Urbanization and population growth have resulted in an imbalance in water supply and demand, and limited water resources have made it more difficult to manage Water Distribution System (WDS). Smart Water Grid (SWG) technology has emerged as a method t...
Urbanization and population growth have resulted in an imbalance in water supply and demand, and limited water resources have made it more difficult to manage Water Distribution System (WDS). Smart Water Grid (SWG) technology has emerged as a method to address these problems, as it provides real-time or near-real-time water-related information to managers and customers in both directions throughout the supply and demand process. In particular, by measuring and utilizing customers water consumption data, it can help improve the management efficiency of the WDS and reduce both energy and water leakage. However, the existing statistical methods are limited in their ability to analyze the collected water consumption data as big data. As an alternative, Machine Learning (ML) algorithms have emerged as modern methods for processing such big data, and they have been shown to be superior to existing statistical methods. This indicates that algorithms can be applied to the water consumption data collected from smart water grids to obtain various analysis results that cannot be obtained using the existing statistical methods.
There are also rapid advancements being made in intelligent water management methods using this, so there is a need for a very accurate short-term WDF for monitoring and controlling the WDS and detecting leaks in the pipeline in real time. Therefore, in the present research, a short-term WDF framework is proposed that applies ML algorithms to the analysis of real-time water consumption data collected from SWM. This analysis used data collected from 00:00 on January 1, 2018, to January 1, 2020, from 527 SWMs installed in a pilot plant built in Block 112 of YeongJong Island, South Korea by the SWGRG. The results are as follows:
First, by using filtering (local minima and Min. Max. allowable discharge) and the one-class SVM model for outlier detection, it is possible to pre-process raw data more scientifically and rationally. It also has the advantage of imputing, as even if all the hourly water consumption datasets by the day are missing, it uses combined k-NN and FTT for imputation of the missing values. Second, the k-means clustering model was proposed as a method of estimating the water consumption using mechanical water meters. With this method, eight clusters were divided according to the usage and the pipe diameter, and the average water demand pattern was estimated properly. Lastly, the MLSTM model that can consider a number of variables (weather, holiday effects, etc.) was applied to the short-term WDF. As comparative models, the avg. WC, ARIMA, ANN, QMMP+, and LSTM models were used. However, the calculation results were poor compared to those of the LSTM model considering water consumption alone, so there was no need to consider some variables in the calculation. Meanwhile, the noise generated during the calculation process can be included in the forecasted results; further, the Kalman filter was applied to increase the reliability, and the effect was excellent in improving the results.
The assessment of the performance of short-term WDF using the SWG and ML algorithms with hydraulic modeling based on water consumption data indicates that it is still difficult to accurately forecast water demand for actual operation. However, it is believed that this proposed framework will be helpful for WDS operation and management as a guiding tool. For future research, as water consumption is closely related to the human life cycle, higher-resolution data are to be collected from each customer's taps and analyzed.