Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Average model”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 793 records · Page 44
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Outflow Boundary Conditions for Turbine-Integrated Rotating Detonation Combustors

This study examines outflow boundary conditions (BCs) in computational fluid dynamics (CFD) simulations of a transition duct with and without guide vanes that converts supersonic flow exiting a rotating detonation combustor (RDC) to subsonic flow to drive a turbine. Since the flow exiting the transition duct has swirling shock waves with significant spatial and temporal variations in pressure, temperature, and Mach number, imposing proper BCs poses a challenge. To ensure all swirling shock waves exit the transition duct without creating non-physical reflected waves at its outlet, this study examined three outflow BCs: (1) the average pressure imposed at the duct’s outlet, (2) a nonreflecting BC (NRBC) with a specified average pressure imposed at the duct’s outlet, (3) the average pressure imposed at the outlet of an extension duct made up of a buffer layer and a sponge layer. This study is based on the three-dimensional, unsteady density-weighted-ensemble-averaged continuity, Navier–Stokes, and energy equations for a thermally perfect gas closed by the realizable k–ε model and “enhanced” wall functions. The results obtained show that imposing an average pressure at the transition duct’s outlet produces spurious waves that degrade the physical meaningfulness of the solution. When the NRBC was applied, swirling shock waves exited the duct’s outlet without creating spurious waves. However, its usage requires the gas to be thermally, as well as calorically, perfect, which this study shows could be a concern. By imposing the average pressure at the outlet of an extension duct, the gas does not need to be calorically perfect. The results obtained show the effects of the sponge layer’s length and coarsening ratio on damping nonuniformities in non-physical reflected waves to ensure the flow exiting the transition duct’s outlet can do so as if there are no boundaries present and has the desired average pressure—even though the BC is applied at the extension duct’s outlet.

gas turbines↗
DOE OSTI2023

Machine learning from RANS and LES to inform coarse grid simulations

Nuclear system thermal hydraulic analysis has historically relied on computationally inexpensive 1D codes. However, such tools are unable to capture multiscale multidimensional effects in large nuclear reactor enclosures. On the other hand, simulations with higher fidelity can be too expensive for such purposes. One of the ways to reduce computational cost is to perform simulations on a coarse grid, which, unfortunately, introduces large discretization errors. In this paper, two high-to-low data-driven approaches are investigated: (1) a coarse grid turbulence model to predict eddy viscosity and (2) correction of errors in coarse grid velocity fields. The approaches aim to reduce grid- and turbulence model-induced errors in coarse grid Reynolds-averaged Navier–Stokes (RANS) simulations. Two sources of high-fidelity data, RANS and large eddy simulations (LES), are explored. To extract the eddy viscosity from the LES data, an inverse optimization problem is solved. However, the LES eddy viscosity is shown to be comparable to the RANS eddy viscosity in terms of error reduction. Therefore, the directly available RANS eddy viscosity was used to develop a coarse grid data-driven turbulence model. Additionally, error correction in velocity is used to reduce the remaining uncertainties and bring the results closer to reality. In conclusion, the performance of the frameworks is demonstrated for a scaled upper plenum of a gas-cooled reactor facility.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2025

Evaluating mesoscale model predictions of diurnal speedup events in the Altamont Pass Wind Resource Area of California

Mesoscale model predictions of wind, turbulence, and wind energy capacity factors are evaluated in the Altamont Pass Wind Resource Area of California (APWRA), where the diurnal regional sea breeze and associated terrain-driven speedup flows drive wind energy production during the summer months. Results from the Weather Research and Forecasting model version 4.4 using a novel three-dimensional planetary boundary layer (3D PBL) scheme, which treats both vertical and horizontal turbulent mixing, are compared to those using a well-established one-dimensional (1D) scheme that treats only vertical turbulent mixing. Each configuration is evaluated over a nearly 3-month-long period during the Hill Flow Study, and due to the recurring nature of the observed speedup flows, diurnal composite averaging is used to capture robust trends in model performance. Both model configurations showed similar overall skill. The general timing and direction of the speedup flows is captured, but their magnitude is overestimated within a typical wind turbine rotor layer. Both also fail to capture a persistent observed near-surface jet-like flow, likely due to the limited grid resolution that is typical of mesoscale models. However, the 3D PBL configuration shows several minor improvements over the 1D PBL configuration, including improved wind speed and turbulence kinetic energy profiles during the accelerating phase of the speedup events, as well as reduced positive wind speed bias at surface stations across the APWRA region. Using a mesoscale wind farm parameterization, modeled capacity factors are also compared to monthly data reported to the US Energy Information Administration (EIA) during the study period. Although the monthly trend in the data is captured, both model configurations overestimate capacity factors by roughly 7 %–11 %. Through model evaluation, this study provides confidence in the 3D PBL scheme for wind energy applications in complex terrain and provides guidance for future testing.

