In this paper, we tackle the challenge of improving both user capacity and power allocation in wireless networks with sub-channel assignment constraints. We start by generating channel data using the Shannon capacity formula and use it to train a multiple linear regression model. This model incorporates randomly generated power, noise, and fading values as input features. We then create new test data to predict sub-channel capacities and employ these predictions to solve our optimization models. In our first model, we include the regression equations as constraints, treating power and capacity as variables while maintaining the accuracy of the model. In the second formulation, we use the predicted values as parameters to optimize the network. Preliminary numerical results show that the first model offers greater flexibility, providing optimal or near-optimal solutions with reduced computational time. We believe this approach holds promise for future wireless networks like 5G, 5G+, and 6G.