Decadal Prediction and Predictability of Extremes in Ocean Eddy Resolving Coupled Models
Final Project Report
SEARCH · Search NASA
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.
Final Project Report
In the high level waste tanks at the Savannah River Site (SRS), hydrogen is produced continuously by interaction of the radiation in the tank with water in the waste. Consequently, the vapor spaces of the tanks are purged to prevent the accumulation of H{sub 2} and possible formation of a flammable mixture in a tank. Personnel at SRS have developed an empirical model to predict the rate of H{sub 2} formation in a tank. The basis of this model is the prediction of the G value for H{sub 2} production. This G value is the number of H{sub 2} molecules produced per 100 eV of radiolytic energy absorbed by the waste. Based on experimental studies it was found that the G value for H{sub 2} production from beta radiation and from gamma radiation were essentially equal. The G value for H{sub 2} production from alpha radiation was somewhat higher. Thus, the model has two equations, one for beta/gamma radiation and one for alpha radiation. Experimental studies have also indicated that both G values are decreased by the presence of nitrate and nitrite ions in the waste. These are the main scavengers for the precursors of H{sub 2} in the waste; thus the equations that were developed predict G values for hydrogen production as a function of the concentrations of these two ions in waste. Knowing the beta/gamma and alpha heat loads in the waste allows one to predict the total generation rate for hydrogen in a tank. With this prediction a ventilation rate can be established for each tank to ensure that a flammable mixture is not formed in the vapor space in a tank. Recently personnel at Hanford have developed a slightly different model for predicting hydrogen G values. Their model includes the same precursor for H{sub 2} as the SRS model but also includes an additional precursor not in the SRS model. Including the second precursor for H{sub 2} leads to different empirical equations for predicting the G values for H{sub 2} as a function of the nitrate and nitrite concentrations in the waste. The difference in the two models has led to the questions of how different are the results predicted by the two models and which model predicts the more conservative (larger) G values. More conservative G values would predict higher H{sub 2} generation rates that would require higher ventilation rates in the SRS tanks. This report compares predictions based on the two models at various nitrate and nitrite concentrations in the SRS HLW tanks for both beta/gamma and for alpha radiation. It also compares predicted G values with those determined by actually measuring the H{sub 2} production from four SRS HLW tanks (Tanks 32H, 35H, 39H, and 42H). Lastly, the H{sub 2} generation rates predicted by the two models are compared for the 47 active SRS high level waste tanks using the most recent tank nitrate and nitrite concentrations and the beta/gamma and alpha heat loads for each tank. The predictions of the models for total H{sub 2} generation rates from the 47 active SRS waste were, for the most part, similar. For example, the predictions for both models applied to 25 tanks agreed within {+-}10% of each other. For the remaining 22 tanks, the SRS prediction was more conservative for 9 tanks (maximum 29% higher) and the Hanford prediction was more conservative for 13 tanks (maximum 19% higher). When comparing G values predicted by the equations presuming only alpha radiation or only beta/gamma was present the results were somewhat different. The results of predictions for alpha radiation, at the 47 current nitrate and nitrite concentrations in the SRS tanks indicated that all the SRS predictions were higher (up to 30%) than the Hanford predictions and thus more conservative. For beta/gamma radiation the predictions for both models agreed to {+-}10% for 18 of the combinations, the Hanford model predicted higher values (11 up to 17%) for 25 of the concentrations considered, and the SRS model predicted higher G values for the remaining two combinations (12 and 17%). For the four SRS tanks, where we compared measured G values to those predicted by the two different models, the results for two tanks (Tanks 35 and 39) were in good agreement with predictions from both models. For the other two tanks (Tanks 32 and 42) the predictions of both models were conservative. The predictions were 3 to 4X higher than the measured G values for H{sub 2} production.