17 WIND ENERGY↗
DOE OSTI2020

Air Dispersion Modeling for the Idaho National Laboratory Permit to Construct P-2015.0023 Facility Emission Cap Revision Request

The U.S. Department of Energy Idaho Operations Office (DOE-ID) is seeking to modify permit to construct (PTC) P-2015.0023 issued to U S Dept. of Energy – Idaho National Laboratory (INL) on January 11, 2018 by the Idaho Department of Environmental Quality (DEQ). This PTC allowed the INL to become a synthetic minor source of air pollutants using a facility emission cap (FEC). Air dispersion modeling was performed as part of the permit revision application process to demonstrate that the modifications will comply with the National Ambient Air Quality Standards (NAAQS). This report documents the modeling methodology and complete results for the air dispersion impact analysis and is supplied as an appendix to the permit revision request, Idaho National Laboratory Permit to Construct P-2015.0023 Facility Emission Cap Revision Request (DOE-ID 2020). All CAPs regulated under Section 109 of the Clean Air Act were modeled with the exception of lead (Pb) and ozone which are not required to be modeled by DEQ. Consistent with the original permit application, modeling was not performed for toxic air pollutants (TAPs) as uncontrolled emissions did not exceed screening emission levels for carcinogenic and non-carcinogenic TAPs. Acceptable model design values (concentrations) for each CAP and averaging period were summed with average background concentrations and compared to NAAQS. The results demonstrate the impacts of requested FEC limits for CAPs are less than applicable standards and demonstrate emissions up to FEC limits will not cause a violation of ambient air quality standards.

99 GENERAL AND MISCELLANEOUS↗
DOE OSTI2023

Uncertainty Aware Deep Learning for Fault Prediction Using Multivariate Time Series Signals

The superconducting radio-frequency cavities are a crucial component of the Continuous Electron Beam Accelerator Facility (CEBAF) at Jefferson Lab. When a cavity faults, beam delivery to experimental end users is disrupted. Prediction of cavity faults prior to onset is essential to reduce operation and maintenance costs. In this work, a parallel long short-term memory (LSTM)-convolution neural network (CNN)-based deep learning (DL) model is proposed to predict impending faults using pre-fault signals. Further, we introduce an uncertainty quantification approach using Monte Carlo dropout with the LSTM-CNN model to ascertain confidence in the prediction. The model was tested using multivariate time series signals from stable cavity operations and before faults. Initial results show that on the test dataset, the model can identify impending faults before their onset with an average 10-fold cross validation accuracy of 97.39% and a standard deviation of 0.12% using a 100-ms time window. It is also observed that the model performs better as the prediction time moves closer to the fault onset. For additional context, we compare the performance of the model with three machine-learning-based (ML) fault prediction models. Our proposed parallel LSTM-CNN-based DL method shows better performance than the ML-based methods.

Rahman, Md Monibor↗
DOE OSTI2023

A phenomenological wobbling model for isolated pulsars and the braking index

ABSTRACT An isolated pulsar is a rotating neutron star possessing a very high magnetic dipole moment, thus providing a powerful radiating mechanism. These stars loose rotational energy E through various processes, including a plasma wind originating from a highly magnetized magnetosphere and the emission of magnetic dipole radiation (MDR). Such phenomena produce a time decreasing angular velocity Ω(t) of the pulsar that is usually quantified in terms of its braking index. Although these mechanisms are widely acknowledged as the primary drivers of the spin evolution of isolated pulsars, it is plausible that other contributing factors influencing this effect have yet to be comprehensively investigated. Most of young isolated pulsars present a braking index different from that given by the MDR and plasma wind processes. Working in the weak field (Newtonian) limit, we take in this work a step forward in describing the evolution of such a system by allowing the star’s shape to wobble around an ellipsoidal configuration as a backreaction effect produced by its rotational deceleration. It is assumed that an internal damping of the oscillations occurs, thus introducing another form of energy loss in the system, and this phenomenon may be related to the deviation of the braking index from the models based on $\dot{E} \sim -\Omega ^4$ predictions. Numerical calculations suggest that the average braking index for typical isolated pulsars can be thus explained by a simple phenomenological model.

