Zpryme Trendz Volume IX

This week, I will dive a little bit deeper into artificial intelligence (AI) and how it can be applied to utility business models. Adoption of AI across all industries is still low at 20% amongst AI-aware companies, but that number is most certainly lower within the utility industry. This should not be the case since, based on Mckinsey research [PDF], AI will add $5.8 trillion worth of value to industry. As much as we talk about these advancements, we aren’t doing the work of embedding them into our operations to create the future we desire. I am not suggesting that we become AI-first companies, but I am suggesting that, having seen what this technology can do for businesses in other industries (specifically in healthcare and insurance), there is no reason why the utility industry cannot adopt this technology at scale. Other industries and the utility industry are similar in that there are tons of historic and real-time data sets to train the AI models that will provide greater insights than we can currently get from our basic models.

Some AI approaches and examples of how they can be applied in the utility industry are below. I explain the approach and provide a use case in bold.

1. Reinforcement learning: learning by trial and error, the most common way for humans to learn. This is what Google’s DeepMind used in its data center cooling optimization outcomes. i) The intelligent autonomous entity, otherwise known as the agent, observes the current state of a digital environment (historic energy usage data and simulations). ii) The agent then takes actions to maximize the accrual of a set long term rewards (energy usage at an optimal price for example).  iii) The agent uses the feedback it receives to determine whether the action taken was progressive or a hindrance toward the set reward. iv) Depending on the outcome the agent continues to search for the optimal set of steps to achieving the goal, until it achieves the goal.
2. Generative Models/GANs: Agents learn a probability (high dimensional) distribution over the training data. The agent can then output new models similar to the training data. GANs offer a path towards unsupervised learning and is a hot segment of AI right now. GANs have two neural networks: generator, which takes random noise as input and synthesizes content, and discriminator, which is already trained/learned to recognize what real signals look like and can determine if the generators’ content is real or fake. A use case in the utility industry is in bots that utilize natural language and context-based semantics in conversational interfaces to, for example, help a low-income customer address a billing or payment issue.
3. Networks with Memory: Neural networks represent one of the techniques of machine learning. Traditional neural networks forget sequential task learning due to changing weights between training for different tasks (different weights apply for task A than for task B and the neural network forgets task A once it is trained in task B). This flaw in neural networks can be solved with: i) recurrent neural network variants which process and predict time-series ii) differentiable neural computers which navigate complex data structures by combining several neural networks and memory systems to learn (and not forget) iii) elastic weight consolidation algorithms which slow down learning on certain weights based on how important the weights were to previous tasks performed, and iv) progressive neural networks which, for a new task, learn lateral connections between task specific models and extracts useful features from previously learned networks. An example of application for the utility industry is in time series predictions for the behavior of IoT devices and for control and actuation of grid based devices to prevent failure.
4. Transfer learning: To avoid the huge amount of training data requirements that most AI have, there is the growing trend towards transfer learning where the AI improves the learning of a new task by transferring knowledge acquired from a previous task. Creating and training shallow networks using the deep learning networks of, say, another company. A use case is in modeling the failure or optimal behavior state of a digital twin of a turbine and consequently the physical asset state.
5. Hardware for training and inference: Graphics processing units enable the performance of multiple actions in parallel/concurrently, a different approach to CPUs which most companies currently use for the performance of data analysis tasks (limited). Companies that can acquire or develop GPUs can train models more efficiently and faster, create larger sets and improve model performance through optimization which allows them to embed the improvements in improving business operations. Data analysis for price and market states is the most obvious use case here.
6. Simulation environments: Simulation environments are the use of test beds to measure and train AI systems to simulate the actual physical behavior of elements of that system. This approach allows the simulation of the performance of a power plant, smart city or smart home.

What does this mean for your organization? Regardless of where your business lies in the utility industry value chain, there is an artificial intelligence approach that can serve the needs you have to ensure your business can compete with the upstart technology-first companies that are utilizing these approaches.

How do you want to respond to this? We’ll be holding the first workshop ‘Using The Supermind (AI + HI) For Corporate Strategic Planning’ in San Antonio on Nov 20th. CPS Energy is providing the space, and we’ll focus on how to gather all the insights and signals from your historic data and mine tour structured and unstructured data from within and without your company to determine your strategic plans, possibilities, and actions. Reach out to Seyi@asha-labs.com or Jason.rodriguez@zpryme.com.

Till next week.