Source-linked AI summary
Massive MIMO Channel Prediction: Kalman Filtering vs. Machine Learning
Hwanjin Kim, Sucheol Kim, Hyeongtaek Lee, Chulhee Jang, Yongyun Choi, Junil Choi
TL;DR
Outdated CSI can degrade massive-MIMO performance, while prior predictors may rely on theoretical channel models that deviate from realistic channels. The paper develops mobility-aware VKF and ML predictors using realistic SCM channels and compares their accuracy, data-rate, and complexity behavior. Both improve over outdated-channel baselines, while ML has lower operational complexity after training despite higher total complexity.
Problem
Outdated CSI degrades massive-MIMO performance, and prior predictors may rely on analytical channel models that poorly represent realistic channels.
Method
The paper uses SCM channels, a low-complexity mobility estimator, a Yule-Walker-based VKF predictor, and an MLP predictor using LMMSE noise preprocessing.
Results
Both predictors substantially outperform outdated and extrapolation-based channels in prediction accuracy and sum-rate; ML preprocessing achieves nearly the same NMSE as VKF.
Takeaways & Limitations
ML prediction has higher total complexity than VKF but becomes far less complex operationally after offline training.
Abstract
from arXiv · showhide
This paper focuses on channel prediction techniques for massive multiple-input multiple-output (MIMO) systems. Previous channel predictors are based on theoretical channel models, which would be deviated from realistic channels. In this paper, we develop and compare a vector Kalman filter (VKF)-based channel predictor and a machine learning (ML)-based channel predictor using the realistic channels from the spatial channel model (SCM), which has been adopted in the 3GPP standard for years. First, we propose a low-complexity mobility estimator based on the spatial average using a large number of antennas in massive MIMO. The mobility estimate can be used to determine the complexity order of developed predictors. The VKF-based channel predictor developed in this paper exploits the autoregressive (AR) parameters estimated from the SCM channels based on the Yule-Walker equations. Then, the ML-based channel predictor using the linear minimum mean square error (LMMSE)-based noise pre-processed data is developed. Numerical results reveal that both channel predictors have substantial gain over the outdated channel in terms of the channel prediction accuracy and data rate. The ML-based predictor has larger overall computational complexity than the VKF-based predictor, but once trained, the operational complexity of ML-based predictor becomes smaller than that of VKF-based predictor.
I. INTRODUCTION
The paper addresses outdated CSI in massive MIMO by evaluating channel prediction with realistic SCM channels rather than only theoretical models. It develops and compares VKF- and ML-based predictors, with mobility estimation used to select prediction complexity.
- Outdated CSI caused by UE mobility can degrade massive MIMO performance, motivating prediction from past CSI without additional training overhead.
- Prior Wiener- and Kalman-filtering methods commonly assume analytical channel models, while ML methods learn channel characteristics from data without prior channel-model assumptions.
- The paper uses the 3GPP-adopted SCM to represent realistic wireless environments and develops a low-complexity mobility estimator for selecting predictor complexity.
- The developed predictors are a VKF using Yule-Walker-estimated vector AR parameters and an MLP using LMMSE-preprocessed noisy received signals.
- The BS estimates mobility, chooses a complexity order, receives pilot-based signals, and predicts the channel for subsequent communication processing.
B. General framework of channel prediction
The prediction framework selects complexity according to estimated UE mobility and predicts the next channel from previous measurements using a tractable predictor.
- The BS first estimates UE mobility, then selects a prediction complexity order to balance prediction complexity and accuracy.
- The next channel is predicted from previous channel measurements using an arbitrary predictor whose order is linked to estimated mobility.
- Because realistic SCM channels make the resulting optimization highly nonlinear, the paper develops tractable Kalman-filtering and machine-learning predictors.
III. MOBILITY ESTIMATION
The SATC-based mobility estimator exploits spatial averaging in massive MIMO to estimate UE mobility from very few channel snapshots, supporting complexity selection for prediction.
- The estimated mobility determines the complexity order used by the developed channel predictors, reflecting a performance–complexity trade-off.
- The paper proposes a SATC-based mobility estimator for time-varying channels.
- The estimator requires only two channel snapshots by spatially averaging temporal correlation across massive-MIMO antennas.
- Mobility can be estimated across different UE geometries using two snapshots, although accurate Doppler-frequency estimation remains outside the paper’s scope.
IV. KALMAN FILTER-BASED PREDICTION
The paper models SCM channel variation with a vector AR process and estimates its parameters using Yule–Walker equations for VKF-based prediction. Levinson–Durbin recursion and regularization address computational and numerical issues in parameter estimation.
- VKF prediction: The VKF-based channel predictor is summarized as a Kalman-filter algorithm using the estimated AR parameters and covariance quantities.The supplied algorithm passages list prediction-error, Kalman-gain, and minimum-MSE matrix steps.
- AR model and parameter estimation: The VKF predictor represents SCM channel variation with a vector AR(p) model and estimates AR matrices and innovation covariance using Yule–Walker equations.The AR model is selected over ARMA because parameter estimation requires linear rather than nonlinear equations, while prior evidence indicates similar Kalman prediction performance.
- Computational implementation: Levinson–Durbin recursion avoids the large matrix inversion required by the Yule–Walker equations.