Araujo, E. C. A. (ORCID:0000000213520879)↗
DOE OSTI2020

Intermodel Spread in the Pattern Effect and Its Contribution to Climate Sensitivity in CMIP5 and CMIP6 Models

Radiative feedbacks depend on the spatial patterns of sea surface temperature (SST) and thus can change over time as SST patterns evolve—the so-called pattern effect. In this study, we investigate intermodel differences in the magnitude of the pattern effect and how these differences contribute to the spread in effective equilibrium climate sensitivity (ECS) within CMIP5 and CMIP6 models. Effective ECS in CMIP5 estimated from 150-yr-long abrupt4×CO2 simulations is on average 10% higher than that estimated from the early portion (first 50 years) of those simulations, which serves as an analog for historical warming; this difference is reduced to 7% on average in CMIP6. The (negative) net radiative feedback weakens over the course of the abrupt4×CO2 simulations in the vast majority of CMIP5 and CMIP6 models, but this weakening is less dramatic on average in CMIP6. For both ensembles, the total variance in the effective ECS is found to be dominated by the spread in radiative response on fast time scales, rather than the spread in feedback changes. Using Green’s functions derived from two AGCMs shows that the spread in feedbacks on fast time scales may be primarily due to differences in atmospheric model physics, whereas the spread in feedback evolution is primarily governed by differences in SST patterns. Intermodel spread in feedback evolution is well explained by differences in the relative warming in the West Pacific warm-pool regions for the CMIP5 models, but this relation fails to explain differences across the CMIP6 models, suggesting that a stronger sensitivity of extratropical clouds to surface warming may also contribute to feedback changes in CMIP6.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2020

Effects of groundwater pumping on ground surface temperature: A regional modeling study in the North China Plain

Over-exploitation of groundwater (GW) in the North China Plain (NCP) since the 1960s has many environmental consequences. However, understanding of the dominant mechanisms remains limited, particularly at regional scale. In this study, the coupled ParFlow.CLM model representing subsurface and land-surface processes and their interactions was applied in the NCP at high spatio-temporal resolutions. The model was validated using the water and energy fluxes reported in previous studies and from the JRA-55 reanalysis. Numerical experiments were designed to examine the relative impacts of GW pumping and irrigation on the ground surface temperature (GST). Results show significant effects of GW pumping on GST in the NCP. Generally, the subsurface acts as a buffer to temporal variations in heat fluxes at the land-surface, but long-term pumping can gradually weaken this buffer, resulting in increases in the spatio-temporal variability of GST, as exemplified by hotter summers and colder winters. Considering that changes of water table depth (WTD) can significantly affect land surface heat fluxes when WTD ranges roughly between 0.01–10 m, the 0.5 m/year increase of WTD simulated by the model due to pumping can continue to increase the average WTD and hence, disturb GST, for about 20 years from the pre-pumping WTD in the NCP, before the WTD exceeds 10 m. The variations of GST are expected to be faster initially and gradually slow down due to the nonlinear behaviors of GST with WTD. The findings from this study in the NCP may also have implications for other regions with GW depletion.

Yang, Chen↗
DOE OSTI2021

PARIS: Predicting application resilience using machine learning

The traditional method to study application resilience to errors in HPC applications uses fault injection (FI), a time-consuming approach. Furthermore, while analytical models have been built to overcome the inefficiencies of FI, they lack accuracy. In this paper, we present PARIS, a machine-learning method to predict application resilience that avoids the time-consuming process of random FI and provides higher prediction accuracy than analytical models. PARIS captures the implicit relationship between application characteristics and application resilience, which is difficult to capture using most analytical models. We overcome many technical challenges for feature construction, extraction, and selection to use machine learning in our prediction approach. Our evaluation on 16 HPC benchmarks shows that PARIS achieves high prediction accuracy. PARIS is up to 450x faster than random FI (49x on average). Compared to the state-of-the-art analytical model, PARIS is at least 63% better in terms of accuracy and has comparable execution time on average.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Analytical Modeling of Metal Foam Composite Phase Change Materials (PCM) in Thermal Energy Storage Using Asymptotic Analysis

