SEARCH · Search NASA

Results for “Deep uncertainty”

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.

SourceResearch connectionFrom yearThrough year

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

At least 217 records · Page 12

All sources NASA NTRS DOE OSTI

DOE OSTI2023

Efficient Bayesian inference with latent Hamiltonian neural networks in No-U-Turn Sampling

When sampling for Bayesian inference, one popular approach in the computational field is to use Hamiltonian Monte Carlo (HMC) and specifically the No-U-Turn Sampler (NUTS), which automatically decides the end time of the Hamiltonian trajectory. However, HMC and NUTS can require numerous numerical gradients of the target density and can prove slow in practice when relying on computationally expensive forward models. We propose Latent Hamiltonian neural networks (L-HNNs) with HMC and NUTS for solving Bayesian inference problems. Once trained, L-HNNs do not require numerical gradients of the target density during sampling, and hence numerous evaluations of the forward computational model. Moreover, L-HNNs satisfy important properties such as perfect time reversibility and Hamiltonian conservation, making them well-suited for use within HMC and NUTS because stationarity can be shown. We also propose the integration of L-HNNs in an online error monitoring scheme, in which numerical gradients of the target density are used for a few samples whenever the L-HNNs prediction errors are large. This online error monitor scheme prevents sample degeneracy in regions of low probability density and ensures robust uncertainty quantification. We demonstrate L-HNNs in NUTS with online error monitoring on several analytical examples involving complex, heavy-tailed, and high-local-curvature probability densities. We then demonstrate the applicability of L-HNNs in NUTS to two computational case studies, namely the Allen-Cahn stochastic partial differential equation and an elliptic partial differential equation with 25 and 50 inference parameters, respectively. Overall, the L-HNNs in NUTS with online error monitoring satisfactorily inferred these probability densities. In conclusion, compared to traditional NUTS, L-HNNs in NUTS with online error monitoring required 1–2 orders of magnitude fewer numerical gradients of the target density and improved the effective sample size (ESS) per gradient (which is a measure of both the sampling quality and the computational expense) by an order of magnitude.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2024

Convergence in simulating global soil organic carbon by structurally different models after data assimilation

Abstract Current biogeochemical models produce carbon–climate feedback projections with large uncertainties, often attributed to their structural differences when simulating soil organic carbon (SOC) dynamics worldwide. However, choices of model parameter values that quantify the strength and represent properties of different soil carbon cycle processes could also contribute to model simulation uncertainties. Here, we demonstrate the critical role of using common observational data in reducing model uncertainty in estimates of global SOC storage. Two structurally different models featuring distinctive carbon pools, decomposition kinetics, and carbon transfer pathways simulate opposite global SOC distributions with their customary parameter values yet converge to similar results after being informed by the same global SOC database using a data assimilation approach. The converged spatial SOC simulations result from similar simulations in key model components such as carbon transfer efficiency, baseline decomposition rate, and environmental effects on carbon fluxes by these two models after data assimilation. Moreover, data assimilation results suggest equally effective simulations of SOC using models following either first‐order or Michaelis–Menten kinetics at the global scale. Nevertheless, a wider range of data with high‐quality control and assurance are needed to further constrain SOC dynamics simulations and reduce unconstrained parameters. New sets of data, such as microbial genomics‐function relationships, may also suggest novel structures to account for in future model development. Overall, our results highlight the importance of observational data in informing model development and constraining model predictions.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2021

The development of rainfall retrievals from radar at Darwin

Abstract. The U.S. Department of Energy Atmospheric Radiation Measurement program Tropical Western Pacific site hosted a C-band polarization (CPOL) radar in Darwin, Australia. It provides 2 decades of tropical rainfall characteristics useful for validating global circulation models. Rainfall retrievals from radar assume characteristics about the droplet size distribution (DSD) that vary significantly. To minimize the uncertainty associated with DSD variability, new radar rainfall techniques use dual polarization and specific attenuation estimates. This study challenges the applicability of several specific attenuation and dual-polarization-based rainfall estimators in tropical settings using a 4-year archive of Darwin disdrometer datasets in conjunction with CPOL observations. This assessment is based on three metrics: statistical uncertainty estimates, principal component analysis (PCA), and comparisons of various retrievals from CPOL data. The PCA shows that the variability in R can be consistently attributed to reflectivity, but dependence on dual-polarization quantities was wavelength dependent for 1 10mmh-1. Rainfall estimates during these conditions primarily originate from deep convective clouds with median drop diameters greater than 1.5 mm. An uncertainty analysis and intercomparison with CPOL show that a Colorado State University blended technique for tropical oceans, with modified estimators developed from video disdrometer observations, is most appropriate for use in all cases, such as when 1 10mmh-1 (deeper convective rain).

