Comparative analysis of neural network techniques for predicting water consumption time series

doi:10.1016/j.jhydrol.2010.01.005

Journal of Hydrology

Volume 384, Issues 1–2, 15 April 2010, Pages 46-51

https://doi.org/10.1016/j.jhydrol.2010.01.005 Get rights and content

Abstract

Monthly water consumption time series have been predicted using a series of Artificial Neural Network (ANN) techniques including Generalized Regression Neural Networks (GRNN), Cascade Correlation Neural Network (CCNN) and Feed Forward Neural Networks (FFNN). One hundred and eight data sets for the city of Izmir, Turkey are used for a number of ANN modeling exercises. Several ANN models depending on the combination of antecedent values of water consumption records are constructed and the best fit input structure is investigated. The performance of ANN models in training and testing stages are compared with the observed water consumption values to identify the best fit forecasting model based upon a number of selected performance criteria.

Introduction

Water demand forecasting is essential for the planning and management of water resources. There are several factors affecting the demand for water. These factors include population, rainfall, humidity, temperature, industrial and commercial conditions. A reliable forecast is expected to include the effects of such socio-economic and climatic factors. The conventional approaches such as regression techniques and time series models have been used to forecast future water demand. Lahlou and Colyer (2000) used multiple regression demand models to forecast water demand incorporating water conservation activities. Froukh (2001) used mathematical and heuristic approaches for long-term water demand forecasts in a decision-support system framework integrating demand management into demand forecasting. Wong and Mui (2007) constructed a mathematical model to determine flushing water demands for high-rise residential buildings in Hong Kong. Babel et al. (2007) developed a multiple coefficient water demand forecast and management model for the domestic sector considering various socio-economic, climatic and policy-related factors. Jain et al. (2001) used autoregressive models to forecast peak weekly water demand. Zhou et al. (2000) proposed time series models to forecast daily water consumptions based on trend, seasonality, and climatic correlation and autocorrelation components.

Artificial Intelligence techniques including Artificial Neural Networks (ANN) and Fuzzy Inference Systems (FIS) have been in use for some time as an alternative tools for predicting complicated water systems. Some specific ANN applications to water resources systems include rainfall-runoff modeling (Kumar et al., 2005), hydrologic time series modeling (Jain and Kumar 2007), runoff and sediment yield modeling (Turan and Yurdusev, 2009, Agarwal et al., 2006), water demand forecasting in irrigation delivery system (Pulido-Calvo et al., 2007). Unlike many hydrological applications, it is observed that there are limited applications of ANN to water consumption modeling. Liu et al., 2003, Griñó, 1992 applied ANN techniques to forecasting water demand in urban areas. Bougadis et al. (2005) investigated the relative performance of regression, time series analysis and ANN models for short-term peak water demand forecasting. Firat et al. (2009a) used several ANN techniques to model monthly water consumptions based on several socio-economic and climatic factors. The same data is also used by Yurdusev and Firat (2009) to develop an Adaptive Neuro Fuzzy Inference System. There appeared to be a need to model the monthly water consumption values used in Firat et al. (2009a) as time series. Firat et al. (2009b) performed a comparative modeling exercise using two Fuzzy Inference Systems. There left to be an obvious need to compare the ANN techniques for modeling monthly water consumption time series based on how well they perform. This study attempts to perform this task. The ANN techniques considered includes Generalized Regression Neural Network (GRNN), Cascade Correlation Neural Network (CCNN) and Feed Forward Neural Network (FFNN). The models based on the combinations of the antecedent water consumption records are constructed and the best fit input structure is investigated. The performance of ANN models in training and testing sets are compared with the observations and the best fit model forecasting model is identified.

Section snippets

Artificial Neural Networks (ANN)

ANN inspired by using studies of biological neural system is composed of processing elements called neurons or nodes. In literature, different types of ANN methods are used for forecasting and modeling of engineering problems. Three ANN techniques are considered in this study, for which brief descriptions are provided as follows.