The use of phase change materials (PCMs) for thermal energy storage can release or absorb a significant amount of latent heat during the freezing or melting process, offering a higher energy storage density. One of the main drawbacks of PCMs is their low thermal conductivity, resulting in poor thermal performance. Recent research has attempted to enhance heat transfer and increase the thermal conductivity of PCMs, including the use of metal foams. However, modeling the metal foam composite PCM using conventional methods is computationally expensive. This paper proposes an asymptotic solution for a Stefan-like problem subject to a convective boundary for outward solidification in a hollow cylinder, capable of predicting the freeze-melt cycle of the metal foam composite PCM. Specifically, three temporal regimes and four spatial layers are considered in the asymptotic analysis for each phase change process. The thermal conductivity is calculated by a theoretical three-dimensional tetrakaidecahedron model, while other thermophysical properties are obtained using the method of volume averaging. The results are verified with numerical data and validated against experimental data in the literature. The presented analytical modeling framework could have the potential to be applied to other types of composite PCMs with considerably lower computational costs compared with conventional methods.

analytical model↗
DOE OSTI2025

Standardising the “Gregory method” for calculating equilibrium climate sensitivity

The equilibrium climate sensitivity (ECS) – the equilibrium global mean temperature response to a doubling of atmospheric CO 2 – is a high-profile metric for quantifying the Earth system's response to human-induced climate change. A widely applied approach to estimating the ECS is the “Gregory method” (Gregory et al., 2004), which uses an ordinary least squares (OLS) regression between the net radiative flux, N, and surface air temperature anomalies, ΔT, from a 150 year experiment in which atmospheric CO 2 concentrations are quadrupled. The ECS is determined by extrapolating the linear fit to N=0, i.e. the ΔT-intercept, indicating the point at which the system is back in equilibrium. This method has been used to compare ECS estimates across the CMIP5 and CMIP6 ensembles and will likely be a key diagnostic for CMIP7. Despite its widespread application, there is little consistency or transparency between studies in how the climate model data is processed prior to the regression, leading to potential discrepancies in ECS estimates. We identify 32 alternative data processing pathways, varying by differences in global mean weighting, net radiative flux variable, anomaly calculation method, and linear regression fit. Using 44 CMIP6 models, we systematically assess the impact of these choices on ECS estimates and calculate uncertainty ranges using two bootstrap approaches. While the inter-model ECS range is insensitive to the data processing pathway, individual outlier models exhibit notable differences. Approximating a model's native grid cell area (if irregular) with cosine of the latitude can decrease the ECS by 11 %, the choice of N-variable can change the ECS by 6 %, and some anomaly calculation methods can introduce spurious temporal correlations in the processed data. Beyond data processing choices, we also evaluate an alternative linear regression method – total least squares (TLS) – which has a more statistically robust basis than OLS. However, for consistency with previous literature, and given TLS may reduce the ECS compared to OLS (by up to 24 %), thereby making a known bias in the Gregory method worse, we do not feel there is sufficient clarity to recommend a transition to TLS in all cases. To improve reproducibility and comparability in future studies, we recommend a standardised Gregory method: weighting the global mean by cell area, using the top of the atmosphere (as opposed to the top of model) N-variable, and calculating anomalies by first applying a rolling average to the preindustrial control timeseries then subtracting from the raw CO 2 quadrupling experiment. This approach accounts for model drift while reducing noise in the data to best meet the pre-conditions of the linear regression. While CMIP6 results of the multi-model mean ECS appear insensitive to these processing choices, similar assumptions may not hold for CMIP7, underscoring the need for standardised data preparation in future climate sensitivity assessments.

