Predictive analytics

Learn catchment behaviour from data

Explore where deep learning can assist inflow prediction and where physical interpretation still matters.

Application guide4 min readEdition: 25 September 2026

A different modelling task

Rainfall–runoff models estimate streamflow from meteorological and catchment information. Their outputs can support a water-resource study, but streamflow prediction is not the same as estimating water available after allocation, storage and operating rules.

What sequence models contribute

Long short-term memory networks can represent dependencies over time. The 2018 Kratzert study tested this approach across a collection of catchments. Evaluate relevance to your basin, measurement quality and intended flow regime before transferring a method.

Test what matters for supply

  • Check low-flow and drought-period performance, not only a whole-record score.
  • Examine bias in cumulative volume and the duration of dry sequences.
  • Keep future climate and land-use changes separate from performance on historical data.
  • Compare with a suitable hydrological baseline and examine uncertainty.

Carry uncertainty downstream

Run plausible inflow sequences through storage and allocation models rather than assuming a single trace is certain. Probabilistic hydrological research provides methods to examine predictive distributions. Translate these into the service outcomes of the planning study.

Define the intended hydrological use

A rainfall–runoff model can support streamflow simulation, short-term forecasting or scenario analysis. Those uses have different information boundaries. A retrospective model that uses observed rainfall throughout an event is not equivalent to an operational forecast that must use uncertain future rainfall.

For water security, low-flow periods and cumulative inflow through drought can matter more than matching flood peaks. For flood response, timing and high-flow performance may dominate. Define the purpose before choosing a training objective or announcing that one model is better.

Prepare catchment data and a fair split

Check rainfall coverage, missing gauges, evaporation inputs, flow rating changes and time alignment. Distinguish natural catchment response from regulated releases and abstractions. If observed streamflow is strongly influenced by operations that are omitted from the inputs, a model may struggle for reasons unrelated to algorithm quality.

Use chronological holdouts for prediction at a known catchment. If claiming transfer to an ungauged catchment, hold out the relevant catchment groups and ensure their observations did not influence fitting. Regional similarity, shared gauges and overlapping records can create hidden leakage.

Keep water balance and low-flow errors visible

Volume bias = [ΣQsim − ΣQobs] / ΣQobs

Use a consistent time step or integrate flows with their actual interval durations.

An overall goodness-of-fit score can be dominated by high flows. Inspect cumulative volume, seasonal behaviour, recession periods, zero-flow occurrence where relevant and the timing of recovery after drought. Define how missing or censored observations are handled.

A model can reproduce familiar historical patterns yet respond poorly to a changed climate or operating regime. Do not treat a fitted relationship as a physical guarantee outside its evidence range. Compare with an appropriate conceptual model and test sensitivity to the inputs and conditions that drive the planning result.

Evaluate uncertainty and downstream consequences

Separate uncertainty in rainfall forecasts, observations, model parameters and model structure where useful to the decision. A predictive interval should be tested for coverage and width across the important regimes, including low flow. Wide intervals can achieve high coverage while offering little operational value.

Pass plausible inflow errors or ensembles into the resource model and examine storage and shortfall consequences. A small mean flow bias can accumulate over a long dry sequence. The same hydrological error can have different consequences depending on initial storage, abstraction rules and demand.

Use the research as evidence with a defined scope

The linked LSTM and uncertainty papers are research examples with particular datasets, methods and evaluation questions. Their results motivate methods to investigate; they do not validate a local catchment model or establish a universal advantage over conceptual modelling.

Retain the training data boundary, split design, baseline comparison, regime-specific errors and model version. Connect the chosen inflow series to a closing resource-system balance and the reliability metrics. For a runnable introduction to chronological learning, use the demand-forecasting exercise.

Sources & further reading

Source findings are distinguished from editorial interpretation. Apply current local criteria and project evidence when making engineering decisions.