According to data, the total cost of water losses globally amounts to $15 billion, with two-thirds of these water losses being from low and middle-income countries. A lot of this could be attributed to leaks in the water distribution network, unauthorised consumption, and poor metering, which cost utilities over $4 billion each year. In fact, due to these losses, it has been estimated that the demand of water for Mumbai city will go up 71% from the current value, which could supply the needs of 16 cities like Mumbai.
Thus, in an attempt to identify and reduce these water losses, Smartterra, has built an AI-powered operational intelligence system with the help of classic machine learning algorithms and geospatial data. Led by co-founders Gokul Krishna Govindu and Giridharan Sengaiah, this solution has been the first pilot in India not only to reduce water losses but also ensure water availability in developing cities. The company further aims to reduce utilities’ carbon footprints, which will also help in minimising the electricity bills of consumers.
The Hurdles Faced
While developing the system, Smartterra, firstly, combined the disparate data sources of SCADA, CRMs, meter data management systems, geospatial data stores, maintenance records, and revenue/billing systems to extract patterns from data, which was previously impossible. “However, with the help of the developed algorithm, it was easy for us to identify subtle patterns of loss in both the network and in water consumption,” said Gokul.
Secondly, there were critical data gaps in these sources, which required to be addressed before feeding them into the AI algorithms. The company leveraged a combination of analytical techniques and subject matter expertise to fill these data gaps.
Thirdly, the company had to integrate external data sources like satellite imagery, open-source mapping with traffic, urban activity, soil analysis to gain a 360-degree view of the problem being faced.
Explaining further, Gokul stated, “Often, analyses of this nature are only carried out periodically since it is a time-consuming and expensive exercise. We wanted to do things differently, and therefore build a solution that would allow cities to continuously evaluate their networks, using feedback to up the accuracy of our models and provide predictions of improving quality over time.”
The Tech Behind
To address the challenges, Smartterra took a three-step process. “At the topmost layer, we have a customised AI-powered analytics and monitoring system, which targets deficiencies in the data, identifying connections and network elements like pipes, pumps, valves, and stores to investigate anomalous behaviour,” explained Gokul. To facilitate this, the company had to address the critical feature of contextual cohorts to compare the identified anomalies with similar “normal” elements.
Case in point — once the artificially intelligent machine tags a faulty meter, the company offers a comparison to “similar” consumers with working meters. This enables users to easily see which ones deviate most from their cohorts and therefore need to be prioritised. Keeping the explainability factor in place for improving decision-making, the company used techniques like SHAP, CORELS, and RuleFit to either directly build explainable models or explain existing models.
Underneath the analytics layer is the unified data model, which consumes data from utilities periodically on various aspects of their operations, like supply, consumption, connection details, geospatial maps, and network information. Internally, the inputs are matched with custom data, adding necessary context for the city or region covered by the utility.
“We also have custom data cleaning and preprocessing modules that ensure the quality of the data. This holistic view of the data allows us to use different sources effectively,” said Gokul. “We incorporate urban dynamics through land use, satellite imagery, and terrain/soil data, which helps us capture the nuances in city activity, socio-economics, and local variability.” Leveraging geospatial data provides an edge in the interpretation of water consumption behaviour, demand projections, and network planning.
Finally, the company leveraged MEAN (Mongo – Express – Angular – Node) stack to design workflow engines so that the users can now create tasks for on-ground verification of the given predictions. This allows analysts to push results to field operators who can determine whether a fix is required or not. Further, these operators provide additional data, as part of a continuous learning system, where feedback from on-ground investigations is put back to models in order to improve them. The workflow engine ensures that the AI system is a full-cycle from data to models to on-field verification and back to data.
The company used a variety of models to build the solution, where they started with classic machine learning algorithms that have been demonstrated to work well, such as RandomForest (classifiers and regressors), support vector machines, and gradient boosted trees. Additionally, Gokul is also exploring using long short-term memory and recurrent neural network models, with TensorFlow and Keras, to take advantage of time-series data.
Benefits & Future Plan
With the initial deployed pilot project of the AI-powered analytics, the company was able to identify several obstacles to precise water usage, like incorrect billing to customers, faulty meters and mismatches in the data. The AI system showed that there was a 3.2% potential increase in monthly revenue using just “soft” fixes that were relatively inexpensive. Also, “the results were validated by an on-ground survey that confirmed that predictions made by the AI models were 80% accurate,” stated Gokul.
Broadly, these results were aimed to help operators reduce water losses in their networks, resulting in more availability of water to required citizens. The solution has been designed to ensure that the residents of the developing cities receive consistent, high-quality water supply so that they can focus on building better lives. The company is also in talks with various utilities in India and South-East Asia to develop solutions to reduce their losses. “We hope to have several pilot projects underway in the coming months,” concluded Gokul.