Deterministic Prediction Research Articles

Adaptive soft sensing of river flow prediction for wastewater treatment operation and risk management

Many wastewater utilities have discharge permits directly tied with the receiving river flow, so it is critical to have accurate prediction of the hydraulic throughput to ensure safe operation and environment protection. Current empirical knowledge-based operation faces many challenges, so in this study we developed and assessed daily-adaptive, probabilistic soft sensor prediction models to forecast the next month's average receiving river flowrate and guide the utility operations. By comparing 11 machine-learning methods, extra trees regression exhibits desired deterministic prediction accuracy at day 0 (overall accuracy index: 3.9 × 10−3 1/cms2) (cms: cubic meter per second), which also increases steadily over the course of the month (e.g., MAPE and RMSE decrease from 41.46% and 23.31 cms to 3.31% and 2.81 cms, respectively). The overall classification accuracy of three river flow classes reaches 0.79 at the beginning and increases to about 0.97 over the course of the predicted month. To manage the uncertainty caused by potential false negative classification as overestimations, a probabilistic assessment on the predictions based on 95% lower PI is developed and successfully reduces the false negative classification from 17% to nearly zero with a slight sacrifice of overall classification accuracy.

Cluster-Based Analogue Ensembles for Hindcasting with Multistations

The Analogue Ensemble (AnEn) method enables the reconstruction of meteorological observations or deterministic predictions for a certain variable and station by using data from the same station or from other nearby stations. However, depending on the dimension and granularity of the historical datasets used for the reconstruction, this method may be computationally very demanding even if parallelization is used. In this work, the classical AnEn method is modified so that analogues are determined using K-means clustering. The proposed combined approach allows the use of several predictors in a dependent or independent way. As a result of the flexibility and adaptability of this new approach, it is necessary to define several parameters and algorithmic options. The effects of the critical parameters and main options were tested on a large dataset from real-world meteorological stations. The results show that adequate monitoring and tuning of the new method allows for a considerable improvement of the computational performance of the reconstruction task while keeping the accuracy of the results. Compared to the classical AnEn method, the proposed variant is at least 15-times faster when processing is serial. Both approaches benefit from parallel processing, with the K-means variant also being always faster than the classic method under that execution regime (albeit its performance advantage diminishes as more CPU threads are used).

Deterministic and uncertainty crude oil price forecasting based on outlier detection and modified multi-objective optimization algorithm

Crude oil price forecasting provides reference for stabilizing the energy economic, significantly improving the investment and operation decision-making. However, owing to the complexity and significant fluctuations of crude oil market, it is difficult to achieve accurate and reliable crude oil price forecasting. In this study, we develop a novel hybrid crude oil price forecasting model. Firstly, the Hampel identifier is employed to remove outliers in crude oil price series. Subsequently, complete ensemble empirical mode decomposition is used to reduce noise. Then a novel modified multi-objective water cycle algorithm is proposed to optimize the parameters of echo state network. Finally, deterministic and uncertainty prediction is conducted to verify the model performance. The results reveal that the proposed hybrid model outperforms various contract models in deterministic and interval predictions, as well as in daily and weekly forecasting. Therefore, the proposed hybrid model is a reliable tool for crude oil price forecasting and serve as a reference for decision making in energy economic market.

Scenario prediction for power loads using a pixel convolutional neural network and an optimization strategy

Accurate and reliable prediction of power load is critical to ensure the economy and stability of power systems. However, deterministic point prediction can scarcely be accurate due to the fluctuating and stochastic behavior of power load series, resulting in high risks for the system operation. Scenario prediction is a widely used method to model stochastic behavior by generating a group of possible power load scenarios rather than deterministic point predictions, so that system operators can account for the uncertainty of power loads. In this paper, a new deep generative network-based method is proposed for scenario prediction of power loads, in which structure and parameters are redesigned on the original pixel convolutional neural network (PixelCNN). An optimization model is presented to search for a range of power load scenarios with similar shapes, temporal dependency, and probability distribution as the real ones. Numerical simulations on a real-world power load dataset show that the PixelCNN outperforms other generative networks for the scenario prediction of power loads.

