Uncertainty-aware improved LSTM flood forecasting in a snowmelt-dominated arid basin

An improved Long Short-Term Memory model reproduced the dominant shape, peak discharge and timing of nine flood events in China’s Ili River Basin, according to a study in Frontiers in Water that used hourly observations from June to September 2024. Under an observation-updated, one-hour-ahead forecasting protocol, the model achieved a mean Nash–Sutcliffe efficiency, or NSE, of 0.9633 and a mean coefficient of determination, R², of 0.9015 across held-out events. Its average error in predicted peak discharge was 3.90%, while its mean absolute error in peak timing was 2 hours. These results applied to one arid, snowmelt-influenced basin and a nine-event dataset, rather than to a continuous regional record or a broad set of extreme floods.

The Ili basin lies in the western Tianshan Mountains of northwestern Xinjiang and encompasses elevations from 362 to 6,274 metres. Seasonal snow and glacier storage in its high mountains connect changes in temperature to delayed runoff, while rainfall can add a faster input to the hydrograph. The study used nine independent 24-hour flood records containing hourly precipitation, air temperature and measured discharge. The source gives the minimum observed peak flow as 250.80 cubic metres per second in its data summary, although a later event table lists 320.4 cubic metres per second for Event 8; the maximum was 892.90 cubic metres per second. Only two events included recorded rainfall within the analysed window: Event 1 received 0.5 millimetres and Event 3 received 12.0 millimetres, including a maximum of 7.5 millimetres per hour.

The forecasting task required the model to predict discharge at hour t + 1 using only precipitation, temperature and antecedent discharge available at or before hour t. Each prediction was therefore updated whenever a new discharge observation became available, making this a direct, one-hour-ahead forecast rather than a recursive multi-step prediction. To reduce temporal leakage, the researchers withheld one complete flood event at a time for outer testing. Scaling, feature screening, window construction, input perturbation, model fitting and hyperparameter selection were performed exclusively on the other eight events. Predictions were scored from target hours 8 through 24, while the first seven hours supplied antecedent history for the longest candidate input window. This structure was important because a randomly divided collection of hourly windows could otherwise allow information from the same event to appear in both development and testing.

The improved workflow combined several complementary procedures. Pearson correlation screening retained temperature, precipitation and antecedent discharge as the principal predictors. Chronological sliding windows reconstructed training sequences without temporal interpolation, while low-level Gaussian noise with a standard deviation of 0.02 was added only to standardized training inputs; validation and test data were left unchanged. A Dropout layer with a probability of 0.2, L2 weight decay of 0.0001 and early stopping were used to limit excessive fitting. Nested leave-one-event-out validation selected among look-back periods of 1, 3, 5 or 7 hours and networks with 2, 4 or 8 hidden units. Five models using fixed random seeds were averaged for each outer-fold forecast. The model used one LSTM layer followed by a fully connected output layer, with Adam optimisation, mean squared error as the loss function and an initial learning rate of 0.001.

Event-level results showed consistently high hydrograph agreement, although performance was not uniform. NSE ranged from 0.9162 for Event 9 to 0.9884 for Event 7, and six events achieved NSE values of at least 0.95. R² was more variable, spanning 0.6353 to 0.9844. The two lowest R² values occurred in Events 8 and 9, whose predictions were smoother or temporally displaced relative to the measured records. Absolute relative error in event-total discharge remained between 2.46% and 4.11%, averaging 3.43%. Peak-discharge errors ranged from 2.50% to 8.01%, and peak timing differed by 0 to 7 hours. Event 1 reproduced the peak hour exactly but had the largest peak-magnitude error, whereas Event 8 had a smaller magnitude error but a seven-hour timing discrepancy. The comparison shows why a single overall efficiency score cannot describe every aspect of flood forecasting.

Compared with a baseline LSTM evaluated on the same events and forecast window, the integrated improvements produced better results on all four reported summary indicators. Mean NSE rose from 0.8718 to 0.9633. Mean event-total relative error fell from 4.19% to 3.43%, mean peak-discharge error declined from 5.48% to 3.90%, and mean absolute time-to-peak error decreased from 3.11 to 2 hours. The baseline reached an NSE of at least 0.90 in only three events, whereas all nine events exceeded 0.90 with the improved configuration. Because the two systems differed in several operations at once, however, the comparison measures the combined workflow and does not isolate the contribution of feature screening, noise injection, regularization, early stopping or parameter selection.

