Pretty general question, requiring more than a stack overflow response. The first thing I’d consider is setting up a predictive algorithm like the linear regression you spoke of. You can also add a constant to it, as in mx+b where the B is the known quantity of food for reservations. So you would run linear regression, and then add a constant to the finalized prediction approximating the impact of reservations. As you get more data, you could start to incorporate reservations as a variable to your model. From there, you would want to build another model for estimating the amount to buy, because you are going to have a cost function that places more emphasis on having extra vs too little. You would have to know the cost vs. the profit to develop an algorithm for calculating the risk associated with too much food vs. too little, but it would not be difficult. You might want to research profit curves: https://en.wikipedia.org/wiki/Profit_maximization
Hopefully that’s enough to get you started!
solved What is the appropriate machine learning algorithm for a restaurant’s sales prediction? [closed]