Abstract Background Driven by deep learning, inter-residue contact/distance prediction has been significantly improved and substantially enhanced ab initio protein structure prediction. Currently, most of the distance prediction methods classify inter-residue distances into multiple distance intervals instead of directly predicting real-value distances. The output of the former has to be converted into real-value distances to be used in tertiary structure prediction. Results To explore the potentials of predicting real-value inter-residue distances, we develop a multi-task deep learning distance predictor (DeepDist) based on new residual convolutional network architectures to simultaneously predict real-value inter-residue distances and classify them into multiple distance intervals. Tested on 43 CASP13 hard domains, DeepDist achieves comparable performance in real-value distance prediction and multi-class distance prediction. The average mean square error (MSE) of DeepDist’s real-value distance prediction is 0.896 Å 2 when filtering out the predicted distance ≥ 16 Å, which is lower than 1.003 Å 2 of DeepDist’s multi-class distance prediction. When distance predictions are converted into contact predictions at 8 Å threshold (the standard threshold in the field), the precision of top L/5 and L/2 contact predictions of DeepDist’s multi-class distance prediction is 79.3% and 66.1%, respectively, higher than 78.6% and 64.5% of its real-value distance prediction and the best results in the CASP13 experiment. Conclusions DeepDist can predict inter-residue distances well and improve binary contact prediction over the existing state-of-the-art methods. Moreover, the predicted real-value distances can be directly used to reconstruct protein tertiary structures better than multi-class distance predictions due to the lower MSE. Finally, we demonstrate that predicting the real-value distance map and multi-class distance map at the same time performs better than predicting real-value distances alone.
Accurately predicting Li-ion battery capacity trajectories using early-life data can dramatically improve battery-life understandings and be used to rapidly evaluate design/cost/performance trade-offs when developing new battery materials. Accurate early-life predictions enable researchers to quickly iterate over cell designs and material precursor properties without consistently cycling cells to failure. To this end, we present a toolbox that uses a combined Gaussian Process and Bayesian regression approach that capitalizes on signals other than just capacity (e.g., dQ/dV, voltage drops) to rapidly predict capacity-fade trajectories. The prediction tool uses Bayesian regression to fit functional forms, e.g., power law, sigmoids, etc., to predict capacity-fade dynamics. By fitting functional forms, the capacity fade can be interrogated at any point in the future, allowing for early cell-failure prediction. Additionally, Bayesian regression allows for accurate uncertainty estimates that account for cell-to-cell variability (aleatoric uncertainty) and the lack of observation data (epistemic uncertainty). By only using early cycle data to predict the capacity fade trajectory, uncertainty bounds at end-of-life can be extremely large. The large uncertainty bounds are further exacerbated because there is no systematic way to define the prior distribution of the functional forms' parameters. We improve our the predicted trajectory confidence interval of our predicted trajectory using two methods. First, we shows that a small amount of held-out cycling data is sufficientuse some train cells, that have been cycled to failure to derive information regarding the appropriate prior distributions for the functional forms' parameters of the functional form, effectively leading to data-driven priors.. We propose constructing the data-driven priors by first running a Bayesian regression starting with uninformed priors to generate intermediate cell-specific posterior parameter distributions. These posterior distributions are combined using a Ggaussian mixture model for each parameter to create the data-driven priors. These mixture models serve as the data-driven prior distributions for the parameters for. Second, we derive multiple features, e.g., C_dchg 0.5 DoD 0.5, log (|mean(dQ/dV_(w_3-w_0 ) (V)|), etc., from the train cellsheld-out cycling data, identify which the features are that best predicting capacity at early/mid-life cycles, and then create Ggaussian process regression models that are used for predicting capacity at early/mid-life cycles for the test cells (see blue dots with error bars in Fig 1b). Finally, these predicted data-points are used in addition to the actual early cycle data capacity fade to construct the Bayesian regression trajectory for the test cell s. Notably. We note that these two methods are complementary and can be combined with each other. We evaluate the performance of our proposed method on an testing open-source dataset from Iowa State University and Iowa Lakes Community College (ISU-ILCC). This dataset comprises of 251 nickel-manganese-cobalt/graphite Lithium-ion cells that are cycled under 63 different conditions. We compute the mean average percentage error (MAPE) and negative log predictive density (NLPD) to quantify the efficacy of our method. Our initial findings suggest that, when only few observations are available, for test cells, when using only Bayesian regression with uninformed priors, a power law functional provides the most accurate predictions. with very few data points. However, asHowever, a the number of data points increases, a twin sigmoidal function becomes more accurate as the number of observations further increases. We also find that using as little as 10% of the data set towards generating data-driven priors can lead to significant improvement in prediction accuracy when using early cycle data. Lastly, we found that augmenting early-cycle data with Gaussian process-predicted capacity data for Bayesian regression greatly improves the prediction accuracy. We will present a comprehensive comparison of our methods to other methods available in the literature and apply this method to additional battery datasets.
