by Amy Vredevoogd
In municipal water management, maintaining aging infrastructure while predicting and meeting the ever-evolving needs of communities presents an ongoing challenge. Traditional approaches often struggle to keep pace with the complexities inherent in such a dynamic environment. These approaches frequently rely on the use of spreadsheets, physical filing cabinets, knowledge of experienced but aging workers, and siloed systems. However, amidst these challenges emerges hope in the form of Artificial Intelligence (AI).
Communities that embrace an AI strategy will witness the coming together of their knowledge bases, resulting in unified data repositories. These unified data repositories, combined with AI, provide insights and access to data that were previously difficult to achieve. Leaders will gain the ability to make informed decisions, leveraging the full spectrum of data to drive strategic initiatives and propel their water systems forward.
In this article, we embark on a journey to explore the potential of AI to revolutionize the water industry, starting by laying the foundation with a primer on AI basics. By equipping ourselves with this knowledge, we aim to discern how AI can be effectively applied within our water municipalities. Then, through a lens of identifying assets, data analysis, and data-driven predictions, we illuminate how AI can reshape the future of water management and offer opportunities for improved efficiency, sustainability, and resilience.
Figure 1. A neural network model, as illustrated here, is comprised of nodes organized into layers with the data flowing unidirectionally.
Smart Software Versus AI
Let's begin with the fundamentals: smart software operates within a deterministic framework, characterized by rules-based systems, predefined decision trees, scripted automation, and reliance on traditional statistical methods such as regression analysis. While this software can be intricate, it falls short of true AI. Unlike its smart counterpart, AI endeavors to emulate human cognitive processes. AI models possess the remarkable ability to discern patterns, make decisions, and devise solutions autonomously. What sets AI apart is its capacity to learn from data inputs and experiences, thereby dynamically altering outputs on its own. This transformative capability distinguishes AI as a groundbreaking paradigm in computing.
Machine Learning
Machine learning (ML) is one of the pivotal subsets of AI. It is defined as “the use and development of computer systems that are able to learn and adapt without following explicit instructions, by using algorithms and statistical models to analyze and draw inferences from patterns in data.” [1] In slightly simpler terms, ML enables a machine to learn without human intervention by ingesting and analyzing data, improving the outcome over time.
ML Architecture
A neural network is a computing model inspired by the neurons in the human brain. This neural network model (Figure 1) is comprised of nodes organized into layers, including an input layer, hidden layers, and an output layer. Data flows unidirectionally, entering the input layer, traversing through the hidden layers, and culminating in the output layer. Within each layer, nodes can receive input from multiple preceding nodes and transmit output to subsequent nodes. The connections between nodes depicted in the diagram represent the flow of information. Inputs undergo successive transformations as they pass through hidden layers, guided by initially randomized weights and thresholds. Through iterative adjustments based on input-output comparisons, these weights and thresholds are fine-tuned to achieve desired outcomes.
While most models feature between one and 10 hidden layers, neural networks with more than three layers are termed deep learning systems or deep neural networks. The fundamental architecture discussed here is a Feedforward neural network, the simplest type. However, various architectures exist to address diverse problem types. Some incorporate feedback loops, where output from hidden layers is fed back as input, facilitating complex algorithms for sequential data handling and consideration of previous model inputs.
Figure 2. Developing a robust ML model entails a structured process encompassing gathering data, preprocessing, model selection, training and testing, and evaluation.
ML Process
The usage of ML models permeates everyday applications such as mapping, translation apps, song recommendations, and facial recognition. However, developing a robust ML model entails a structured process (Figure 2) encompassing setup, selection, training, and refinement.
