The data underpinning the study come from a high-frequency (HF) radar system operated by BMKG, which measures the eastward and northward components of surface flow across 291 active grid points spaced roughly one kilometre apart on a 21-by-21 grid. After rigorous quality control, including speed thresholds, temporal consistency checks and spatial outlier filtering, the record spans December 2022 to February 2026, yielding 28,321 hourly timesteps with about 85 percent temporal coverage. Missing values were filled using tidal harmonic predictions from the UTide software, ensuring that every forecasting model received gap-free input sequences. The data were then split chronologically into training, validation and test periods, with the test window covering July 2025 through February 2026, so that all methods were judged on conditions they had never seen.

The first phase of the evaluation compared twelve methods for one-step-ahead prediction, spanning a deliberately wide ladder of sophistication: naive persistence, tidal harmonic analysis, classical time-series models such as ARIMA and exponential smoothing, a reduced-rank spatio-temporal statistical model known as EOF-VAR, shallow machine learning, and several deep learning architectures. Evaluated consistently across all 291 grid cells and the full test period, the deep learning models achieved the lowest root-mean-square errors for both current components. A convolutional neural network (CNN) reached 11.31 centimetres per second for the zonal component, a skill score of 0.50 relative to persistence, while a hybrid CNN-gated recurrent unit (CNN-GRU) achieved 15.44 centimetres per second for the meridional component, a skill score of 0.42. All three neural networks posted correlation coefficients above 0.97 for the zonal component.

Perhaps the most scientifically revealing result of the first phase was not that deep learning won, but why the classical baselines lost. The pointwise statistical models, which treat each grid cell in isolation, fell well behind, with ARIMA managing skill scores of only around 0.12 to 0.13. Yet when the researchers allowed a purely linear statistical model to share information across space, through the EOF-VAR approach that decomposes the current field into empirical orthogonal functions and models the leading modes with a vector autoregression, it recovered most of the gap to the neural networks, reaching skill scores of 0.39 and 0.35. The authors conclude that the principal limitation of the classical baselines is their neglect of spatial structure, not their lack of nonlinearity. The neural networks nonetheless retained a consistent edge, which the researchers attribute to their nonlinear representation of the coupled spatio-temporal field.

The second phase probed how much past information the models need, retraining the three best deep learning architectures with lookback windows of three, six and twelve hours. Because the dominant semidiurnal M2 tide in the Sunda Strait has a period of roughly 12.42 hours, a twelve-hour window provides approximately one full tidal cycle of phase information. Only the hybrid CNN-GRU improved monotonically with longer lookbacks, edging its errors down to 11.21 and 15.41 centimetres per second, while the standalone CNN and GRU showed no consistent benefit. The interpretation is architectural: the GRU component can integrate tidal phase information from ordered sequences, while the CNN supplies spatial context at each timestep, whereas the standalone CNN stacks frames along the channel dimension and loses temporal ordering, and the standalone GRU flattens the spatial fields and loses spatial structure.

The third phase extended the task to six-hour nowcasting, comparing direct multi-step prediction against autoregressive decoding. Three architectures were tested: a ConvLSTM encoder-decoder that generates each future frame conditioned on its previous prediction, a bidirectional encoder-forecaster (BiEF) that processes the input sequence in both directions, and a direct multi-step CNN-GRU that emits all six future frames in a single forward pass. Under a carefully controlled comparison, with all five trainable models retrained across five random seeds on identical hardware and training budgets, accuracy turned out to be governed primarily by model capacity, plateauing near one million parameters. The one-million-parameter CNN-GRU-MS-Small achieved the best results, 18.39 and 22.07 centimetres per second, with skill scores of 0.77 and 0.72, and at matched capacity the direct and autoregressive strategies were statistically indistinguishable.

Where the architectures did differ sharply was computational cost, and this proved decisive for operational deployment. The direct multi-step model trains about 2.5 times faster and runs roughly 4.5 times faster at inference, because it requires a single forward pass rather than sequential decoding of the six-hour horizon. A single direct model also covers all 291 grid cells, whereas the strongest classical baseline would require fitting and maintaining 291 separate per-cell ARIMA models. On this basis the researchers recommend the compact CNN-GRU-MS-Small for operational nowcasting, noting that while all models are fast enough in absolute terms for hourly updates, the direct approach scales better to larger grids, higher update frequencies and constrained hardware.

The study also uncovered a striking diurnal rhythm in forecast errors. All five neural network models showed elevated errors during afternoon and evening local hours, and concurrent wind measurements from three BMKG weather stations at Merak, Ciwandan and Bakauheni revealed a clear sea-breeze signal peaking near 15:00 local time. The zonal prediction errors correlated positively with this afternoon wind enhancement, with a Spearman rank correlation of 0.48 across the 24 diurnal hours, consistent with sea-breeze-driven currents injecting variability that current-only models cannot anticipate. The meridional errors, by contrast, followed a distinct diurnal pattern not explained by wind speed, and the evidence points instead to tidal-phase effects. Seasonal stratification added a further layer: errors rose by 20 to 26 percent during the northwest monsoon, when stronger and more variable winds introduce non-tidal dynamics that no model in the comparison fully captured.

