Using Machine Learning and the Water Evaluation and Planning model to Evaluate Climate Change Impacts on Surface Water Allocation in the Upper Awash Sub-Basin, Ethiopia

Climate change, rapid urbanisation, and population growth are placing escalating pressure on surface water resources in Ethiopia’s Upper Awash Sub-Basin, a region already grappling with rising agricultural demand and limited water availability. This study presents an integrated modelling framework that combines Machine Learning (ML)techniques with the Water Evaluation and Planning (WEAP) system to evaluate andoptimise surface water allocation under future climate and demographic scenarios. Historical climate patterns from 1948 to 2010 were examined using both observed and satellite-based datasets (Princeton datasets). A Random Forest (RF) algorithm trained on these data achieved high predictive accuracy (R² = 0.97 for training; 0.96 for testing; RMSE = 0.12°C). The analysis revealed a historical warming trend of 0.04 to 0.05°C per year and a gradual decline in precipitation, ranging from 0.22 to 0.40 mm per year. Future climate projections based on CMIP6 under the Shared Socioeconomic Pathways (SSP4.5 and SSP8.5) for the period 2025 to 2075 estimate temperature increases between 0.9°C and 1.6°C per year. Projected precipitation changes range from a 15.5% increase (SSP4.5) between 2025 and 2075 to a 6.3% decline (SSP8.5), indicating heightened uncertainty and variability in regional water availability. These climatic changes are expected to intensify hydrological stress across the basin, reducing baseflow by approximately 46.9 mm per year, increasing evapotranspiration to between 890 mm and 1010 mm annually, and causing soil moisture fluctuations from +16.4 mm to −124.9 mm per year. Simultaneously, the population, projected using linearregression and exponential growth models, is expected to grow from 6.3 million in 2025 to nearly 39 million by 2075, substantially increasing water demand across all sectors. The ML-enhanced WEAP model demonstrated improved forecasting capabilities, with the RF model outperforming the Long Short-Term Memory (LSTM) network (MAE: 0.41 vs. 0.46). SHapley Additive exPlanations (SHAP) analysis identified lagged population growth and unmet demand as the most influential predictors, alongside temperature and drought-related variables. To support adaptive water governance, a Non-Dominated Sorting Genetic Algorithm II (NSGA-II) was employed to generate Pareto-optimal water allocation strategies. Two trade-off solutions were identified: (1) an equity-oriented strategy, allocating 38% of water to urban use, 37% to agriculture, and 25% to industry (Gini index = 0.22; demand penalty = 0.10). Here, the Gini index, a measure of distributional fairness across sectors, indicates relatively high equity, while the demand penalty, a measure of unmet demand relative to total demand, remains low. (2) An efficiency-focused strategy, favouring urban (43%) and industrial (25%) sectors over agriculture (32%), offers a 9% economic gain at the cost of a slightly higher demand penalty (0.12), reflecting greater overall efficiency but less balanced distribution. Overall, this research demonstrates the potential of a hybrid ML–WEAP approach to improve long-term water allocation planning in data-scarce, climate-sensitive regions. The framework offers a replicable, evidence-based tool for enhancing climate resilience and promoting equitable and efficient water resource management in the Upper Awash Sub-Basin and beyond.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Hirko, Deme Betele

Advisors dc:contributor.advisor

• Du Plessis, Jakobus Andries
• Bosman, Adele

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/136061

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/136061