📄 earth_science

From forecast skill to planning value: multi-horizon data-driven hydrological forecasting for sustainable management of the Jucar River Basin District

This study presents a data-driven multi-horizon forecasting pipeline for the Jucar River Basin that successfully generates monthly predictions of aquifer levels, reservoir storage, and river discharge up to 48 months ahead by separating endogenous memory from exogenous climate and demand drivers, demonstrating high skill for groundwater and reservoirs while identifying systematic bias as the primary limitation for long-range river forecasts under various future scenarios.

Original authors: Alejandro Garcia Monteagudo, Marina Arnaldos Orts, Miguel Angel Pardo Picazo

Published 2026-07-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alejandro Garcia Monteagudo, Marina Arnaldos Orts, Miguel Angel Pardo Picazo

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict the future of a river, a lake, or an underground water tank. This is the job of hydrologists, scientists who study how water moves through our world. But predicting water isn't like guessing the weather for tomorrow; it's more like trying to guess how full a bathtub will be next month, knowing that someone is constantly turning the faucet on and off, and that the tub has a slow, leaky drain. In places like the Mediterranean, where summers are scorching and rain is unpredictable, getting these predictions right is a matter of survival for farmers, cities, and ecosystems. To do this, scientists use "models"—mathematical recipes that take past data (like how much rain fell last month) and try to guess what happens next. The big question is: how far into the future can we trust these guesses before they start to drift off course?

This paper tackles that exact problem for the Jucar River Basin in eastern Spain, a region where water is tightly managed by dams and pumps. The researchers built a smart, data-driven system to forecast three different types of water storage: underground aquifers (like giant sponges under the ground), reservoirs (man-made lakes), and rivers. They didn't just look at one type of water; they realized that a river, a lake, and a sponge all behave differently. A river reacts fast to rain, a lake changes slowly, and a sponge takes a long time to soak up or dry out. The team tested their "crystal ball" against a very simple trick: just assuming next month will look exactly like this month. They found that while this simple trick works surprisingly well for some things, their fancy new system is needed for others, but only up to a certain point in time.

The study reveals a clear "personality" for each part of the water system. For the underground aquifers, the water is so stubborn and slow-moving that the simplest guess (assuming things stay the same) is almost perfect. The researchers' complex model did slightly better, but the aquifer is so predictable that even a year or four years from now, the forecast is still very reliable. It's like predicting the temperature of a giant block of ice; it won't change much no matter what happens outside.

However, rivers are a different story. They are like a hyperactive dog that reacts instantly to every treat (rain) or command (water usage). The researchers found that their model could predict river flow accurately for about 12 months. But if you try to guess the river's flow 24 or 48 months ahead, the model starts to get confused. It begins to drift, slowly overestimating how much water will be there. This happens because the math used to build the model has a tiny bias—it tends to guess "a little bit more" every single month, and over years, those tiny guesses add up to a huge error. The model can't see the future perfectly; it just gets better at guessing the next step than the step after that.

The team also ran "what-if" scenarios, imagining a future where the climate gets hotter and people use more or less water. They found that for rivers, these different futures start to look very different from each other within that 12-month window. But for the underground aquifers and the big reservoirs, the different futures look almost identical for a long time. The water in the ground and the big lakes is so buffered by their size and the way they are managed that the immediate future doesn't change much, regardless of whether the weather gets slightly hotter or cooler.

Crucially, the authors warn that when they show a spread of different future scenarios, it doesn't mean they have a perfect probability of what will happen. It just means "if the weather acts like this, the water will look like this." It's a tool for stress-testing plans, not a crystal ball that tells you exactly what will occur. The study concludes that for sustainable water management, we need to know exactly how far we can trust our forecasts: 48 months for underground water, 36 months for reservoirs, and only 12 months for rivers. Beyond those limits, the best we can do is look at different possibilities and prepare for the worst, rather than trying to predict a single, specific future.

Technical Summary: From Forecast Skill to Planning Value

Problem Statement
Regulated Mediterranean river basins face increasing challenges from non-stationary hydroclimatic conditions, including recurrent droughts and high-intensity rainfall events. Effective water management requires monthly forecasts that are not only skillful but also transparent regarding their limits. Current data-driven hydrological forecasting faces three primary limitations when applied to regulated basins:

Scope: Most studies focus on streamflow at near-natural gauges, neglecting reservoir storage and groundwater levels, which possess distinct persistence structures.
Attribution: Models often conflate autoregressive memory with exogenous climate and demand forcing, making it difficult to isolate the marginal value of external drivers over simple persistence.
Operational Interpretation: Scenario projections are frequently presented without linking them to horizon-dependent reliability limits, leading to potential misinterpretation of scenario spread as calibrated predictive uncertainty rather than sensitivity to forcing assumptions.

Methodology
The study develops a data-driven, multi-horizon forecasting framework for the Jucar River Basin District (eastern Spain), targeting three typologies: aquifer piezometric levels, reservoir storage, and river discharge.

