Time Dependent Loss Function Resolves Systematic Underestimation in Short Term Wind Forecasts

Accurate short-term wind speed forecasting is essential for the reliable integration of wind energy into power grids, yet modern artificial intelligence systems frequently predict weaker winds than actually occur. In an extensive empirical evaluation conducted across Northeast China, atmospheric scientists and computational modelers demonstrated that standard deep learning architectures systematically underestimate scalar wind velocities over multi-hour horizons when trained using conventional symmetric error objectives like root mean squared error. To resolve this persistent skew without introducing heavier neural network architectures, the researchers introduced a lightweight mathematical framework called Time-Dependent Weighted Quantile Loss. The specialized objective dynamically scales asymmetric penalty weights as the prediction horizon extends into the future, actively countering error accumulation across extended lead times. Published findings reveal that this single objective adjustment substantially suppresses negative prediction bias, refines error distributions, and sharpens detection capabilities for critical high-velocity wind events across multiple leading meteorological neural network architectures.

The structural vulnerability of conventional machine learning in atmospheric modeling stems directly from the statistical properties governing boundary layer wind dynamics. In most geographic regions worldwide, near-surface wind speed behaves according to a right-skewed, heavy-tailed Weibull distribution wherein mild and moderate breezes constitute the vast majority of continuous observations, while operationally vital high-speed conditions occur infrequently. When standard neural networks minimize symmetric error metrics such as mean squared error or mean absolute error, parameter updates are overwhelmingly dominated by the abundant low-velocity samples. This gradient imbalance draws predictions toward the central density of the distribution, creating a conservative forecasting bias that routinely blunts moderate-to-high velocity peaks. In grid operations, such underestimation can cause power generation shortfalls, dispatch instability, and unexpected curtailment. While data resampling and model expansion have previously been explored to counter sample imbalance, these interventions substantially increase computational burdens and model complexity, leaving an architectural gap for targeted loss formulation.

To investigate the problem under operational conditions, the research team focused on a large geographic domain spanning from thirty-eight degrees north to fifty-four degrees north latitude and from one hundred sixteen degrees east to one hundred thirty-six degrees east longitude, enclosing the northeastern region of China. This expansive territory features dense deployments of utility-scale wind farms situated within a representative mid-latitude monsoon climate bordered by maritime waters and complex topography. The modeling framework ingested hourly historical atmospheric data from the European Centre for Medium-Range Weather Forecasts Reanalysis version 5, covering six consecutive years from 2018 through 2023. Grid inputs incorporated five physical fields at a spatial resolution of zero point twenty-five degrees: the zonal wind component at one hundred meters, the meridional wind component at one hundred meters, geopotential height at the one thousand hectopascal pressure level, atmospheric temperature at one thousand hectopascal, and surface elevation derived from the global relief model.

The experimental sample construction relied on a continuous sliding-window sequence spanning forty-eight consecutive hours, where the first twenty-four hours served as multi-channel input context to predict the subsequent twenty-four hours of horizontal vector wind evolution. To prevent temporal data leakage while preserving realistic sample diversity, the investigators split the records from 2018 through 2022 chronologically into training and validation sets using an approximate nine-to-one temporal ratio. Crucially, they introduced a temporal buffer by eliminating sixteen consecutive overlapping samples across the split boundary, enforcing a forty-eight-hour separation that prohibited predictive information from bleeding across the partition. The complete annual sequence of 2023 was isolated as an independent evaluation set. Training samples utilized a sliding stride of three hours to maximize pattern diversity, whereas testing routines adopted a six-hour stride consistent with World Meteorological Organization synoptic reporting intervals, ensuring that verification reflected practical operational deployment schedules.

The core innovation behind the Time-Dependent Weighted Quantile Loss formulation lies in its deliberate coupling of asymmetric error penalization with forecast lead time. While conventional quantile regression penalizes underestimation more severely than overestimation when setting the quantile parameter above zero point five, standard quantile loss applies an unvarying penalty scalar across all prediction steps. In multi-step deep learning forecasts, however, atmospheric predictability degrades and uncertainty compounds systematically as lead time progresses from one hour to twenty-four hours. The proposed formulation introduces an exponential scaling function that gradually elevates the target quantile parameter from an initial lower bound at the first forecast hour to an upper bound at the final twenty-fourth hour. Controlled by a growth rate parameter, this progression applies gentle asymmetric adjustments during early forecast hours when model skill is naturally higher, before enforcing aggressive penalties against underestimation during later stages when accumulated negative drift typically peaks.