Commencing with data gathering, this crucial phase involves compiling diverse datasets ranging from images, databases, and text files of Internet of Things (IoT) data or freely available repository data, tailored to the specific project requirements. Preprocessing or cleaning the data follows suit, ensuring the integrity and quality of the input. Just as the efficacy of any software hinges on the quality of its data, these models demand meticulous data scrubbing to eliminate missing values, duplicates, outliers, or inconsistencies. Various strategies can address missing data, including removal, imputing median values, or prediction based on dataset characteristics. Establishing a pristine dataset ensures the model commences training with consistent input, laying a robust foundation for future stages.
The subsequent step entails selecting the most suitable model for the project and dataset, categorized into three main types shown in the table below.
Once the ideal model architecture is determined, the next phase involves training the model using the designated dataset, often partitioned into training and test sets. The training set teaches the algorithm, while the test set evaluates the model’s performance on unseen data. Following model training, the model is evaluated and parameters are fine tuned to optimize accuracy and precision. Finally, the tested and tuned model is ready for deployment and can be used wherever needed.
ML in Practice
Now that we’ve established the fundamentals of AI, let’s explore how these algorithms are pertinent to the water industry. Using the model types described above, here are a few ways they can be applied, shown in the table below.
Using AI as a Tool in the Water Industry
AI can easily be leveraged by water managers and operators, using existing or gathered data, to help implement desired solutions. These solutions may either entail off-the-shelf AI products from vendors or custom solutions tailored to specific circumstances. To understand the potential uses of AI in the water industry, we will break it down into three key stages: 1. Identifying assets, 2. Analysis of systems and data, and 3. Predictions.
ASSETS
Identifying
Figure 3. In this 1946 tie card, key fields are delineated by the green boxes. The system extracts information such as material type, address, size, and date directly from the tie card. Leveraging this digitized data, municipalities can now perform comprehensive searches, sorting, and filtering operations with ease.
AI technology presents a dynamic solution for asset identification within photographs and similar imagery. Specific AI models can discern the classification of objects such as pump stations, fire hydrants, or water towers using a probability score. Then, by detecting object edges, the software delineates the precise boundaries of those objects. This approach offers a means of extracting value from a municipality’s existing image repositories. After the collection of all relevant images, the next step involves processing the images through an AI recognition software. Once the model has processed all the asset images, users gain access to a wealth of data. For instance, municipalities can swiftly retrieve all images featuring hydrants or implement sorting and filtering mechanisms. Additional attributes like age or location can be appended as tags to the images, allowing for refined searches or filters based on specific criteria. The ability to use this data extends as far as one’s images and tagging capabilities allow, enabling detailed insights into asset distribution, age demographics, or geographic concentrations.
Optical Character Recognition (OCR), combined with AI for data extraction and categorization, is an indispensable tool for water utilities. This combination of technologies facilitates the conversion of text from images into machine-readable formats, offering a practical solution for data extraction. This process enables the scanning and digitization of textual information embedded within images, which can then be sorted, filtered, integrated into databases, or exported into spreadsheets for further analysis. A compelling municipal application of OCR combined with AI lies in the digitization of tie cards (Figure 3), a repository of valuable information typically housed in filing cabinets or boxes. By employing this technology, municipalities can unlock a wealth of data previously inaccessible due to its paper-based format.
When implementing this application, the initial phase involves collecting and scanning all tie cards, followed by configuring the software to identify key fields such as street, owner, and register number. After all the fields are identified, you will begin to train the model. This starts with processing a small subset of tie cards and then evaluating the model's accuracy in identifying the fields. If discrepancies are found, adjustments to the identification parameters can be made before processing a larger sample of tie cards. This process continues until satisfactory results are achieved. Upon completion, a municipality gains access to a digitized repository of tie card data, allowing for data analysis and management. Users can filter and sort records based on predefined fields, enabling targeted searches for specific criteria such as addresses associated with materials like lead. This technology empowers municipalities to extract insights from previously underutilized data sources, enhancing efficiency and informing the decision-making process.
