With the explosive increase of the types and quantity of devices in wireless networks, it is necessary to design a fast and energy-efficient decision-making strategy for the resource allocation to maintain the efficient operation of the wireless system. However, traditional resource allocation strategies designed based on optimization methods are of high complexity and poor real-time performance, which are not conducive to online decision-making. In this paper, deep learning approaches are introduced to design the optimal beamforming for Sink nodes in the SWIPT-enabled Sensor-Cloud system. First, the energy-efficiency maximization problem is first formulated with a “return/pay” form. In order to realize the application of deep learning for problem solving, a high-dimensional solvable mathematical expression is transformed from the maximization problem, and then a SWIFT-WMMSE algorithm iteratively is designed to obtain the optimal beamforming vector. At the same time, the convergence of the SWIPT-WMMSE algorithm is proved. Second, the approachability of the neural network to approximate the SWIPT-WMMSE algorithm is discussed. Furtherly, based on the error propagation in DNN approaching process, the criteria for the DNN scale design is deduced, and the approximation to the SWIPT-WMMSE algorithm is realized through the training of DNN, which is operated as an alternative algorithm to solve the optimal beamforming vectors for Sink nodes. Finally, the simulation results verify the effectiveness of SWIPT-WMMSE algorithm and DNN model, as well as the approximation effect of DNN model and its advantages in improving the system performance, especially in the aspect of complexity reducing and time saving.