Uncertainty Quantification of WRF Model for Rainfall Prediction over the Sichuan Basin, China

The mesoscale Weather Research and Forecasting (WRF) model has been widely employed to forecast day-ahead rainfalls. However, the deterministic predictions from the WRF model incorporate relatively large errors due to numerical discretization, inaccuracies in initial/boundary conditions and parameterizations, etc. Among them, the uncertainties in parameterization schemes have a huge impact on the forecasting skill of rainfalls, especially over the Sichuan Basin which is located east of the Tibetan Plateau in southwestern China. To figure out the impact of various parameterization schemes and their interactions on rainfall predictions over the Sichuan Basin, the Global Forecast System data are chosen as the initial/boundary conditions for the WRF model and 48 ensemble tests have been conducted based on different combinations of four microphysical (MP) schemes, four planetary boundary layer (PBL) schemes, and three cumulus (CU) schemes, for four rainfall cases in summer. Compared to the observations obtained from the Chinese ground-based and encrypted stations, it is found that the Goddard MP scheme together with the asymmetric convective model version 2 PBL scheme outperforms other combinations. Next, as the first step to explore further improvement of the WRF physical schemes, the polynomial chaos expansion (PCE) approach is then adopted to quantify the impacts of several empirical parameters with uncertainties in the WRF Single Moment 6-class (WSM6) MP scheme, the Yonsei University (YSU) PBL scheme and the Kain-Fritsch CU scheme on WRF rainfall predictions. The PCE statistics show that the uncertainty of the scaling factor applied to ice fall velocity in the WSM6 scheme and the profile shape exponent in the YSU scheme affects more dominantly the rainfall predictions in comparison with other parameters, which sheds a light on the importance of these schemes for the rainfall predictions over the Sichuan Basin and suggests the next step to further improve the physical schemes.

Ensemble wind speed prediction system based on envelope decomposition method and fuzzy inference evaluation of predictability

As an indispensable part of the current global power system, wind energy has always been the focus of research over the world. In the production process of wind power, wind speed is a crucial factor and the requirements for wind speed prediction accuracy are increasing in practical applications. Therefore, as the main contribution of this paper, a novel decomposition-recognition-integration-prediction system (DRIPS) is proposed based on a newly developed predictive difficulty index (PDI) that synthesizes complexity, chaos, and long-term dependence of time series data. PDI comprehensively quantifies the basic characteristics and prediction difficulty of the series, filling the gap in the existing intuitionistic evaluation method. To verify the predictive ability and effectiveness of DRIPS, we select two American on-shore wind sites (Nolan and Kern) as the site of the experiments. At each site, the 10-minute-interval wind speed for three months in 2018 is collected as experimental samples. The simulation results show that DRIPS can provide excellent performance in the accuracy of wind speed prediction. In terms of deterministic prediction and uncertainty prediction, DRIPS performs less than 2% on the mean absolute percentage error for point prediction and less than 0.5 on the predictive interval score for interval prediction. Such performance is significantly better than that of the common prediction models such as BP, SVM, etc. Moreover, By comparing the experimental results of different integration strategies, the integration strategy based on PDI can improve the prediction accuracy significantly.

The predictability study of the two flavors of ENSO in the CESM model from 1881 to 2017

In this study, we evaluated the predictability of the two flavors of the El Niño Southern Oscillation (ENSO) based on a long-term retrospective prediction from 1881 to 2017 with the Community Earth System Model. Specifically, the Central-Pacific (CP) ENSO has a more obvious Spring Predictability Barrier and lower deterministic prediction skill than the Eastern-Pacific (EP) ENSO. The potential predictability declines with lead time for both the two flavors of ENSO, and the EP ENSO has a higher upper limit of the prediction skill as compared with the CP ENSO. The predictability of the two flavors of ENSO shows distinct interdecadal variation for both actual skill and potential predictability; however, their trends in the predictability are not synchronized. The signal component controls the seasonal and interdecadal variations of predictability for the two flavors of ENSO, and has larger contribution to the CP ENSO than the EP ENSO. There is significant scope for improvement in predicting the two flavors of ENSO, especially for the CP ENSO.

Bayesian Spatio-Temporal grAph tRansformer network (B-STAR) for multi-aircraft trajectory prediction

Multi-Agent Trajectory Prediction is a critical and challenging component across different safety–critical engineering applications, e.g., autonomous driving and flight systems. Trajectory prediction tools are required for the next-generation air transportation system (NextGen). In practice, the prediction of aircraft trajectories needs to consider the impact of various sources, such as environmental conditions, pilot/controller behaviors, and potential conflicts with nearby aircraft. Huge uncertainties associated with these factors lead to the untrustworthiness of a deterministic trajectory prediction model. Moreover, the safety assurance in the near-terminal area is of specific interest due to the increased airspace complexity, where the instrument/visual flight rules are applied. In this work, we propose the Bayesian Spatio-Temporal grAph tRansformer (B-STAR) architecture to model the spatial and temporal relationship of multiple agents under uncertainties. It is shown that the proposed B-STAR achieves state-of-the-art performance on the ETH/UCY pedestrian dataset with UQ competence. Then, multi-aircraft near-terminal interactive trajectory prediction model is trained and validated with real-world flight recording data. The sensitivity study on the prediction/observation horizon and the graph neighboring distance threshold are performed. The code is available at https://github.com/ymlasu/para-atm-collection/tree/master/air-traffic-prediction/MultiAircraftTP.

