I don't think so. Modern botnets are mostly made of devices that are operating 24/7 already, such as compromised IP cameras, set top boxes, SOHO routers, IoT devices, etc.
The energy spent for TCP/IP stack usage is negligible at best, even when pushing those embedded CPUs to 100%.
Power consumption fluctuations need to be up in the billions of watts before power companies generally care and must do something about it. Wifi routers are limited to 1W output power, so you'd need a lot more than just the hundreds of millions of wifi routers bleating out TCP packets at the top of their lungs to take down the power grid.
Also, what the power companies really care about are changes in consumption; once they've adjusted the grid parameters to compensate for an increase in power consumption, they're happy until the consumption drops off. Using wifi or any internet traffic to destabilize the grid is just not going to work because there just isn't enough raw drain available, even if the attackers could get their timing absolutely flawlessly perfect so every wifi model popped on at once.
Hm, in my experience the difference between idle and 100% CPU usage on a modern ARM processor (e.g. Allwinner H3) is around 1 Watt. That's more or less what an LCD monitor in standby draws.
I wouldn't call that significant (as in, impacting the global energy consumption significantly) even if thousands of devices started the attack at the same time.
Would you call that detectable? Eespecially en masse. Perhaps a smart grid could detect these attacks in some way and dynamically adjust power to compromised devices?
I wouldn't. The signal ratio to noise very likely just isn't there.
Remember, people regularly operate toaster ovens, microwaves, hairdryers, etc on a fluxuating basis, and THOSE tend to consume more like 1200+ Watts for a /single/ device.
There is still real cost to moving those unsolicited bits at the target though. At the receiving end a server that that has all cores operating at capacity has a higher power consumption than a server that is somewhat idle - lower P states or even a C state. Power consumption is fairly dynamic in a datacenter chassis with Xeons. In addition there is an increased cost of cooling this increased heat dissipation as well.
The energy spent for TCP/IP stack usage is negligible at best, even when pushing those embedded CPUs to 100%.