A FFNN consists of at least three layers, input, output and hidden layers. The input signals presented to the system in input layer are processed in forward through to

Study area

The values of the metropolitan area of Izmir, Turkey are used for modeling monthly water consumptions. Based on her population and industrial development, Izmir is the third largest city of Turkey. Her population is rapidly growing due to the domestic population movement to the city from the disadvantageous parts of the country. The water resources of the city are scarce as it is located in the west coast of the country where the average temperature is significantly high whereas the rainfall is

Data set

For the development of the prediction models, the total 108 monthly data records of water consumption were collected in the period 1997–2005 for the city of Izmir, Turkey (Fig. 2). The data set was divided into two subsets, training and testing sets. The statistical parameters such as minimum value (x_min), maximum value (x_max), mean ( $\bar{x}$ ), standard deviation (s_x) and skewness coefficient (c_sx) for training and testing data sets are calculated and given in Table 1 to provide a comparison of the

Results and discussion

The models given in Table 2 were trained and tested using associated data sets by three different ANN methods, GRNN, CCNN and FFNN. The performances of models are shown in Fig. 4.

Comparing the results of models, it is seen that performances of M5 models consisting of the combination of five antecedent values of water consumption records are better than other models. The NRMSE and AARE values of M5 GRNN, CCNN and FFNN models are higher than those of other models. According to these criteria, M5

Conclusions

The capability of Artificial Neural Networks to model water consumption time series is tested in this study. The ANN techniques considered include Generalized Regression Neural Networks (GRNN), Cascade Correlation Neural Networks (CCNN) and Feed Forward Neural Network (FFNN). For this purpose, six models based on several combinations of previous monthly water consumption values are constructed. Comparison of model results have shown that the performance of the M5 model consisting of five

References (20)

A. Agarwal et al.
Simulation of runoff and sediment yield using artificial neural networks
Biosystems Engineering
(2006)
M. Firat et al.
Comparative analysis of fuzzy inference systems for water consumption time series prediction
Journal of Hydrology
(2009)
A. Jain et al.
Hybrid neural network models for hydrologic time series forecasting
Applied Soft Computing
(2007)
B. Kim et al.
Prediction of plasma etching using a randomized generalized regression neural network
Vacuum
(2004)
J. Liu et al.
Forecast of water demand in Weinan City in China using WDF–ANN model
Physics and Chemistry of the Earth
(2003)
I. Pulido-Calvo et al.
Linear regressions and neural approaches to water demand forecasting in irrigation districts with telemetry systems
Biosystems Engineering
(2007)
M.E. Turan et al.
River flow estimation from upstream flow records by artificial intelligence methods
Journal of Hydrology
(2009)
L.T. Wong et al.
Modeling water consumption and flow rates for flushing water systems in high-rise residential buildings in Hong Kong
Building and Environment
(2007)
M.A. Yurdusev et al.
Adaptive neuro fuzzy inference system approach for municipal water consumption modeling: an application to Izmir, Turkey
Journal of Hydrology
(2009)
S.L. Zhou et al.
Forecasting daily urban water demand: a case study of Melbourne
Journal of Hydrology
(2000)

There are more references available in the full text version of this article.

Cited by (88)