Two further analyses strengthened confidence in the findings. A tidal-residual decomposition showed that the deep learning models reduce non-tidal, likely wind-driven variability by 16 to 20 percent relative to persistence, while confirming that ARIMA’s modest advantage over persistence derives entirely from tracking the smooth tidal oscillation rather than from any skill at predicting the sub-tidal residual that dominates total error. And when the researchers mapped errors across the strait, they found the same geography for every method, classical and neural alike: errors lowest in the central domain and highest near the northwest and southeast boundaries where tidal amplification and spatial gradients are steepest, with per-cell error maps correlating at 0.96 to 0.98 between ARIMA and the deep networks. This indicates that the error distribution is set by the underlying tidal physics rather than by model class.

Relative errors of 3 to 7 percent of the observed speed range are of the same order as those reported in far calmer open-sea environments such as the Gulf of Thailand and Monterey Bay, though the authors caution that this normalisation partly reflects the strait’s exceptionally large speed range and that cross-site comparisons are limited by differences in resolution and horizon. The message, they stress, is not that deep learning is superior everywhere, but that it does not break down in energetic strait settings. For the thousands of vessels transiting the Sunda Strait each year, and for the growing number of coastal regions worldwide where HF radar networks monitor fast-changing waters, the study offers a practical blueprint: hybrid architectures that exploit both space and time, lookback windows long enough to capture a full tidal cycle, and direct multi-step designs that trade nothing in accuracy while cutting computational cost dramatically. As machine learning continues its advance through the geosciences, this benchmark demonstrates that the winning formula lies as much in honest, capacity-matched comparison against strong baselines as in the sophistication of the models themselves.

Subject of Research: Comparative evaluation of statistical and deep learning methods for forecasting high-frequency radar surface currents in the tidally energetic Sunda Strait

Article Title: Comparative evaluation of statistical and deep learning methods for high-frequency radar surface current forecasting in a narrow tropical strait

Article References: Gautama, D., & Putra, A. A. (2026). Comparative evaluation of statistical and deep learning methods for high-frequency radar surface current forecasting in a narrow tropical strait. Advances in Statistical Climatology, Meteorology and Oceanography, 12(2), 221-242. https://doi.org/10.5194/ascmo-12-221-2026

Image Credits: AI Generated

DOI: 10.5194/ascmo-12-221-2026

Keywords: ocean surface currents, high-frequency radar, deep learning, Sunda Strait, tidal forecasting, CNN-GRU, ConvLSTM, nowcasting, maritime safety, operational oceanography, machine learning, Indonesia

The research, carried out by Chenchen Wang, Keqiang Miao, Tong Xin, Weiqun Fan, Chunyan Hu, and Wei Li of the Institute of Engineering Thermophysics at the Chinese Academy of Sciences, addresses a long-standing headache in altitude ground testing. These facilities simulate the temperature, pressure, and flow conditions of flight by feeding an engine through a sophisticated intake system that mixes hot and cold gas streams. In steady operation, standard proportional–integral controllers can hold the inlet pressure steady. But the moment the engine accelerates, its rotational speed climbs and the intake mass flow demand changes rapidly, dragging the inlet pressure along an unpredictable, nonlinear trajectory. Add unmeasured disturbances — fluctuations in the mixing of gas streams, downstream cabin pressure shifts, and the coupled thermal dynamics of a massive piping network — and the problem becomes one that classical control theory has struggled to solve cleanly.

Previous attempts to tame the transient have followed two broad paths. On the control side, engineers have deployed gain-scheduled PI schemes, fuzzy controllers, and increasingly sophisticated active disturbance rejection control (ADRC) and its linear variant, LADRC, including cascade and fixed-time formulations designed for high-flow intake systems. These approaches estimate and cancel disturbances in real time, but they react to deviations after they begin to develop. On the modeling side, researchers have built data-driven simulations of engine transients, recognizing that physics-based models of the full facility are too slow and too uncertain for real-time prediction. What has been missing, the authors argue, is a prediction tool that both understands the character of an acceleration or deceleration event as it unfolds and produces forecasts accurate enough to serve as feedforward compensation — telling the control system what is about to happen before it happens.

The new method attacks the problem with an architecture that mirrors the physics of the transient itself. At its core sits a CNN–GRU sequence encoder: a convolutional neural network layer stack that scans short windows of sensor history to pick out local, sharp features — the telltale signature of a throttle slam or a rapid deceleration — followed by a gated recurrent unit network that integrates those features across time, capturing how the event’s dynamics evolve. This pairing is well suited to the task because acceleration and deceleration are fundamentally event-like: they have sharp onsets, distinctive local patterns, and temporal dependencies that a pure convolutional model or a pure recurrent model tends to blur. The convolutional front end acts like a pattern detector for the early moments of a transient, while the recurrent backend maintains the running memory needed to track the pressure trajectory seconds into the future.

