Forecasting next-window wireless link quality is difficult when both the links that will be observed and their qualities are unknown at decision time. We formulate this setting as panel-first prequential forecasting: the complete candidate panel is emitted before any same-window availability or quality is read. The proposed method treats a 58,946-parameter stable predictor as the model-class identity and permits only one bounded conditional-median residual proposal. A single dataset-blind chronological gate grants or withholds that proposal; rejection returns the stable prediction exactly. We establish pre-target measurability, samplewise output and loss-change bounds, exact fallback, and active-edge linear storage. With one fixed configuration, the proposal passes later confirmation segments on FlockLab, Outdoor LoRa, and LoED and improves held-out MAE/RMSE by 1.53%/0.25%, 0.91%/0.07%, and 1.27%/1.26%, respectively. A chronology-tiered external ledger preserves every outcome without changing the rule: Sálvora returns identity; Antwerp and Funchal are registered discovery results; Montería fails the prespecified structural screen; and Brisbane is amendment-registered replication. The three positive external evaluations achieve 1.34%–3.28% held-out joint-objective gains. A prospectively frozen Strasbourg hardware confirmation fixed the complete protocol and implementation before target reservation and generated its 72-window test only after the admission decision and a manifest cryptographically binding the serialized models had been independently timestamped. Proposal B passed confirmation and improved sealed-test MAE, RMSE, and the joint objective by 4.12%, 3.02%, and 3.47%, with 5/5 seed and 4/4 prespecified time-block wins. The 90% interval for joint-objective gain, [3.12%, 3.77%], exceeded a predeclared 1% author-policy/statistical-resolution margin. Relative-gain sensitivity and measured parameter, storage, and CPU costs separate predictive improvement from deployment value. B is therefore used only when the common gate passes and its measured cost is acceptable for the target application.
This paper addresses the trade-off between throughput and energy efficiency in Internet of Things (IoT) systems. A joint multi-level adaptive vacation and delay strategy is proposed for relay assisted Short Packet Communication (SPC). The relay is modeled as operating in four states: transmission, delay, vacation, and idle. Based on this model, closed-form expressions are derived for steady-state queue length, average durations of different system states, average throughput, and energy efficiency. Numerical results show that the benefit of delay strategy is condition dependent and is jointly affected by transmission reliability, system load, and vacation parameters. Research shows that the delay-assisted vacation strategy exhibits stronger robustness in improving throughput than in enhancing energy efficiency, with its performance gains jointly constrained by system load and link reliability. The strategy is only stably effective under high-load and high-reliability scenarios; thus, it should be deployed differently according to specific system conditions and optimization objectives rather than being enabled globally without distinction. Further comparison with multi-vacation strategies in the literature demonstrates that the proposed approach achieves a better balance between throughput energy efficiency over a broad feasible load–reliability range under the evaluated vacation settings. The results provide theoretical guidance for strategy design and parameter optimization in low-power IoT systems.
With the rapid development of edge IoT and IIoT networks, intrusion detection systems must handle heterogeneous high-dimensional traffic under strict resource constraints. Existing lightweight detection pipelines often struggle to jointly balance detection effectiveness, feature compactness, and deployment efficiency, especially when feature selection and hyperparameter tuning interact under dynamic search conditions. To address these issues, we propose TGWO-DCC, a lightweight intrusion detection framework with dynamic cross-layer collaboration. First, a Topology-Guided Walrus Optimizer is developed to improve feature elimination efficiency and search stability in high-dimensional discrete spaces. Second, a Dynamic Cross-layer Collaboration Mechanism coordinates outer-layer hyperparameter exploration and inner-layer feature selection through symbiotic memory and state-aware intervention, thereby suppressing ineffective exploration and mitigating cross-layer oscillation during joint optimization. Experiments on BoT–IoT, Edge–IIoTset, and UNSW-NB15 show that TGWO-DCC maintains competitive detection performance while reducing feature dimensionality and per-sample inference cost. Deployment on a Raspberry Pi 4B further shows that TGWO-DCC maintains stable detection performance with fewer input features across multiple backend model families and constrained runtime profiles.
