IWA Summary of AI in Utilities
Artificial intelligence (AI) has become a ubiquitous plot point across sectors when digital revolution leaders speak about the future of their industry. Whether it was first introduced to you via pop culture or as a real-world solution to pressing global challenges, AI is now part of our digital vocabulary. For the water sector, implementing AI for water will, amongst other benefits, allow water professionals to efficiently utilise their data deluge to make better decisions.
This blog series aims to build understanding of AI in its current market ready state, introduce examples of applications across the water sector, and question the future of AI for water. The recent publications about Decision Intelligence, Digital Water, and Intelligent Water, gives us an opening to take a step back and build capacity about AI basics for water professionals. This technology will be fundamental to the future of the sector, beyond the transition to smart water and digital water. As a first step, it is helpful to explore the technology behind the hype – here are some AI basics and examples relating to the water sector.
To explain Artificial Intelligence, it is useful to recognise what we consider to be “intelligence”. We generally describe intelligence as the ability to collect information from the environment, learn from that information to make decisions, and take action based on those information-based decisions.
As a simple example, imagine you are home and you hear a dripping sound. Looking outside, you see that it is not raining. Your memory kicks in and you remember having heard this sound before. You deduce that the dripping sound signifies that there is likely a leak somewhere in your home. Being a water professional, you make the responsible decision to find the leak and fix it. All of this happens somewhat unconsciously in your brain, but if you break down the steps, they are: collect information, learn from it, make a decision, and take action. In the field of AI, the question is whether a machine could undertake these steps or similar ones with minimal human intervention.
Artificial intelligence, as a working definition, is intelligence exhibited by machines or computers, allowing them to perform tasks such as understanding, learning, reasoning, planning and more. Generally, when we talk of AI in its current applications, we are speaking about systems that are able to rationally solve complex problems, predict outcomes, and take action in real-world situations to achieve goals. While the goal of some AI professionals is to create Artificial General Intelligence (AGI), the current state of the technology on the market is Artificial Narrow Intelligence (ANI), which is an AI that specialises in one area. A typical example of ANI is Google Deepmind’s AlphaGo which surpasses human ability at the board-game Go. ANI includes language processing, predictive modelling, image recognition, and classification tools. This technology currently underpins many services in the public sphere like spam filtering, shopping recommendations, voice assistants, education, and scientific research.
Machines that learn
Today, when we speak about applied AI, we are referring to algorithms that can learn. The current state of the technology allows us to automate perception. In other words, we have the computational power that can see or hear to predict outcomes. This is otherwise known as Machine Learning (ML) which is the subset of applied AI that helps derive meaning from data generated by people, phones, devices, smart systems, etc. Increasingly, the volume of data collected is surpassing the ability of humans to make sense of it and use it efficiently. We need automated systems that learn from changing data to adapt to a constantly evolving environment.
Simply, Machine Learning uses data to answer questions. A predictive model is trained using data, and it can later create predictions or answer questions given previously unseen data. As more data is gathered, the model can be improved over time and new predictive models can be deployed.
Building on the above leak example, it is possible to train a model to notify you when and where a leak is detected. For humans, our data collection is built on years of experience of sounds, sights, and events. If we wanted to find the source of the leak, we would likely be able to gauge where it is through trial and error, walking around and listening intently. In the absence of learning about the doppler effect or the physics that underpin it, we know from experience that sounds are louder when they are closer. Similarly, ML is the concept of learning from experience. Machine Learning is how we abstract models of the environment without explicitly giving it rules or physics based expressions.
One form of ML is called supervised learning – this is when a system is trained with large amounts of labeled data to classify or plot a requested outcome. In supervised learning, the data inputs and outputs are labelled. An easy example to wrap our heads around is training a model to discern between two images, for example a pump or a pipe, using labeled data. Through ML, the model begins to recognize the patterns that constitute a pump compared to a pipe and becomes able to classify these images. Through iteration and learning, the model is more accurately able to discern between the two images to the extent that when the model is fed with a previously unseen image of a pipe, it will identify it as such, ultimately differentiating pipes and pumps. Another example would be for a ML model that learns to predict based on numerical datasets. For example, researchers can use data about water use, water supplies, population increase, demographic data, and more, to build a dynamic model that predicts areas of conflict and crisis before it occurs. Supervised learning hinges on accurately labelled and structured data – it is the key to unlocking supervised learning.
Unsupervised learning, another form of ML, is able to create outputs from unlabelled data. With this method, the machine learns hidden patterns in the training dataset to make predictions about the output. This type of machine learning is useful for discovering patterns in data, clustering problems, and anomaly detection. When fed with unlabelled images of pumps and pipes, an unsupervised machine learning model would cluster the pumps together, separate from the pipes, based on patterns that it finds in the images but would not label the images. Similarly, an unsupervised machine learning model could, for example, find trends and associations in customer payment data. This type of ML opens up new possibilities of data analytics that can extract insights from unlabelled datasets.
Other types of machine learning are being trialled across industries, allowing for enhanced predictions and will be explored in the next blog of this series. Advances in AI will continue to expand opportunities for the water sector to undergo transformation towards a water wise future.
AI for water utilities
We are seeing technology companies providing AI based solutions that allow utilities to make the most of the data deluge that results from the digitisation of the water sector. Far from replacing human operators, AI is being used to enhance decision making by extending people’s cognitive powers. Operators are required to make decisions with increasing amounts of complex data inputs. At a minimum, ML enables operators to make informed decisions based on actionable insights. The predictive capabilities of ML are based on processing collected information, learning from the data, and providing possible outcomes. ML technology tackles the predictive and prescriptive phases of AI transformation. The decision making and action taking is in the hands of human operators who can decide what actions to take based on intelligently repurposed and distilled information. In the future, we could expect automation to be added to the sequence of AI transformation, enabling people to shift into higher value added jobs within utilities.
AI is currently used in water utilities for intelligent control, process optimisation, asset monitoring and proactive management, event detection, and infrastructure planning. Predictive modelling and prescriptive suggestions enables a shift from reactive and static to proactive and dynamic management. Providing operators with AI augmented decision making for intelligent control of their systems leads to optimised scheduling that can drastically reduce energy costs, chemical inputs, and water use, as well as enable better allocation of staff time.
This level of intelligence is crucial in smart cities where increasing amounts of data and connectivity between systems and sensors are becoming standard. As computing costs decrease and the technology becomes more accessible, AI will quickly become business as usual as the next phase of technological innovation, in the same way as the internet, spreadsheets, and sensors have. The opportunity of AI for the water sector is to accelerate innovation and optimisation, adding efficiency to organisations in tasks where machines are better than humans at recognising patterns and predicting outcomes.
In a sector where deep institutional knowledge is being lost to the silver tsunami of retirement; environmental pressure from climate change is shifting steady states; impacts of urbanisation are affecting system dynamics; and financial constraints are limiting, AI is a solution that can liberate precious resources. The power of AI is to uncover connections that result in transformative insights for companies. Up to now, we have asked whether our utilities had a digital strategy. Now, we need to ask ourselves whether our water utilities have an AI strategy?