Geosciences↗
DOE OSTI2020

A substantial role of soil erosion in the land carbon sink and its future changes

Realistic representation of land carbon sink in climate models is vital for predicting carbon climate feedbacks in a changing world. Although soil erosion that removes land organic carbon has increased substantially since the onset of agriculture, it is rarely included in the current generation of climate models. Using an Earth system model with soil erosion represented, we estimated that on average soil erosion displaces 5% of newly fixed land organic carbon downslope annually in the continental US. In the lower Mississippi river basin and the Cascades, the fraction can be as large as 40%. About 12% of the eroded organic carbon is eventually exported to inland waters, which is equal to 14% of the simulated net carbon gain by terrestrial ecosystems. By comparing the eroded organic carbon export to rivers with the particulate organic carbon export to oceans, we demonstrated that a large fraction of the carbon export to rivers could have been mineralized in inland waters. Importantly, with a direct comparison of eroded and exported soil organic carbon and land net carbon uptake, we found that Earth system models that ignore soil erosion likely offset the erosional carbon loss by increasing heterotrophic respiration implicitly. But as soil erosion and heterotrophic respiration respond differently to a warming climate, this unrealistic compensation would lead to biased predictions of future land carbon sink.

Tan, Zeli↗
DOE OSTI2024

The SAGA Survey. V. Modeling Satellite Systems around Milky Way–Mass Galaxies with Updated UniverseMachine

Abstract Environment plays a critical role in shaping the assembly of low-mass galaxies. Here, we use the U niverse M achine (UM) galaxy–halo connection framework and Data Release 3 of the Satellites Around Galactic Analogs (SAGA) Survey to place dwarf galaxy star formation and quenching into a cosmological context. UM is a data-driven forward model that flexibly parameterizes galaxy star formation rates (SFRs) using only halo mass and assembly history. We add a new quenching model to UM, tailored for galaxies with m ⋆ ≲ 10 9 M ⊙ , and constrain the model down to m ⋆ ≳ 10 7 M ⊙ using new SAGA observations of 101 satellite systems around Milky Way (MW)–mass hosts and a sample of isolated field galaxies in a similar mass range from the Sloan Digital Sky Survey. The new best-fit model, “UM-SAGA,” reproduces the satellite stellar mass functions, average SFRs, and quenched fractions in SAGA satellites while keeping isolated dwarfs mostly star-forming. The enhanced quenching in satellites relative to isolated field galaxies leads the model to maximally rely on halo assembly to explain the observed environmental quenching. Extrapolating the model down to m ⋆ ∼ 10 6.5 M ⊙ yields a quenched fraction of ≳30% for isolated field galaxies and ≳80% for satellites of MW-mass hosts at this stellar mass. Spectroscopic surveys can soon test this specific prediction to reveal the relative importance of internal feedback, cessation of mass and gas accretion, satellite-specific gas processes, and reionization for the evolution of faint low-mass galaxies.

Wang, Yunchong (ORCID:000000018913626X)↗
DOE OSTI2025

Crack Identification and Characterization in Deformed Nb3Sn Rutherford Cable Stacks Using Machine Learning

An investigation of instance segmentation of cracks in Nb3Sn 4-stack 40-strand Rutherford cables using machine learning is presented. Three samples were uniaxially and biaxially loaded before metallographic inspections were performed. The Mask R-CNN model was used in the Detectron2 framework with pre-trained weights but fine-tuned to detect and segment cracks. The model detected cracks with bounding box and mask average precisions (AP) of 42.8 and 27.9, respectively, and was used for instance segmentation of all cracks in the three samples. More cracks were found in the sample pre-loaded along the z-axis (i.e., along the cable length). Pre-loading along the x-axis (i.e., on the cables edges) reduced the number of cracks and changed the crack orientation distribution, away from being highly aligned with the y-axis (i.e., normal to the cables broad faces), i.e., the direction with the highest applied load. Fine-tuning of the Segment Anything Model (SAM) was also studied but performed poorly without human-provided prompts. However, the zero-shot capability of SAM showed high promises to accelerate the image annotation process for applications beyond this study.

Croteau, Jean-Francois↗
DOE OSTI2020

Multi-Frequency Analysis of Simulated versus Observed Variability in Tropospheric Temperature

