پیش‌بینی کوتاه‌مدت تقاضای آب در شبکه توزیع شهری با استفاده از مدل Attention-BiGRU و مهندسی ویژگی‌های زمانی: مطالعه موردی شهر اصفهان

Abstract
Objective: Accurate urban water demand forecasting plays a vital role in water resources management and the efficient operation of urban water distribution systems. This study aims to develop an Attention-BiGRU-based model integrated with temporal feature engineering for short-term forecasting of urban water demand in Isfahan, Iran.
Method: Monthly water demand data, together with demographic, climatic, and hydrological variables, were preprocessed and enhanced through temporal feature engineering before being used to train the proposed model. The forecasting performance of the proposed approach was compared with ARIMA, SARIMA, Random Forest (RF), XGBoost, and LSTM models using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²).
Results: The Attention-BiGRU model outperformed the other forecasting models, achieving higher prediction accuracy, lower error values, and a higher coefficient of determination. Moreover, the proposed model more accurately captured temporal trends and seasonal patterns, showing better agreement between the observed and predicted water demand than the benchmark models.
Conclusion: The findings indicate that integrating an attention mechanism into the BiGRU network and applying temporal feature engineering provide an effective approach to improving short-term urban water demand forecasting. The proposed framework can serve as a reliable decision-support tool for intelligent water distribution system management and operational planning.

Extracted Abstract
Introduction
Urban water demand forecasting plays a vital role in sustainable water resources management and the efficient operation of urban water distribution systems. Accurate short-term forecasts help optimize water production, reduce operational costs, improve energy efficiency, and ensure a reliable water supply. However, water demand is affected by various demographic, climatic, and seasonal factors, resulting in complex nonlinear patterns that are difficult to predict accurately.
Traditional statistical models, such as ARIMA and SARIMA, have been widely used for water demand forecasting but often fail to capture nonlinear relationships and long-term temporal dependencies. Machine learning methods, including Random Forest (RF) and XGBoost, improve prediction accuracy by modeling nonlinear interactions; however, their ability to learn sequential temporal information remains limited.
Deep learning models, particularly Long Short-Term Memory (LSTM) and Bidirectional Gated Recurrent Unit (BiGRU), have shown superior performance in time-series forecasting. The BiGRU architecture captures information from both past and future sequences, while the attention mechanism further enhances forecasting performance by emphasizing the most relevant historical information.
This study proposes an Attention-BiGRU model integrated with temporal feature engineering for short-term urban water demand forecasting. Monthly water demand data from Isfahan, Iran, together with demographic, climatic, and hydrological variables, were used to develop and evaluate the proposed framework. Its performance was compared with ARIMA, SARIMA, RF, XGBoost, and LSTM using standard evaluation metrics to assess its effectiveness for intelligent urban water management.
 
Method
This study proposes an Attention-BiGRU framework combined with temporal feature engineering for short-term urban water demand forecasting. Monthly water demand data from Isfahan, Iran, together with demographic, climatic, and hydrological variables, including population, air temperature, rainfall, relative humidity, evaporation, reservoir storage level, river flow, and drought index, were used to develop the prediction model.
The dataset was preprocessed by handling missing values and normalizing all variables to improve model performance. Temporal feature engineering was then applied to extract informative time-dependent features, enabling the model to better capture seasonal patterns and temporal dependencies.
The proposed model employs a Bidirectional GRU architecture with an attention mechanism to learn sequential information and automatically emphasize the most relevant historical observations. Its performance was compared with five widely used forecasting models: ARIMA, SARIMA, RF, XGBoost, and LSTM.
Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²), providing a comprehensive assessment of forecasting accuracy.
 
Results
The comparative analysis showed that the proposed Attention-BiGRU model outperformed all benchmark models in short-term urban water demand forecasting. Compared with ARIMA and SARIMA, it more effectively captured nonlinear relationships and seasonal variations, while temporal feature engineering further improved its ability to learn hidden temporal patterns.
Random Forest and XGBoost achieved better performance than traditional statistical models but were less effective than deep learning approaches in modeling long-term temporal dependencies. Although the LSTM model produced satisfactory results, the proposed Attention-BiGRU generated predictions that were more consistent with the observed water demand throughout the study period.
The combination of bidirectional sequence learning and the attention mechanism enabled the proposed model to focus on the most informative historical observations, resulting in more accurate prediction of both long-term trends and seasonal fluctuations. Accordingly, the Attention-BiGRU model achieved the lowest MAE, RMSE, and MAPE values and the highest R² among all evaluated models.
Overall, the results demonstrate that integrating temporal feature engineering with an Attention-BiGRU architecture significantly improves short-term urban water demand forecasting accuracy and provides a reliable tool for intelligent urban water management.
 
Conclusion
This study presented an Attention-BiGRU-based framework combined with temporal feature engineering for short-term urban water demand forecasting. By integrating demographic, climatic, and hydrological variables with engineered temporal features, the proposed model effectively captured the complex nonlinear relationships and temporal dependencies inherent in urban water demand data.
The comparative analysis demonstrated that the proposed approach achieved superior forecasting performance compared with conventional statistical models, machine learning algorithms, and the LSTM network. The combination of bidirectional sequence learning and the attention mechanism enabled the model to focus on the most relevant historical information, resulting in more accurate and stable predictions. These findings indicate that the proposed framework provides an effective solution for improving short-term water demand forecasting under varying environmental and operational conditions.
Accurate demand forecasting can support water utilities in optimizing water production, improving operational efficiency, reducing energy consumption, and enhancing decision-making for sustainable water resource management. Therefore, the proposed model has significant potential for practical implementation in intelligent urban water distribution systems.
Although the proposed framework demonstrated promising performance, this study was limited to monthly data from a single urban area. Future research may extend the proposed approach by incorporating higher-resolution temporal data, additional socioeconomic variables, and advanced deep learning architectures such as transformer-based models. Furthermore, evaluating the proposed framework in different climatic regions would provide additional evidence of its robustness and generalizability.