Short-term wave power forecasting with hybrid multivariate variational mode decomposition model integrated with cascaded feedforward neural networks
2024, Renewable Energy
Wave power is an emerging renewable energy technology that has not reached its full potential. For wave power plants, a reliable forecast system is crucial to managing intermittency. We propose a novel robust short-term wave power (Pw) forecasting method, MVMD-CFNN, based on a multivariate variational mode decomposition hybridized with cascaded feedforward neural networks. By using cross-correlation, we were able to determine the significant input predictor lags. To overcome the non-linearity and non-stationarity issues, the proposed MVMD method is then used to demarcate the significant lags into intrinsic mode functions (IMFs). To forecast the short-term PW, the MVMD-CFNN model incorporated the IMFs into cascaded feedforward neural networks. Validation and benchmarking of the MVMD-CFNN model at two stations in Queensland, Australia has been conducted against standalone cascaded feedforward neural networks, boosted regression trees, extreme learning machines, and hybrid models, MVMD-BRT and MVMD-ELM. According to the results, the MVMD-CFNN predicts PW accurately against the benchmark models. The outcomes of this research can contribute to the application and implementation of clean energy worldwide for sustainable energy generation.
Boruta extra tree-bidirectional long short-term memory model development for Pan evaporation forecasting: Investigation of arid climate condition
2024, Alexandria Engineering Journal
In this study, two deep learning approaches, bidirectional long short-term memory (BiLSTM) and long short-term memory (LSTM), were used along with adaptive boosting and general regression neural network to forecast multi-step-ahead pan evaporation in two arid climate stations in Iran (Ahvaz and Yazd). Lagged time series of meteorological data and pan evaporation data were input to the machine learning models. Two feature selection methods, i.e., the Boruta extra tree and XGBoost, were used to select significant inputs to reduce the number of inputs and model complexity. Different statistical metrics were used to investigate the model performance. The results demonstrated that Boruta-extra-tree-based models were more accurate than XGBoost-based models. Compared with the machine learning techniques, the combination of Boruta extra tree and BiLSTM enabled more accurate one-day-ahead forecasting of pan evaporation for both sites (Root Mean Square Error (RMSE) = 1.6857, for the Ahvaz station, and RMSE = 1.3996 for the Yazd station). The proposed model was used to forecast up to 30 days ahead of pan evaporation in both stations. The results showed that the Boruta-BiLSTM model could accurately forecast the pan evaporation for up to 30 days in both stations.
An automatic energy saving strategy for a water dispenser based on user behavior
2022, Advanced Engineering Informatics
Nowadays, numerous public buildings provide water dispensers to supply drinking water which causes more energy consumption. A typical water dispenser periodically heats and cools the water to ensure that hot, warm, and cold water are always available for the user. However, this mechanism is inefficient because the users do not request hot and cool water continuously. Ideally, the boiling and cooling schedule should follow the demand pattern to save electricity consumption. When no demand, a water dispenser can enter a sleep mode.
Therefore, this study presents an automatic energy-saving strategy for a water dispenser based on user behavior. The proposed system allows the water dispenser to automatically determine the appropriate time to heat, boil, and enter sleep mode based on user behavior. The proposed control strategy involves several steps. First, it collects historical data, analyzes water consumption behavior. The sensors installed in the water dispenser collect water consumption data. Second, this study applies Recurrent Neural Networks with Long-Short Term Memory to predict future water consumption. Finally, the proposed system utilizes the prediction result to determine heating, cooling, and sleep mode schedule.
This study uses a water dispenser on a university campus as a prototype to test the proposed system. The effectiveness of the proposed system is measured by two factors, namely electricity consumption, and customer satisfaction. These two parameters are chosen because the proposed system should reduce electricity consumption while maintaining hot and cold water availability whenever needed. According to the simulation results, the proposed controlling strategy can reduce electricity consumption up to 28% monthly while maintaining a service level of 97%. This result shows that the proposed system is a good control system for water dispensers. By applying this controlling system, public buildings could reduce their energy bills without sacrificing their provision of drinking water.
Estimation of air-flow parameters and turbulent intensity in hydraulic jump on rough bed using Bayesian model averaging
2021, Applied Soft Computing
Citation Excerpt :
For more details about the experimental setup, refer to Felder and Chanson [4]. Three widely-used multilayer perceptron (MLP), support vector regression (SVR) and generalized regression neural network (GRNN) classifier models as individual models were implemented [e.g. 26–31]. First, the architecture of each model was determined based on trial-and-error.
