نوع مقاله : مقاله پژوهشی
نویسندگان
گروه مهندسی عمران، دانشکده فنی مهندسی، دانشگاه آیت ا... بروجردی، بروجرد، ایران
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Water supply management had become a key challenge for utilities like regional water authorities and wastewater agencies due to Iran’s arid climate. With rising water supply costs, the accurate estimation of water demand has become increasingly critical. Climate change, further complicates this task by indirectly increasing consumption. Therefore, making traditional trend-based projections unreliable. This study leverages the latest Shared Socioeconomic Pathways (SSP) climate scenarios to model future climatic variables and project urban water demand in Borujerd City. A hybrid Radial Basis Function-Genetic Algorithm (RBF-GA) machine learning model has been employed to analyze correlations between climatic factors and water consumption. Our methodology and findings are benchmarked against Iran’s national water demand guidelines (Standard 117-3 first revision). Results show that the 117-3’s estimates are systematically Overestimated. On the other hand, the RBF-GA model outperforms both standalone RBF and the 117-3 approach in case of RMSE and R^2 accuracy measures. Projections indicate that mid-term (2022–2036) climate change will raise daily water demand by an average of 60.462 m³/day under SSP2-4.5 scenario and 61.768 m³/day under SSP5-8.5 scenario respectively. These projections exceed even the overstated estimations of 117-3’s approach.
کلیدواژهها [English]
Access to drinking water is a critical societal need. Inadequate water supply planning has led to irreversible consequences, including mass migration to cities and socioeconomic crises. Hundreds of Iranian villages have already been abandoned due to water scarcity, as shown by Bastani (2016), who linked droughts to the depopulation of 14 villages in Darab County. Current water demand estimates rely on Publication 117-3, a mandatory national standard for water infrastructure projects. However, the first revision of this guideline ignores climate change effect, risking inaccurate projections.
Publication 117-3 calculates water demand using fixed tables based on population, climate zones, and geology, but its static approach fails to account for climatic shifts. Studies (e.g., Khalifeh et al., 2013; Yazdi et al., 2016) reveal its limitations, such as underestimating peak demand coefficients by 27.3 percent (Moradi Sabzehkouhi et al., 2014). Traditional methods also require extensive data collection, which is costly and impractical for large cities. Machine learning (ML) offers a solution by processing complex variables without predefined biases. For example:
Tabesh et al. (2015) used artificial neural networks (ANNs) to predict water demand.
Chen et al. (2017) employed random forests with historical consumption data.
Koutrou et al. (2008) applied ANNs for mid-term urban water forecasts.
Yet, most prior studies treated water demand as a function of population or past usage, neglecting temperature’s indirect effects. Recent work (Sharghi et al., 2024) shows that even weakly correlated variables (e.g., temperature) improve machine learning model accuracy in ground water fluctuation modelling, if participated. Few studies (e.g., Lin et al., 2019; Rafsanjani & Haghighat, 2020) have integrated older climate models (CMIP5/RCPs) to project water demand.
This study addresses two gaps:
Most prior work uses outdated CMIP5 climate models; we employ CMIP6/SSP scenarios for higher accuracy.
Hybrid ML models (e.g., RBF-GA) remain underexplored for urban water forecasting.
By combining these innovations, the study provides actionable insights for cost-efficient water resource allocation. Overestimation (Publication 117-3) wastes funds, while underestimation triggers water crises. Comparing RBF-GA outputs with traditional methods will quantify these biases and support evidence-based policy updates.
This study aims to project urban water demand in Borujerd City, Iran, moving beyond conventional observational methods by incorporating climate change impacts. Particularly the indirect influence of rising temperatures on consumption patterns. It also evaluates the performance of the proposed model against traditional water estimation methods (Publication 117-3), highlighting how climate change significantly alters long-term projections and underscores the need to revise outdated approaches.
Location: Borujerd City (33°54′N, 48°45′E), Iran; area: 2,642 km²; elevation: 1,600 m.
Climate: Mediterranean based on Publication 117-3, with cold winters and dry summers.
Population: 326,452 (2016 census); growth rate: 2.35 percent.
Water Consumption data were extracted from Borujerd’s Water Authority consist of Monthly data (m³/month) from (2003-2022), using subscriber counts (vs. total population) for higher correlation with usage.