Inspecting
AI software presents a solution for processing images and video footage of assets, enabling the identification of features and potential defects in these assets. Through analysis of diverse imagery sources such as ground-level shots, drone videos, and satellite imagery, AI algorithms can detect subtle differences. Additionally, these systems can interpret high-resolution video streams and closed-circuit television (CCTV) feeds, expanding their use across different media types.
A municipal application of this technology involves employing AI software to analyze video footage of underground pipes, effectively identifying defects with remarkable accuracy. These AI models leverage extensive training data, often uncovering issues that a person might not see. This software can search for one specific type of defect or identify multiple types at once. A sample of the assessable defects includes infiltration, cracks, sags, misalignment, tree root intrusions, collapses, and more. This capability streamlines the identification of specific pipe issues, facilitating targeted interventions rather than whole pipe replacements. Furthermore, by conducting periodic analyses of the same pipes, this technology provides a dynamic view of the pipes' structural integrity and evolving issues over time. This real-time insight into the condition of your pipes provides invaluable support for repair initiatives, future project planning, and overall asset management.
Representing
Utilizing AI software models offers a new approach to representing physical assets through innovative tools. One application is the creation of virtual replicas or mirrors of municipalities, constructed through multiple layers. At its base layer is the Digital City (Figure 4), a virtual rendition of a cityscape crafted from diverse datasets including aerial imagery, streetscape data, satellite imaging, and indoor mapping, each asset located with latitude and longitude coordinates. Progressing beyond static representation, the evolution continues with the Connected City. This transition necessitates the deployment of IoT devices onto tangible assets, enabling remote monitoring, analysis, control, and fostering an interconnected network of infrastructure. Culminating in the Intelligent City, the combination of AI software and the two foundational layers creates new capabilities for a municipality. This tier facilitates data visualization, analysis of both real-time and historical data streams, deployment of AI driven models created for urban contexts, and the execution of multivariate simulations, collectively harnessing the power of digital transformation. Through this progression, a Digital Twin of the municipality is created—a dynamic, digital counterpart ready to enhance the decision-making processes, improve operational efficiency, and forecast potential scenarios and pave the way for informed, proactive management strategies.
ANALYSIS
Optimize the Distribution Network
Innovative applications of AI, specifically employing ML reinforcement models, provide potential in optimizing municipal water distribution networks. Remember, we learned above that reinforcement models are provided the goal, but not how to achieve it. In this case, parameter, structure, or system optimization would be the goal. You would provide the AI model a collection of measurable parameters such as water usage, flow rates, system pressure, and water quality, as well as static details like structural and system information. Using these datasets, the model would run repeatedly, learning over time, and find areas for optimization and offer recommendations for enhancing the system. There are multiple benefits to this type of AI-driven optimization, including ensuring the efficient distribution of water, minimizing water loss, enhancing overall operational efficiency to reduce energy consumption and costs, evaluating the feasibility of new construction projects, and strategizing risk mitigation measures to improve network resilience and reliability.
Figure 4. A Digital Twin is a virtual rendition of a cityscape crafted from the layers of a Digital City, a Connected City, and an Intelligent City.
Figure 5. Incorporating AI empowered software as a component in leak detection offers the advantage of real-time awareness of potential issues.
Leak Detection
Incorporating AI-empowered software as a component in leak detection within water systems (Figure 5) offers the advantage of real-time awareness of potential issues. This approach involves the deployment of IoT devices with sensors across the water infrastructure. Once installed and operational, these specific sensors continuously collect data, furnishing a steady stream of real-time information concerning water flow rates and pressure. By utilizing a ML learning model capable of learning over time from the constant incoming data, the system gains the ability to distinguish between regular operations and anomalies indicative of leaks or other issues. This adaptive software effectively identifies deviations in water flow patterns, thus enhancing detection efficiency, curbing water loss, mitigating infrastructure damage, and ultimately reducing repair costs while promoting more sustainable water consumption practices.
Water Resource Management
The integration of AI into water resource management software presents a path for maximizing the return on investment in water resource management. By strategically deploying IoT devices with sensors for measuring water levels and other variables throughout watersheds, organizations can harness a wealth of data to drive informed decision-making. Through AI-powered analysis, such as illustrated in Figure 6, municipalities can gauge the influence of weather behavior and usage changes on their water network. Armed with this intelligence, they can dynamically optimize water allocation and consumption practices. This adaptive approach fosters agile management strategies, ensures the efficient utilization and preservation of vital water resources, and reduces costs as an added benefit.
Figure 6. Through AI-powered analysis, municipalities can gauge the influence of weather behavior and usage changes on their water network.
Water Quality Monitoring
Figure 7. Deploying water quality software using AI within water systems empowers municipalities to proactively detect threats to water quality before they escalate.
Deploying water quality software using AI within water systems empowers municipalities to proactively detect threats to water quality before they escalate (Figure 7). This approach involves deploying IoT devices equipped with sensors capable of measuring various parameters essential to water quality assessment, including dissolved oxygen, temperature, pH levels, and the presence of microorganisms and chemicals. Once integrated into the system, these devices enable simultaneous monitoring of contaminant levels across multiple locations. Leveraging AI-driven analytics, managers and supervisors can extract invaluable insights from the collected data, facilitating the timely identification of emerging risks. For instance, AI algorithms can detect patterns indicative of potential increases in harmful algal blooms, prompting preemptive action from system operators. By quickly addressing such threats, rather than reacting after they reach critical levels, municipalities can effectively safeguard water quality and ensure the uninterrupted delivery of safe drinking water to their communities.
Asset Management
The capabilities of asset management tools have grown when combined with advancements in AI. Leveraging asset management tools integrated with ML enables conducting risk-based analyses to predict the likelihood of asset failure. These ML models (Figure 8) employ standard asset data like size, material, age, and previous failures and combine them with a wealth of additional variables to create a smarter probability of failure. This data analysis plays a pivotal role in determining the criticality score of assets, enabling municipalities to effectively prioritize infrastructure management efforts. By leveraging such software, municipalities can optimize asset utilization to proactively manage and prioritize infrastructure assets to maximize their lifespan while minimizing costs through strategic repair and replacement initiatives.
Figure 8. ML models employ standard asset data like size, material, age, and previous failures and combine them with a wealth of additional variables to conduct analysis and make recommendations.
PREDICTIVE
Water Demand
Forecasting water demand presents a complex challenge for municipalities, necessitating the utilization of historical data, water resource information, and meteorological patterns for accurate conclusions. While manual methods may work, leveraging ML’s unsupervised learning models offers opportunities to identify patterns within large and multiple sourced datasets. As we learned above, unsupervised learning models are not provided the desired outcome - they conclude that by iterative learning over time. The data for this water demand learning model starts with the collection of current and historical water usage and resource data, which is then combined with diverse public sources encompassing weather patterns and precipitation data. After the data sources are provided to the model, these models generate a multitude of simulation scenarios, showing potential water demand outcomes under varied conditions such as droughts, water bans, alternative water sourcing, and inclement weather events. By examining these simulations, a municipality gains insights into the impacts on water resources, demand dynamics, revenue changes, and associated costs across different response strategies. Empowered by these predictive capabilities, a municipality can engage in informed decision-making, enabling proactive management, strategic planning, and timely interventions to meet future water demands sustainably and efficiently.
Figure 9. Leveraging ML models, this software excels at detecting subtle changes and patterns in water quality parameters, allowing it to predict potential changes over time.