- AR model and parameter estimation: The Yule–Walker system uses aggregated channel auto-correlation matrices, which can be estimated from sampled measurement vectors when channel statistics remain stable over several coherence intervals.The sampled auto-correlation matrix is formed from measurement data under the stated stationarity assumption.
- Numerical conditioning: At high AR order p, the sampled auto-correlation matrix can become ill-conditioned, so the method uses a small diagonal regularization term.The simulations use the regularized matrix rather than the unregularized sample estimate.
B. Kalman filter-based prediction
The vector AR(p) model is converted into an equivalent first-order state-space model for sequential Kalman filtering. The reformulated measurement equation and state transition enable the VKF to produce the next-channel prediction from current observations.
- State-space formulation: The vector AR(p) process is rewritten as an equivalent first-order state equation with a stacked order-p state vector.The state transition matrix contains the AR parameter matrices and identity-shift blocks.
- State-space formulation: The measurement equation is reformulated from the received-signal model and combined with the state equation for sequential Kalman filtering.
- Channel prediction: The predicted channel at time n+1 is obtained from the first MrN elements of the predicted state vector.
V. MACHINE LEARNING-BASED PREDICTION
The paper develops an ML-based predictor for SCM channels using an MLP, preceded by LMMSE-based preprocessing of received signals. The section presents the neural-network structure, preprocessing method, and training procedure as the components of the proposed predictor.
- ML-based prediction: The ML-based SCM channel predictor uses an MLP, LMMSE-based received-signal preprocessing, and a dedicated training method.
A. MLP structure
The MLP predicts the next channel vector from a configurable sequence of LMMSE-preprocessed inputs. Its real-valued implementation uses fully connected layers and trains against preprocessed channel targets rather than unavailable true channels.
- Network architecture: The MLP contains input, output, and hidden layers, with the hidden layer composed of L fully connected layers.
- Input-output structure: The MLP maps a sequence of LMMSE-preprocessed channel estimates to the predicted channel at the next time slot.The input order I balances MLP complexity and prediction performance.
- Preprocessing: LMMSE preprocessing is presented as a simple practical alternative to higher-complexity deep CNN denoising approaches.
- Network architecture: The real-valued implementation separates real and imaginary parts at the input and output, then reconstructs the complex-valued predicted channel vector.
- Training: Training uses ADAM and mean squared error between the predicted channel and the noise-preprocessed channel rather than the true channel.The true channel is unavailable in practice for the training loss.
VI. COMPLEXITY ANALYSIS AND NUMERICAL RESULTS
The section analyzes predictor complexity and then compares VKF-based and MLP-based predictors through numerical studies of accuracy and achievable sum-rate.
- Numerical studies determine predictor parameters and compare VKF-based and MLP-based predictors in prediction accuracy and achievable sum-rate.
A. Complexity analysis
The complexity analysis separates VKF estimation and filtering costs from MLP training, preprocessing, and prediction costs. Although MLP training makes its total complexity higher, offline training leaves its prediction phase less complex than VKF prediction.
- VKF complexity includes AR-parameter estimation and Kalman filtering, each involving matrix inversion.
- MLP complexity comprises training and prediction phases, with LMMSE preprocessing adding to the prediction cost.
- The MLP-based predictor has much higher total complexity than the VKF-based predictor because NepochNtrain ≫ MrN.
- After offline training, MLP prediction complexity becomes lower than VKF prediction complexity, mitigating the MLP predictor’s high overall cost.
B. Numerical results
The numerical results tune predictor orders and compare prediction accuracy and achievable sum-rate under SCM settings. Mobility-based order selection, preprocessing, and both predictors’ performance are evaluated across SNR, samples, and prediction time.
- The simulations use SCM urban-microcell settings with a 3 km/h UE, 2.3 GHz carrier, 40 ms slots, and an 8×8 BS UPA.
- The MLP uses L = 2 hidden layers and 512 nodes per layer because two layers suffice for prediction and reduce training complexity.
- Higher UE mobility requires higher AR-order and input-order to achieve the same prediction accuracy, creating an accuracy–complexity trade-off.
- The effective AR-order is approximately AR-order ≈ 0.3 · Mobility [km/h] under the NMSE < −20 dB and SNR = 20 dB criterion.
- The VKF-based predictor requires fewer samples than the MLP-based predictor to reach the same prediction accuracy at all tested SNR values.
- The developed predictors outperform outdated and extrapolation-based channels, while preprocessed MLP achieves almost the same NMSE as VKF and gains about 5 dB at 0 dB SNR.
- VKF and MLP achieve almost identical sum-rates, and adaptive order selection provides additional gain for both predictors.
VII. CONCLUSION
The paper develops mobility-aware VKF- and MLP-based predictors for time-varying massive MIMO channels and compares their accuracy, sum-rate, and complexity. Both outperform outdated-channel prediction, while the MLP offers lower post-training complexity despite higher total complexity.
- A low-complexity mobility estimator sets the proper prediction complexity order for time-varying massive MIMO channels.
- The VKF predictor estimates autoregressive parameters from spatial channel model data, while the MLP predictor uses LMMSE noise pre-processing.
- Both predictors achieve substantial gains over the outdated channel in prediction accuracy and sum-rate.
- The MLP-based predictor has much higher total complexity than the VKF-based predictor.
- After training, the MLP-based predictor has much lower operational complexity than the VKF-based predictor, making offline learning important for practical adoption.