Data and Preprocessing: The framework utilizes monthly observations from 1988 to 2023 (with reservoir data ending in 2016, requiring a reconstructed bridge for 2017–2023). Target variables are modeled in increment space ( $\Delta y_t = y_t - y_{t-1}$ ) to stabilize learning across heterogeneous magnitudes, with levels reconstructed recursively.
Feature Engineering: Two predictor families are evaluated:
- ENDO (Endogenous): Encodes system memory via lagged states ( $y_{t-1}, y_{t-12}$ ) and seasonal terms (month, sin/cos).
- EXOG (Exogenous): Augments ENDO with bias-adjusted climate predictors (precipitation, $T_{max}, T_{min}$ ) and sectoral demand drivers.
Model Architecture: Four tabular regression models (Bayesian Ridge, ARD Regression, Gradient Boosting, XGBoost) were evaluated. No neural networks were used. Model selection was separated from performance reporting using an inner validation window.
Validation Protocol: A strict leakage-aware rolling-origin pseudo-future protocol was employed. Training data strictly precedes evaluation data to avoid optimistic skill estimates in autocorrelated series. Forecasts are generated recursively up to 48 months.
Scenario Construction: Three scenario families (Base, Favorable, Unfavorable) were constructed for 2024–2030 using a hybrid delta-change and stochastic weather generation approach based on CMIP6/ScenarioMIP logic. These serve as stress-testing pathways, not probabilistic predictions.
Metrics: Performance is assessed using Kling–Gupta Efficiency (KGE) and its decomposition into correlation ( $r$ ), variability ratio ( $\alpha$ ), and bias ratio ( $\beta$ ). Skill is benchmarked against a naïve persistence model using the RMSE-based skill score ( $SS_{RMSE}$ ). A reliability horizon ( $h_{rel}$ ) is defined as the maximum lead time where median KGE $\ge 0.6$ .

Key Results
The study reveals that forecast skill is strongly typology-dependent, governed by the physical mechanisms of each system component:

Aquifers: Exhibit extremely high predictability (median KGE > 0.97) across all horizons due to strong autocorrelation. However, the ENDO configuration outperforms EXOG at long horizons; adding exogenous climate/demand data actually degrades RMSE performance relative to persistence, likely because groundwater signals evolve too smoothly for monthly climate noise to add value.
Reservoirs: Show a distinct skill pattern where models substantially outperform naïve persistence from six months onward. In the long-range bucket (24–36 months), the framework achieves significant skill scores ( $SS_{RMSE} \approx +0.46$ to $+0.50$ ). While ENDO yields slightly higher KGE, EXOG reduces absolute errors more effectively, though both confirm that the framework adds value where persistence degrades.
Rivers: Skill is confined to a 12-month reliability horizon. Beyond this, performance drops below the persistence benchmark.
- Degradation Mechanism: KGE decomposition identifies the bias ratio ( $\beta$ ) as the dominant degradation mechanism for rivers at longer leads. This is driven by the recursive accumulation of positively skewed monthly discharge increment errors (right-skewed distribution), leading to systematic overestimation.
- Role of EXOG: Exogenous forcing partially mitigates this bias drift by constraining the long-run mean of the trajectory, though it does not significantly improve temporal correlation.

Scenario Findings

Divergence: Scenario spread is strongest for rivers within the reliability horizon, while remaining muted for aquifers and storage-conditioned for reservoirs.
Interpretation: For storage-dominated components (aquifers/reservoirs), near-term scenario differences (2024–2030) are driven primarily by demand pathway assumptions rather than detectable climate change signals, as realized temperature deltas in this window are small relative to interannual variability.
Uncertainty: Scenario spread represents sensitivity to forcing assumptions, not calibrated predictive uncertainty.

Significance and Claims
The paper claims to advance the field by shifting the focus from raw forecast skill to planning value through a typology-aware framework. Its primary contributions are:

Typology-Specific Reliability: It establishes explicit, physically grounded reliability horizons (48 months for aquifers, 36 for reservoirs, 12 for rivers), arguing that model outputs should be used differently depending on the lead time and system type.
Mechanistic Attribution: By decomposing KGE and benchmarking against persistence, the study isolates why models succeed or fail (e.g., identifying bias accumulation in rivers vs. memory dominance in aquifers), moving beyond black-box performance metrics.
Operational Clarity: It clarifies that in regulated basins, high KGE scores in storage-dominated systems often reflect inertia rather than predictive power, and that scenario spread should not be misinterpreted as probabilistic uncertainty.
Transferability: The framework is designed to be transferable to other regulated or semi-arid basins where heterogeneous operational datasets (monthly resolution) are available, offering a scalable alternative to process-based models for near-term planning.

The authors maintain a modest stance, acknowledging limitations such as the use of synthetic forcing for scenarios, the lack of formal evaluation under pre-defined hydroclimatologically contrasted periods, and the uncertainty introduced by the reservoir reconstruction bridge. They emphasize that the framework is a tool for stress-testing and comparative planning rather than a deterministic predictor of future climate states.

More like this