 | Vol 1., No. 6 May 31, 2023 Scott Clearwater, Gridmetrics, Inc.® s.clearwater-c@gridmetrics.io |
The recent introduction of ChatGPT to the general public has generated an enormous outburst of creativity from eager users. Large language models are already changing the dynamics of entire industries such as education, software development, and customer care. However general the results of ChatGPT and its generative AI ilk may seem to be, there are significant limitations. We now look into three of these limitations and the Gridmetrics counterpart.
ChatGPT, when asked (by me) how to exploit non-public information ChatGPT responded with:
As an AI language model, I am programmed to adhere to strict ethical guidelines and cannot assist or advise on any unethical, illegal, or harmful activities, including exploiting non-public information.
So there is the first limitation: no access to non-public information. Almost by definition non-public information can be extremely valuable by its very nature because it enables understandings not possible otherwise. This impediment to ChatGPT’s responses closes off huge areas of potentially lucrative insights. By contrast, Gridmetrics data consists entirely of non-public information: namely, grid voltages at some 300,000 neighborhood-level locations in the US.
ChatGPT was trained on text and can only work indirectly with numerical data, such as building programming models with guidance from the prompter. That is the second limitation. In distinction, Gridmetrics data is entirely numerical meaning it can be analyzed using myriad statistical and other powerful numerical techniques directly. Gridmetrics has used this data to construct a number of indexes relating to different qualities of grid resilience which we discuss more below.
ChatGPT in particular was trained with data only up to September 2021. This means that there is an unavoidable lack of real-time situational awareness. That is the third limitation. By comparison, the data from Gridmetrics is sampled every five minutes providing a near real-time hyper-local view of the grid down to 1km x 1km cells (which often contain multiple sensors).
These three limitations (summarized in the table below) bring us to the point of this blog: exploiting non-public numerical data in the service of improving grid resilience due to natural or man-made disturbances via a deep understanding of the grid down to the neighborhood level. Presently, and for the foreseeable future, there is only one source of national-level grid awareness: Gridmetrics. Gridmetrics has over three hundred thousand sensors with about half the US population within a kilometer of one of its sensors—and Gridmetrics continues to increase its already formidable footprint.
| Characteristic | ChatGPT | Gridmetrics |
| Data source | public | non-public |
| Type of data | text | numerical |
| Timeframe | historical only | historical and real-time |
Now let’s look more specifically into the limitations of ChatGPT versus Gridmetrics for the grid resilience problem. When asked (again by me) where the most volatile areas of the US power grid were ChatGPT responded with the Northeastern US, the Gulf Coast region, California, and rural areas. The first two locations were due to hurricanes especially, and for California due to wildfires, and for rural areas because of the difficulty in maintaining power over such long distances from the power source. These are very general answers and not useful for other than very cursory information.
When asked (from now on assume I’m asking) about the historical frequency of outages in various areas in the US the best ChatPGT could do was to give some examples of severe storms like hurricanes and freezes and the number of people affected, information available from many news sources. On the other hand, most utilities do publish their current outage maps but ChatGPT apparently never accessed that public data as it could not provide any outage frequency analysis, even as I pressed it for more details.
Now we come to the unique data provided by Gridmetrics: 300,000 sensors over most of the US sampled every five minutes—more than 85 million measurements per day. We can begin to see the power of Gridmetrics raw data and the derivatives of it.
First, the real-time data. What can be done with the raw data is to provide near-real time voltage level and outage information that could be used in literally life-saving situations where power is out such as near senior-care facilities. A timely response by emergency management could send generators to those locations or provide other critical infrastructure to save lives.
In a more business-oriented scenario think of a parent company with a chain of locations and knowing that particular locations are without power. This could cause the parent to shift order flow to one or more of its nearby facilities for timely order fulfillment. For example, a drug store chain or pizza delivery chain that reprioritizes order fulfillment based on the power status of its assets. Thus, real-time power information becomes a valuable tool for supply chain management.
From a cybersecurity perspective, Gridmetrics data can inform appropriate government agencies if something unusual is occurring in the grid that may be due nefarious actors. Only by knowing what is “typical” or “normal” from historical data can outlier data be tagged for further investigation. And only with a synchronized national footprint can such local or national-level real-time data be accessed.
Now, how to exploit this historical data. From a planning perspective knowing which areas have poorer power resilience could result in a more efficient use of funds to achieve greater power equity between communities. From a regulatory perspective, agencies could compel utilities or jurisdictions to upgrade their facilities because of the measured historical performance over various dimensions of power resilience, such as those provided by the PowerEventNotificationSystem Indexes. From the business perspective for example, insurance companies can verify business disruption claims more quickly resulting in more efficient underwriting and profitability. Also, real estate companies could add a “Power Sustainability Index” to the description of a property thereby giving prospective buyers an important factor to consider—and not something to be taken for granted.
Presently, Gridmetrics offers a number of indexes of power resilience measures derived from its raw data and these are summarized in the table below. Each of these indexes can be computed for any region within the Gridmetrics coverage area and over any times within Gridmetrics’ historical data. Relative voltage in the table means that each sensor’s voltage is normalized by an historical reference value from that sensor. Note that each of these indexes captures different aspects of the voltage time series and together provide a comprehensive measure of the resilience of electrical power in a region that can be analyzed by itself or in comparison to other regions and timeframes. See https://gridmetrics.io/resources/ for a more detailed explanation of some of the PENS Indexes.
| PENS Resilience Indexes | Function (over a region and time span) |
| PENS Outage Index | Fraction of regions with an outage |
| PENS Reliability Index | Probability with confidence level that relative voltage is within a narrow range |
| PENS Stability Index | Standard deviation of relative voltages |
| PENS Quality Index | Non-linear function of deviation from reference voltage |
| PENS Volatility Index | Fraction of time voltage does not change |
| Sustainability Index | Amalgam of the preceding indexes |
To conclude this contrast with ChatGPT we give ChatGPT’s own answer about the limitations of current AI models which definitely come into play in applications of Gridmetrics data:
Also, keep in mind that interpreting the results and making “inferences” in a broader sense often requires human understanding and context, especially for complex and high-stakes decisions.
Over a hundred years ago electrical power changed human society forever. Today, ChatGPT is changing everything it touches—but what it doesn’t touch is also ripe for profound transformation. In this blog article we’ve shown how Gridmetrics data and its derivatives provide unprecedented insights into the electrical grid and thereby offer impactful opportunities for progress in such diverse areas as public safety, power equity, business efficiency, and cyber security. In a forthcoming blog article I will give some specific examples of the PENS Indexes in action and demonstrate the wealth of insights that are possible. Cutting-edge applications of this data invite your creativity.
Learn more about Gridmetrics actionable power intelligence solutions.
