Search Results (1,565)

Surface soil moisture forecasting is important for drought monitoring, irrigation management, and land–atmosphere process analysis but remains challenging because near-surface soil moisture is jointly influenced by antecedent precipitation, atmospheric drying, soil properties, vegetation conditions, and irregular multi-source observations. This study proposes an Antecedent-Precipitation-Informed Surface Soil Water Balance and Time-Aware Mamba–Mixture-of-Experts (API-SWB-Mamba-MoE) framework for forecasting in situ volumetric soil moisture at approximately 5 cm depth using only information available before the target time. The framework combines a process-guided API-SWB physical prior, a time-aware Mamba temporal encoder, a context-conditioned MoE residual decoder, and gated residual fusion. The U.S. source-domain stations were evaluated using five independently repeated station-level random holdout splits, with approximately 70%, 15%, and 15% of the stations assigned to training, validation, and testing in each repetition, respectively. The five resulting U.S.-trained models were further applied without target-domain retraining or fine-tuning to six German and French stations for zero-shot transfer evaluation. Across the U.S. test prediction–observation pairs pooled from the five repetitions, the proposed model achieved a Pearson correlation coefficient (R) of 0.934, a root mean square error (RMSE) of 0.035 cm³ cm⁻³, a Kling–Gupta efficiency (KGE) of 0.922, and a mean bias error (MBE) of −0.001 cm³ cm⁻³. Pooled zero-shot predictions yielded RMSE values of 0.036 and 0.038 cm³ cm⁻³ and KGE values of 0.885 and 0.917 for Germany and France, respectively. The pooled U.S. test RMSE was 7.9–22.2% lower than that of the ablation variants and 20.5–32.7% lower than that of the benchmark models. These results suggest that combining a process-guided prior with time-aware sequence modeling and context-conditioned expert routing offers a promising approach for station-scale soil moisture forecasting under irregular multi-source observations. The external results provide preliminary evidence of zero-shot transferability at the six selected sites, although broader regional validation remains necessary. Full article

Button mushrooms (Agaricus bisporus) are highly perishable, making rapid and automated postharvest freshness evaluation crucial for cold-chain logistics. Visible-near-infrared (Vis-NIR) hyperspectral imaging is promising for non-destructive food quality assessment, yet mining highly redundant spatial–spectral data remains challenging. This study proposes a novel artificial intelligence approach, the patch-aligned multimodal interaction fusion network (PAMIF-Net), to monitor postharvest mushroom freshness. Using a Vis-NIR dataset of 400 A. bisporus caps over a 9-day refrigerated storage period, this deep learning architecture dynamically fuses global spectral features (indicating internal physicochemical shifts) with localized spatial morphological features (capturing surface deterioration) using a gated attention mechanism. Extensive evaluations across 10 independent trials demonstrated that PAMIF-Net achieved optimal classification accuracy (up to 100% on the current test set) and a minimal mean absolute error for five-class storage time recognition. Furthermore, it exhibited superior computational efficiency and significantly lower inference latency compared to classical machine learning and standard deep learning backbones. This multimodal spatial–spectral deep learning framework demonstrates the feasibility of combining HSI and deep learning for the specific task of automated storage time recognition of A. bisporus under controlled refrigerated conditions. Full article

