The first step is to ascertain "reasonable" configurations of the facility being modeled. Configuration includes physical characteristics (types of machines, number of machines of each type, typical availability v. down time, etc.) as well as information about the demand patterns (types of jobs, arrival rates of jobs, typical due dates, etc.). For random quantities, such as job arrivals or machine failures, you would look for an appropriate distribution (e.g., exponential interarrival times for jobs) and "average" or "typical" parameters (mean arrival rate, mean time between arrivals, ...). If you were modeling an existing (real-world) facility, you would just use the configuration data from that facility and best estimates of the demand data.
Since you are not modeling an actual facility, you could try looking for literature in which similar facilities are modeled, finding ones that are comparable, and parameterizing yours similarly to what has been published. Note that the published problems do not need to be identical. So, for instance, if your problem includes machine breakdowns and repairs and you find a published instance A with most of your features but not breakdowns, and another instance B that is somewhat similar to your problem and includes breakdowns but not some other thing you have, you could try mixing the breakdown rates from B with the other characteristics of A.
An alternative approach is to build a discrete simulation model for a facility of the type you are modeling. You would pick "dimensions" (number of machine types, number of machines etc.) that you think are reasonable, assign plausible distributions to random quantities (time between job arrivals, time between machine breakdowns, ...), then tune parameters such as mean interarrival time or mean time between breakdowns by running the simulation with different parameter combinations until you get a few combinations that result in reasonable system behavior (the facility is not starved for work, the queue does not explode, ...). That will give you some specific facility profiles you can use for testing your scheduling model/algorithm.
If your goal is to publish results, you need to be careful that your test problems look "reasonable" to a reviewer and do not look as if they were cherry-picked to produce good results with your model/algorithm. That's why basing your problems on a real-world facility is most desirable, and basing them on published instances for comparable models is second most desirable.