(Power put on lines - expected transmission losses) * billed cost of power - customer payments.

I hope for the sake of their billing department they're using a database and some customized software for the data, but could just as likely be in a large and painful excel spreadsheet.

Power theft is not particularly interesting to the power company unless they're not getting paid for it. If it's just the wrong person paying, that's their problem.

You can measure power movements at a macro scale. So you can see how much energy is being produced by power stations, and how that’s being fed into distribution networks, plus you can sum all the power usage going through end user meters.

All that together give you enough data to pretty accurately estimate stolen power. But not enough data to figure out with any precision where it being stolen. You might know that certain parts of a city are worse that others, but you couldn’t identify the house or even road where it’s happening.

I reckon you can chuck some machine learning at the problem. Sum of all meter readings on the sub network and then measure the power going in. Other variables might be ambient temperature, underground temp, etc.