We have experienced something similar.
When adding an external water level to one outlet in a model the remaining outlets will have a starting condition where an "imaginary" external water level is set to the invert level of the outlet. We had a case with an emergency outlet which essentially is in a backward slope towards the network. When starting our calculations and we had an external water level enabled in another part of the system the whole model was filled with water due to that the "imaginary" water level at the emergency outlet made water flow into the system and fill it. Maybe you can check if the water level you initially get in the system can be linked to any outlet with a backward slope into the system. We used a weir to describe the emergency outlet instead and the problem is reported.