Purpose: Respiratory-gated radiotherapy and dynamic tracking employ real-time imaging and surrogate motion-monitoring methods with tumor motion prediction in advance of real-time. This study investigated respiratory motion data length on prediction accuracy of tumor motion. Methods: Predictions generated from the algorithm are validated against a one-dimensional surrogate signal of amplitude versus time. Prediction consists of three major components: extracting top-ranked subcomponents from training data matching the last respiratory cycle; calculating weighting factors from best-matched subcomponents; fusing data proceeding best-matched subcomponents with respective weighting factors to form predictions. Predictions for one respiratory cycle (∼3-6seconds) were assessed using 351 patient data from the respiratory management device. Performance was evaluated for correlation coefficient and root mean square error (RMSE) between prediction and final respiratory cycle. Results: Respiratory prediction results fell into two classes, where best predictions for 70 cycles or less performed using relative prediction and greater than 70 cycles are predicted similarly using relative and derivative relative. For 70 respiratory cycles or less, the average correlation between prediction and final respiratory cycle was 0.9999±0.0001, 0.9999±0.0001, 0.9988±0.0003, 0.9985±0.0023, and 0.9981±0.0023 with RMSE values of 0.0091±0.0030, 0.0091±0.0030, 0.0305±0.0051, 0.0299±0.0259, and 0.0299±0.0259 for equal, relative, pattern, derivative equal and derivative relative weighting methods, respectively. Respectively, the total best prediction for each method was 37, 65, 20, 22, and 22. For data with greater than 70 cycles average correlation was 0.9999±0.0001, 0.9999±0.0001, 0.9988±0.0004, 0.9988±0.0020, and 0.9988±0.0020 with RMSE values of 0.0081±0.0031, 0.0082±0.0033, 0.0306±0.0056, 0.0218±0.0222, and 0.0218±0.0222 for equal, relative, pattern, derivative equal and derivative relative weighting methods, respectively. Respectively, the total best prediction for each method was 24, 44, 42, 30, and 45. Conclusion: The prediction algorithms are effective in estimating surrogate motion in advance. These results indicate an advantage in using relative prediction for shorter data and either relative or derivative relative prediction for longer data.
Purpose: The objective of this work was to develop a comprehensive knowledge-based methodology for predicting achievable dose–volume histograms (DVHs) and highly precise DVH-based quality metrics (QMs) in stereotactic radiosurgery/radiotherapy (SRS/SRT) plans. Accurate QM estimation can identify suboptimal treatment plans and provide target optimization objectives to standardize and improve treatment planning. Methods: Correlating observed dose as it relates to the geometric relationship of organs-at-risk (OARs) to planning target volumes (PTVs) yields mathematical models to predict achievable DVHs. In SRS, DVH-based QMs such as brain V{sub 10Gy} (volume receiving 10 Gy or more), gradient measure (GM), and conformity index (CI) are used to evaluate plan quality. This study encompasses 223 linear accelerator-based SRS/SRT treatment plans (SRS plans) using volumetric-modulated arc therapy (VMAT), representing 95% of the institution’s VMAT radiosurgery load from the past four and a half years. Unfiltered models that use all available plans for the model training were built for each category with a stratification scheme based on target and OAR characteristics determined emergently through initial modeling process. Model predictive accuracy is measured by the mean and standard deviation of the difference between clinical and predicted QMs, δQM = QM{sub clin} − QM{sub pred}, and a coefficient of determination, R{sup 2}. For categories with a large number of plans, refined models are constructed by automatic elimination of suspected suboptimal plans from the training set. Using the refined model as a presumed achievable standard, potentially suboptimal plans are identified. Predictions of QM improvement are validated via standardized replanning of 20 suspected suboptimal plans based on dosimetric predictions. The significance of the QM improvement is evaluated using the Wilcoxon signed rank test. Results: The most accurate predictions are obtained when plans are stratified based on proximity to OARs and their PTV volume sizes. Volumes are categorized into small (V{sub PTV} < 2 cm{sup 3}), medium (2 cm{sup 3} < V{sub PTV} < 25 cm{sup 3}), and large (25 cm{sup 3} < V{sub PTV}). The unfiltered models demonstrate the ability to predict GMs to ∼1 mm and fractional brain V{sub 10Gy} to ∼25% for plans with large V{sub PTV} and critical OAR involvements. Increased accuracy and precision of QM predictions are obtained when high quality plans are selected for the model training. For the small and medium V{sub PTV} plans without critical OAR involvement, predictive ability was evaluated using the refined model. For training plans, the model predicted GM to an accuracy of 0.2 ± 0.3 mm and fractional brain V{sub 10Gy} to 0.04 ± 0.12, suggesting highly accurate predictive ability. For excluded plans, the average δGM was 1.1 mm and fractional brain V{sub 10Gy} was 0.20. These δQM are significantly greater than those of the model training plans (p < 0.001). For CI, predictions are close to clinical values and no significant difference was observed between the training and excluded plans (p = 0.19). Twenty outliers with δGM > 1.35 mm were identified as potentially suboptimal, and replanning these cases using predicted target objectives demonstrates significant improvements on QMs: on average, 1.1 mm reduction in GM (p < 0.001) and 23% reduction in brain V{sub 10Gy} (p < 0.001). After replanning, the difference of δGM distribution between the 20 replans and the refined model training plans was marginal. Conclusions: The results demonstrate the ability to predict SRS QMs precisely and to identify suboptimal plans. Furthermore, the knowledge-based DVH predictions were directly used as target optimization objectives and allowed a standardized planning process that bettered the clinically approved plans. Full clinical application of this methodology can improve consistency of SRS plan quality in a wide range of PTV volume and proximity to OARs and facilitate automated treatment planning for this critical treatment site.
Purpose: To demonstrate knowledge-based 3D dose prediction for external beam radiotherapy. Methods: Using previously treated plans as training data, an artificial neural network (ANN) was trained to predict a dose matrix based on patient-specific geometric and planning parameters, such as the closest distance (r) to planning target volume (PTV) and organ-at-risks (OARs). Twenty-three prostate and 43 stereotactic radiosurgery/radiotherapy (SRS/SRT) cases with at least one nearby OAR were studied. All were planned with volumetric-modulated arc therapy to prescription doses of 81 Gy for prostate and 12–30 Gy for SRS. Using these clinically approved plans, ANNs were trained to predict dose matrix and the predictive accuracy was evaluated using the dose difference between the clinical plan and prediction, δD = D{sub clin} − D{sub pred}. The mean (〈δD{sub r}〉), standard deviation (σ{sub δD{sub r}}), and their interquartile range (IQR) for the training plans were evaluated at a 2–3 mm interval from the PTV boundary (r{sub PTV}) to assess prediction bias and precision. Initially, unfiltered models which were trained using all plans in the cohorts were created for each treatment site. The models predict approximately the average quality of OAR sparing. Emphasizing a subset of plans that exhibited superior to the average OAR sparing during training, refined models were created to predict high-quality rectum sparing for prostate and brainstem sparing for SRS. Using the refined model, potentially suboptimal plans were identified where the model predicted further sparing of the OARs was achievable. Replans were performed to test if the OAR sparing could be improved as predicted by the model. Results: The refined models demonstrated highly accurate dose distribution prediction. For prostate cases, the average prediction bias for all voxels irrespective of organ delineation ranged from −1% to 0% with maximum IQR of 3% over r{sub PTV} ∈ [ − 6, 30] mm. The average prediction error was less than 10% for the same r{sub PTV} range. For SRS cases, the average prediction bias ranged from −0.7% to 1.5% with maximum IQR of 5% over r{sub PTV} ∈ [ − 4, 32] mm. The average prediction error was less than 8%. Four potentially suboptimal plans were identified for each site and subsequent replanning demonstrated improved sparing of rectum and brainstem. Conclusions: The study demonstrates highly accurate knowledge-based 3D dose predictions for radiotherapy plans.