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2025

Anomalous electroweak physics unraveled via evidential deep learning

The ever-growing ecosystem of beyond standard model (BSM) calculations and parametrizations has motivated the development of systematic methods for making quantitative cross-comparisons over the wide range of possible models, especially with controllable uncertainties. In this setting, the language of uncertainty quantification (UQ) furnishes useful metrics for assessing statistical overlaps and discrepancies among BSM and related models. In this study, we leverage recent machine learning (ML) developments in evidential deep learning (EDL) for UQ to separate data (aleatoric) and knowledge (epistemic) uncertainties in a model-discrimination setting. We construct several potentially BSM-motivated scenarios for the anomalous electroweak interaction (AEWI) of neutrinos with nucleons in deep inelastic scattering ( v DIS). These scenarios are then quantitatively mapped, as a demonstration, alongside Monte Carlo replicas of the CT18 PDFs used to calculate the $\varDelta \chi ^{2}$ statistic for a typical multi-GeV v DIS experiment, CDHSW. Our framework effectively highlights areas of model agreement and provides a classification of out-of-distribution (OOD) samples. By offering the opportunity to quantitatively understand model overlaps, the approach presented in this work can help facilitate efficient BSM model exploration and exclusion for future New Physics searches.

AI↗

DOE OSTI2022

Sparse Data Machine Learning Integration with Theory, Experiment and Uncertainty Quantification: Process-Structure-Property-Performance of Friction Deformation Processing

Computer vision and deep learning tools that advance the ability to establish processing-structure-property-performance (PSPP) relations are presented. The Bayesian binning method for image segmentation enables quantitative analysis of microstructural features in an automated way, while the analysis of shapes and relative orientation of these features reveals local deformation maps indicative of both, material flow and residual stresses due to materials processing. The deep learning method leads to the previous knowledge agnostic mapping of empirically observed microstructural zones in friction stir welding (FSW) process and synthetic microstructure generation capability that is statistically equivalent to experimentally collected data.

97 MATHEMATICS AND COMPUTING↗

NASA NTRS2004

Uncertainties in estimates of the risks of late effects from space radiation

Methods used to project risks in low-Earth orbit are of questionable merit for exploration missions because of the limited radiobiology data and knowledge of galactic cosmic ray (GCR) heavy ions, which causes estimates of the risk of late effects to be highly uncertain. Risk projections involve a product of many biological and physical factors, each of which has a differential range of uncertainty due to lack of data and knowledge. Using the linear-additivity model for radiation risks, we use Monte-Carlo sampling from subjective uncertainty distributions in each factor to obtain an estimate of the overall uncertainty in risk projections. The resulting methodology is applied to several human space exploration mission scenarios including a deep space outpost and Mars missions of duration of 360, 660, and 1000 days. The major results are the quantification of the uncertainties in current risk estimates, the identification of factors that dominate risk projection uncertainties, and the development of a method to quantify candidate approaches to reduce uncertainties or mitigate risks. The large uncertainties in GCR risk projections lead to probability distributions of risk that mask any potential risk reduction using the "optimization" of shielding materials or configurations. In contrast, the design of shielding optimization approaches for solar particle events and trapped protons can be made at this time and promising technologies can be shown to have merit using our approach. The methods used also make it possible to express risk management objectives in terms of quantitative metrics, e.g., the number of days in space without exceeding a given risk level within well-defined confidence limits. Published by Elsevier Ltd on behalf of COSPAR.