To establish rigorous comparative baselines, the researchers evaluated the new loss function across three distinct spatiotemporal deep learning paradigms: a classical three-dimensional convolutional neural network known as Conv3D, a purely convolutional encoder-translator-decoder network termed SimVP, and the advanced multivariate fusion network MFWPN, which combines convolutional layers with spatial and temporal Transformer modules. All networks were trained using Adam optimization on an NVIDIA A100 graphics processing unit with a fixed learning rate of zero point zero zero one and a batch size of four. The iteration count spanned thirty epochs for the lightweight Conv3D and fifty epochs for SimVP and MFWPN. The total optimization objective balanced component-wise horizontal wind vector losses with the scalar wind speed quantile term using an empirical scaling factor of two. Demonstrating remarkable computational efficiency, the integration of the time-dependent loss altered only the optimization landscape without introducing extra neural layers, increasing per-epoch training durations by less than one point twenty-five percent.

Comparative benchmarking across alternative loss formulations demonstrated that simple weighting and robust error metrics fail to cure distributional underestimation. Using the Conv3D backbone as a standardized testing platform over the full twenty-four-hour horizon, weighted mean squared error and log-cosh formulations reduced root mean squared error by only two point sixty-four percent and two point forty-eight percent relative to standard error functions, while trimming absolute bias by eleven point twenty percent and nine point ninety-three percent. Standard static quantile loss demonstrated much stronger efficacy, decreasing root mean squared error by thirteen point eighty-seven percent and suppressing absolute bias by fifty-four point eighty-seven percent. The newly proposed time-dependent quantile loss achieved the highest performance across all continuous metrics, producing a sixteen point thirty-two percent reduction in root mean squared error and diminishing absolute negative bias by seventy-six point eighty-seven percent, confirming that time-varying penalization matches the compounding nature of sequence forecast errors.

Spatial and temporal breakdowns across all three deep learning architectures revealed that the mitigation of forecast bias was widespread across the study domain. Under standard symmetric loss training, baseline models developed severe, spatially coherent negative biases that worsened dramatically over time, producing domain-wide underestimations approaching negative one meter per second in advanced architectures and negative four meters per second in localized Conv3D forecasts, particularly across offshore marine sectors and the China-Mongolia border. Introducing time-dependent quantile loss markedly suppressed these negative anomalies. For the state-of-the-art MFWPN architecture, root mean squared error dropped by four point forty-five percent to eight point eighty-six percent between the eighth and twenty-fourth forecast hours, while mean absolute error dropped by three point sixty-four percent to nine point zero percent from the tenth hour onward. Regionally, the sharpest error reductions exceeded eight percent along northern inland frontiers, though gains were more moderate across complex coastal terrain.

A stratified evaluation across observed wind speed percentiles highlighted how the asymmetric objective strategically redistributes forecast precision across velocity regimes. In the lowest velocity bracket spanning the zeroth to twenty-fifth percentiles, the optimized MFWPN model experienced a minor controlled degradation, with root mean squared error and mean absolute error increasing by sixteen point two six six percent and sixteen point four zero two percent alongside a shift toward slight positive bias. However, this trade-off unlocked decisive improvements across operationally vital thresholds. Within moderate velocities between the fiftieth and seventy-fifth percentiles, root mean squared error and mean absolute error fell by ten point seven five nine percent and twelve point zero one four percent, while absolute bias decreased by fifty-five point three two seven percent. In the highest ranges above the seventy-fifth percentile, error reductions exceeded twelve percent, while the overall underestimation ratio dropped across every velocity bin.

The practical value of the formulation was further validated through categorical tracking of extreme wind occurrences, defined as velocities exceeding the ninetieth percentile threshold of approximately nine point forty-four meters per second. The time-dependent loss produced substantial gains in the probability of detection across all lead times. Although this heightened responsiveness induced a modest rise in the false alarm ratio, comprehensive skill metrics demonstrated decisive net improvements. The Critical Success Index for extreme events improved by five point seventy-one percent to twenty-three point twenty-one percent across the eight-to-twenty-four-hour lead window, while the composite F1 score rose by four point ninety-three percent to seventeen point four percent beyond the fourteenth hour. Seasonal evaluations confirmed identical performance trajectories across winter, spring, summer, and autumn, establishing that adjusting quantile loss dynamically across forecast lead times offers an efficient, architecture-agnostic solution for operational wind energy prediction.

Subject of Research: Deep learning optimization for short-term meteorological wind speed forecasting

Article Title: Mitigating systematic underestimation in short-term wind speed forecasting via time-dependent weighted quantile loss (TDWQLoss)

Article References: Wang, Z., Cheng, D., Wang, Z., Wei, J., & Song, G. (2026). Mitigating systematic underestimation in short-term wind speed forecasting via time-dependent weighted quantile loss (TDWQLoss). Geoscience Letters, 13(1), Article 71. https://doi.org/10.1186/s40562-026-00521-x

Image Credits: AI Generated

DOI: 10.1186/s40562-026-00521-x

Keywords: wind speed forecasting, deep learning, quantile loss, meteorological modeling, renewable energy, spatiotemporal prediction, atmospheric science, ERA5 reanalysis, Mitigating, systematic, underestimation, short-term