Accurate stereo visual–inertial measurement is essential for mobile robots operating in Global Navigation Satellite System (GNSS)-denied and structurally complex environments. In stereo visual–inertial odometry (VIO), pose and trajectory outputs depend strongly on the point measurements delivered by the visual front end before sensor-fusion update. In sparse-texture but structurally regular scenes, tracked point features may exhibit poor persistence, uneven spatial distribution, and local tracking noise, even when informative line structures are present. Existing point–line VIO methods can improve positioning accuracy by introducing line landmarks or line residuals, but they usually modify the estimator state, measurement model, and Jacobian treatment. We present a scene-adaptive line-aware visual measurement conditioning method for stereo VIO front ends with point-measurement updates. The method uses 2-D image-line segments as lightweight structural priors and applies bounded normal-direction conditioning to reliable point measurements before a fixed-interface VIO back-end update. A sparse pruning safeguard removes only highly inconsistent long-lived tracks under strong structural support, while a scene-level confidence gate attenuates the intervention when line evidence is weak or unstable. The method is instantiated and evaluated in an S-MSCKF pipeline. On the reported EuRoC MAV sequences, it reduces the sequence-averaged absolute trajectory error (ATE) RMSE by approximately 13% relative to S-MSCKF, with 3–27% reductions on machine-hall sequences. On three real-world robot measurement sequences with an RTK-aided inertial reference, the mean Sim(2)-aligned planar position error decreases from 8.72 m to 7.31 m, and the mean yaw error decreases from $8.{02}^{\circ }$ to $6.{76}^{\circ }$; an additional scale-preserving SE(2) evaluation reveals sequence-dependent planar behavior and residual metric-scale sensitivity. Candidate-level stereo-consistency diagnostics show subpixel mean and 95th-percentile image-domain perturbations without systematic vertical-stereo bias, while the final reliability-weighted primary-view update is analytically bounded by approximately 0.221 pixels in the reported implementation. Runtime profiling reports an average front-end time of 33.34 ms on the tested CPU platform, close to the 33.3 ms frame period of the 30 Hz stereo input, although the $\mu +3\sigma$ runtime of 47.23 ms exceeds a strict frame-by-frame 30 Hz budget. These results suggest that line-aware front-end conditioning can improve visual measurement quality in structured stereo visual–inertial sensing without modifying the evaluated back-end interface. Full article

This study proposes Uncertainty-Aware Staged Trust-Adaptive Residual Physics-Informed Machine Learning (U-STAR-PIML), a recursive one-step dynamics model for fixed-wing unmanned aerial vehicles, evaluated with the JSBSim C172x as a surrogate simulation benchmark. It combines a learnable compact six-degree-of-freedom prior, history-dependent temporal residual correction, trust-aware state-dependent residual gating, hard kinematics, a heteroscedastic one-step uncertainty head, and staged optimization. The protocol separates one-step accuracy, recursive rollout, predictive-interval behavior, and physically distinct distribution shifts. Across five training seeds, the full U-STAR-PIML model (E5) achieves a mean one-step root-mean-square error (RMSE) of $0.005515\pm 0.000009$ and the lowest mean rollout-position RMSE of $18.61\pm 2.15$ m; the data-driven baseline has the lowest 20 s all-state RMSE of $2.923\pm 0.623$. Recursive rankings remain seed-sensitive, without a universal winner. For the representative E5 cross-condition evaluation, the exact wind vector used in the JSBSim simulation is supplied to the model at every prediction step, i.e., perfect wind information is assumed. Under this assumption, wind out-of-distribution (OOD) conditions cause the largest degradation, with rollout-position RMSE reaching 72.76 m; wind-estimation error is not evaluated. An external zero-shot evaluation on 10 independent IDF-DS Ranger 2400 real-flight logs reduces pooled one-step all-state RMSE from 0.12374 for Persistence to 0.03379, although improvements are not uniform across dynamic state groups. The uncertainty head yields 95% empirical coverage of 96.68–100%, with conservative over-coverage under most conditions. These results support simulation-based prediction and an initial cross-airframe transfer diagnostic but do not establish same-airframe sim-to-real transfer, recursive real-flight stability, or operational validity. Full article

Non-invasive blood pressure (BP) monitoring using photoplethysmography (PPG) has significant potential, yet accurately predicting systolic (SBP) and diastolic (DBP) blood pressure using photoplethysmogram (PPG) and electrocardiogram (ECG) signals remains challenging. This work proposes a novel dual-stream 1D encoder–decoder architecture for cuffless BP estimation from raw photoplethysmography (PPG) and electrocardiography (ECG) signals. The model incorporates multi-scale temporal feature extraction via dilated residual convolutions, cross-signal feature modulation at the bottleneck using Feature-wise Linear Modulation (FiLM), learned scale-wise fusion across encoder levels, attention-gated skip connections, and a hybrid mean-attention pooling regression head. Evaluated under calibration-based conditions on the PulseDB dataset (with selected data values of SBP ∈ [57, 180] mmHg and DBP ∈ [25, 100]), the proposed model achieved a mean absolute error (MAE) of 2.03 mmHg, mean error (ME) of 0.38 ± 3.13 mmHg, and an R² of 0.93 for diastolic BP (DBP) and an MAE of 3.81 mmHg, ME of −0.28 ± 5.54 mmHg, and an R² of 0.92 for systolic BP (SBP). The results meet the British Hypertension Society (BHS) and IEEE-1708 standard and achieved an “A” Grade. ECG alone provides lower prediction errors than PPG alone under the evaluated conditions, while their combination yields the best performance. Gender- and age-stratified analyses reveal consistent model behavior across demographic subgroups, with prediction error increasing modestly in older cohorts, particularly among women. This study provides an accurate, calibration-based, cuffless BP estimation and highlights its potential for non-invasive BP monitoring applications. Full article

Ground-based Global Navigation Satellite System (GNSS) stations provide continuous atmospheric remote sensing through electromagnetic propagation delays. Precise point positioning (PPP) yields zenith tropospheric delay (ZTD), and ZTD gradients provide a proxy for off-farm tropospheric structure that is unavailable to supervisory control and data acquisition (SCADA)-only forecasts. We propose a model combining a long short-term memory (LSTM) backbone, GNSS conditioning, and a graph neural network (GNN), denoted LSTM+GNN+GNSS, for 4 h wind-power forecasting. Historical PPP-derived ZTD and quality indicators condition a shared temporal representation; a ZTD-gradient direction proxy, turbine geometry, and observation confidence guide a gated graph-residual correction at 15–90 min. On 666 common Yandun test origins, we compare LSTM, LSTM+GNN, and LSTM+GNN+GNSS. The complete system achieves a normalized mean absolute error (nMAE) of 4.53% (4.527 ± 0.132% across three power-model seeds), reducing nMAE by 7.57% relative to LSTM+GNN and 8.23% relative to LSTM. Paired moving-block 95% confidence intervals support both comparisons, while Bonferroni-adjusted lead-wise tests agree from 30 to 225 min. These results demonstrate the incremental predictive value of the complete GNSS-conditioning pathway under the chronological holdout protocol. Full article

Interaction-level activation patching is used to infer nonlinear cooperation in neural networks, but its numerical validity has not been systematically audited. In two pretrained checkpoints—GPT-2-medium and Pythia-410M, audited on one CUDA backend with a single templated cloze task family—we identified two independent failure modes. First, exact finite pair interactions subtract four $O\left(1\right)$ forward evaluations while the target scales as $O\left({\alpha }^2\right)$, creating a precision floor. In matched float32/float64 experiments, at a descriptive $3\times$ floor, 13.0% and 51.2% of full-displacement interactions fell below the threshold—across reasonable $2\times$–$10\times$ cutoffs, these fractions ranged from 8.6 to 33.0% and from 39.5 to 82.1%, respectively—with sign disagreements of 2.9% and 13.2%; deterministic repetitions showed zero spread and therefore failed to expose the error. Second, even in float64, endpoint cross-Hessian estimates diverged from finite interactions as displacement increased, reaching 41–68% error at the full-replacement scale commonly used in patching; this displacement relationship is fitted on only these two checkpoints, and the Pythia-410M fit is visibly less stable. The corrupted prompts contain neither candidate answer, and the fixture’s behavioral contrast was not serialized at measurement time. A post hoc audit of the exact deposited fixture now confirms the intended contrast (the answer beats the distractor on all clean prompts in both models; the clean-minus-corrupt contrast is positive on $20/20$ and $17/20$ prompts). Thus, the quoted constants are properties of this behaviorally supported fixture and backend; we expect the audit procedure, not the constants, to transfer. We introduce an inexpensive ${\alpha }^2$ scaling audit that detects cancellation floors and hidden mixed-precision bottlenecks; it exposed a hardcoded float32 softmax path inside nominal-float64 GPT-NeoX inference. A motivating negative result—that local logical gate topology does not predict attribution-patching error—was obtained under an author-held internal specification that is not independently timestamped and only in ten small synthetic transformers; it is untested at pretrained scale. These results show that reproducibility alone does not establish measurement validity and that precision error and endpoint approximation error must be audited separately. The proposed checks provide a practical validation standard for interaction-level mechanistic interpretability claims. Full article