Non-NASA Center↗

NASA NTRS2011

Reliability Impacts in Life Support Architecture and Technology Selection

Equivalent System Mass (ESM) and reliability estimates were performed for different life support architectures based primarily on International Space Station (ISS) technologies. The analysis was applied to a hypothetical 1-year deep-space mission. High-level fault trees were initially developed relating loss of life support functionality to the Loss of Crew (LOC) top event. System reliability was then expressed as the complement (nonoccurrence) this event and was increased through the addition of redundancy and spares, which added to the ESM. The reliability analysis assumed constant failure rates and used current projected values of the Mean Time Between Failures (MTBF) from an ISS database where available. Results were obtained showing the dependence of ESM on system reliability for each architecture. Although the analysis employed numerous simplifications and many of the input parameters are considered to have high uncertainty, the results strongly suggest that achieving necessary reliabilities for deep-space missions will add substantially to the life support system mass. As a point of reference, the reliability for a single-string architecture using the most regenerative combination of ISS technologies without unscheduled replacement spares was estimated to be less than 1%. The results also demonstrate how adding technologies in a serial manner to increase system closure forces the reliability of other life support technologies to increase in order to meet the system reliability requirement. This increase in reliability results in increased mass for multiple technologies through the need for additional spares. Alternative parallel architecture approaches and approaches with the potential to do more with less are discussed. The tall poles in life support ESM are also reexamined in light of estimated reliability impacts.

Lange, Kevin E.↗

DOE OSTI2021

Uncertainty quantification for Multiphase-CFD simulations of bubbly flows: a machine learning-based Bayesian approach supported by high-resolution experiments

In this paper, we developed a machine learning-based Bayesian approach to inversely quantify and reduce the uncertainties of multiphase computational fluid dynamics (MCFD) simulations for bubbly flows. The proposed approach is supported by high-resolution two-phase flow measurements, including those by double-sensor conductivity probes, high-speed imaging, and particle image velocimetry. Local distributions of key physical quantities of interest (QoIs), including the void fraction and phasic velocities, are obtained to support the Bayesian inference. In the process, the epistemic uncertainties of the closure relations are inversely quantified while the aleatory uncertainties from stochastic fluctuations of the system are evaluated based on experimental uncertainty analysis. The combined uncertainties are then propagated through the MCFD solver to obtain uncertainties of the QoIs, based on which probability-boxes are constructed for validation. The proposed approach relies on three machine learning methods: feedforward neural networks and principal component analysis for surrogate modeling, and Gaussian processes for model form uncertainty modeling. The whole process is implemented within the framework of an open-source deep learning library PyTorch with graphics processing unit (GPU) acceleration, thus ensuring the efficiency of the computation. The results demonstrate that with the support of high-resolution data, the uncertainties of MCFD simulations can be significantly reduced. The proposed approach has the potential for other applications that involve numerical models with empirical parameters.

42 ENGINEERING↗

DOE OSTI2025

Precision beam diagnostics at the NuMI facility using muon monitor observations

The Neutrinos at the Main Injector (NuMI) facility at Fermilab delivers an intense neutrino beam for multiple experiments by producing pions that decay into neutrinos, muons, and other particles. Magnetic horns—the primary pion focusing elements in the NuMI beamline—exhibit predominantly linear optics, enabling a predictable relationship between the proton beam and the resulting pion and muon phase spaces. This study has two primary objectives: first, to evaluate and confirm the linearity of the horn focusing mechanism using analytical models and numerical simulations; and second, to demonstrate that key beam parameters—such as proton beam intensity, beam position on target, and horn current—can be extracted from muon monitor observations within this linear optics framework. Using a machine learning model trained on spill-by-spill muon monitor data, we infer the horn current with a precision of ±0.05%, the beam intensity with ±0.1%, and the beam position on target with ±0.018⁢ mm horizontally and ±0.013⁢ mm vertically. This approach provides a reliable cross-check of beam parameters, helping to reduce systematic uncertainties that are critical for future experiments such as the Deep Underground Neutrino Experiment, which will rely on the neutrino beam produced by the Long-Baseline Neutrino Facility.

Beam control↗

DOE OSTI2023

Comparative Analysis of Confidence Metrics for Nuclear Criticality Safety

Nuclear criticality safety standards provide guidance on the requirements and recommendations to establish confidence in computerized model results used to support operation with fissionable materials. By design, the guidance is not prescriptive, leaving the analysts free to determine how various sources of uncertainties are to be statistically aggregated. This report compares the analyses and key assumptions behind four notable methodologies documented in the nuclear criticality safety literature: the parametric, nonparametric, Whisper, and TSURFER methodologies. Because of the involved use of statistics entangled with heuristic recipes, the results of these methodologies are often difficult to interpret. Also, they are augmented by additional large administrative margins, eliminating the incentive to understand their differences. With the new resurgent wave of advanced nuclear systems focused on economizing operation—including advanced reactors, fuel cycles, and fuel concepts—there is a strong need to develop a clear understanding of uncertainties and their fusion methodologies to reduce uncertainties in a scientifically defensible manner. This report offers a deep dive into the various assumptions of the four noted methodologies, their adequacy, and their limitations, to provide guidance on developing confidence for the emergent nuclear systems. These systems are expected to be challenged by the scarcity of experimental data.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗

DOE OSTI2026

The SPT-3G+ Experiment on the South Pole Telescope

Observations of the cosmic microwave background (CMB) offer an unparalleled opportunity to advance our understanding of fundamental physics. SPT-3G+ is an upgraded receiver for the arcminute-resolution South Pole Telescope (SPT) that plans to deploy in late 2028. SPT-3G+ will increase the CMB mapping speed of SPT by nearly an order of magnitude over the currently installed SPT-3G receiver. SPT-3G+ will have ~24,000 transition-edge sensor (TES) bolometers in two frequency bands with center frequencies at 95 GHz and 150 GHz that will be read out with microwave multiplexing. SPT-3G+ will measure the CMB lensing spectrum and galaxy clusters to constrain the growth of structure, dark matter, and dark energy. SPT-3G+ will also reach critical thresholds on inflationary constraints by combining data with BICEP/Keck, forming the South Pole Observatory (SPO). BICEP/Keck has deep degree-angular scale measurements but is currently delensing-limited, while SPT-3G+ will provide deep lensing measurements. Forecasts show that SPO will reach an uncertainty on the tensor-to-scalar ratio $r$ of $\sigma(r) \sim 1.2\times 10^{-3}$ by 2034. A detection at these levels would provide evidence of inflation and probe new physics at grand unified theory energy scales, while no detection would exclude large classes of models and shift the scientific paradigm describing the early universe. I will give an overview of SPT-3G+ including its design and current status.

Simon, Sara M. [Fermilab] (ORCID:0009000006683584)↗

DOE OSTI2024

Neural Network Prediction of Strong Lensing Systems with Domain Adaptation and Uncertainty Quantification

Modeling strong gravitational lenses is computationally expensive for the complex data from modern and next-generation cosmic surveys. Deep learning has emerged as a promising approach for finding lenses and predicting lensing parameters, such as the Einstein radius. Mean-variance Estimators (MVEs) are a common approach for obtaining aleatoric (data) uncertainties from a neural network prediction. However, neural networks have not been demonstrated to perform well on out-of-domain target data successfully - e.g., when trained on simulated data and applied to real, observational data. In this work, we perform the first study of the efficacy of MVEs in combination with unsupervised domain adaptation (UDA) on strong lensing data. The source domain data is noiseless, and the target domain data has noise mimicking modern cosmology surveys. We find that adding UDA to MVE increases the accuracy on the target data by a factor of about two over an MVE model without UDA. Including UDA also permits much more well-calibrated aleatoric uncertainty predictions. Advancements in this approach may enable future applications of MVE models to real observational data.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

DOE OSTI2024

Bayesian reduced-order deep learning surrogate model for dynamic systems described by partial differential equations

We propose a reduced-order deep-learning surrogate model for dynamic systems described by time-dependent partial differential equations. This method employs space–time Karhunen–Loève expansions (KLEs) of the state variables and space-dependent KLEs of space-varying parameters to identify the reduced (latent) dimensions. Subsequently, a deep neural network (DNN) is used to map the parameter latent space to the state variable latent space. An approximate Bayesian method is developed for uncertainty quantification (UQ) in the proposed KL-DNN surrogate model. The KL-DNN method is tested for the linear advection–diffusion and nonlinear diffusion equations, and the Bayesian approach for UQ is compared with the deep ensembling (DE) approach, commonly used for quantifying uncertainty in DNN models. It was found that the approximate Bayesian method provides a more informative distribution of the PDE solutions in terms of the coverage of the reference PDE solutions (the percentage of nodes where the reference solution is within the confidence interval predicted by the UQ methods) and log predictive probability. The DE method is found to underestimate uncertainty and introduce bias. For the nonlinear diffusion equation, we compare the KL-DNN method with the Fourier Neural Operator (FNO) method and find that KL-DNN is 10% more accurate and needs less training time than the FNO method.

97 MATHEMATICS AND COMPUTING↗

NASA NTRS2007

Space Radiation Cancer Risks

Space radiation presents major challenges to astronauts on the International Space Station and for future missions to the Earth s moon or Mars. Methods used to project risks on Earth need to be modified because of the large uncertainties in projecting cancer risks from space radiation, and thus impact safety factors. We describe NASA s unique approach to radiation safety that applies uncertainty based criteria within the occupational health program for astronauts: The two terrestrial criteria of a point estimate of maximum acceptable level of risk and application of the principle of As Low As Reasonably Achievable (ALARA) are supplemented by a third requirement that protects against risk projection uncertainties using the upper 95% confidence level (CL) in the radiation cancer projection model. NASA s acceptable level of risk for ISS and their new lunar program have been set at the point-estimate of a 3-percent risk of exposure induced death (REID). Tissue-averaged organ dose-equivalents are combined with age at exposure and gender-dependent risk coefficients to project the cumulative occupational radiation risks incurred by astronauts. The 95% CL criteria in practice is a stronger criterion than ALARA, but not an absolute cut-off as is applied to a point projection of a 3% REID. We describe the most recent astronaut dose limits, and present a historical review of astronaut organ doses estimates from the Mercury through the current ISS program, and future projections for lunar and Mars missions. NASA s 95% CL criteria is linked to a vibrant ground based radiobiology program investigating the radiobiology of high-energy protons and heavy ions. The near-term goal of research is new knowledge leading to the reduction of uncertainties in projection models. Risk projections involve a product of many biological and physical factors, each of which has a differential range of uncertainty due to lack of data and knowledge. The current model for projecting space radiation cancer risk relies on the three assumptions of linearity, additivity, and scaling along with the use of population averages. We describe uncertainty estimates for this model, and new experimental data that sheds light on the accuracy of the underlying assumptions. These methods make it possible to express risk management objectives in terms of quantitative metrics, i.e., the number of days in space without exceeding a given risk level within well defined confidence limits. The resulting methodology is applied to several human space exploration mission scenarios including lunar station, deep space outpost, and a Mars mission. Factors that dominate risk projection uncertainties and application of this approach to assess candidate mitigation approaches are described.

Cucinotta, Francis A.↗

DOE OSTI2025

Accelerating Hamiltonian Monte Carlo for Bayesian inference in neural networks and neural operators

Hamiltonian Monte Carlo (HMC) is a powerful and accurate method to sample from the posterior distribution in Bayesian inference. However, HMC techniques are computationally demanding for Bayesian neural networks due to the high dimensionality of the network’s parameter space and the non-convexity of their posterior distributions. Therefore, various approximation techniques, such as variational inference (VI) or stochastic gradient MCMC, are often employed to infer the posterior distribution of the network parameters. Such approximations introduce inaccuracies in the inferred distributions, resulting in unreliable uncertainty estimates. In this work, we propose a hybrid approach that combines inexpensive VI and accurate HMC methods to efficiently and accurately quantify uncertainties in neural networks and neural operators. The proposed approach leverages an initial VI training on the full network. We examine the influence of individual parameters on the prediction uncertainty, which shows that a large proportion of the parameters do not contribute substantially to uncertainty in the network predictions. This information is then used to significantly reduce the dimension of the parameter space, and HMC is performed only for the subset of network parameters that strongly influence prediction uncertainties. This yields a framework for accelerating the full batch HMC for posterior inference in neural networks. We demonstrate the efficiency and accuracy of the proposed framework on deep neural networks and operator networks, showing that inference can be performed for large networks with tens to hundreds of thousands of parameters. Finally, we show that this method can effectively learn surrogates for complex physical systems by modeling the operator that maps from upstream conditions to wall-pressure data on a cone in hypersonic flow.

