Joint Beamforming Design of Double-RIS aided SWIPT Network using Deep Learning

This paper explores a double Reconfigurable Intelligent Surface (RIS)-assisted Simultaneous Wireless Information and Power Transfer (SWIPT) downlink communication system. It aims to maximize the weighted sum rate for all SWIPT users by jointly optimizing transmit beamforming at the access point (AP), passive beamforming at each RIS, and power splitting (PS) ratio at each user. The optimization works within constraints including transmit power, energy harvesting at each user, and unit modulus at the RISs. The study first examines a perfect channel state information (CSI) system with continuous phase shift, using multiple antennas at both AP and users to enhance spectral efficiency through spatial multiplexing. Due to the non-convex nature of the problem, it is reformulated and solved using block coordinate descent (BCD), alternating between transmit beamforming and phase shifts. While transmit beamformers and power splitting ratios can be obtained in closed form when phase shifts are fixed, sequential rank one constraint relaxation (SROCR) is utilized to obtain the optimal phase shifts. For imperfect CSI and discrete phase shifts, we develop a deep learning quantization network (DLQN) to design and evaluate active and passive beamforming and PS ratio at each user. Simulation results show that the proposed double-RIS scheme achieves up to 11.6% higher weighted Sum Rate (WSR) than single-RIS systems and up to 33.3% over random phase schemes. The DLQN achieves within 2.8% of the Alternating Optimization (AO) solution while reducing inference complexity by over 8000 times at N = 60 elements per surface and maintains robust performance within 6.7% of the perfect-CSI bound at normalized Mean Square Error (MSE), e = 0.75. The DLQN also outperforms a black-box unsupervised DNN benchmark, as the generic network saturates once multi-user interference dominates, and it surpasses analytical-KKT and DNN-based multi-RIS SWIPT techniques at low-to-moderate power, with these baselines converging only at high transmit power through their deterministic RIS array gain. The AO algorithm converges within 5 (minimum) outer iterations across all tested configurations, and both algorithms satisfy the minimum energy harvesting constraint at 100% rate under the proposed power splitting framework.