*) preceding the failure was assessed using a non-stationary Generalized Extreme Value (GEV) model with linear temporal trends in scale and location parameters. Similarly, the return period of cumulative rainfall (K) for the preceding 5 and 30 days was evaluated using a non-stationary Gamma distribution. The joint return periods for (A*, K5*) and (A*, K30*) were determined using bivariate copulas. The analysis compared exceedance probabilities at the time of failure versus the time of dam construction (available for 497 dams). To examine recent trends, the 1979–2021 CPC-CONUS land precipitation dataset was used to analyze trends in 10- and 100-year return period events for daily maximum rainfall (A) and preceding 5- and 30-day rainfall (K5 and K30). The study focused on large dams (height > 12.2 m or storage capacity > 1.23 million m³) classified as high or significant hazard dams. The Mann-Kendall test was employed to assess trends in annual dam incident numbers and dependence between A and K. GEV distributions were tested under four assumptions (no temporal trend, linear trend in μA, linear trend in σA, and linear trends in both μA and σA), choosing the model with the lowest Bayesian Information Criterion (BIC). The k-day antecedent precipitation was modeled using a 2-parameter gamma distribution. The dependence structure between A and K was modeled using copulas, considering the temporal variation in the correlation between the cumulative distribution functions of A and K.

The study found that the maximum 1-day rainfall associated with dam failures (A*) was often a moderate event (return period under 10 years), much less extreme than the PMF used in design criteria. However, the joint occurrence of A* and preceding multi-day rainfall (K5* or K30*) was far rarer. Clusters of dam failures were observed in specific regions, such as South Carolina (October 2015) and New Jersey/New York (Hurricane Irene, 2011), characterized by persistent heavy rainfall. Comparing exceedance probabilities at the time of failure versus the year of dam construction did not reveal a significant widespread change, suggesting that the compound events, while rare at failure, were also rare at design. However, regional variations were present in exceedance probabilities for A*, K5*, and K30*. Analysis of 1979–2021 CPC-CONUS data revealed widespread statistically significant trends in the joint events of 100-year A and 10-year K for both 5 and 30 days. These joint events show an increased occurrences in areas with a significant number of large, high-hazard dams. Thus, despite the increase in the frequency and severity of compound precipitation events, the joint occurrences associated with dam failures often had very high return periods even accounting for recent trends.

The findings emphasize the importance of considering antecedent conditions and compound precipitation events in dam design, as many recent failures occurred under more moderate conditions than anticipated by traditional criteria. The increase in the frequency and severity of compound events highlights the urgent need to improve prediction of these conditions, acknowledging the role of persistent atmospheric circulation patterns and climate variations. Further research is needed to improve the predictability of these mechanisms. Assessing reservoir storage at the time of events is also challenging, requiring more comprehensive data on reservoir operations and conditions. The high percentage of unrated or unknown dam conditions in the NID data limits a comprehensive assessment of additional failure factors.

This study demonstrates that compound rainfall events are a significant factor in recent US dam failures. Traditional design criteria need revision to incorporate the risk of these compound events. Future research should focus on improving the prediction of compound precipitation events, incorporating reservoir operational strategies into risk assessments, and developing a portfolio risk analysis of all US dams considering climate, fragility, and operational risks.

The study's reliance on precipitation as a proxy for soil moisture is a limitation. Access to long-term, comprehensive soil moisture data is needed for a more robust assessment. The availability of dam construction dates for only 497 of the 552 dams limits the temporal analysis. The study excludes snowmelt-related failures due to data limitations. The study also does not fully incorporate the role of reservoir operation in dam failure risk.