Decomposing Non-Stationary Time Series for Hydrological Forecasting

25 concepts24 papers1 proposal

Hydrological systems are notoriously difficult to predict because they are governed by highly non-linear and non-stationary processes. In a stationary system, statistical properties like the mean, variance, and volatility remain constant over time. In hydrology, however, these properties are constantly shifting due to seasonal cycles, climate change, and human interventions such as dam regulations (Li et al.).

Furthermore, the relationship between environmental inputs and outputs is highly state-dependent. As noted by Kan et al., a small fluctuation in daily rainfall can yield vastly different runoff results depending on the catchment's underlying state, such as soil moisture saturation. When an intense rainfall event hits an already saturated basin, it triggers a chaotic, "flashy" hydrograph that standard sequence models struggle to predict (Nearing et al.). Traditional data-driven models often fail in these contexts because they assume that historical correlations will remain constant over time (Yuan et al.).

The Intuition of Signal Decomposition

To overcome non-stationarity, a prominent research direction leverages a "divide-and-conquer" strategy: breaking down a single, chaotic time series into multiple simpler, more stable sub-signals. This is primarily achieved through signal decomposition techniques such as Ensemble Empirical Mode Decomposition (EEMD) or Variational Mode Decomposition (VMD).

Why does this work? A raw streamflow or runoff time series is actually a composite of multiple overlapping physical phenomena. It contains low-frequency, highly predictable components (like annual seasonal cycles or long-term climate trends) mixed with high-frequency, volatile components (like sudden storm events or daily industrial dam releases).

By applying a method like VMD, the authors isolate these distinct frequency bands into separate sub-signals, known as Intrinsic Mode Functions (IMFs) (Tang et al.; Chen et al.). Each individual IMF exhibits significantly lower non-stationarity and variance than the original raw data, making them far easier for machine learning models to learn and predict.

Hybrid Architectures for Temporal Modeling

Once the signal is decomposed, a hybrid architecture leverages specialized neural networks to forecast each sub-signal independently before recombining them into a final prediction.

Raw Non-Stationary Signal ───► [ Signal Decomposition ] ───► IMF 1 (Low Freq)  ───► [ LSTM ] ───┐
                                                       ───► IMF 2 (Mid Freq)  ───► [ LSTM ] ───┼─► Combined Forecast
                                                       ───► IMF 3 (High Freq) ───► [ CNN  ] ───┘

Long Short-Term Memory (LSTM) networks serve as the backbone for this temporal modeling step. Because the low-frequency sub-signals capture long-term dependencies, LSTMs can easily track their smooth transitions over time (Yuan et al.). However, high-frequency sub-signals contain sharp spikes and noise. To handle this, researchers often integrate Convolutional Neural Networks (CNNs) or attention mechanisms.

For instance, Ghimire et al. utilize CNNs to extract localized features from the data before passing them to an LSTM for temporal analysis. Similarly, Long et al. introduce frequency-aware attention mechanisms to help the network distinguish between predictable environmental patterns and random noise.

This decomposition-hybrid paradigm is highly versatile, proving successful not only in streamflow and runoff prediction but also in modeling daily evaporation rates (Fu et al.) and even non-stationary maritime traffic flow (Li et al.; Wang et al.).

Adapting to Spatial and Environmental Extremes

While decomposing a single station's time series provides significant accuracy gains, real-world hydrological systems are also spatially interconnected. Runoff from an upstream basin directly impacts downstream river stages hours or days later.

To capture these broader spatial dynamics under global climate change, researchers are moving toward adaptive, graph-based frameworks. For example, Zhang et al. developed an interactive local-global model where independent neural networks simulate individual stations, and a hydrological graph iteratively adjusts their predictions based on spatial relationships. By combining localized temporal modeling with global spatial constraints, these frameworks maintain robust predictive power even when environmental baselines shift permanently.

Key Foundational Work

An improved hybrid data-driven model and its application in daily rainfall-runoff simulation (2016) — contributed 1 concept elements to this direction.
Dam Inflow Forecasting for Short Term Flood Based on Neural Networks in Nakdong River Basin (2004) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_non_stationary_time_series_forecasting_16b91d1b,
  title = {non-stationary time series forecasting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-stationary-time-series-forecasting/rdr_16b91d1be89f4ecd9915bcaa48b79624},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_non_stationary_time_series_forecasting_16b91d1b_v1,
  title = {non-stationary time series forecasting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-stationary-time-series-forecasting/rdr_16b91d1be89f4ecd9915bcaa48b79624/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake