Channel State Information Prediction Accuracy Indicator

Description

CSI-PAI (Channel State Information Prediction Accuracy Indicator) is a Rel-19 feature introduced in 3GPP TS 38.212 and TS 38.214. It provides a quantitative or qualitative measure of how accurate a predicted CSI report is relative to the actual channel conditions. CSI prediction is a technique where the UE or gNB estimates future channel state information based on past measurements, allowing the scheduler to pre-emptively adapt transmission parameters such as modulation and coding scheme (MCS), precoding, and beamforming. Without an accuracy indicator, the network cannot know whether the predicted CSI is reliable, which can lead to link adaptation errors, throughput loss, or unnecessary retransmissions.

The CSI-PAI operates within the CSI reporting framework. The UE computes a predicted CSI based on a configured prediction model or algorithm and also calculates an accuracy metric. This metric may be based on the difference between the predicted CSI and a reference or actual CSI measured at a later time. The indicator is then reported to the gNB via uplink control information (UCI) on PUCCH or PUSCH, multiplexed with the CSI report. The gNB uses the CSI-PAI to decide whether to trust the predicted CSI for scheduling, to adjust the prediction horizon, or to fall back to non-predictive CSI reporting.

Key components include the CSI prediction configuration, which defines the prediction horizon, the reference signals used (e.g., CSI-RS), and the reporting parameters. The accuracy indicator can be a scalar value, a quantized level, or a set of metrics such as normalized mean square error (NMSE) or correlation. TS 38.212 specifies the encoding of the CSI-PAI in the CSI report, while TS 38.214 defines the UE procedure for computing and reporting it. The indicator may be reported periodically, semi-persistently, or aperiodically, depending on the CSI reporting configuration.

In the network architecture, CSI-PAI is a physical layer enhancement that supports advanced features like mobility prediction, beam management, and dynamic spectrum sharing. It is particularly relevant for high-mobility scenarios, where channel conditions change rapidly, and for mmWave bands, where beam alignment is critical. By providing feedback on prediction accuracy, CSI-PAI enables closed-loop adaptation of prediction algorithms and improves the robustness of link adaptation.

CSI-PAI is expected to be used in conjunction with other Rel-19 enhancements such as AI/ML-based CSI prediction and network energy savings. It allows the network to balance the trade-off between prediction gain and prediction error, ensuring that predicted CSI is used only when beneficial. The indicator also facilitates interoperability between different UE implementations, as it provides a standardized way to report prediction quality.

Purpose & Motivation

CSI-PAI was created to address the lack of standardized feedback on the reliability of predicted CSI in NR. Prior to Rel-19, CSI prediction was mostly implementation-specific, and the network had no means to assess whether a predicted CSI report was accurate enough for scheduling. This led to conservative link adaptation or performance degradation when predictions were inaccurate. CSI-PAI solves this by introducing a standardized accuracy indicator that the UE reports alongside predicted CSI, enabling the gNB to make informed decisions about using the prediction.

The motivation stems from the growing interest in AI/ML-based channel prediction for 5G-Advanced and 6G. As networks move toward higher frequencies and more dynamic environments, accurate CSI prediction becomes essential for maintaining reliable connections. However, prediction algorithms vary in quality, and without a feedback mechanism, the network cannot adapt. CSI-PAI provides that feedback, allowing the network to dynamically adjust prediction parameters, switch between prediction and conventional CSI, and optimize resource allocation.

Historically, CSI reporting in NR (Rel-15/16/17) focused on current channel conditions, with no support for prediction accuracy. Rel-18 introduced some AI/ML-based CSI enhancements, but the accuracy indicator was not standardized. Rel-19 formalizes CSI-PAI, making it a key enabler for predictive scheduling and beam management. It also aligns with the broader trend of integrating AI/ML into air interface design, where trust and reliability metrics are crucial for deployment.