The research, led by Daqian Liu along with Wenshuai Jiang, Yuntao Shi, Zhenwu Lei, and Yingying Wang, addresses a fundamental tension in vehicle-to-everything (V2X) communication. Intelligent vehicles increasingly offload compute-heavy tasks to proximal edge servers or cloud infrastructures because onboard processors cannot always meet the computational and energy demands of modern automotive workloads. But the decision of where to offload hinges on knowing — in advance — what the end-to-end latency of each candidate destination will be. This is easier said than done. V2X scenarios are inherently complex and stochastic: vehicles move at high speeds, wireless channels fluctuate with fading and interference, base station loads shift minute by minute, and propagation conditions change as cars pass through tunnels, urban canyons, and open highways. Precise network latency forecasting under such volatility has long frustrated engineers, and point predictions from conventional models routinely miss the true delay by wide margins when traffic conditions deviate from training data.

The Beijing team’s solution rests on a two-stage signal processing pipeline that borrows techniques from renewable energy forecasting, a field that faces remarkably similar problems of noisy, non-stationary time series. In the first stage, the raw latency signal is decomposed using variational mode decomposition, or VMD. Unlike empirical mode decomposition, which is recursive and prone to mode mixing, VMD formulates decomposition as a variational problem: it simultaneously seeks a set of intrinsic mode functions — sub-signals, each with a compact frequency band around its own center frequency — such that the sum of the modes reconstructs the original signal while the total bandwidth of each mode is minimized. The practical effect is that the volatile, noise-laden raw delay measurements are split into several smoother, more predictable components. A neural predictor can then model each component separately, capturing slow drifts, periodic load fluctuations, and fast transients independently, which dramatically enhances overall prediction accuracy. This decomposition step effectively mitigates the noise in raw data and reduces the stochastic fluctuations that typically derail single-stage forecasting models.

The second stage confronts a problem that point predictions cannot solve on their own: uncertainty. Even a highly accurate point forecast tells an offloading system nothing about how confident it should be in that number. The researchers therefore model the residual error distributions using kernel density estimation, a non-parametric statistical technique that constructs a smooth probability density from the observed prediction errors without assuming any particular distributional form, such as Gaussianity. From these density estimates, the system generates confidence intervals around each point prediction — for example, an interval within which the true delay is expected to fall 90 percent of the time. This transforms the output from a single brittle number into a range with a quantified reliability, allowing an offloading controller to reason about worst-case latency rather than merely average latency. For applications like cooperative driving or remote vehicle control, where exceeding a latency budget can trigger failure, that distinction is not academic; it is the difference between a usable prediction and a dangerous one.

To integrate these prediction capabilities into a real cellular architecture, the team designed a centralized framework built on the network data analytics function, or NWDAF, a standardized component of 5G core networks defined by 3GPP in its Release 16 specifications. NWDAF is the 5G system’s designated hub for collecting network data and providing analytics services to other network functions. In the proposed architecture, NWDAF aggregates propagation delay measurements and relevant network state information, runs the VMD-plus-prediction pipeline, and exposes latency forecasts with their confidence intervals to offloading decision modules. Vehicles or edge orchestration functions can then query NWDAF before dispatching a computational task, using the predicted E2E latency for each candidate destination — edge server or cloud — to select the one most likely to satisfy the task’s deadline. This centralized design avoids duplicating prediction machinery on every vehicle and keeps the computational burden of forecasting off the resource-constrained mobile side.

The experimental evaluation was conducted in simulation using Simu5G, an OMNeT++ library specifically developed for end-to-end performance evaluation of 5G networks, which allowed the researchers to generate realistic 5G propagation delay traces under varying vehicular conditions. The results are striking. The proposed method achieved a mean absolute percentage error, or MAPE, of just 1.721 percent. By comparison, a long short-term memory network — the recurrent architecture that has been the workhorse of network traffic prediction for years — reached only 2.800 percent, and a standard deep neural network fared worse still at 3.097 percent. In a domain where offloading decisions must discriminate between candidate servers whose latencies may differ by only a few percent, cutting the prediction error by nearly 40 percent relative to LSTM translates directly into better destination selection and fewer deadline violations.

The reliability gains from interval prediction proved equally significant. When the kernel density estimation module produced 90 percent confidence intervals, the prediction interval coverage probability — PICP, the standard metric for how often the true value actually falls within the predicted bounds — reached 94.97 percent. In other words, the intervals were slightly conservative, covering the true latency more often than the nominal level required, which is precisely the behavior a safety-oriented system wants. A coverage probability at or above the nominal confidence level means the offloading controller can plan against the upper bound of the interval with confidence that it represents a genuine worst case, rather than an optimistic estimate that fails under stress.

What makes this work particularly timely is its placement in the broader trajectory of connected and automated driving research. Prior efforts at quality-of-service prediction for 5G connected driving, latency estimation in vehicular mobile edge computing, and deep reinforcement learning based offloading have all grappled with the same underlying stochasticity, but typically with point predictions and generic architectures. The Beijing group’s contribution is to combine signal decomposition, non-parametric uncertainty quantification, and standards-aligned network architecture into a single coherent pipeline aimed squarely at the offloading problem. The approach also echoes a broader trend in network engineering: techniques proven in wind speed and wind power forecasting — where VMD and KDE-based interval prediction have become standard tools — migrating into telecommunications as networks themselves become objects of statistical prediction.

There are, of course, limits that future work must address. The study reports simulation results rather than field trials on live 5G networks, and while Simu5G is a well-regarded evaluation platform, real deployments introduce effects — hardware variability, actual handover behavior, operator-specific scheduling policies — that simulations approximate imperfectly. The centralized NWDAF-based design also concentrates both computational load and trust in a single network function, raising questions about scalability across very large vehicle fleets and about the latency of the prediction service itself. The researchers note that no external datasets were generated or analyzed during the study, and the work was supported by China’s National Key Research and Development Program, the National Natural Science Foundation of China, and the Beijing Natural Science Foundation.

Nevertheless, the implications are considerable. As automotive manufacturers and network operators move toward commercial deployment of vehicle-to-everything services — from high-definition map updates and infotainment offloading to sensor sharing and cooperative perception — the need for dependable, uncertainty-aware latency forecasts will only intensify. 6G research roadmaps already envision AI-native network architectures in which prediction and resource allocation are deeply intertwined, and methods of exactly this kind, combining decomposition-based forecasting with principled statistical bounds, are likely candidates for that role. For now, the Beijing team’s results demonstrate that even a hard problem like V2X latency prediction yields to the right combination of signal processing discipline and statistical honesty. A 1.7 percent error and a 95 percent interval coverage do not merely look good on a benchmark table; they are the kind of numbers that allow an offloading algorithm to commit a safety-relevant computation to a network path and know, with quantified confidence, that the answer will come back in time.

The study also highlights how the intellectual geography of machine learning research is shifting. Techniques that once lived in energy systems journals — VMD for wind speed series, KDE for wind power intervals — are now core components of vehicular networking papers, and the network operators who run tomorrow’s 5G and 6G infrastructure will need engineers fluent in both domains. If the line between communications engineering and statistical signal processing continues to blur at this pace, the connected vehicles of the coming decade may owe their smooth, responsive intelligence not to any single breakthrough, but to this kind of quiet cross-pollination between fields.