Bayesian inference↗

DOE OSTI2026

Deep Spectroscopy with DESI for Photometric Redshift Training and Calibration

Deep spectroscopic samples can improve photometric redshift (photo-z) estimates and reduce uncertainties on redshift distributions. Such improvements can increase the cosmological constraining power of large imaging-based experiments such as the Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST) and mitigate what may be a limiting systematic effect. We present results from the “DESI-Deep pilot” program, which was designed to assess the capability of the Dark Energy Spectroscopic Instrument (DESI) on the 4m Mayall telescope to measure redshifts of galaxies as faint as expected lensing samples for early LSST data (m i ≤ 24.5). We find that DESI is remarkably efficient at this task, with redshift success rates comparable to the results of observations from 10 m class telescopes with only ∼2 × longer integration time (rather than ∼8 × longer as would be expected from aperture-area scaling), while simultaneously achieving ∼30 times larger multiplexing. We also find that the signal-to-noise ratio of the spectra scales as expected for background-limited observations even for the longest exposure times (∼7 hr) and faintest targets in the program. These results demonstrate that DESI could provide the definitive redshift sample for the early years of LSST with a modest investment of observing time. Based upon the results of this program, we provide updated predictions for the time required to collect benchmark samples for photo-z training and calibration using a variety of spectroscopic facilities. Finally, we describe a potential “DESI-Deep” survey designed to train and calibrate photo-z’s for imaging experiments, and provide forecasts of its impact on cosmological inference.

Dey, Biprateep [University of Toronto; University ↗

DOE OSTI2022

Optimizing radar scan strategies for tracking isolated deep convection using observing system simulation experiments

Abstract. Optimizing radar observation strategies is one of the most important considerations in pre-field campaign periods. This is especially true for isolated convective clouds that typically evolve faster than the observations captured by operational radar networks. This study investigates uncertainties in radar observations of the evolution of the microphysical and dynamical properties of isolated deep convective clouds developing in clean and polluted environments. It aims to optimize the radar observation strategy for deep convection through the use of high-spatiotemporal cloud-resolving model simulations, which resolve the evolution of individual convective cells every 1 min, coupled with a radar simulator and a cell tracking algorithm. The radar simulation settings are based on the Tracking Aerosol Convection Interactions ExpeRiment (TRACER) and Experiment of Sea Breeze Convection, Aerosols, Precipitation and Environment (ESCAPE) field campaigns held in the Houston, TX, area but are generalizable to other field campaigns focusing on isolated deep convection. Our analysis produces the following four outcomes. First, a 5–7 m s−1 median difference in maximum updrafts of tracked cells is shown between the clean and polluted simulations in the early stages of the cloud lifetimes. This demonstrates the importance of obtaining accurate estimates of vertical velocity from observations if aerosol impacts are to be properly resolved. Second, tracking of individual cells and using vertical cross section scanning every minute capture the evolution of precipitation particle number concentration and size represented by polarimetric observables better than the operational radar observations that update the volume scan every 5 min. This approach also improves multi-Doppler radar updraft retrievals above 5 km above ground level for regions with updraft velocities greater than 10 m s−1. Third, we propose an optimized strategy composed of cell tracking by quick (1–2 min) vertical cross section scans from more than one radar in addition to the operational volume scans. We also propose the use of a single-RHI (range height indicator) updraft retrieval technique for cells close to the radars, for which multi-Doppler radar retrievals are still challenging. Finally, increasing the number of deep convective cells sampled by such observations better represents the median maximum updraft evolution with sample sizes of more than 10 deep cells, which decreases the error associated with sampling the true population to less than 3 m s−1.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2020

Building thermal load prediction through shallow machine learning and deep learning

Building thermal load prediction informs the optimization of cooling plant and thermal energy storage. Physics-based prediction models of building thermal load are constrained by the model and input complexity. In this study, we developed 12 data-driven models (7 shallow learning, 2 deep learning, and 3 heuristic methods) to predict building thermal load and compared shallow machine learning and deep learning. The 12 prediction models were compared with the measured cooling demand. It was found XGBoost (Extreme Gradient Boost) and LSTM (Long Short Term Memory) provided the most accurate load prediction in the shallow and deep learning category, and both outperformed the best baseline model, which uses the previous day's data for prediction. Then, we discussed how the prediction horizon and input uncertainty would influence the load prediction accuracy. Major conclusions are twofold: first, LSTM performs well in short-term prediction (1 h ahead) but not in long term prediction (24 h ahead), because the sequential information becomes less relevant and accordingly not so useful when the prediction horizon is long. Second, the presence of weather forecast uncertainty deteriorates XGBoost's accuracy and favors LSTM, because the sequential information makes the model more robust to input uncertainty. Training the model with the uncertain rather than accurate weather data could enhance the model's robustness. Finally, our findings have two implications for practice. First, LSTM is recommended for short-term load prediction given that weather forecast uncertainty is unavoidable. Second, XGBoost is recommended for long term prediction, and the model should be trained with the presence of input uncertainty.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