Abstract Motivation Accurate prediction of residue–residue distances is important for protein structure prediction. We developed several protein distance predictors based on a deep learning distance prediction method and blindly tested them in the 14th Critical Assessment of Protein Structure Prediction (CASP14). The prediction method uses deep residual neural networks with the channel-wise attention mechanism to classify the distance between every two residues into multiple distance intervals. The input features for the deep learning method include co-evolutionary features as well as other sequence-based features derived from multiple sequence alignments (MSAs). Three alignment methods are used with multiple protein sequence/profile databases to generate MSAs for input feature generation. Based on different configurations and training strategies of the deep learning method, five MULTICOM distance predictors were created to participate in the CASP14 experiment. Results Benchmarked on 37 hard CASP14 domains, the best performing MULTICOM predictor is ranked 5th out of 30 automated CASP14 distance prediction servers in terms of precision of top L/5 long-range contact predictions [i.e. classifying distances between two residues into two categories: in contact (<8 Angstrom) and not in contact otherwise] and performs better than the best CASP13 distance prediction method. The best performing MULTICOM predictor is also ranked 6th among automated server predictors in classifying inter-residue distances into 10 distance intervals defined by CASP14 according to the precision of distance classification. The results show that the quality and depth of MSAs depend on alignment methods and sequence databases and have a significant impact on the accuracy of distance prediction. Using larger training datasets and multiple complementary features improves prediction accuracy. However, the number of effective sequences in MSAs is only a weak indicator of the quality of MSAs and the accuracy of predicted distance maps. In contrast, there is a strong correlation between the accuracy of contact/distance predictions and the average probability of the predicted contacts, which can therefore be more effectively used to estimate the confidence of distance predictions and select predicted distance maps. Availability and implementation The software package, source code and data of DeepDist2 are freely available at https://github.com/multicom-toolbox/deepdist and https://zenodo.org/record/4712084#.YIIM13VKhQM. Supplementary information Supplementary data are available at Bioinformatics online.
A new approach to signal prediction and prognostic assessment of spacecraft health resolves an inherent difficulty in fusing sensor data with simulated data. This technique builds upon previous work that demonstrated the importance of physics-based transient models to accurate prediction of signal dynamics and system performance. While models can greatly improve predictive accuracy, they are difficult to apply in general because of variations in model type, accuracy, or intended purpose. However, virtually any flight project will have at least some modeling capability at its disposal, whether a full-blown simulation, partial physics models, dynamic look-up tables, a brassboard analogue system, or simple hand-driven calculation by a team of experts. Many models can be used to develop a predict, or an estimate of the next day s or next cycle s behavior, which is typically used for planning purposes. The fidelity of a predict varies from one project to another, depending on the complexity of the simulation (i.e. linearized or full differential equations) and the level of detail in anticipated system operation, but typically any predict cannot be adapted to changing conditions or adjusted spacecraft command execution. Applying a predict blindly, without adapting the predict to current conditions, produces mixed results at best, primarily due to mismatches between assumed execution of spacecraft activities and actual times of execution. This results in the predict becoming useless during periods of complicated behavior, exactly when the predict would be most valuable. Each spacecraft operation tends to show up as a transient in the data, and if the transients are misaligned, using the predict can actually harm forecasting performance. To address this problem, the approach here expresses the predict in terms of a baseline function superposed with one or more transient functions. These transients serve as signal templates, which can be relocated in time and space against the signal background. One then has the ability to reconstruct a signal regardless of the precise timing of the transients. During operation, one applies the actual start times of spacecraft activities as they occur, and produces a reconstructed, accurate predict in real-time. This general approach is valid under two important conditions. First, the transients themselves must be time-invariant. Second, the transients must be reasonably consistent with respect to different operating points. Both of these assumptions are generally valid, but in the case of a complicated system with numerous types of overlapping transients, this approach may not be effective. Fortunately, there tend to be few transients in spacecraft telemetry of sensor quantities or low-level health and status information because these signals rarely reflect multiple different types of operation. Furthermore, if the predict is at least reasonably close to actual operation, the shift either in time, or in the operating point when the transient occurs - is likely to be small. The proposed approach considers three ways to recognize a transient. The first and most reliable is to use a different signal that identifies operating mode - often the transient will be correlated to a change in operating mode, and this change can usually be detected positively from discrete signals in spacecraft telemetry. If there is no useful mode signal, the second option is to identify the transient template by hand, and detect the onset of the transient using a curve-fitting approach. Finally, if one elects not to choose by hand, there is an option for automatic selection. This technique has been applied to sensor data from several JPL missions and industrial applications, demonstrating an improvement in accurate prediction of future behavior and early detection of system problems. This technique is applicable to practically any time-varying, quantitative sensor measurement.