The event records help explain why sequence fit, peak magnitude and peak timing must be considered separately. Event 3 had the highest observed discharge at 892.9 cubic metres per second, peaked in hour 8 and included the largest rainfall total; the model achieved an R² of 0.9809, an NSE of 0.9815 and a one-hour peak-time error. Event 8 had a listed observed peak of 320.4 cubic metres per second, but its predicted crest was displaced by seven hours. Event 9 rose early, declined through the middle of the sequence and increased again later, producing the lowest NSE despite no recorded rainfall. The observed temperature range across all events was 13.4 to 34.9°C, with precipitation-free floods occurring at both warmer and cooler conditions. These contrasts show that event magnitude alone does not explain forecasting difficulty; the placement and shape of the hydrograph also matter.

The study describes a temperature–runoff lag averaging 11.67 hours across the events, a pattern consistent with cumulative thermal forcing and delayed meltwater accumulation in the basin. In the two rainfall-observed events, temperature–runoff lags were 12 and 8 hours, respectively, and their temperature correlations were lower. The authors interpret rainfall as a short-duration recharge capable of modifying the magnitude or timing of a melt-influenced flood, while the analysis does not quantitatively partition discharge among snowmelt, glacier melt, rainfall or other sources. This distinction matters because the model used temperature, precipitation and discharge, not direct measurements of snow storage, melt rate, radiation or runoff source. The lag and meteorological comparisons are therefore descriptive associations rather than proof of an exclusive runoff mechanism.

A separate sensitivity analysis used Latin hypercube sampling and Monte Carlo experiments to vary selected hyperparameters, including hidden units, learning rate and input time step. Under 1,000 random parameter combinations, aggregate metrics remained within the study’s cited engineering thresholds, although the output statistics showed measurable dispersion. Hidden neurons, learning rate and time step were identified as the most influential parameters in the tested configuration. The framework was therefore interpreted as comparatively stable within the sampled ranges, not as having a universally transferable optimum. This analysis concerns the sensitivity of deterministic forecasts, not calibrated prediction uncertainty: it did not generate event-wise discharge distributions, time-varying confidence intervals, coverage probabilities or exceedance risks. Averaging five random-seed forecasts also did not constitute a probabilistic ensemble with validated interval calibration.

The evidence supports the improved model for the specific task tested, while several limits remain. The dataset contains only nine events from one basin, reconstructed windows do not create additional independent floods, and only two events contain measured rainfall. The model was not compared with persistence, in which the latest observed discharge serves as the next-hour forecast, so its gains cannot yet be separated from the skill of a simple short-term reference. More broadly, the network does not explicitly simulate snow accumulation, energy exchange, runoff generation or source partitioning. The authors identify larger event archives, additional basins, persistence benchmarks, sequential component tests and physical predictors such as snow-water equivalent, snow-covered area, freezing-level elevation and radiation as logical next steps. Probabilistic methods, including deep ensembles, bootstrap resampling or Monte Carlo Dropout, would require separate tests of interval coverage, width and reliability before they could support operational uncertainty estimates.

Subject of Research: Event-scale LSTM flood forecasting and hyperparameter sensitivity in a snowmelt-dominated arid basin

Article Title: Uncertainty-aware improved LSTM flood forecasting in a snowmelt-dominated arid basin

Article References: Xiao, Y., Bi, Y., Zhang, Z., & Tian, D. (2026). Uncertainty-aware improved LSTM flood forecasting in a snowmelt-dominated arid basin. Frontiers in Water, 8, Article 1916396. https://doi.org/10.3389/frwa.2026.1916396

Image Credits: AI Generated

DOI: 10.3389/frwa.2026.1916396

Keywords: LSTM flood forecasting, Ili River Basin, snowmelt runoff, event-based validation, peak discharge, hyperparameter sensitivity, hydrological forecasting, arid basins, machine learning, Xinjiang, Uncertainty-aware, improved