Predictive Channel State Information (CSI) Framework: Evolutional CSI Neural Network (evoCSINet)

In this paper, we present evoCSINet, a predictive channel state information (CSI) framework that learns latent dynamics of radio channel for prediction applications. The framework uses deep neural networks to identify the dynamics representations from radio channel images. The latent dynamics has the potential to enable a recursive multi-step-ahead prediction in latent space. The proposed evoCSINet framework, in turn, provides an accurate multi-step-ahead prediction model under 3GPP CSI feedback mechanism. We demonstrate the effectiveness of the evoCSINet-based CSI prediction through evaluations of channel predictions and investigate its impact on
the performance of feedback-based precoding scheme in multiantenna systems. Our experimental results show that the proposed evoCSINet can guarantee a minimum percentage capacity of 95% of the upper capacity bound, at a UE speed 15km/h, up to a prediction depth of 10 slots ahead.