Western US snowpack—snow that accumulates on the ground in the mountains—plays a critical role in regional hydroclimate and water supply, with 80% of snowmelt runoff being used for agriculture. While climate projections provide estimates of snowpack loss by the end of the century and weather forecasts provide predictions of weather conditions out to 2 weeks, less progress has been made for snow predictions at seasonal timescales (months to 2 years), crucial for regional agricultural decisions (e.g., plant choice and quantity). Seasonal predictions with climate models first took the form of El Niño predictions 3 decades ago, with hydroclimate predictions emerging more recently. While the field has been focused on single-season predictions (3 months or less), we are now poised to advance our predictions beyond this timeframe. Utilizing observations, climate indices, and a suite of global climate models, we demonstrate the feasibility of seasonal snowpack predictions and quantify the limits of predictive skill 8 months in advance. This physically based dynamic system outperforms observation-based statistical predictions made on July 1 for March snowpack everywhere except the southern Sierra Nevada, a region where prediction skill is nonexistent for every predictor presently tested. Additionally, in the absence of externally forced negative trends in snowpack, narrow maritime mountain ranges with high hydroclimate variability pose a challenge for seasonal prediction in our present system; natural snowpack variability may inherently be unpredictable at this timescale. This work highlights present prediction system successes and gives cause for optimism for developing seasonal predictions for societal needs.
In this study, a thorough and definitive evaluation of Predictive Optimal Energy Management Strategy (POEMS) applications in connected vehicles using 10 to 20 s predicted velocity is conducted for a Hybrid Electric Vehicle (HEV). The presented methodology includes synchronous datasets gathered in Fort Collins, Colorado using a test vehicle equipped with sensors to measure ego vehicle position and motion and that of surrounding objects as well as receive Infrastructure to Vehicle (I2V) information. These datasets are utilized to compare the effect of different signal categories on prediction fidelity for different prediction horizons within a POEMS framework. Multiple artificial intelligence (AI) and machine learning (ML) algorithms use the collected data to output future vehicle velocity prediction models. The effects of different combinations of signals and different models on prediction fidelity in various prediction windows are explored. All of these combinations are ultimately addressed where the rubber meets the road: fuel economy (FE) enabled from POEMS. FE optimization is performed using Model Predictive Control (MPC) with a Dynamic Programming (DP) optimizer. FE improvements from MPC control at various prediction time horizons are compared to that of full-cycle DP. All FE results are determined using high-fidelity simulations of an Autonomie 2010 Toyota Prius model. The full-cycle DP POEMS provides the theoretical upper limit on fuel economy (FE) improvement achievable with POEMS but is not currently practical for real-world implementation. Perfect prediction MPC (PP-MPC) represents the upper limit of FE improvement practically achievable with POEMS. Real-Prediction MPC (RP-MPC) can provide nearly equivalent FE improvement when used with high-fidelity predictions. Constant-Velocity MPC (CV-MPC) uses a constant speed prediction and serves as a “null” POEMS. Results showed that RP-MPC, enabled by high-fidelity ego future speed prediction, led to significant FE improvement over baseline nearly matching that of PP-MPC.
Predictive control has gained increasing attention for its ability to reduce energy consumption and improve occupant comfort in buildings. The plug loads prediction is a key component for the predictive building controls, as plug loads is a major source of internal heat gains in buildings. This study proposed a novel method to apply the Long-Short-Term-Memory (LSTM) Network, a special form of Recurrent Neural Network, to predict plug loads. The occupant count and the time have been confirmed to drive the plug load profile and thus selected as the features for the plug load prediction. The LSTM network was trained and tested with ground truth occupant count data collected from a real office building in Berkeley, California. Results from the LSTM network markedly improve the prediction accuracy compared with traditional linear regression methods and the classical Artificial Neural Network. 95% of 1-hour predictions from LSTM network are within ±1 kW of the actual plug loads, given the average plug loads during the office hour is 8.6 kW. The CV(RMSE) of the predicted plug load is 11% for the next hour, and 20% for the next eight hours. Lastly, we compared four prediction approaches with the office building we monitored: LSTM vs. ARIMA, with occupant counts vs. without occupant counts. It was found, the prediction error of the LSTM approach is around 4% less than the ARIMA approach. Using occupant counts as an exogenous input could further reduce the prediction error by 5%-6%. The findings of this paper could shed light on the plug load prediction for building control optimizations such as model-predictive control.
Abstract This study focuses on assessing the representation and predictability of North American weather regimes, which are persistent large-scale atmospheric patterns, in a set of initialized subseasonal reforecasts created using the Community Earth System Model, version 2 (CESM2). The k -means clustering was used to extract four key North American (10°–70°N, 150°–40°W) weather regimes within ERA5 reanalysis, which were used to interpret CESM2 subseasonal forecast performance. Results show that CESM2 can recreate the climatology of the four main North American weather regimes with skill but exhibits biases during later lead times with overoccurrence of the West Coast high regime and underoccurrence of the Greenland high and Alaskan ridge regimes. Overall, the West Coast high and Pacific trough regimes exhibited higher predictability within CESM2, partly related to El Niño. Despite biases, several reforecasts were skillful and exhibited high predictability during later lead times, which could be partly attributed to skillful representation of the atmosphere from the tropics to extratropics upstream of North America. The high predictability at the subseasonal time scale of these case-study examples was manifested as an “ensemble realignment,” in which most ensemble members agreed on a prediction despite ensemble trajectory dispersion during earlier lead times. Weather regimes were also shown to project distinct temperature and precipitation anomalies across North America that largely agree with observational products. This study further demonstrates that unsupervised learning methods can be used to uncover sources and limits of subseasonal predictability, along with systematic biases present in numerical prediction systems. Significance Statement North American weather regimes are large-scale atmospheric patterns that can persist for several days. Their skillful subseasonal (2 weeks or greater) prediction can provide valuable lead time to prepare for temperature and precipitation anomalies that can stress energy and water resources. The purpose of this study was to assess the climatological representation and subseasonal predictability of four key North American weather regimes using a research subseasonal prediction system and clustering analysis. We found that the Pacific trough and West Coast high regimes exhibited higher predictability than other regimes and that skillful representation of conditions across the tropics and extratropics can increase predictability during later lead times. Future work will quantify causal pathways associated with high predictability.
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.
Polycyclic aromatic hydrocarbons (PAHs) are present in the environment as complex mixtures with components that have diverse carcinogenic potencies and mostly unknown interactive effects. Non-additive PAH interactions have been observed in regulation of cytochrome P450 (CYP) gene expression in the CYP1 family. To better understand and predict biological effects of complex mixtures, such as environmental PAHs, an 11 gene input-1 gene output fuzzy neural network (FNN) was developed for predicting PAH-mediated perturbations of dermal Cyp1b1 transcription in mice. Input values were generalized using fuzzy logic into low, medium, and high fuzzy subsets, and sorted using k-means clustering to create Mamdani logic functions for predicting Cyp1b1 mRNA expression. Model testing was performed with data from microarray analysis of skin samples from FVB/N mice treated with toluene (vehicle control), dibenzo[def,p]chrysene (DBC), benzo[a]pyrene (BaP), or 1 of 3 combinations of diesel particulate extract (DPE), coal tar extract (CTE) and cigarette smoke condensate (CSC) using leave-one-out cross-validation. Predictions were within 1 log{sub 2} fold change unit of microarray data, with the exception of the DBC treatment group, where the unexpected down-regulation of Cyp1b1 expression was predicted but did not reach statistical significance on the microarrays. Adding CTE to DPE was predicted to increase Cyp1b1 expression, whereas adding CSC to CTE and DPE was predicted to have no effect, in agreement with microarray results. The aryl hydrocarbon receptor repressor (Ahrr) was determined to be the most significant input variable for model predictions using back-propagation and normalization of FNN weights. - Highlights: ► Tested a model to predict PAH mixture-mediated changes in Cyp1b1 expression ► Quantitative predictions in agreement with microarrays for Cyp1b1 induction ► Unexpected difference in expression between DBC and other treatments predicted ► Model predictions for combining PAH mixtures in agreement with microarrays ► Predictions highly dependent on aryl hydrocarbon receptor repressor expression.
