Short-term water demand predictions coupling an artificial neural network model and a genetic algorithm Open Access

Forecasting urban water demand (UWD) is a crucial issue to ensure the better design, operation, and management of water distribution systems (WDSs). While long-term forecasting is mainly required for planning and design, short-term forecasting is particularly used for operation and management. More specifically, in the most recent applications of real-time control (RTC), knowledge of the near future fluctuations in consumption is required (e.g. Pascual et al. 2013; Kang 2014; Doghri et al. 2020).

The UWD is a complex and nonlinear function of different factors such as time, socioeconomic factors, climatic and meteorological variables, and the cost of the supplied water (Ghiassi et al. 2008; Odan & Reis 2012; Hussien et al. 2016). UWD prediction models can be univariate, where only the UWD records are considered as inputs for the models (e.g. Alvisi et al. 2007; Alvisi & Franchini 2014; Romano & Kapelan 2014). Additionally, the models can be multivariate, with some of the influencing factors mentioned above also being considered as model inputs to predict the UWD (e.g. Zhou et al. 2002; Herrera et al. 2010; Adamowski et al. 2012; Odan & Reis 2012; Tiwari & Adamowski 2013; Bakker et al. 2014; Tian et al. 2016). Both univariate and multivariate models can be based on various modeling approaches, from which the more commonly applied are pattern-based models (e.g. Alvisi et al. 2007; Bakker et al. 2013, 2014; Gagliardi et al. 2017; Pacchin et al. 2017, 2019), regression analyses (e.g. Adamowski & Karapataki 2010; Bakker et al. 2014), classical time series models (e.g. Bougadis et al. 2005; Caiado 2010; Arandia et al. 2015; Chen & Boccelli 2018; Viccione et al. 2019; Guarnaccia et al. 2020; Mu et al. 2020), artificial intelligence (AI) methods, such as artificial neural networks (ANN) (e.g. Cutore et al. 2008; Firat et al. 2010; Romano & Kapelan 2014; Gagliardi et al. 2017; Pacchin et al. 2019; Pesantez et al. 2020), random forest (e.g. Herrera et al. 2010; Mu et al. 2020; Nasser et al. 2020; Pesantez et al. 2020), support vector machines (SVM) (e.g. Candelieri et al. 2019; Pesantez et al. 2020), or fuzzy logic and neuro-fuzzy models (e.g. Vijayalaksmi & Babu 2015; Jithish & Sankaran 2017). More recently, hybrid models combining two of the previously cited approaches have also been proposed (e.g. Quevedo et al. 2010; Alvisi & Franchini 2014; Suhartono et al. 2018). While many factors impact the performance of UWD prediction models, Sebri (2016) found that forecasting accuracy of urban water demand is significantly influenced by demand periodicity, forecast horizon, forecasting method, model specification and some study specific characteristics such as the sample size, the publication year and the development level of the country on which the study was conducted. House-Peters & Chang (2011) and Donkor et al. (2014) presented an overview of the different existing models adapted to UWD forecasts up to 2010, while Ghalehkhondabi et al. (2017) published a review of the papers concerning specifically soft computing methods applied to UWD from 2005 to 2015. It was reported that using many explanatory variables for multivariate UWD prediction models poses a great challenge in terms of collecting and keeping track of the data since explanatory variables must have sufficiently long records if they are to be used as independent variables for developing forecasting models (House-Peters & Chang 2011; Donkor et al. 2014). Ghalehkhondabi et al. (2017) found that although it is still very difficult to pick a single method as the overall best, ANNs have been superior in many cases in short-term UWD forecasting. This can be derived from the inherent capability of ANNs in terms of analyzing the nonlinear data. Also, the limited number of applications of metaheuristics (such as evolutionary algorithms) in water demand forecasting was identified as one of the incentives for potential future research direction (Ghalehkhondabi et al. 2017).

