Two Thermometers, One Ranking

The manager is right. Breaking either temperature column costs this model at most 6 kWh, half of what breaking occupancy costs, so temperature is a minor driver.

Neither temperature rise can be read on its own. While one column is shuffled, the model can still read the temperature from the other. Shuffling both together, with the same reordering, would measure temperature's real weight, and could show a far bigger rise than 6 kWh.

Temperature is worth about 5 + 6 = 11 kWh, just behind occupancy. With twin features, adding the two rises together gives the importance of the pair, the same way twin credits are added together.

If the team retrains without feels_like_temp, the rise for outdoor_temp will stay at about 5 kWh, because permutation importance belongs to the column, whichever model reads it.