Abstract Many problems in science and engineering require making predictions based on few observations. To build a robust predictive model, these sparse data may need to be augmented with simulated data, especially when the design space is multi-dimensional. Simulations, however, often suffer from an inherent bias. Estimation of this bias may be poorly constrained not only because of data sparsity, but also because traditional predictive models fit only one type of observed outputs, such as scalars or images, instead of all available output data modalities, which might have been acquired and simulated at great cost. To break this limitation and open up the path for multi-modal calibration, we propose to combine a novel, transfer learning technique for suppressing the bias with recent developments in deep learning, which allow building predictive models with multi-modal outputs. First, we train an initial neural network model on simulated data to learn important correlations between different output modalities and between simulation inputs and outputs. Then, the model is partially retrained, or transfer learned, to fit the experiments; a method that has never been implemented in this type of architecture. Using fewer than 10 inertial confinement fusion experiments for training, transfer learning systematically improves the simulation predictions while a simple output calibration, which we design as a baseline, makes the predictions worse. We also offer extensive cross-validation with real and carefully designed synthetic data. The method described in this paper can be applied to a wide range of problems that require transferring knowledge from simulations to the domain of experiments.
Abstract Recent advances in explainable artificial intelligence (XAI) methods show promise for understanding predictions made by machine learning (ML) models. XAI explains how the input features are relevant or important for the model predictions. We train linear regression (LR) and convolutional neural network (CNN) models to make 1-day predictions of sea ice velocity in the Arctic from inputs of present-day wind velocity and previous-day ice velocity and concentration. We apply XAI methods to the CNN and compare explanations to variance explained by LR. We confirm the feasibility of using a novel XAI method [i.e., global layerwise relevance propagation (LRP)] to understand ML model predictions of sea ice motion by comparing it to established techniques. We investigate a suite of linear, perturbation-based, and propagation-based XAI methods in both local and global forms. Outputs from different explainability methods are generally consistent in showing that wind speed is the input feature with the highest contribution to ML predictions of ice motion, and we discuss inconsistencies in the spatial variability of the explanations. Additionally, we show that the CNN relies on both linear and nonlinear relationships between the inputs and uses nonlocal information to make predictions. LRP shows that wind speed over land is highly relevant for predicting ice motion offshore. This provides a framework to show how knowledge of environmental variables (i.e., wind) on land could be useful for predicting other properties (i.e., sea ice velocity) elsewhere. Significance Statement Explainable artificial intelligence (XAI) is useful for understanding predictions made by machine learning models. Our research establishes trustability in a novel implementation of an explainable AI method known as layerwise relevance propagation for Earth science applications. To do this, we provide a comparative evaluation of a suite of explainable AI methods applied to machine learning models that make 1-day predictions of Arctic sea ice velocity. We use explainable AI outputs to understand how the input features are used by the machine learning to predict ice motion. Additionally, we show that a convolutional neural network uses nonlinear and nonlocal information in making its predictions. We take advantage of the nonlocality to investigate the extent to which knowledge of wind on land is useful for predicting sea ice velocity elsewhere.
This study quantifies the state of the art in the rapidly growing field of seasonal Arctic sea ice prediction. A novel multimodel dataset of retrospective seasonal predictions of September Arctic sea ice is created and analyzed, consisting of community contributions from 17 statistical models and 17 dynamical models. Prediction skill is compared over the period 2001–20 for predictions of pan-Arctic sea ice extent (SIE), regional SIE, and local sea ice concentration (SIC) initialized on 1 June, 1 July, 1 August, and 1 September. This diverse set of statistical and dynamical models can individually predict linearly detrended pan-Arctic SIE anomalies with skill, and a multimodel median prediction has correlation coefficients of 0.79, 0.86, 0.92, and 0.99 at these respective initialization times. Regional SIE predictions have similar skill to pan-Arctic predictions in the Alaskan and Siberian regions, whereas regional skill is lower in the Canadian, Atlantic, and central Arctic sectors. The skill of dynamical and statistical models is generally comparable for pan-Arctic SIE, whereas dynamical models outperform their statistical counterparts for regional and local predictions. The prediction systems are found to provide the most value added relative to basic reference forecasts in the extreme SIE years of 1996, 2007, and 2012. SIE prediction errors do not show clear trends over time, suggesting that there has been minimal change in inherent sea ice predictability over the satellite era. Overall, this study demonstrates that there are bright prospects for skillful operational predictions of September sea ice at least 3 months in advance.