Hardware noise and finite sampling perturb the fidelity estimates forming a quantum-kernel Gram matrix. We measured how far three hardware-reconstructed Gram matrices depart from an exact statevector reference for one frozen four-qubit ZZ feature map on N = 24 observation windows from an indoor air-quality time series. The corresponding circuits were executed on ibm_fez at 1024 shots per circuit in three separate, non-interleaved jobs—one per configuration: baseline, dynamical decoupling alone, and gate twirling alone. All were complete, finite, and positive-semidefinite. Off-diagonal root-mean-squared error (RMSE) against the reference was 0.0878, 0.0864, and 0.0427; full-matrix centered kernel alignment (CKA) ranged from 0.933 to 0.989, and the post hoc diagonal-excluded (U-centered) CKA ranged from 0.816 to 0.986. The gate-twirled job deviated least on every reported geometry axis; its baseline contrasts are deletion-stable for the Spearman, mean-absolute-error, RMSE, and full-matrix CKA diagnostics, while the Pearson and diagonal-excluded contrasts fall just below that convention. Dynamical decoupling was not separated from the baseline. The observed error exceeded both finite-shot reference scales, so under those sampling-only models, sampling does not explain it. Centered kernel–target alignment did not track reconstruction fidelity and stayed at or below each label-permutation reference: implementation fidelity and task relevance are distinct diagnostic axes. All configuration-level statements describe three realized jobs on one backend; no mitigation-efficacy, classifier-superiority, forecasting, or quantum-advantage claim is made. Full article

Near-infrared (NIR) video-based remote photoplethysmography (rPPG) enables physiological measurement under low-light conditions using monochrome facial videos. NIR videos lack chrominance-based cues, while weak cardiac variations remain entangled with motion, illumination, and imaging noise. To address these limitations, we propose a wavelet-based spectral feature gating (WSFG) module for NIR-based rPPG estimation. The WSFG module applies a two-level stationary wavelet transform (SWT) to decompose intermediate temporal features without temporal downsampling. Specifically, learnable spectral gates regulate each sub-band contribution and reconstruct branch-specific temporal features through inverse SWT. Three specialized branches preserve cardiac variations, suppress nuisance components, and model relative optical attenuation from low-frequency reference features. Experiments on egoPPG-DB dataset demonstrate that the proposed method achieves the lowest mean absolute error (MAE) and mean absolute percentage error (MAPE) of 8.56 bpm and 9.86%, respectively. Further analysis highlight the effectiveness of frequency-aware feature modeling for NIR-only rPPG estimation. Full article

Fact checking is a key task in the field of natural language processing, and aims to judge the authenticity of a given claim by using external evidence. Existing multi-agent-based fact-checking systems achieve significant improvements on multiple benchmark datasets, but still have two core limitations: firstly, they lack a full-process faithfulness-verification mechanism to test the faithfulness of intermediate products at each stage, resulting in the amplification of hallucination errors cascading through the pipeline; secondly, when both supportive and refuting evidence exist, the system often relies on the implicit preferences of the model to make judgments. To address these issues, this paper proposes a multi-agent fact-checking framework called FRAME (Faithful, Reproducible, Agent-based, Multi-stage, and Evidence perception). FRAME addresses these limitations through two core designs: firstly, it designs a three-stage faithfulness-gating mechanism, embedding a faithfulness-verifier agent into the pipeline for faithfulness verification, and triggering targeted repairs when unfaithfulness is detected; secondly, it constructs a Bayesian-inspired structured perception framework for evidence-conflict perception, classifying the retrieved evidence, performing multi-attribute reweighting and weighted synthesis, and, finally, outputting a structured decision with a complete reasoning chain. FRAME is systematically evaluated on multiple benchmark datasets, and the experimental results demonstrate the advantages, generalization ability, and fault tolerance of FRAME when dealing with different types of datasets. Full article