Water Quality
As was discussed in the “Analysis – Water Quality Monitoring” section above, water quality monitoring can benefit from deploying IoT devices equipped with specialized water sensors at points within the water distribution network. Now moving from analysis to predictions, these data streams, complemented by current and historical imagery from satellites and comprehensive weather data, serve as inputs for software with AI-driven predictive models of water quality. Leveraging ML models, this software excels at detecting subtle changes and patterns in water quality parameters, even amidst vast datasets or diverse data sources, enhancing detection sensitivity beyond human abilities (Figure 9). By analyzing these patterns, the software generates forecasts of water quality parameter fluctuations over time, providing a municipality with invaluable insights into potential shifts in water quality. Armed with these predictive capabilities, a municipality can proactively respond to emerging challenges, ensuring the continued delivery of safe and high-quality water.
Flood Risk
An increasingly vital application of AI in the water industry is the prediction of flood events, a capability that can enhance disaster preparedness and response efforts. Advanced predictive ML models, as previously outlined, harness a diverse array of data sources tailored for the specific use. For flood prediction, these encompass real-time and historical data spanning weather forecasts, land dynamics, river morphology, drainage characteristics, and regional topography. By combining and analyzing these datasets, predictive models can forecast impending flood events with remarkable accuracy, often up to seven days in advance. In addition, these models can provide insights into the anticipated floodwater levels and the expected duration until water levels recede. Equipped with such predictive capabilities, a municipality can proactively prepare for potential flood events, implementing mitigation measures and mobilizing resources to safeguard lives and property. By leveraging AI-driven flood prediction models, regions can enhance their resilience to natural disasters and minimize the impacts of flooding on their communities.
Figure 10. Providing models with large amounts of data allows it to identify anomalies that may indicate potential issues such as abnormal pump behavior.
Maintenance for Water Supply
Echoing the applications we've already explored for using ML’s unsupervised learning model, it can also create valuable insights into predictive maintenance for water supplies. Providing these models with large amounts of data from sensors, assets, and monitoring systems allows the model to identify anomalies that may indicate potential issues such as abnormal pump behavior (Figure 10). By continuously examining real-time data streams and comparing them against historical data patterns, this ML-enabled software can predict failures and recommend preemptive maintenance measures. This proactive approach empowers managers to schedule maintenance tasks with precision, intervening when necessary to minimize downtime and mitigating the risk of costly issues. By facilitating decision-making grounded in data-driven insights, AI-driven predictive maintenance strategies optimize asset performance and enhance the resilience of water supply infrastructure.
Impact - Water Clarity Project Cape Cod
Let's explore a compelling real-life illustration of AI's impact on the water industry at a large scale. This case exemplifies how raw field data, collected over many years, can be harnessed to train AI models effectively.
Since 2001, data on water clarity, measured by Secchi disk depth (SDD), have been intermittently gathered from 217 ponds across Cape Cod (Figure 11). Supplementary data on maximum pond depth further enriched the dataset. Leveraging this extensive historical field data from 154 ponds, a sophisticated random forest ML model was meticulously trained. After rigorous testing and refinement, the model was employed to extrapolate SDD values from satellite imagery.
Long-term trends spanning from 1984 to 2022 were extrapolated for 157 ponds, revealing that 149 ponds exhibited improvements in water clarity, while eight ponds experienced deterioration. Short-term variations between 2021 and 2022 showcased a contrasting picture, with 96 ponds witnessing deteriorating water clarity, while no ponds displayed improvement.[2]
Figure 11. Since 2001, data on water clarity have been intermittently gathered from 217 ponds across Cape Cod. Image courtesy Journal of Environmental Management (see footnote for link).
Risks and Opportunities
Cost
Cost remains a significant factor influencing the adoption of AI-enabled software, with pricing often fluctuating and initial investment levels still relatively high. This dynamic landscape can pose challenges for organizations, necessitating a careful assessment of the potential benefits against the associated costs. It is imperative to weigh the potential rewards of AI implementation against its upfront expenses to determine its viability in specific contexts.
Security
It is crucial to recognize that data input into public AI systems may be utilized for model training, raising concerns about data security and privacy. In contrast, closed-loop internal systems offer greater control and may provide enhanced security measures.
