In this work, we considered power allocation and request scheduling in mobile ad hoc networks (MANET) clusters, which are generally modeled as optimization problems with constraints. Optimization algorithms often incur significant time complexity, which creates significant discrepancies between theoretical results and real-time processing. We aim to provide a novel machine-learning-based perspective to address this challenge. We use a deep neural network (DNN) to approximate the nonlinear mapping between the inputs and outputs of an optimization algorithm. If a nonlinear mapping can be learned accurately by a DNN, then optimization tasks can be performed more efficiently. We propose SPCDNet as a scheduling and power control deep network mapping method. A key challenge in training a DNN for resource allocation problems is a lack of ground-truth data, meaning the optimal power allocation between the time slots of each transmitter is unknown. To address this issue, we designed an optimal solver based on linear programming optimization methods and used its solutions to train SPCDNet. Simulation results demonstrate that SPCDNet can serve as a computationally inexpensive replacement for traditionally expensive optimization algorithms in real-time tasks and provide very good approximation solutions, where the average run time of SPCDNet for each network size is very low compared to the hundreds of seconds used by an optimal solver, whose time complexity increases exponentially with the input size.