Historical data include Maximum temperature data from Borujerd Synoptic Station (https://data.irimo.ir). Future Projections were obtained using CMIP6 outputs (ACCESS-ESM1-5 model, SSP2-4.5/SSP5-8.5 scenarios).
This study utilizes CMIP6 model to incorporate future temperature fluctuations in machine learning algorithm. CMIP6 models Improved radiative forcing calculations vs. CMIP5, integrating socioeconomic factors (SSPs). However, raw GCM outputs have curse resolution. Therefore, this study applies LARS-WG downscaling technique to convert coarse GCM outputs to local-scale projections
This study utilizes a hybrid machine learning algorithm consist of a Radial Basis Function coupled whit Genetic Algorithm (RBF-GA) to learn the complex relation between climate variables and water consumption.
The architecture of RBF consists of input layer (temperature, subscriber count), hidden layer (Gaussian radial basis functions), output layer (linear). Training and test data were divided 80 percent and 20 percent respectively. Main limitations of RBF are constant radius of Gussian functions and hard to find optimum amount of hyperparameters. GA Integration enhance the performance of model in many ways it helps incorporate a variety of radiuses and finding the optimum set of hyperparameters. It also helps to avoid manual trial and error to fined optimum hyperparameters.
Publication 117-3 use a traditional method of geometric growth. But population Estimation for future projections were obtained using (MATLAB) exponential curve fitting generate monthly subscriber projections (2022–2037).
The research methodology is structured as follows: First, temperature variations in the study area under climate change conditions for the period 2022–2037 are projected using statistical downscaling under SSP scenarios. Subsequently, the machine learning model is trained and optimized using observational data, with the training dataset comprising 80 percent of the total observed data. This dataset includes two input variables monthly maximum temperature, the number of subscribers and the target vector, which is the urban water consumption in Borujerd. The machine learning model, RBF-GA, establishes a function between these two variables (temperature and population) to predict urban water consumption.
The downscaled temperature projections indicate a clear increase in maximum temperatures in the mid-term future (2022–2036) under both SSP2-4.5 and SSP5-8.5 scenarios. This upward trend is evident in the monthly averages. Based on the results, the annual mean temperature is expected to rise by 1.1°C and 1.43°C under the SSP2-4.5 and SSP5-8.5 scenarios, respectively.
The hybrid RBF-GA model outperformed the standard RBF model in validation accuracy, which suffered from overfitting despite strong training performance. RBF-GA’s genetic algorithm optimizes neuron selection through iterative evaluation of unseen data, improving predictive stability. While the standard model produced erratic results, the hybrid approach maintained alignment with observational trends and minimized deviations, demonstrating its suitability for robust long-term forecasting under climate change scenarios.
The RBF-GA model outperformed Publication 117-3 in estimating water demand, achieving markedly lower RMSE (824 vs. 11,779 m³/day). While Publication 117-3’s overestimations provide safety margins, they lack precision for near-term climate-adaptive planning. RBF-GA projections for 2036 under SSP scenarios exceed the publication’s estimates, reflecting escalating climate-driven demand. The study concludes that Publication 117-3’s static framework is inadequate for short-term water resource management under climate change, advocating for adaptive machine learning approaches to enhance accuracy while maintaining operational safeguards.
This study underscores the necessity of climate-adaptive strategies in water resource management, leveraging downscaled CMIP6 data and a hybrid RBF-GA model to project future demand. The model outperformed conventional methods, revealing Publication 117-3’s limitations in addressing climate-induced demand increases. A proposed global warming coefficient could enhance traditional standards, aligning safety margins with dynamic climatic realities.
Ali Sharghi: Investigation, Formal analysis, Software, Writing- Original draft, Visualization
Mehdi Komasi: Data curation, Conceptualization, Methodology, Validation, Reviewing and Editing final draft.
Datasets of GCMs outputs related to this article can be found at https://esgf-node.llnl.gov/search/cmip6/ (accessed 21 December 2024), powered by [ESGF and COG], hosted by [ Lawrence Livermore National Laboratory, Department of Energy].
Datasets of the synoptic station of Silakhor can be found at https://data.irimo.ir/ (accessed 21 December 2024), Iran Meteorological Organization (IRIMO).
LARS WG weather simulator model can be downloaded at https://www.rothamsted.ac.uk/ (accessed 21 December 2024).
The study was approved by the Ethics Committee of the Samimnoor https://www.samimnoor.ir/ The authors avoided data fabrication, falsification, plagiarism, and misconduct.
)