https://www.sciencedirect.com/science/article/abs/pii/S2352467719300748

TL;DR: Math

Unsupervised algorithm for disaggregating low-sampling-rate electricity consumption of households [...] Our method is implemented as an algorithm combining NILM and load profile simulation. This algorithm, based on a Markov model, allocates an activity chain to each inhabitant of the household, deduces from the whole-house power measurement and statistical data the appliance usage, generate the power profile accordingly and finally returns the share of energy consumed by each appliance category over time.

It depends on the kind of meter that your utility uses and how they have it configured.

Smart meters can measure that data at higher frequencies (up to several kilohertz) and from that data you can detect signatures that identify devices inside your house and even, in some cases, what they're doing. For example, when your washer turns on it runs a pump (which draws a specific load) for a set amount of time and then goes through a cycle of running a motor to agitate the load, which draws energy in a specific way as it turns back and forth. When you turn on an LED light, it runs at a steady rate and draws the same amount of energy. When your AC runs it draws a different amount of current than the washer or LED.

With enough data, over time, you can determine which devices are in the house and when compared to a database of known signatures you can classify the device. Ex: all Samsung Refrigerator Model 23e4234 work the same way so once you identify the signature of one you can identify others.

Here's some articles talking about it:

https://www.sciencedirect.com/science/article/abs/pii/S2352467719300748

https://energyinformatics.springeropen.com/articles/10.1186/s42162-019-0096-9

Also, just because your meter is only normally reporting every few seconds it doesn't mean that it isn't capable of recording the data faster for diagnostic purposes ("Why is this house suddenly using 10x the power?!") or law enforcement purposes ("What house in this neighborhood is using HID lighting?").

Not all meters have this capability. Old style meters with a disk don't record data at all and some of the older smart meters can only sample at lower frequency. You can do the same math on the lower frequency data but if you can't measure fast transient events you lose some of the more specific capabilities (like knowing the exact model number of your refrigerator).

Smart meters can measure, instantaneously, I and V at high frequencies (up to several kilohertz) and by looking at long term and transient event signatures in that data it is possible to classify the loads as coming from specific kinds of devices (down to individual model numbers for known devices).

Even if you have an older smart meter (not the analog ones) that can only sample at a few hz you can still do the same kinds of things but with less accuracy:

https://energyinformatics.springeropen.com/articles/10.1186/s42162-019-0096-9

Electrical load signatures have been demonstrated to contain a great information content. This bears promising potential for the application of signal processing algorithms to extract relevant high-level (i.e., abstract) features from the possibly large volume of consumption data. One prominent example for load signature analysis is NIALM, first introduced by George Hart in (Hart 1985; 1992). A range of approaches to detect activities and identify the causing appliances have been presented in literature, e.g., in (Zeifman and Roth 2011), and numerous companies have added disaggregation products to their portfolio in recent years. The process of inferring appliance activity through NIALM is composed of three major steps: Data acquisition, feature extraction, and load identification (Zoha et al. 2012). All of which have been extensively investigated in research, resulting in a large set of proposed algorithms, methods, and features, e.g., in (Jin et al. 2011; De Baets et al. 2017; Leeb et al. 1995; Bergés et al. 2011; Kahl et al. 2017). Event detection is commonly a part of the feature extraction step and used to detect changes in appliance operation from the data.

Event detection algorithms can be categorized by their analysis of steady state or transient information. Algorithms relying on steady state information, such as power consumption readings during the periods before and after a state transition, are well-suited to detect events of appliances with a constant power consumption in each of their modes of operation. The second option is to operate on transient signatures, i.e., the power consumption changes that can be observed during an event. They allow for the characterization of a device and its mode of operation by the unique shape of its power consumption during state changes (Zoha et al. 2012; Anderson et al. 2012).