LSTM-based Power Delay Profile Predictions for Intra-bus Wireless Propagation

Rajeev Shukla 1 , Atharva Verma 1 Aniruddha Chandra 1 , Ondřej Zelený 2 , Radek Zavorka 2 , Jiri Blumenstein 2 , Aleš Prokeš 2 , Jarosław Wojtuń 3 , Jan M. Kelner 3 , Cezary Ziolkowski 3 , Domenico Ciuonzo 4 · 2026

Read LSTM-based Power Delay Profile Predictions for Intra-bus Wireless Propagation by Rajeev Shukla 1 , Atharva Verma 1 Aniruddha Chandra 1 , Ondřej Zelený 2 , Radek Zavorka 2 , Jiri Blumenstein 2 , Aleš Prokeš 2 , Jarosław Wojtuń 3 , Jan M. Kelner 3 , Cezary Ziolkowski 3 , Domenico Ciuonzo 4 (2026) 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-2603-19707.txt.

Introduction

Shukla and co-authors use a long short-term memory network to predict wireless channels inside a bus. The model learns how the channel transfer function changes with transmitter and receiver spots. Their tests use the 60 GHz band, where signals fade fast. From each predicted function they build a power delay profile. The profile taps came out with an average error under ten percent. The authors present this as a light tool for planning in-vehicle links.

Catalog record

CREATORRajeev Shukla 1 , Atharva Verma 1 Aniruddha Chandra 1 , Ondřej Zelený 2 , Radek Zavorka 2 , Jiri Blumenstein 2 , Aleš Prokeš 2 , Jarosław Wojtuń 3 , Jan M. Kelner 3 , Cezary Ziolkowski 3 , Domenico Ciuonzo 4
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE13562
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.19707

