CNIT (Consorzio Nazionale Interuniversitario per le Telecomunicazioni), ITALY
Prof. Francesca Cuomo, Prof. Ioannis Chatzigiannakis, Prof. Ilenia Tinnirello, Ph.D Domenico Garlisi, Mr. Gabriele Restuccia.
Water management and LPWAN integration
Optimization and digitalization of Water Distribution Networks (WDNs) are becoming key objectives in our modern society. Global water consumption is constantly increasing, with a huge demand increment each year. On the other side, the world is facing a global water deficit, which is foreseen to be about 40% by 2030. Climate change is going to make the situation worse. Current changes in temperatures, storms and rain behaviors urgently require a better management of water distribution between different regions. WDNs are typically old, worn and obsolete. These inadequate conditions of the infrastructures lead to significant water losses due to leakages inside pipes, junctions and nodes. In Europe, it has been measured that the mean value of wasted water is equal to 26% of the total consumption [1]. Leakage control in current WDNs is typically passive: leaks are repaired only when they are visible.
Low Power Wide Area Network (LPWAN) technologies, and especially LoRaWAN [2], are becoming consolidated connectivity solutions for deploying applications devised to monitor and control large scale smart systems. Moreover, most of the current Smart WDNs (SWDNs) solutions just collect measurements from the end devices and then send the data to the cloud servers in order to execute the desired analyses, in a centralized way. Thanks to the ELEGANT paradigm we can design new solutions to improve monitoring, administration, leak management and leak prediction, by exploiting edge processing capabilities inside SWDNs.
The ELEGANT paradigm for Smart Water Systems
The ELEGANT paradigm fits the requirements of Smart Water Systems in terms of energy efficiency, continuum computing and intelligent analysis.
Figure 1 depicts the considered architecture where smart sensors are deployed throughout the WDN in certain points with strategic importance such as water tanks, junctions, pumps or end users where the water demand happens. These smart sensors are wirelessly interconnected via LoRaWAN to some Gateways (GW) that in turn are interconnected to network servers. Both at the GW and the network servers follow the ELEGANT paradigm seamlessly combining Big Data and Edge computing functionality to support the smart water distribution management.
Figure 1 – LoRaWAN platform and its integration in the ELEGANT framework for water management
Enabling Stream-based Machine learning analysis
The data collected from the smart sensors are used to assist in performing critical management tasks including leak detection, water flow prediction and redirection. The execution of the ML solutions is done continuously on the stream of data arriving from the sensors. Following the ELEGANT paradigm, the optimal point of execution of the ML algorithms is selected automatically by the ELEGANT framework. The optimal selection point is based on the processing, storage and energy resources available throughout the network infrastructure.
To demonstrate the benefits of the ELEGANT approach we evaluate the performance of a pre-built network offered from the Open Water Analytics's community public repository for testing [3]. We consider three different sub-networks named A, B, and C with an increasing amount of total nodes (83, 1038, and 2099) to measure the scalability of the resulting system. Figure 2 depicts the three subnetworks where blue bullets represent nodes of the network and circles represent the coverage area of each GW included in the LoRaWAN infrastructure.
Figure 2 – Water Distribution Nodes and their LoRaWAN coverage.
We considered the leakage detection case and we tested different classifiers: KNeighborsClassifier, LinearSVM, RBFSVM, DecisionTree, RandomForest, AdaBoostClassifier, GaussianNaiveBayes. Performance results are shown in Figure 3 in terms of accuracy achieved by different classification schemes as well as of the considered time period of the training data set. On the left (Figure 3.a), we have the case of a shorter time period (one week), while on the right (Figure 3.b) we have a longer time (one month). We observe that even for the more complex network topologies B and C, longer traces are able to resolve the majority of topological ambiguities in identifying the leakage location and achieve very high accuracy (as high as 98% for a model based on Decision Trees).
Figure 3 – Leakage detection Accuracy in different networks and with different time periods training (left one week, right one month).
Reference
[1] Europe’s Water in Figures An overview of the European drinking water and waste water sectors 2021 edition EurEau ISBN 978-2-9602226-3-0
[2] LoRa Alliance Technical Committee. LoRaWAN 1.1 Specification. https://lora-alliance.org/resource-hub/lorawantm-specification-v11, 2017.
[3] https://raw.githubusercontent.com/OpenWaterAnalytics/epanet-example-networks/master/epanet-tests/large/NW_Model.inp