Sîrbu et al. (2026) Short-Term Streamflow Forecasting for River Management, Using ARIMA Models and Recurrent Neural Networks

Identification

• Journal: Hydrology
• Year: 2026
• Date: 2026-03-04
• Authors: Nicolai Sîrbu, Andrei-Mihai Rugină
• DOI: 10.3390/hydrology13030082

Research Groups

• Department of Hydrotechnical Engineering, Technical University of Civil Engineering, Bucharest, Romania

Short Summary

This study conducts a controlled comparison between SARIMA and stacked LSTM models for 7-day-ahead daily water-depth forecasting using synthetic hydrographs across normal, drought, and flood regimes, concluding that both approaches exhibit statistically comparable median performance.

Objective

• To assess whether increased model complexity (LSTM) leads to systematically improved short-term forecasting skill compared to SARIMA under univariate settings.
• To explore regime-dependent performance patterns across contrasting hydrological conditions using repeated rolling forecasts.
• To highlight the role of evaluation design and metric selection in shaping comparative conclusions.

Study Configuration

• Spatial Scale: Synthetic hydrological time series representing generalized river water-depth dynamics, not tied to a specific geographical location.
• Temporal Scale: Daily time step; 7-day forecast horizon; evaluation over a two-year period (104 weekly forecasts) from a five-year simulated dataset.

Methodology and Data

• Models used: Autoregressive Integrated Moving Average (ARIMA), Seasonal Autoregressive Integrated Moving Average (SARIMA), Long Short-Term Memory (LSTM) recurrent neural networks (stacked architecture).
• Data sources: Stochastically generated synthetic daily water-depth hydrographs, designed to represent normal, drought, and flood hydrological regimes. Data was preprocessed using a 15-day trailing moving average (MA15).

Main Results

• SARIMA and LSTM models demonstrated statistically comparable median performance for 7-day-ahead water-depth forecasting under normal and drought conditions, with Wilcoxon signed-rank test p-values exceeding 0.05.
• For the single flood event represented, LSTM models (Global Forecast Skill Index (GFSI) 2.652–2.685) showed lower GFSI peaks than SARIMA (GFSI 4.002), but this finding is illustrative due to the limited sample size (n=1).
• LSTM models with longer historical input windows (60–120 days) produced smoother and more stable predictions, while shorter windows (30 days) allowed for faster responses to local variations but exhibited greater instability.
• Both SARIMA and LSTM models tended to underestimate the maximum magnitude of extreme flood events, with prediction error increasing with event severity.
• A 15-day trailing moving average (MA15) was selected as the optimal smoothing method for the synthetic hydrographs, achieving the lowest total score across noise variance, roughness, and mean absolute deviation.
• The SARIMA(2,1,4)(1,0,1)m=37 model was identified as the most suitable classical statistical model based on the Akaike Information Criterion (AIC).

Contributions

• Provides a controlled and transparent comparison between classical statistical (SARIMA) and deep learning (LSTM) models for short-term water-depth forecasting using synthetic daily hydrographs.
• Employs a systematic experimental framework across controlled normal, drought, and flood regimes, mitigating confounding factors common in real-world data.
• Utilizes a robust rolling-origin evaluation strategy to generate multiple overlapping 7-day predictions, reducing bias from short validation windows.
• Introduces and applies the Global Forecast Skill Index (GFSI) as a comprehensive, integrated metric for a balanced assessment of forecast skill across diverse hydrological regimes.
• Highlights the importance of rigorous evaluation design and metric selection in comparative forecasting studies, positioning classical and deep learning models as complementary tools.

Funding

• Internal grant program ARUT of the Technical University of Civil Engineering Bucharest, Romania (Contract No. 1414/07.02.2025, Identifier 11).

Citation

@article{Sîrbu2026ShortTerm,
  author = {Sîrbu, Nicolai and Rugină, Andrei-Mihai},
  title = {Short-Term Streamflow Forecasting for River Management, Using ARIMA Models and Recurrent Neural Networks},
  journal = {Hydrology},
  year = {2026},
  doi = {10.3390/hydrology13030082},
  url = {https://doi.org/10.3390/hydrology13030082}
}