Pattern-based models rely on the identification of periodic patterns that characterize UWD over different periods of time, while classical time series models do not necessarily take this periodicity into account explicitly. The most popular univariate classical time series models are the autoregressive moving average (ARMA) model and its derivatives, such as autoregressive (AR), autoregressive integrated moving average (ARIMA), seasonal ARIMA, periodic ARMA, threshold AR, and fractionally integrated ARMA models (Adamowski & Karapataki 2010). As for AI models, they do not presume a specific model structure and are thus said to ‘learn’ from the data. Among those AI models, ANN models have been used in numerous studies to predict UWD during the last decade (see some examples above). ANN models have proven to be powerful for mapping the nonlinear trends of UWD, even in the case of possibly noisy multivariate time series (Ghalehkhondabi et al. 2017). Studies carried out in this field highlight the dominance of ANN over conventional techniques (Babel & Shinde 2011). Bougadis et al. (2005) showed that ANN models outperform different regression and time series models for short-term peak water demand forecasting for data from the city of Ottawa, Canada. Similar results were reported by Adamowski & Karapataki (2010) and Adamowski et al. (2012) in favor of ANNs and wavelet transforms coupled with ANNs, compared to other conventional techniques for short-term water demand forecasts. The performance of ANN models is dependent on the set of explanatory variables fed into the model as input and on the forecasting horizon. Prediction accuracies over 98% were achieved by Babel & Shinde (2011) when using only the historic daily demand as the explanatory variable to forecast short-term UWD for the city of Bangkok, Thailand. However, they showed that meteorological, water utility and socioeconomic variables have a greater influence on medium-term (e.g. monthly) predictions. The benefit of univariate models was also reported by Odan & Reis (2012) where their ANN models for short-term UWD prediction did not require the use of weather variables, resulting in a simpler and faster model to train. Also, size of the data sets of case studies can affect the ANN model performance. As Gagliardi et al. (2017) showed, an ANN model applied to small districts, with a low number of users and more variability in water demands, can outperform a pattern-based model while, for districts that contain a large number of users, the pattern-based model tends to be more efficient than the ANN one.

Evolutionary algorithms (Bäck 1996) have been used along with AI techniques for UWD prediction, either for the optimization of training algorithms (Rangel et al. 2016) or optimization of model hyperparameters (Chen 2009). Romano & Kapelan (2014) suggested optimization of the hyperparameters (e.g., number of hidden neurons and number of training cycles) and input structure (e.g. number of past demand values and additional explanatory variables to be used) of their ANN model with an evolutionary algorithm for the prediction of UWD 1 to 24 h ahead, and showed that this approach generates accurate forecasts for UWD. In their application of ANNs for the prediction of hourly UWD, Herrera et al. (2010) compared the growing and sliding window approaches to train their ANN model. For their case study, they found that the growing window approach led to better results.

In terms of UWD, the definition of short-term predictions varied among authors. Although many of them consider predictions 1 h ahead as being short term (e.g. Shvartser et al. 1993; Zhou et al. 2002), others qualify the daily, weekly or even monthly predictions as being short term (e.g. Jain & Ormsbee 2002; Bougadis et al. 2005; Yurdusev & Firat 2009). The forecast horizon depends on the purpose of the model application (Ghiassi et al. 2008; Bakker et al. 2013). More specifically for real-time control applications, such as active pressure management, demand predictions should be provided a few times an hour (e.g. Creaco 2017; Doghri et al. 2020).

Few applications of UWD predictions at time steps lower than 1 h have been presented in the literature: ANN, support vector regression (SVR) and random forest for 10 min predictions in Nasser et al. (2020); a pattern-based model for 15 min predictions in Bakker et al. (2013); a seasonal ARIMA model for 15 min predictions in Arandia et al. (2015); and a pattern-based model combined to an ARIMA model for 10 min predictions in Quevedo et al. (2010). A few studies have compared the performances of prediction models on making short-term UWD forecasts (daily in Adamowski et al. 2012; Tiwari & Adamowski 2013; Bai et al. 2014; and hourly in Ghiassi et al. 2008; Herrera et al. 2010). A thorough search of the scientific literature showed that only two studies comparing the performances of different UWD prediction models for time steps lower than 1 h have already been published (Mu et al. 2020; Nasser et al. 2020). Mu et al. (2020) used a long short-term memory (LSTM) model to predict short-term UWD based on data with time steps of 15 min, 1 and 24 h. The performance of the LSTM-based model was compared with ARIMA, SVR, and random forest models. Nasser et al. (2020) compared the performance of different models for 10 min UWD predictions. However, this last study was limited to a few households (granular water demand) and only compares AI models among them. Short-term (10- or 15-min) time steps describe the UWD dynamics in details and can be required for some real-time applications such as optimizing the exact timing of pump switch (Bakker et al. 2013) or for dynamic pressure management (Creaco 2017; Doghri et al. 2020). These can help to provide the required amount and pressure of water to urban areas at the lowest operation cost. Furthermore, a short-term 15-min time step UWD prediction can be used for leakage and energy analysis (Mu et al. 2020).