Responsible AI
Ensuring the responsible development and deployment of AI systems is paramount. AI model creators and software vendors utilizing AI models can enhance accountability by subjecting their models to third-party evaluations. Organizations like Model Evaluation and Threat Research (METR), a non-profit entity, offer impartial assessments to ascertain whether AI systems present any potential threats. Collaborating with reputable evaluators such as METR can instill confidence in the reliability and ethical integrity of AI solutions. Notably, METR is currently engaged in evaluating systems developed by industry leaders such as Anthropic (Claude) and OpenAI (ChatGPT, DALL-E), contributing to the establishment of best practices and standards in AI governance.
Implementation Insights
We have explored identifying assets, analyzing data, and predicting outcomes as discrete subjects. Within each subject we dissected various applications tailored to municipal settings. However, it is crucial to recognize that these topics and applications need not be approached in isolation. Many AI software packages cover multiple subjects and several applications of the topics. One example is an AI water distribution analysis software package that may also do predictions, since you have already collected and analyzed the pertinent data. Such a tool may extend its functionality to proactively identify and predict potential leaks within the water system, leveraging the interconnected nature of the collected data.
Challenges and Considerations
While the potential benefits of AI in the water industry are vast, several challenges must be addressed to realize its full potential. Data quality and availability remain significant hurdles, as AI algorithms rely on large volumes of high-quality data to achieve optimal performance. Water utilities must invest in data infrastructure and data management practices to ensure the reliability and integrity of data collected from sensors and other sources.
Moreover, the deployment of AI technologies in the water sector raises concerns about cybersecurity and data privacy. As water utilities become increasingly interconnected and reliant on digital systems, they must implement robust cybersecurity measures to safeguard against cyber threats and unauthorized access to sensitive data. Additionally, utilities must address ethical and regulatory considerations surrounding the use of AI to ensure transparency, accountability, and fairness in decision-making processes.
Conclusion
AI has the potential to revolutionize the water industry by enhancing efficiency, sustainability, and resilience. From optimizing water resource management to improving infrastructure maintenance and water quality monitoring, AI-driven technologies offer innovative solutions to address the challenges facing the water sector. By harnessing the power of AI, water utilities can improve service delivery, mitigate risks, and ensure the availability of clean and safe water for future generations.
However, realizing the full potential of AI in the water industry requires collaboration between stakeholders, investment in technology and infrastructure, and a commitment to address challenges related to data, cybersecurity, and ethics. As we continue to embrace AI-driven innovations, we have an unprecedented opportunity to transform the water industry and create a more sustainable future for all.
AI Definitions
- AI: Artificial Intelligence
- Deep Learning: Models the brains neural network using multiple layers.
- Digital Twin: Virtual representations of physical objects. Can have sensors on the object to simulate the behavior and monitor operations.
- Likelihood of Failure (LOF): The probability that an asset can fail based on associated properties, spatial interaction with other GIS layers, and asset data that correlates with observed failures in the local system.
- Machine Learning (ML): Uses data to train model to make predictions.
- Machine Vision: Using computers to “see.”
- Model Evaluation & Threat Research (METR): A non-profit that assesses if AI systems pose a threat. Currently working with Anthropic and OpenAI to evaluate their systems.
- Natural Language Processing (NLP): The branch of AI that attempts to understand text and spoken language.
- Optical Character Recognition (OCR): Pulling text from images.
[1]Oxford Languages, Retrieved May 14, 2024
[2] Coffer, M. M., Nezlin, N. P., Barlett, N., Pasakarnis, T., Lewis, T. N., & DiGiacomo, P. M. (2024a, February 29). Satellite imagery as a management tool for monitoring water clarity across freshwater ponds on Cape Cod, Massachusetts. Journal of Environmental Management. https://www.sciencedirect.com/science/article/pii/S0301479724003207
Amy Vredevoogd is a Staff Software Engineer at Weston & Sampson in Reading, Massachusetts.
Published in NEWWA, June 2024.