Background: Large language models (LLMs) can produce clinically plausible recommendations that are not adequately supported by authoritative evidence. Objectives: We introduce GuidelineGuard, a modular multi-agent retrieval-augmented generation pipeline in which a separate Auditor verifies claim–sentence support before a recommendation is surfaced. Methods: The original evaluation used a 73-sentence guideline corpus and GG-Bench-60, with replication across three open-weight backbones. In response to reviewer concerns about benchmark size and selective evaluation, we added a source-traceable GG-Bench-200 stress test and the complete 500-case held-out PQA-L test split of PubMedQA. The revision experiments compare single-pass RAG, a paired multi-agent no-Auditor ablation, and GuidelineGuard; the paired runner is designed to share the Planner–Retriever–Clinician draft so that the Auditor is the only intervention. Checkpoint verification confirmed an identical observable pre-audit state for all 200 GG-Bench cases and 496/500 PubMedQA cases; four PubMedQA cases were regenerated after quota-interrupted resumption and were correct commitments in both arms. Because the originally used hosted Llama endpoints became unavailable after the initial experiments, the expanded runs use openai/gpt-oss-20b for generation and openai/gpt-oss-120b for the Auditor. Results: On GG-Bench-200, single-pass RAG achieved 0.970 operational accuracy, while the no-Auditor and GuidelineGuard arms achieved 0.955 and 0.925, respectively. GuidelineGuard committed on 186/200 cases (coverage 0.930) and was correct on 185/186 commitments (selective accuracy 0.995); all 186 commitments cited at least one gold evidence identifier. Relative to the paired no-Auditor arm, the gate rejected six otherwise-correct commitments and no incorrect commitment. On PubMedQA-500, single-pass RAG achieved 0.644 operational accuracy at 0.950 coverage, the no-Auditor arm 0.638 at 0.896 coverage, and GuidelineGuard 0.550 at 0.736 coverage. Selective accuracy increased across those operating points from 0.678 to 0.712 to 0.747. Within the 496 PubMedQA cases with verified-identical observable pre-audit state, the gate rejected 36 incorrect and 45 correct pre-audit commitments, demonstrating both error enrichment and a substantial false-rejection cost. Conclusions: The expanded results support GuidelineGuard as a selective claim–evidence verification mechanism, not as a universally more accurate generator. Its value is the explicit, auditable coverage–risk trade-off; the appropriate verification threshold is task- and cost-dependent and requires prospective clinical validation. Full article

Multi-frame fusion is standard for converting frame-by-frame license plate character detections into stable, reliable readings. Character Time-Series Matching (CTM), a leading approach, associates characters across frames using the Hungarian algorithm with a fixed Euclidean distance threshold and a translation-only motion model, reporting 96.7% accuracy on the UFPR-ALPR dataset. In this work, we demonstrate that this high performance is protocol-dependent: when ground-truth static plate crops and pre-segmented tracks are used, CTM performs strongly. However, in real-world scenarios involving moving cameras (such as drone-mounted cameras, helmet-mounted cameras, and mobile platforms) where inter-frame geometry changes dynamically, baseline multi-frame association frameworks that combine fixed spatial gates with unconstrained translation propagation fail. In these cases, temporal fusion provides no benefit and degrades plate recognition performance below the single-frame baseline. Indeed, under its own Intersection over Union (IoU) tracker, this literature method correctly reads only 15.1% of plates in traffic videos recorded with a real moving camera (86 human-verified tracks). To address this vulnerability, we propose Geo-CTM (Geometry-Constrained CTM), an association pipeline integrating height-scaled adaptive matching gates, inter-frame similarity estimation via Random Sample Consensus (RANSAC), transform-guided character coasting, and co-occurrence-constrained duplicate track elimination. Systematic motion-model ablation demonstrates that while the complete association pipeline provides the primary foundation for robustness (raising mean accuracy from 85.40% to over 91.5%), estimating a similarity transform (91.82%) delivers the most physically grounded and identifiable representation on planar plates without estimation degeneration. While our method performs comparably to CTM on ideal data when using the same detector and detections, it minimizes performance loss under geometric distortion conditions where CTM is inadequate. For instance, a statistically significant improvement is achieved under a 0 → 60° perspective change; in real traffic videos, with the tracker held fixed so that the fusion layer is the only variable, performance rises from 15.1% to 26.7% under the IoU tracker of the original system and from 16.3% to 29.1% under ByteTrack ($+11.6$ and $+12.8$ points; exact McNemar $p=0.021$ and $p=0.013$), whereas changing the tracker alone while holding the fusion layer fixed moves accuracy by only 1–2 points and is not statistically significant. Finally, our error taxonomy analysis demonstrates that on the undistorted benchmark all residual errors correspond to zero-evidence cases beyond the reach of decision-level fusion, while under dynamic perspective distortion errors are dominated by association misalignment, highlighting the specific development areas that future performance improvements must target. Full article