The main objective of this paper was to evaluate how the optimization of the hyperparameters of an ANN model and the use of a growing window approach for its training could improve the 15 min UWD predictions made with this model. The performance of the ANN model was also compared with those of a classical time series model (ARIMA) and of a pattern-based model (Bakker et al. 2013). To the best of the authors’ knowledge, this is the first time that the performance of these three types of models in making predictions of UWD for time steps shorter than 1 h has been compared. Data sets provided by two Canadian cities are used for these evaluations.

The datasets analyzed in this study were collected from two cities in the province of Quebec (Canada). The data consisted of 15 min time step records of drinking UWD. The datasets were divided into training, validation and testing data for the ANN model. The training data were used to train the model and update the parameters. The validation data were used to select model parameters and stop the algorithm early, based on minimum validation error which efficiently avoids model overfitting or underfitting. The testing data were used to evaluate the performance of the ANN model to the unseen data. The same testing data were used to evaluate the performance of the pattern-based model.

For the dataset from City 1, the records present the total drinking water produced in the treatment plant for a period of 5 years (2009–2013). The first 4 years of observation were used as training and validation samples for the ANN models (from 1 January 2009, 00:00 to 3 August 2012, 17:45 for training and from 3 August 2012, 18:00 to 31 December 2012, 23:45 for validation), and the remaining year (from 1 January 2013, 00:00 to 31 December 2013, 23:45) was used as a test set to assess the accuracy of all prediction models (ANN, ARIMA and pattern-based). The average recorded consumption for this dataset is about 14,560 m³/d with a standard deviation of about 3,090 m³/d for 15 min time steps. The dataset from City 2 presents the total water provided to a district metered area and covers a period of 23 months. The first period of observations was used as a training and validation set (from 1 September 2012, 00:00 to 1 November 2013, 23:45 for training and from 2 November 2013, 00:00 to 31 December 2013, 23:45 for validation) for the ANN models and the remaining 7 months (from 1 January 2014, 00:00 to 21 July 2014, 23:15) were used as a test set to assess the accuracy of all prediction models (ANN, ARIMA and pattern-based). The average consumption, in this case, is about 887 m³/d with a standard deviation of about 311 m³/d for 15 min time steps. The time series for the two cities are shown in Figure 1 where training, validation and test sets are identified for each city, while their mean daily patterns are shown in Figure 2.

To remove the outliers from both databases, the filloutliers function in MATLAB (ver. R2019a) has been used, which defines outliers as points outside three standard deviations from the mean and replace the outlier with the nearest element that is not an outlier. Finally, 128,167 data points were used as training/validation sets and 35,028 as test set for City 1. Whereas for City 2, in total 66,145 data points were divided into 46,752 and 19,393 data points for training/validation and test sets, respectively. As illustrated in Figure 1, there were about 1,200 time steps in the dataset for City 2 with constant values (from time step 8,800 to time step 10,000). As these data were used only for the training of the ANN models and rather negligible compared to the 46,752 data points, they were not excluded from the data set. Also, our preliminary results indicated that performance indices for the ANN models are not dependent on the presence or exclusion of this part of the data set.

ANNs have been proven to have excellent predictive ability in various domains (Sun & Zhang 2002; Nastos et al. 2013; Mislan et al. 2015; Cheng & Bao 2020).

One of the most popular ANNs in the scope of UWD prediction are feedforward neural networks (FNNs) (Hamed et al. 2004). The layers of the ANN can be fully or partially connected and, for the purpose of forecasting, the weights should be adapted accordingly in a process called training.

A number of optimization algorithms can be used for the training process including gradient descent and Levenberg–Marquardt (Hamed et al. 2004; Ghalehkhondabi et al. 2017). The Levenberg–Marquardt algorithm is often utilized for multilayer perceptron neural networks due to its faster convergence as it adopts the method of approximate second derivative (Singh et al. 2010).

In this work, three ANN models are developed. In all cases, datasets were decomposed to the trend and cyclical components, and were used as inputs to the ANN models. For each time step prediction, the two last time steps were utilized. The first model is based on initially performing one time training, using the entire training and validation data set, and then obtaining the performance results on the whole test data set with the trained network (Single ANN). The second model considers for each time step in the test data set, all the data before that time step as a training and validation data set to predict that specific time step. In other words, the growing window concept is utilized in which all the data from the first to current time steps are used to train a network for predicting the next 15 min time step (Multiple ANN). Finally, as a third approach, a genetic algorithm (GA) is used to optimize the ANN hyperparameters.