A hydraulic jump is an abrupt transition between subcritical and supercritical flows which is associated with energy dissipation, air entrainment, spray, splashing, and surface waves. Both physical and numerical modeling were largely applied to study hydrodynamics, turbulence and air-entrainment in the hydraulic jump, while the literature about the application of classifier models is quite limited. Determining air-flow parameters and turbulent intensity has been merely performed by costly and time-consuming experimental methods, while this study is the first attempt to estimate the mentioned parameters using a computer-based methodology with desired precision. In the present study, air-flow parameters including void fraction (C) and bubble count rate (F), as well as turbulent intensity (Tu) on rough bed were estimated using Bayesian model averaging (BMA) and three multilayer perceptron (MLP), support vector regression (SVR) and generalized regression neural network (GRNN) as classifier models. To develop the stated models, the experimental data from Felder and Chanson (2016) were divided into four classes based on longitudinal distance from the jump toe. Results highlighted that the MLP and GRNN models have more accurate results compared to the SVR model. For F and Tu, the GRNN model and for C, the MLP model showed better performance than other models in four classes. The average acceptance rate between 15 and 30% of the BMA model performance for all classes proved the accuracy and efficiency of the proposed methodology. The average RMSE value of BMA results and the bests classifier models were 0.41 and 0.42, respectively, for the estimation of all three parameters. Results revealed that the BMA model by weighting individual classifier models could be able to estimate parameters with better accuracy than the best classifier model in each class. The significant outcome of this study is that the proposed model is able to render accurate results in a complex system such as hydraulic jump.
Air compressor load forecasting using artificial neural network
2021, Expert Systems with Applications
Air compressor systems are responsible for approximately 10% of the electricity consumed in United States and European Union industry. As many researches have proven the effectiveness of using Artificial Neural Network in air compressor performance prediction, there is still a need to forecast the air compressor electrical load profile. The objective of this study is to predict compressed air systems' electrical load profile, which is valuable to industry practitioners as well as software providers in developing better practice and tools for load management and look-ahead scheduling programs. Two artificial neural networks, Two-Layer Feed-Forward Neural Network and Long Short-Term Memory were used to predict an air compressors electrical load. Compressors with three different control mechanisms are evaluated with a total number of 11,874 observations. The forecasts were validated using out-of-sample datasets with 5-fold cross-validation. Models produced average coefficient of determination values from 0.24 to 0.94, average root-mean-square errors from 0.05 kW - 5.83 kW, and mean absolute scaled errors from 0.20 to 1.33. The results indicate that both artificial neural networks yield good results for compressors using variable speed drive (average R² = 0.8 and no naïve forecasting), only the long short-term memory model gives acceptable results for compressors using on/off control (average R² = 0.82 and no naïve forecasting), and no satisfactory results are obtained for load/unload type air compressors (models constituting naïve forecasting).
Forecasting municipal solid waste quantity using artificial neural network and supported vector machine techniques: A case study of Johannesburg, South Africa
2021, Journal of Cleaner Production
Detailed prediction of the amounts of municipal solid waste (MSW) is very crucial for planning and management of MSW in a sustainable manner. Forecasting of MSW quantity is usually very challenging owing to unavailability of data in the low-income countries (LCs) and where data are available, they are often unreliable. The aim of this study is to forecast MSW generated in the City of Johannesburg (CoJ), South Africa with the projection period in continuing guesstimates by using machine learning approach. Two of machine learning algorithms namely: artificial neural network (ANN) and supported vector machine (SVM) were employed to forecast the quantity of MSW that would be generated in the CoJ. The forecast was based on historical data obtained from Statistics South Africa (STATS SA) and the projection was made up to 2050. The data pre-testing and incorporation structure was built in MATLAB simulation software to generate datasets having satisfactory information capacity and characteristic designed for modeling. From the result obtained, it was observed that machine learning algorithm is effective for the development of models for MSW forecasting. In the ANN models, the 10 neurons structure (ANN10) performed best with a determination coefficient (R2) of 99.9%, while in the SVM models, the linear model performed best with R2 of 98.6%. From the results obtained from the ANN10 model, the total amount of MSW generated per year in the City of Johannesburg is envisaged to get to 1.95 × 106 tonnes in 2050 with an average annual waste of 1.78 x 106 tonnes.

View all citing articles on Scopus

View full text

Comparative analysis of neural network techniques for predicting water consumption time series

Abstract

Introduction

Section snippets

Artificial Neural Networks (ANN)

Study area

Data set

Results and discussion

Conclusions

Biosystems Engineering

Journal of Hydrology

Applied Soft Computing

Vacuum

Physics and Chemistry of the Earth

Biosystems Engineering

Journal of Hydrology

Building and Environment

Journal of Hydrology

Journal of Hydrology