Article Preview
TopIntroduction
The rapid expansion of urban areas underscores the increasing necessity for intelligent urban water supply systems, which are crucial for the development of modern smart cities (Wu et al., 2018). A key component of such systems is the prediction of urban water demand, a task that is both significant and challenging. Accurate predictions of water demand are essential for optimal operation of valves and pumps in waterworks and for identifying potential leakages in the supply network. Effective water demand forecasting can not only lower water supply costs but also ensure that urban water needs are met. However, few models are practically applicable to urban water demand prediction. This task is complex due to the multifaceted influencing factors and inherent unpredictability of urban water demand (Du et al., 2021).
The accurate forecasting of urban water demand is a critical aspect of sustainable urban development and efficient resource management. As urban populations grow and climate patterns change, the need for robust forecasting models becomes even more pronounced. In recent years, the integration of machine learning and deep learning techniques in this field has shown promising results, leading to more accurate, adaptive, and efficient forecasting models (Khozani et al., 2022; Xu et al., 2022). Urban water demand forecasting is not just a technical challenge but also a key component in addressing broader issues such as water conservation, infrastructure development, and environmental sustainability (Butler & Memon, 2005; Donkor et al., 2014). Accurate forecasting helps in reducing wastage, optimizing distribution, and planning for future needs, particularly in the context of climate change and urban expansion (Zubaidi et al., 2020). The deep learning models commonly used in the field are as follows (Adamowski & Karapataki, 2010): Linear Regression models are the most basic form of predictive analysis and are used for their simplicity and interpretability (Pulido-Calvo et al., 2007; Sebri, 2016). However, they often fail to capture complex nonlinear relationships inherent in water demand data. ARIMA (auto-regressive integrated moving average) is a popular traditional statistical model for time series forecasting, which is effective for data with a clear trend or seasonal pattern but struggles with non-linear data (Pulido-Calvo et al., 2007). Support vector machines (SVMs) are effective for classification and regression challenges. SVMs work well with high-dimensional data but can be computationally intensive and less effective with very large datasets (Peña-Guzmán et al., 2016). The random forest is an ensemble learning method that is robust to overfitting and can handle both linear and nonlinear data (Chen et al., 2017). However, it can be less interpretable and may require a significant amount of data for optimal performance. The LSTM (long short-term memory network) is a type of recurrent neural network (RNN) particularly well-suited for time series data due to its ability to capture long-term dependencies. However, LSTMs can be complex to tune and train and may require substantial computational resources (Kühnert et al., 2021).
Banjac et al. (2015) developed a water demand prediction system utilizing ANNs, which was a significant step towards understanding the complex dynamics of water demand series. Further, Yin et al. (2018) demonstrated the superiority of ANNs over traditional statistical models in water-energy demand forecasting, highlighting the enhanced predictive accuracy and capability of ANNs in handling complex datasets. Additionally, other machine learning methods have also been explored and applied in the field of water demand forecasting, signifying a shift towards more sophisticated, data-driven approaches in this area. However, these methods, particularly those with shallow architectural designs commonly found in classic intelligent models, face limitations in efficiently handling large-scale data. This restriction hinders their ability to effectively mine and interpret complex data features, presenting a significant challenge in the field of predictive analytics, where understanding and leveraging intricate data patterns is crucial (Du et al., 2021).