Indeed, the performance of an ANN model depends on the optimization of its hyperparameters, which define the topology and learning options of a neural network. The numbers of hidden layers and neurons in each hidden layer, learning rate, cost function, regularization parameter, learning algorithm and maximum validation failure are considered as ANN hyperparameters. In the third approach presented here, a GA is utilized for the hyperparameter optimization of the neural networks.

A genetic algorithm is a strategy of evolutionary computation search algorithms which states that individuals in a population who are best fitted are more likely to survive and reproduce.

In GAs, a chromosome is a set of parameters that defines a proposed solution to the problem that the GA is trying to solve by searching through the space of possible chromosome values. In this work, the hyperparameters of a network are the chromosome of one individual. The major steps of a GA are described in Whitley (1994).

In the case studies presented here, the three hyperparameters (genes) of chromosomes and their range of values are: (i) the number of hidden neurons (1 to 20), (ii) the Levenberg–Marquardt (LM) algorithm parameter (μ) (0 to 1), and (iii) the maximum validation failures (1 to 20).

The relative root mean square error (RRMSE) is used as the fitness function, and three selection operators are employed to select the most fitted individuals as the first and second parents to go through the crossover, namely Roulette wheel, Tournament, and Random selection (Zhong et al. 2005). A crossover percentage of 0.8 and a mutation percentage of 0.3 are used. The number of individuals in an initial population is set to 20. Also, the maximum number of iterations is used as stoppage criterion for GA optimization and is set to 50. The flowchart of the novel ANN model with GA optimization for network hyperparameters is presented in Figure 3.

The autocorrelation function (ACF) and the partial autocorrelation function (PACF), as defined by Box & Jenkins (1976), were used to identify the most appropriate time series model for the dataset (Yang et al. 2013; Arandia et al. 2015). For both case studies presented in this paper, the ACFs decayed slowly with increasing time lags. The PACFs showed a large spike in the first lags and cutoff to 0 after lags 25 and 17 for Cities 1 and 2, respectively. The above observations suggested a non-stationary process for the two datasets. An example of the ACF for 672 lags (time step of data equal to 15 min) is presented in the Supplementary Material (Figure S-1) for the dataset from City 2, in which the periodic behaviour of the water consumption can be seen. With a cycle of positive and then negative values every 96 lags, the observations show the correlation between the data and exhibit the daily seasonality of the water consumption data.

Values of p₁, d and p₂ were determined by analysing the data of the training and validation sets. Through the differentiation process, the trend was removed from the autocorrelation functions and the datasets transformed into stationary series. Indeed, ACFs and PACFs of the differentiated series tend more quickly to reach a value near zero than those of the original series (see Figures S-2 to S-5 in the Supplementary Material). Various orders of models have been tested for the AR and MA processes, however, results are not presented here for brevity. It was concluded that ARIMA(2,1,1) provides satisfying results for both case studies (City 1 and City 2) and was further adopted for the following studies. The test set used to evaluate the performance of the ANN model is the same as the one used for the other models (see section below). Values of the and parameters were calibrated at each time step using the previous 192 data (i.e. a total of 2 days of data) with the estimate function in MATLAB, which is based on the maximum likelihood.

The model developed by Bakker et al. (2013) is the pattern-based model chosen for the applications presented here. This forecasting method combines the daily average estimation of the UWD with the demand pattern to provide 15 min predictions over the following 24 h (only the 15 min ahead predictions are presented and discussed in this paper). The model analyzes the historical data series to determine different factors (f_dotw,typ,i: typical day of the week factor and f_qtr,typ,i.j: typical 15 min time step factor). By exploring the available database, the method defines the specific factors for the seven ordinary days of the week and for each particular day of the year, and the 15 min demand pattern corresponding to each one of these days.

The model was coded using MATLAB 2014a software by considering, as most as possible, the steps and parameters described in Bakker et al. (2013) without any specific consideration of the sprinkle demand. The same method as the one described in Bakker et al. (2013) was applied for the factors calculation. The model will thereafter be called the fully adaptive forecasting (FAF) model. The test set used to evaluate the performance of the ANN model is the same as the one used for the other models (see Performance indicators below).

