Memory Beats Speed: LSTM Outperforms Rival Neural Networks in Two-Week River Flow Forecasts

Forecasting the flow of a river two weeks into the future, hour by hour, has long been one of hydrology’s most stubborn challenges. Now a new study published in the journal Water Resources Management suggests that the secret to cracking it lies not in inventing exotic new artificial intelligence architectures, but in understanding precisely why some neural networks remember and others forget. Researchers at Federation University Australia, working with Snowy Hydro Limited, put three classic deep learning models through a grueling head-to-head comparison and found that the long short-term memory network, or LSTM, delivered forecasts of remarkable accuracy over a 336-hour prediction horizon, while its two rivals collapsed into systematic error.

The study, led by Sandhya Eswara of the Future Regions Research Centre, focused on the Snowy River above Guthega in New South Wales, Australia. The catchment covers just 67.9 square kilometers but packs an extraordinary amount of hydrological complexity into that area. It sits in the headwaters of the Snowy River, beginning just below Mount Kosciuszko, Australia’s highest peak, and lies entirely above the seasonal winter snowline, making it one of the few catchments in the country with that characteristic. Average annual precipitation at Guthega measures 1477 millimeters, with a seasonal peak between June and September. That combination of snow accumulation, snowmelt, rainfall-driven runoff and seasonal transitions means the dataset spans a broad spectrum of hydrological behaviors rather than a single homogeneous climate, which is exactly what a rigorous test of forecasting models demands.

The team trained and tested three architectures with progressively different temporal capabilities: a feed-forward artificial neural network, or FFANN, which processes each input independently with no memory of what came before; a conventional recurrent neural network, or RNN, which feeds its hidden state back into itself to retain recent information; and the LSTM, which adds memory cells and gating mechanisms designed to preserve information over long sequences. The researchers were explicit that the goal was not to propose a new architecture or crown a universal champion, but to conduct a controlled comparison under identical conditions. More fashionable options such as gated recurrent units and transformer models were deliberately excluded, partly because transformers require substantially larger datasets and computational resources than catchment-scale hydrological applications can typically provide.

The input data were as rich as the setting. Hourly observations of streamflow, precipitation in the form of rain or snow, air temperature, wind speed, snow depth, snow density and snow water content all served as predictor variables, with streamflow data collected upstream of Guthega Pondage and meteorological data gathered adjacent to Lake Guthega. Missing values, which constituted only a tiny fraction of the dataset and appeared as isolated observations rather than long gaps, were filled using the mean of the ten nearest non-missing values. All variables were then rescaled to the range between zero and one using min-max normalization, a step that prevents variables with large numerical ranges from dominating the learning process. Spearman correlation analysis confirmed that every predictor had adequate correlation with streamflow, with the snow-based variables showing the strongest relationships and air temperature displaying a negative correlation, reflecting the fact that streamflow rises in winter and during precipitation events when temperatures fall.

Model architecture and input configuration were settled through careful validation rather than guesswork. After extensive experimentation, a network of three hidden layers containing nine, three and one neurons produced the most consistent performance across validation folds. A sensitivity analysis over input lag sizes showed that smaller lags starved the models of historical context while larger lags added complexity without improving accuracy, so a lag of ten time steps was adopted for all experiments. Hyperparameters were tuned through an iterative manual approach, varying one parameter at a time, a strategy chosen for computational efficiency and applied identically to all three models to keep the comparison fair. Each model was trained three times to confirm consistency, and a five-fold time series cross-validation scheme, which preserves the chronological order of observations rather than shuffling them, confirmed that none of the models was overfitting its training data.

The first forecasting test, known as open loop or one-step-ahead prediction, was almost too easy. Given the true weather and streamflow values from the previous ten hours, all three models predicted the next hour’s flow with high coefficients of determination near one and root mean squared errors below one. The LSTM edged out its rivals slightly and, notably, captured peak flow events more effectively than the FFANN and RNN, which tended to over-predict the highest flows. But as the authors point out, forecasting a single hour ahead is not particularly difficult, and its practical applications are limited. The real test came next.

That test was closed loop, or recursive, forecasting, in which each predicted streamflow value is fed back into the model as an input for the next prediction, hour after hour, for two full weeks. This approach is attractive because a single trained model can generate forecasts over very long lead times with modest computational effort, but it carries a dangerous flaw: any error in one prediction contaminates the inputs for the next, and errors can accumulate exponentially. The results were stark. The FFANN produced a root mean squared error of 7.55 and a coefficient of determination of negative 0.87, while the RNN fared little better at 6.41 and negative 0.35. Negative R-squared values mean the models performed worse than simply guessing the average observed flow, and large negative bias values showed both were heavily under-predicting streamflow as the horizon stretched out.

The LSTM, by contrast, held its accuracy across the entire two-week window, achieving a coefficient of determination of 0.87 and a root mean squared error of just 1.99, with near-zero bias and Nash-Sutcliffe efficiency values close to one. Its cumulative forecast volumes closely tracked the observed totals, while the FFANN and RNN drifted into consistent skew. The LSTM even reproduced roughly 80 percent of the observed peak flow magnitudes, a feat the authors attribute to its gated memory mechanism, which preserves meaningful hydrological information while filtering out accumulated prediction noise. In an alpine catchment where streamflow depends not just on immediate rainfall but on snow accumulation, snowmelt and the gradual release of water from catchment storage, that ability to integrate information over extended periods is decisive. The conventional RNN, vulnerable to the vanishing gradient problem during long sequence learning, simply cannot hold onto those long-term dependencies.

The study also offers a subtle lesson about inputs. Wind speed, a variable seldom investigated in streamflow forecasting, proved valuable alongside the more obvious hydrological drivers, and the analysis of the forecast period showed sharp streamflow peaks following strong precipitation events, with streamflow responding to rainfall after a characteristic lag. The authors acknowledge that measurement errors and observational noise remain unavoidable, and that results should be interpreted in light of data quality, but they argue the work represents a significant advance: to their knowledge, no previous study has achieved accurate hourly streamflow forecasts two weeks ahead. The operational implications are considerable, since reliable two-week forecasts would give reservoir operators more lead time to optimize storage releases, improve hydropower scheduling and issue earlier flood warnings. The team’s next ambition is to push the forecast window beyond one month while keeping errors within acceptable bounds, a goal that seems less fanciful now that the value of long-term memory has been demonstrated so clearly.

Subject of Research: Comparative performance of deep learning architectures for multivariate short-term streamflow forecasting

Article Title: Performance Analysis of Deep Learning Algorithms for Short Term, Multi-Variate Streamflow Forecasting

Article References: Eswara, S., Barton, A., Choudhury, T., & Chubb, T. (2026). Performance Analysis of Deep Learning Algorithms for Short Term, Multi-Variate Streamflow Forecasting. Water Resources Management, 40(13), Article 538. https://doi.org/10.1007/s11269-026-04902-x

Image Credits: AI Generated

DOI: 10.1007/s11269-026-04902-x

Keywords: streamflow forecasting, deep learning, LSTM, recurrent neural networks, feed-forward neural networks, hydrology, time series forecasting, Snowy River, snowmelt, recursive forecasting, water resources management, cross-validation