Large language models (LLMs) can turn a flood of cross-layer industrial logs into a fluent incident narrative, but a narrative that cites only real, resolvable events can still be wrong in every relation that matters: the login came from a different workstation, the write command occurred after the physical change it supposedly caused, the action fell inside a planned maintenance window, and the controller does not even actuate the affected process. A cited event is not necessarily supporting evidence. When such a narrative drives automated response, the error propagates into isolating the wrong controller or revoking a legitimate operator. We present EviGuard, a system that decides when an LLM’s understanding is trustworthy enough to act on. EviGuard stores auditable cross-layer evidence in a provenance graph, lets the LLM propose only hypotheses, compiles each hypothesis into atomic machine-checkable claims in an Incident Claim Language, and has an ensemble of deterministic verifiers label every claim supported, contradicted, or unknown against the graph—honoring interval time, event-time policy and credential versions, network reachability, and physical control dependencies. A response gate forbids any high-impact action whose critical preconditions are not all supported. On EviCPS-Bench (42 hardware-in-the-loop attack chains, 9600 claim-level labels, $\kappa =0.87$), EviGuard cuts the unsupported-claim rate from 12.6% to 1.7%, raises relation-edge $F_1$ from 0.64 to 0.89, holds prompt-injection success to 0.4%, and executes zero unverified high-impact actions across 3200 response decisions, at a median end-to-end latency of 0.44 s. Full article

Radiation fields around collimated or shielded sources exhibit strong gradients whose accurate delineation is critical for worker protection and emergency response. Mobile robots can survey such fields, but they sample sparsely and irregularly along their trajectories, and it remains unclear which reconstruction method can be trusted, and where. Using a single dominant collimated source in a two-dimensional indoor setting, this study shows that the answer depends decisively on sampling geometry, and proposes a physics-guided Gaussian process (GP) that performs reliably under trajectory-constrained sampling. A tracked robot combining light detection and ranging (LiDAR)-based simultaneous localization and mapping (SLAM) with a $\gamma$ dose-rate detector surveyed a collimated Cs-137 field in seven independent runs, and all methods were evaluated under both random hold-out (interpolation near visited locations) and spatial block cross-validation (extrapolation into unvisited regions); truth-referenced evaluation against a dense reference field is provided by Poisson-sampled simulations, while experimental accuracy is cross-validated on held-out measurements. Under uniform sampling, a multilayer perceptron (MLP) robustly outperformed GP variants ($R^2=0.95$, stable across 18 seed combinations); under trajectory sampling, its advantage vanished at visited locations and reversed catastrophically in unvisited regions. The proposed physics-guided GP, which uses a fitted collimated-beam template as the GP mean with a Matérn 3/2 residual process, achieved the highest extrapolation $R^2$ (median $0.61$; best baseline $0.31$), reduced the extrapolation error by 32–69% relative to all eight baselines, recovered interpretable source parameters, and provided predictive intervals with approximately calibrated region-level coverage (point-wise error ranking remains weak); a runtime fit-quality gate further renders the correctness of the embedded prior an observable quantity, so the method flags when its own assumptions fail. These results offer quantitative guidance for method selection in robotic radiation mapping under the as low as reasonably achievable (ALARA) principle. Full article

Accurate station-scale weather forecasts support renewable-energy and transportation operations, yet gridded numerical weather prediction (NWP) is affected by representativeness errors and regional biases when transferred to irregularly distributed stations. We propose TF-STNet, a time–frequency dual-branch spatiotemporal network for NWP-to-station bias correction. A station-centered K-nearest-neighbor (KNN) operator retains multiple local NWP trajectories. The time-domain pathway separately encodes recent observations and future NWP, aligns them over the forecast horizon using gated dilated causal convolutions, and propagates lead-resolved states through a coordinate-conditioned directed station graph. The frequency-domain pathway learns spectral weights and applies separate attention to amplitude and phase across neighboring NWP cells. Prediction-level fusion combines the two station forecasts by variable, station, and lead time. The evaluation uses hourly data for wind speed, pressure, relative humidity, and temperature from 455 stations in Hebei, Shandong, Fujian, and Sichuan. Across five independent runs on 16 region–variable tasks, TF-STNet achieves the lowest mean absolute error (MAE) in 15 tasks and the highest Pearson correlation coefficient (PCC) in 15 tasks; its pressure MAE reduction relative to the strongest learned comparator ranges from 16.2% to 57.4% across the four regions. It has lower MAE than raw NWP in seven of eight high-wind or rapid-change event tests and than simple pressure model-output-statistics corrections in all four regions. The Shandong–temperature task and the Hebei–high-wind case illustrate the limits of the present point-forecast formulation. Full article