Today, I did a bit of research on neural network based temperature controllers, and found this article, which describes the design and training of a neural network for controlling a lab water bath. A critical feature of the network described within is that it doesn't output just a binary on/off, but instead a more linear power level.

There are a lot of potential drawbacks to implementing a design that's only on/off. While it may initially seem like rapid on-off would be a perfectly fine PWM method to control heat production, it's really only a fine method if you're using solid-state components. With a mechanical relay, you're going to be in for a lot of noise, and shortened relay lifespan, if you attempt to control PWM with a high frequency. Depending on the cookware you're working with, a very-low frequency PWM behavior may be fine, but really only if there isn't a lot heat loss and the heating element is lower powered (thus you can create a slow sin-wave sort of temperature fluctuation around the set point and avoid relay switching too often).

Given the drawbacks, I decided to look into a more appropriate hardware implementation. Looking at the PowerSwitch Tail product site, I noticed the PSSR/ZC Tail, which combines a solid state relay and zero-crossing detection into one product.

Perfect!

One problem though, the PSSR/ZC Tail is only rated to 450 watts. So, with an ammeter in hand, I hooked up an 8-quart slow-cooker, and discovered that it uses 1.03, 1.94, and 3.0 amps at keep-warm, low, and high settings. This translates to roughly 118, 223, and 345 watts, respectively. The wattage is within the range for the PSSR/ZC Tail.

Given what I've learned, and the drawbacks to binary operation, I'll be changing my approach to us a neural network directly to determine a power output level on a linear scale.