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

LSTM-based Power Delay Profile Predictions for Intra-bus Wireless Propagation Rajeev Shukla 1 , Atharva Verma 1 Aniruddha Chandra 1 , Ondřej Zelený 2 , Radek Zavorka 2 , Jiri Blumenstein 2 , Aleš Prokeš 2 , Jarosław Wojtuń 3 , Jan M. Kelner 3 , Cezary Ziolkowski 3 , Domenico Ciuonzo 4 Abstract Long/short-term memory (LSTM) is a deep learning model that can capture long-term dependencies of wireless channel models and is highly adaptable to short-term changes in a wireless environment. This paper proposes a simple LSTM model to predict the channel transfer function (CTF) for a given transmitter-receiver location inside a bus for the 60 60 GHz millimetre wave band. The average error of the derived power delay profile (PDP) taps, obtained from the predicted CTFs, was less than 10 % 10\% compared to the ground truth. I Introduction Channel characterisation is critical to the design and optimisation of wireless communication systems. It helps to understand the propagation of signals across the wireless channel and mitigate the effect of fading, interference, delay, and pathloss. Channel characterisation using real-time measurement of channel characteristics (like pathloss, fading, power delay profile, delay spread, etc.) prompts the development of empirical, deterministic, and artificial intelligence (AI)-based channel models. These models are beneficial in the event of unavailability of measuring equipment, limited data sources, and lack of financial means, which makes generalising different channel environments difficult [ 17 ] . Empirical channel models are not dynamic, often neglect site-specific effects, and have limited accuracy in complex environments. AI and machine learning (ML), especially deep learning (DL)–based channel models, offer promising solutions for dynamic and complex channel environments. DL models provide more accurate predictions by learning from complex propagation patterns from real-time data, thereby adapting to changes in complex physical environments [ 14 , 13 ] . II Proposed Methodology Fig. 1 : Flow chart for CTF generation using a simple 2-layer LSTM model. With the advent of more connected and autonomous vehicles, channel characterisation inside a vehicle plays a pivotal role in optimising wireless channel performance. Highly reflective metal surfaces and a cluttered physical environment cause significant multipath effects impacting signal strength and delay spread. There had been a renewed interest in millimetre-wave (mmWave) communication with unconventional use cases emerging recently [ 12 ] . The 60 GHz mmWave band made popular with IEEE 802.11ad/ay standards, allows multi-gigabyte wireless connectivity and uses compact high-gain antennas for wireless connectivity [ 11 ] . However, the 60 GHz band has limited penetration capability, which prevents external interference and thus provides seamless connectivity inside public transport vehicles like buses and trains or autonomous vehicles. Due to frequent signal variations and pronounced multipath components (MPCs), traditional analytical methods fail to capture the complex dynamic channel. Thus, AI-driven models can analyse real-world measurement data to predict PDPs or CTFs and improve the accuracy of tapped delay line (TDL) models or ray-tracing predictions. AI/ML can revolutionise channel characterisation and channel modelling for the modern wireless communication system. The authors in paper [ 5 , 6 ] argued in support of this claim. The authors first provided a comprehensive survey on using ML in channel characterisation and antenna optimisation in these papers. Then, they detailed channel scenario identification and channel modelling. Similar assertions were made in [ 9 ] , where the authors explored the integration of AI and visible light communication to provide assistive solutions to improve navigation, avoid obstacles, and improve the overall quality of life of visually impaired people. The paper [ 8 ] gave an in-depth analysis of ML-based channel estimation for orthogonal frequency division multiplexing (OFDM) systems for limited training data. The findings in [ 15 ] suggested a recurrent neural network (RNN) hybrid with NeuralProphet to enhance the efficiency and reliability of forecasting channel state information. To make the AI applications more transparent and trustworthy, authors in [ 4 ] proposed the integration of explainable AI (XAI) with DL models. The authors suggested an XAI-based channel estimation (XAI-CHEST) to estimate doubly-selective channels. Advancements in intelligent transportation systems, increasing autonomy of vehicles, and high bandwidth requiring user-friendly experience make it crucial to investigate in-vehicle wireless channel characteristics. However, studies on in-vehicle channel characterisation are limited to only a few specific scenarios. The most prominent direction of studies is private vehicles like cars and vans and public vehicles like buses and train wagons. In [ 7 ] , authors presented 5 GHz band channel characterisation inside a bus and a mini-van, whereas [ 1 ] studied the effect of topological impact on wireless sensor networks (WSNs) performance. The authors used an in-house developed 3D ray-tracing tool to study the channel behaviour in the absence and presence of human beings inside the bus for proper optimisation of the transceiver position. In another such use-case scenario, extensive real-time measurement [ 2 ] , the authors measured the channel transfer function (CTF) inside a bus using 55 GHz - 65 GHz millimetre waves. Notably, these studies are confined to empirical and deterministic channel characterisation techniques and cannot thus adapt to any physical changes and are non-generalisable. III LSTM-based CTF from Measurement Data The current authors investigated the behaviour of wireless channels inside a bus in an extensive measurement campaign at Brno University, Czech Republic, and reported it in [ 2 , 3 , 10 ] . Analysis of measurements suggested that traditional methods could not accurately capture the highly clustered intra-vehicular channel environment and thus became less effective in interpolating and extrapolating [ 16 ] PDP beyond the measured data set. LSTMs are designed for sequential data and can capture temporal dependencies. CTFs for mmWaves are highly frequency-dependent due to their highly dispersive nature. LSTMs can learn the correlation between CTFs and frequency and capture variations across different transmitter-receiver ( T_{X} - R_{X} ) distances in the intra-vehicular cluttered channel scenario. Motivated by the need for accurate channel modelling at millimetre-wave (mmWave) frequencies, the authors propose an LSTM-based CTF prediction model to forecast CTFs in the 55–65 GHz mmWave band. This LSTM-based prediction follows a supervised learning approach, utilising measurement data from the authors’ own measurement campaigns reported in [14]–[16]. The dataset comprises CTF measurements collected over various T_{X} - R_{X} distances within the 55–65 GHz frequency range. The predicted CTFs are analysed using the TDL model, which is subsequently employed for bit-error-rate (BER) calculation. for layer1\leftarrow 20 to 100 do 1 for layer2\leftarrow 2 to 9 do 2 Initialization : Random seed , model\leftarrow None , epochs \leftarrow e 3 Create layer1 : Add LSTM layer, activate ReLU, enable return sequence 4 Create layer2 : Add LSTM layer, activate ReLU, disable return sequence 5 Create dense layer : Add output dense layer with 1 unit 6 Compile model : MSE loss, Adam optimiser 7 Set : batch\_size\leftarrow 20 , disable shuffle 8 Train model 9 Calculate : loss\_train , loss\_test 10 end for 11 end for 12 Select : { layer1 , layer2 , e } |_{min:loss\_train,loss\_test} 13 Algorithm 1 Hyperparameter tuning for LSTM model The LSTM model requires sequential data for prediction. The proposed LSTM model takes a simplistic approach to predicting the CTFs using only two physical parameters as input. The first parameter for the LSTM model is

Read the complete text