Studies seeking to identify a human-caused global warming signal generally rely on climate model estimates of the “noise” of intrinsic natural variability. Assessing the reliability of these noise estimates is of critical importance. We evaluate here the statistical significance of differences between climate model and observational natural variability spectra for global-mean mid- to upper-tropospheric temperature (TMT). We use TMT information from satellites and large multimodel ensembles of forced and unforced simulations. Our main goal is to explore the sensitivity of model-versus-data spectral comparisons to a wide range of subjective decisions. These include the choice of satellite and climate model TMT datasets, the method for separating signal and noise, the frequency range considered, and the statistical model used to represent observed natural variability. Of particular interest is the amplitude of the interdecadal noise against which an anthropogenic tropospheric warming signal must be detected. We find that on time scales of 5–20 years, observed TMT variability is (on average) overestimated by the last two generations of climate models participating in the Coupled Model Intercomparison Project. This result is relatively insensitive to different plausible analyst choices, enhancing confidence in previous claims of detectable anthropogenic warming of the troposphere and indicating that these claims may be conservative. A further key finding is that two commonly used statistical models of short-term and long-term memory have deficiencies in their ability to capture the complex shape of observed TMT spectra.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

Large Eddy Simulations of Turbulent and Buoyant Flows in Urban and Complex Terrain Areas Using the Aeolus Model

Fast and accurate predictions of the flow and transport of materials in urban and complex terrain areas are challenging because of the heterogeneity of buildings and land features of different shapes and sizes connected by canyons and channels, which results in complex patterns of turbulence that can enhance material concentrations in certain regions. To address this challenge, we have developed an efficient three-dimensional computational fluid dynamics (CFD) code called Aeolus that is based on first principles for predicting transport and dispersion of materials in complex terrain and urban areas. The model can be run in a very efficient Reynolds average Navier–Stokes (RANS) mode or a detailed large eddy simulation (LES) mode. The RANS version of Aeolus was previously validated against field data for tracer gas and radiological dispersal releases. As a part of this work, we have validated the Aeolus model in LES mode against two different sets of data: (1) turbulence quantities measured in complex terrain at Askervein Hill; and (2) wind and tracer data from the Joint Urban 2003 field campaign for urban topography. As a third set-up, we have applied Aeolus to simulate cloud rise dynamics for buoyant plumes from high-temperature explosions. For all three cases, Aeolus LES predictions compare well to observations and other models. These results indicate that Aeolus LES can be used to accurately simulate turbulent flow and transport for a wide range of applications and scales.

42 ENGINEERING↗
DOE OSTI2023

Predicting oxidation damage of ultra high-temperature carbide ceramics in extreme environments using machine learning

Determining the oxidation resistance of UHTC carbides in extreme environments is challenging theoretically and experimentally due to the high dimensional complexity of influencing variables and intricate testing setups. Herein we demonstrate the use of machine learning (ML) models trained with experimental literature data to predict the oxide thickness of UHTC carbides exposed to air based on composition, mean grain size, relative densification, holding time, and temperature. A multi-dimensional database with 76 occurrences is created containing experimental results of Hf, Zr, and Ta carbides plus additives. In this study, the preprocessed database is then used to train ML models to predict their oxidation behavior. The trained model predicts the oxidation damage in the form of an average oxide thickness in UHTC carbides with a Mean Absolute Error (MAE) of ±65.45 μm for samples in the testing set that developed thicknesses up to 1000 μm. The model successfully predicted oxidation damage for a recession rate lower than 60 μm/min. It is noticed that the ensemble method MAE is increased to ±134.34 μm while forecasting the oxidation of samples with a recession rate higher than the threshold. The unprecedented approach is a novel way to predict the damage through the oxidation of carbide compounds before processing for a smarter design with room for improvement.

36 MATERIALS SCIENCE↗
DOE OSTI2022

WinnowML: Stable feature selection for maximizing prediction accuracy of time-based system modeling

Online deep learning (ODL) has become an important methodology for modeling time-based performance of computer systems. An open problem is the intelligent selection of features from raw workload traces of computer systems. The best methods are overly sensitive to noisy data, causing frequent feature changes and re-training. Using all available features inflates training time and introduces model artifacts if some features should have been dropped. We present WinnowML, a method for automatically determining the most relevant feature subset for a predictive time-series model. WinnowML combines existing feature ranking algorithms and a history of each feature's ranking to iteratively rank a feature set to lower prediction error and maximize long term relevance. From this ranked feature set, the most relevant and stable subset is selected to train a model. Experimentally, we show how WinnowML can lower a model's mean absolute relative error up to 42% on average compared to the closest performing approach. Additionally, we lower the fluctuation in feature ranking and selection up to 65%. We also demonstrate how to combine WinnowML and a model search tool to provide improvements in performance of up to 14.5% when compared to using all the feature available.

Bel, Oceane MS↗