Uplink OFDM Channel Prediction with Hybrid CNN-LSTM for 6G Non-Terrestrial Networks
Bruno De Filippo 1 , Carla Amatetti 1 , Alessandro Vanelli-Coralli 1 · 2025
Read Uplink OFDM Channel Prediction with Hybrid CNN-LSTM for 6G Non-Terrestrial Networks by Bruno De Filippo 1 , Carla Amatetti 1 , Alessandro Vanelli-Coralli 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-09326.txt.
Introduction
Bruno De Filippo, Carla Amatetti, and Alessandro Vanelli-Coralli propose, in twenty twenty-five, a hybrid neural predictor for wireless channels in 6G satellite networks. Satellite links move fast: low Earth orbit ends sweep past at kilometers per second. Multi-carrier signals need pilot symbols to track the channel, which eats throughput. Their scheme drops pilots in every other slot and predicts the missing channel response. An encoder-decoder convolutional network plus a memory layer, with skip connections, forecasts the next slot's channel from the current one.
Simulations in standard tapped delay line channels show effective throughput gains. The model generalizes across Doppler spreads and line-of-sight conditions with minimal loss. The authors also discuss the compute cost of the lightweight design.
Catalog record
| CREATOR | Bruno De Filippo 1 , Carla Amatetti 1 , Alessandro Vanelli-Coralli 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-09326
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Uplink OFDM Channel Prediction with Hybrid CNN-LSTM for 6G Non-Terrestrial Networks Bruno De Filippo 1 , Carla Amatetti 1 , Alessandro Vanelli-Coralli 1 Abstract Wireless communications are typically subject to complex channel dynamics, requiring the transmission of pilot sequences to estimate and equalize such effects and correctly receive information bits. This is especially true in 6G non-terrestrial networks (NTNs) in low Earth orbit, where one end of the communication link orbits around the Earth at several kilometers per second, and a multi-carrier waveform, such as orthogonal frequency division multiplexing (OFDM), is employed. To minimize the pilot overhead, we remove pilot symbols every other OFDM slot and propose a channel predictor to obtain the channel frequency response (CFR) matrix in absence of pilots. The algorithm employs an encoder-decoder convolutional neural network and a long short-term memory layer, along with skip connections, to predict the CFR matrix on the upcoming slot based on the current one. We demonstrate the effectiveness of the proposed predictor through numerical simulations in tapped delay line channel models, highlighting the effective throughput improvement. We further assess the generalization capabilities of the model, showing minimal throughput degradation when testing under different Doppler spreads and in both line of sight (LoS) and non-LoS propagation conditions. Finally, we discuss computational-complexity-related aspects of the lightweight hybrid CNN-LSTM architecture. {bruno.defilippo, carla.amatetti2, alessandro.vanelli}@unibo.it I Introduction In the last decade, artificial intelligence (AI) algorithms, such as neural networks (NNs), have been shown providing significant benefits to the telecommunications field, either in terms of computational complexity [ 1 ] and/or communications-related metrics [ 2 ] . This has been recognized by the 3rd generation partnership project (3GPP), leading to several technical reports and specifications of AI-related aspects, including the application of AI to the radio access network [ 3 ] . Among the proposed use cases, channel prediction has attracted particular interest as a means to obtain information on the status of the propagation channel when its estimation is not practical. This is the case of non-terrestrial networks (NTNs), one of the pillars of the 6th generation (6G) of wireless communication standards [ 4 ] , typically characterized by dynamic channel conditions including time-varying delays and Doppler frequency shifts. In this context, channel prediction has been proposed to counter the channel aging effect, i.e. , the drop in accuracy of estimated channel information due to the change of the propagation conditions ( e.g. , a low Eart orbit (LEO) satellite’s movement), which has been proven to hinder the performance of satellite communication systems [ 5 ] . A long short-term memory (LSTM)-based NN was employed in [ 6 ] to bring up to date channel estimates relayed by user equipments (UEs) in time division duplexing NTNs. This has a particular relevance for NTNs, where channel estimates feedbacks are typically necessary to compute beamforming vectors in multiple input multiple output (MIMO) systems. Channel aging in NTNs was also tackled in [ 2 ] and [ 7 ] using a lightweight deep NN and a convolutional NN (CNN) with LSTM, respectively. On the opposite, the authors of [ 8 ] trained an LSTM to predict a time series of MIMO channel matrices based on the most recent estimates, reducing the normalized mean squared error (NMSE) of the predictions by 7 dB with respect to a recurrent NN. A similar approach was employed in [ 9 ] , with the predictor consisting of a CNN with LSTM and the channel data being augmented with a variational autoencoder. Finally, the authors in [ 10 ] tackled the prediction of single carrier channel coefficients by proposing a CNN-LSTM with an attention layer, resulting in an NMSE improvement with respect to the model introduced in [ 8 ] . As shown by the literature, channel prediction has great potential in NTNs. However, most of the works in the literature focus on the NMSE as a comparison metric, neglecting to assess the impact of the predictors on key communication metrics. In our previous work [ 11 ] , we provided a preliminary evaluation of physical layer metrics, i.e. , bit error rate (BER), block error rate (BLER), and throughput (TP), in 6G NTN systems with downlink channel prediction, assuming orthogonal frequency division multiplexing (OFDM) as a waveform. In particular, a convolution-based encoding-decoding NN was used to predict the upcoming OFDM channel frequency response (CFR) matrix from the current estimates. Building up from [ 11 ] , in this paper we: • Propose a novel lightweight channel predictor based on a hybrid CNN-LSTM with skip connections, leveraging the pattern extraction capabilities of convolution-based encoder-decoder architectures and predicting time series of low-level features with an LSTM layer; • Assess the throughput gains achieved by the proposed algorithm in a LEO-based NTN considering realistic channel models, e.g. , NTN-TDL-C, etc [ 12 ] ; • Evaluate the generalization capabilities of the pre-trained model when tested on different user equipment (UE) speeds and channel models. II System model We consider a set of UEs on ground, equipped with a 6G transceiver and characterized by various degrees of mobility, communicating in uplink with a gNodeB (gNB) hosted on a LEO satellite. The UEs are assumed to be able to pre-compensate the Doppler shift and keep track of the Doppler rate resulting from the mobility of the considered NTN node, with just a residual frequency synchronization error affecting the received cyclic prefix OFDM (CP-OFDM) waveform. The transmission chain at the UEs employs a low density parity check (LDPC) channel encoder, a bit interleaver, an M -ary quadrature amplitude modulation (M-QAM) mapper, and a CP-OFDM multiplexer. Focusing on a single UE, we assume that N_{SC} subcarriers have been allocated for uplink transmission. The resource grid \mathbf{X}\in\mathbb{C}^{N_{SC}\times N_{sym}} is thus generated. We assume that, within a slot, M_{\pi} -QAM pilot symbols are inserted over each subcarrier of the resource grid at OFDM symbol indices \mathcal{I}_{\pi}^{(slot)} . However, adopting the same strategy as in [ 11 ] , we reduce the pilot overhead by removing all pilot symbols from every other slot, i.e. , half of the slots only contains data symbols; thus, we here limit the analysis to N_{sym}=2N_{sym}^{(slot)} OFDM symbols, with N_{sym}^{(slot)}=14 representing the number of OFDM symbols per slot. We note that this approach results in a theoretical peak throughput uplift by a factor N_{sym}/(N_{sym}-\left|\mathcal{I}_{\pi}^{(slot)}\right|) , where \left|\mathcal{I}_{\pi}^{(slot)}\right| represents the number of OFDM pilot symbols per slot. The CP-OFDM waveform carrying the resource grid is transmitted by the UE and propagates through a frequency-selective channel, modeled as a tapped delay line (TDL) with the following N_{taps} -paths channel impulse response: h(t,\tau)=e^{j2\pi\epsilon_{D}t}\sum_{n=1}^{N_{taps}}h_{n}(t)\delta(\tau-\tau_{n}), (1) Fig. 1: Channel prediction framework (blue and orange squares correspond to data and pilot symbols, respectively). where \epsilon_{D}\sim\mathcal{N}(0,\sigma_{D}) is the residual frequency synchronization error component, modeled as a zero-mean Gaussian random variable with standard deviation \sigma_{D} , h_{n}(t) and \tau_{n} represent the value and delay (typically assuming \tau_{1}=0 ) of the n -th tap, respectively, and \delta(\cdot) represents the Dirac’s delta function. At the receiver, additive white Gaussian noise (AWGN) corrupts the waveform; then, the cyclic prefix is removed from the signal and the received symbols are de-multiplexed. Representing with \mathbf{H}=FFT[h(t,\tau)] the CFR matrix, with FFT being the fast Fourier transform algorithm, the OFDM resource grid