The accuracies of the different models were evaluated using the following three statistical indices, namely the RRMSE, the Nash–Sutcliffe model efficiency coefficient (E) and the mean absolute percentage error (MAPE) for the test sets of City 1 and City 2 (i.e. respectively from 1 January 2013, 00:00 to 31 December 2013, 11:45, and from 1 January 2014, 00:00 to 21 July 2014, 23:15; see Figure 1). The measures selected to compare the forecasted and measured values are the ones most commonly used by researchers addressing UWD forecasting (Adamowski et al. 2012; Bakker et al. 2013) and they all generate dimensionless outputs. The equations used to compute the values of these indices are given in Table 1, where N is the total number of forecasted values, C_t is the measured value at time t, is the forecasted value at time t, and is the mean of the measured values.

The performance indices (RRMSE, MAPE and E) of the studied models for the test sets of the two case studies are presented in Table 2. Examples of results are illustrated for two specific days in Figure 4 for City 1 and City 2.

The ANN hyperparameter values obtained by GA optimization were as follows: number of hidden neurons = 18, LM parameter μ = 0.4632, and maximum validation failures = 9 for City 1; and number of hidden neurons = 16, LM parameter μ = 0.4307, and maximum validation failures = 3 for City 2. As can be seen in Table 2, the ANN models provide better 15 min UWD predictions (i.e. lower RRMSE, lower MAPE, and higher E values) than the ARIMA and FAF models for both case studies. Moreover, optimization of the hyperparameters of ANN with GA allows refinement of the predictions. It is also worth mentioning that the statistical indicators for City 1 were always better than those for City 2. It can be thought, as suggested by previous studies (e.g. Maidment & Miaou 1986; Bakker et al. 2013; Gagliardi et al. 2017), that the performances of UWD predictions vary depending on the size of the water distribution area. Indeed, when the size of the area increases the total consumption also increases, and its fluctuations are generally mitigated (see Figure 2), making future values easier to predict. It can additionally be observed that that the multiple ANN training approach has led to better results compared to the single ANN training for City 1 than City 2. Again, it seems that the size of the case study has a direct impact on the performance of the models. However, the GA-optimized ANN showed the best performance in both cities compared to other models and proves its reliability in different circumstances.

The presented results show that ANN models outperformed the classical time series and pattern-based models in forecasting short-term UWD. Results in Figure 4 also show that the ARIMA model provided predictions that are delayed from the measured data. This result is typical of low order ARIMA models, in which the projections represent a weighting of the last observations. Better performances of ANN models over time series and linear and nonlinear regression methods have also been presented by Jain et al. (2001) and Adamowski et al. (2012). There are varying degrees of nonlinearity in UWD data that make them difficult to be handled by linear methods. The strengths of ANN models can rely on their inherent ability to capture the nonlinearities related to the UWD time series.

Another finding of this study was that hyperparameter optimization of the ANN model could enhance its prediction performance. This supports the findings of Romano & Kapelan (2014) for the prediction of UWD 1 to 24 h ahead. These authors reported Nash–Sutcliffe efficiencies higher than 0.9 for their adaptive ANN models for both daily and hourly forecasting. Their optimization procedure included six decision variables, namely the number of hidden neurons, the number of training cycles, the training algorithm regularization factor, the lag size, the time of the day and the day of the week.

The 15 min UWD predictions obtained by different models were compared in this paper, based on data collected from two different Canadian WDSs. Considering the intention of developing real-time control tools: i) the models that were compared were exclusively univariate time series models, using only the records of previous UWD as input data to predict the future demand, and ii) the comparison of the models was based on their ability to provide short-term forecasts of UWD. An original model combining ANN and GA, for optimization of the ANN model hyperparameters, was proposed, showing its superiority in providing more accurate 15 min UWD predictions than ARIMA and pattern-based models. However, although the ANN model provided better 15 min UWD predictions than the ARIMA and the tested pattern-based models for the two presented case studies, many authors showed that pattern-based models provided better predictions for longer lead times (e.g. from about 3 h or more in Doghri 2019). Moreover, the efficiency of the GA optimized ANN model should be verified with other consumption datasets and different forecast horizons, in order to validate the obtained results and to generalize these findings. Finally, since coupled wavelet-neural network models (WA-ANNs) have shown good potential to predict UWD (Adamowski et al. 2012), a comparative study between the GA optimized ANN and WA-ANN models for UWD prediction could be useful.

The data used in this paper were provided by two cities from the province of Quebec, Canada, which are not named here for confidentiality reasons, but are gratefully acknowledged. Financial support to Majid Gholami Shirkoohi and to Mouna Doghri was provided by the National Sciences and Engineering Research Council of Canada (Gholami and Doghri) and the CREATE TEDGIEER program (Gholami).

Data cannot be made publicly available; readers should contact the corresponding author for details.