The second innovation is how the network outputs its answer. Rather than a single regression head, the model uses a dual-head, multi-task structure: one head predicts the pressure increment — the step-to-step change in inlet pressure — while the other predicts the absolute pressure trajectory directly. Predicting increments emphasizes the shape and timing of the transient, which is precisely where errors hurt most, while predicting absolute values anchors the forecast to physically meaningful levels. The two tasks reinforce each other, but the authors go a step further with what they call a dynamic-consistency constraint: a loss term that forces the increment head’s predictions to agree with the temporal differences of the absolute head’s trajectory. If the two views of the same physics diverge, the training process penalizes the inconsistency, effectively making the network internally check its own work. The consistency is weighted dynamically by an event-aware scheme that applies extra weight during periods when engine rotational speed is changing rapidly — exactly the moments when accuracy matters most and when unmeasured disturbances are strongest.

Training and validation relied on real measured data from an aero-engine altitude test facility, covering genuine acceleration and deceleration events with their full complement of nonlinear, time-varying behavior. The team benchmarked their model against a demanding lineup of alternatives, including a standalone one-dimensional CNN, a deep GRU network, a temporal convolutional network with dilated causal convolutions, a Transformer encoder, and hybrid TCN–GRU and CNN–GRU variants with single or dual output heads. Every model saw the same event-level training, validation, and test splits, the same normalization, and the same optimizer configuration — Adam with a learning rate of 0.001, a batch size of 128, up to 500 epochs, and early stopping with a patience of 200 epochs. To guard against lucky random seeds, each network was trained five times with different initializations, and results were reported as means and standard deviations.

The verdict was consistent: the full dynamic-consistency-guided multi-task CNN–GRU model beat every baseline on prediction accuracy, transient trend tracking, and robustness. The ablation structure of the comparison tells an instructive story. Moving from a single-head CNN–GRU to the dual-head multi-task version improves performance by letting the network learn complementary representations of increments and levels. Adding the dynamic-consistency constraint improves it further by eliminating the drift between the two heads that otherwise accumulates during long, fast transients. And the event-aware weighting concentrates the model’s capacity where it counts, sharpening predictions precisely during the steep portions of the pressure curve when engine speed is sweeping through its fastest changes. Metrics based on mean absolute error and root mean square error confirmed that the improvements were systematic rather than artifacts of a single test case.

The practical payoff lies in what predictive accuracy enables downstream. Because the model can characterize transient inlet pressure fluctuations before they fully develop, its output can feed a feedforward compensation loop: the inlet pressure control system receives an anticipatory signal and adjusts the mixing of hot and cold gas flows proactively, rather than scrambling to correct a deviation that has already corrupted a measurement. The authors also position the framework as a design tool for inlet pressure control systems, allowing engineers to evaluate candidate control laws against realistic predicted transients without burning test-cell time. In an era when aero-engine development cycles are compressed and altitude test facilities are booked solid, shaving the uncertainty out of transient test data has direct implications for cost, schedule, and ultimately the safety certification of engines that carry passengers.

Broader significance of the work extends beyond test cells. The core ideas — event-aware temporal weighting, dual-head multi-task prediction, and self-consistency between increment and level forecasts — are architecture-agnostic and could transfer to other industrial transient prediction problems: wind tunnel pressure control, power plant load following, or any process where a fast actuation event disturbs a slow thermal system. The study also joins a growing literature showing that hybrid physics-informed and data-driven models are overtaking both purely physical simulations and generic deep networks for aerospace applications, from engine health management to supersonic isolator flow prediction. The data underpinning the study remain confidential and are available from the corresponding author on request, a reminder that in the world of full-scale engine testing, the hardest data to get are often the most valuable. Supported by funding under China’s Engine Thermophysical Test Apparatus program and built on test data from the Institute of Engineering Thermophysics’ test-cell team, the research signals a shift in how ground test facilities may be operated: not as passive measurement chambers that react to the engine, but as predictive systems that see the transient coming and meet it with computed foresight. As neural networks learn the rhythms of acceleration and deceleration, the boundary between flight and simulated flight grows correspondingly thinner.

Subject of Research: Machine learning prediction of transient inlet pressure in aero-engine altitude simulation testing

Article Title: Prediction of Inlet Pressure During Aero-Engine Acceleration and Deceleration Using a Hybrid Neural Network

Article References: Prediction of Inlet Pressure During Aero-Engine Acceleration and Deceleration Using a Hybrid Neural Network. (n.d.). https://doi.org/10.1007/s42405-026-01265-w

Image Credits: AI Generated

DOI: 10.1007/s42405-026-01265-w

Keywords: aero-engine, altitude simulation tests, inlet pressure prediction, hybrid neural network, CNN–GRU, multi-task learning, transient pressure, active disturbance rejection, feedforward compensation, deep learning, altitude test facility, engine acceleration