Impact of IPCC CMIP6 Climate Scenarios on Basin Runoff Using ANN and Multiple Regression Approaches

Abstract
Objective: The objective of this study is to assess future runoff variations in the Khorramabad basin using up-to-date climate data and data-driven modeling approaches.
Method: To this end, outputs from the Sixth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC) under the Shared Socioeconomic Pathways SSP1-2.6 (optimistic) and SSP5-8.5 (pessimistic) were employed for the future periods 2040–2069 and 2070–2099. Climate variables were downscaled using the change factor method. To simulate future runoff, two data-driven models, namely Artificial Neural Networks (ANN) and Multiple Linear Regression (MLR), were developed.
Results: The evaluation of climate models indicated that CMIP6 models perform more reliably in simulating historical temperature than precipitation, highlighting the higher uncertainty associated with precipitation projections. Furthermore, analysis of the downscaled data revealed a decreasing trend in precipitation and an increasing trend in long-term monthly mean temperature under both climate scenarios. In the hydrological modeling framework, the ANN model outperformed the multiple linear regression model in terms of runoff simulation accuracy and its ability to capture complex nonlinear relationships among climatic variables.
Conclusion: Runoff projections derived from the ANN model indicate a substantial decline in runoff across all future periods and scenarios, suggesting intensified water stress and underscoring the urgent need for targeted planning and adaptive strategies to ensure sustainable water resources management in the studied watershed. 
 
 
 
English Extended Abstract
Introduction
Climate change is a principal driver of hydrological alteration worldwide, and semi-arid catchments are especially vulnerable because reductions in precipitation and increases in temperature translate directly into heightened water‑stress.
The present study quantifies the projected evolution of surface runoff in the Khorramabad River Basin (4250 km², Lurestan Province, Iran) by coupling the latest IPCC-AR6 climate information with data-driven hydrological models.
The primary objectives were to (i) evaluate the performance of the newest CMIP6 general-circulation models (GCMs) for the basin, (ii) downscale the GCM outputs to the basin scale, and (iii) assess the ability of a non-linear Artificial Neural Network (ANN) and a multiple linear regression (MR) model to simulate future runoff under the SSP1-2.6 (optimistic) and SSP5-8.5 (pessimistic) pathways.
 
Method
Historical climate series (1971-2000) for precipitation and air temperature were compiled from local stations and used as the reference for model evaluation. Fourteen CMIP6 GCMs were extracted for the historical period and for two future windows – mid-century (2040-2069) and end‑century (2070‑2099) – under SSP1‑2.6 and SSP5‑8.5.
A simple change‑factor (bias‑correction) approach was applied to translate the coarse-resolution GCM fields to the basin scale; the precipitation factor was defined as the ratio of future to baseline mean monthly precipitation, while the temperature factor was the difference between future and baseline mean monthly temperature.
Model skill was assessed using RMSE, NSE, Kling-Gupta Efficiency (KGE) and the coefficient of determination (R²). IPSL‑CM6A‑LR emerged as the most skillful GCM for precipitation (RMSE ≈ 49 mm yr⁻¹, NSE ≈ 0.82, KGE ≈ 0.86) and ACCESS‑CM2 for temperature (RMSE ≈ 0.83 °C, NSE ≈ 0.99). For hydrological simulation, two data-driven models were constructed:

Multiple Linear Regression (MR): A linear equation linking monthly runoff (Q) to precipitation (P) and temperature (T).
Artificial Neural Network (ANN): A feed-forward network with a 3‑10‑1 architecture (three inputs, ten hidden neurons, one output). The observed dataset was partitioned into training (70 %), validation (15 %), and testing (15 %) subsets. ANN training employed the Levenberg-Marquardt algorithm, and model performance was evaluated with the same statistical metrics as above.

 
Results
Downscaled CMIP6 projections indicate a consistent decline in mean annual precipitation of 5–9 % and a rise in mean monthly temperature of up to ~5 °C by the 2070-2099 horizon, with the magnitude of change markedly larger under the high‑emission SSP5‑8.5 scenario.
The ANN consistently outperformed the MR model across all evaluation periods and metrics. In the independent test set the ANN achieved NSE = 0.35, RMSE ≈ 12 % of the observed runoff range, KGE > 0.60 and a correlation coefficient r ≈ 0.71, whereas the MR model produced NSE values close to zero and substantially larger RMSE, reflecting its inability to capture the non‑linear precipitation‑runoff relationship.
Using the ANN, future runoff is projected to decrease by 15-30 % relative to the historic period across both scenarios, with the greatest reductions occurring under SSP5-8.5. The runoff decline mirrors the simulated precipitation deficits and the temperature-driven increase in evapotranspiration, signalling a pronounced intensification of water scarcity in the Khorramabad basin.
 
Conclusion
Integrating up‑to‑date CMIP6 climate projections with an ANN-based hydrological framework provides a robust, basin-specific assessment of future water availability in a semi‑arid environment.
The analysis reveals a clear trajectory toward reduced surface‑water resources, especially under the pessimistic SSP5‑8.5 pathway, and underscores the urgency of implementing adaptive water‑management strategies such as improved irrigation efficiency, optimized allocation policies, the development of flood‑and‑drought mitigation infrastructure, and early‑warning systems.
Future work should explore hybrid physics‑data models and multi-scenario ensembles to better resolve extreme flow events and to further reduce predictive uncertainty.