It is impossible to design a simple and reliable model.

You have at least to take in account the power input of iron losses (depending, practically, on the square of the Bmax, that is on the square of the rms Voltage) as well as to take in account the Power output of the cooling system (which depends on the thermal K and on the Temperature difference (between ambient and machine)).

If you have a big machine, then you will have various sites with a different temperature and a different thermal K....

You can design approximate models for small devices. But you have to know the reason for your calculation.

For instance, if you are interested in the duration (operating life) of the machine,

you should remember (Montsinger rule) that increasing by 10 °C the operating temperature, will half your electrical insulation life....Gianfranco Coletti

I have developed a neural network model, in which the inputs are:total power, reactive power, intensity, field voltaje, field intensity, wáter inlet temperature (wings), wáter temperature outlet (wings), hydrogen inlet temperature, hydrogen outlet temperature, h2 water cooler inlet, H2 wáter cooler outlet. The model output are the temperatures in the tooth tip and slot bottom in the core end. The first results are very promising.

Your model is very nice (actually, in my previous note I did forget the RIsquare factor: excuse me) , However that articulated modeling has a reason to exist and to be used when dealing with a big machine (order of magnitude 500 - 1000 MVA).

Let me advance 2 notes:

a) you do'nt evidence factors that regard long term behaviour (insulation temperature and electric field...)

b) you have to train that neural network, that means you need previous machines data

experience and....time. That method is OK in your place. In other situations it could be non-suitable...For smaller machines, e.g. machines stressed by a PWM voltage/current.

Anyway, let me thank you. Gianfranco Coletti

That´s correct. The model is only valid in my machine, but the methodology could useful in any other. Actually i am training the neural network with the data collected form the field. In the next setp, the data will continuosly monitored and the model will be run better.

Actually, the model has been developed, trained and validated. The training has been done with data up to 108% aprox. And the neural network was used to extrapolate the values of 113% power. The error was less than 1ºC. The neural network has been retrained with all the data up to 113% in order to extrapolate the values of the final power uprate to 117%. The results of the neural network has been compared with a Finite element model and the error is less than 1ºC.

The publication is on going....

Finally the model has been fully developed and it works.