QRF4P‐NRT: Probabilistic Post‐Processing of Near‐Real‐Time Satellite Precipitation Estimates Using Quantile Regression Forests

AbstractAccurate and reliable near‐real‐time satellite precipitation estimation is of great importance for operational large‐scale flood forecasting and drought monitoring. The state‐of‐the‐art precipitation post‐processing model is based on a deterministic approach to construct relationships between satellites estimates and ground observations. We propose a probabilistic postprocessor, the Probabilistic Post‐Processing of Near‐Real‐Time Satellite Precipitation Estimates using Quantile Regression Forests (QRF4P‐NRT), based on quantile modeling, yielding both deterministic and probabilistic predictions. The experimental design incorporates different solutions of near‐real‐time predictors to further improve the model performance. Using the Integrated Multi‐satellitE Retrievals Early Run for Global Precipitation Measurement Mission (IMERG‐E) product as an example, we illustrate that the proposed method significantly improves the overall quality of the raw IMERG‐E and is also superior to the bias‐corrected product (IMERG Final Run, IMERG‐F) at daily scale in a complex mountain basin. Evaluations of the corrected IMERG‐E, raw IMERG‐E, and IMERG‐F using ground observation show that the corrected IMERG‐E improves correlation coefficients (0.7), mean error (−0.14 mm/day) and root mean square error (3.3 mm/day) relative to the raw IMERG‐E (0.31, −0.72 and 5.5 mm/day) and IMERG‐F (0.34, −0.09 and 6.0 mm/day). The error decomposition further confirms that the QRF4P‐NRT improves on the various deficiencies of the raw IMERG‐E product. The ensemble assessment also demonstrates that the quantile outputs provide reliable prediction spread and sharp prediction intervals. The promising results indicate the great potential of the proposed method for probabilistic post‐processing for near‐real‐time satellite precipitation estimates, and for further applications such as hydrological ensemble forecasting.

Decreasing False-alarm Rates in CNN-based Solar Flare Prediction Using SDO/HMI Data

A hybrid two-stage machine-learning architecture that addresses the problem of excessive false positives (false alarms) in solar flare prediction systems is investigated. The first stage is a convolutional neural network (CNN) model based on the VGG-16 architecture that extracts features from a temporal stack of consecutive Solar Dynamics Observatory Helioseismic and Magnetic Imager magnetogram images to produce a flaring probability. The probability of flaring is added to a feature vector derived from the magnetograms to train an extremely randomized trees (ERT) model in the second stage to produce a binary deterministic prediction (flare/no-flare) in a 12 hr forecast window. To tune the hyperparameters of the architecture, a new evaluation metric is introduced: the “scaled True Skill Statistic.” It specifically addresses the large discrepancy between the true positive rate and the false positive rate in the highly unbalanced solar flare event training data sets. Through hyperparameter tuning to maximize this new metric, our two-stage architecture drastically reduces false positives by ≈48% without significantly affecting the true positives (reduction by ≈12%), when compared with predictions from the first-stage CNN alone. This, in turn, improves various traditional binary classification metrics sensitive to false positives, such as the precision, F1, and the Heidke Skill Score. The end result is a more robust 12 hr flare prediction system that could be combined with current operational flare-forecasting methods. Additionally, using the ERT-based feature-ranking mechanism, we show that the CNN output probability is highly ranked in terms of flare prediction relevance.

Development of a disaggregated multi-level factorial hydrologic data assimilation model

A disaggregated multi-level factorial hydrologic data assimilation model (FHDA) is proposed for exploring not only the direct effects from individual uncertainties but also, more importantly, the composite ones from multi-layer and multi-parameter interactions among multiple uncertainties in hydrologic data assimilation systems. Based on a disaggregated multi-level factorial analysis method, the proposed FHDA examined the contributions of multiple uncertainty sources, including data assimilation scheme, sample size, forcing data error, and observed data error. Three parameter assimilation schemes of data assimilation i.e., Ensemble Kalman Filter (EnKF), Standard Kernel Smoother (EnKFSKS), and Kernel Smoother with location shrinkage (EnKFKSLS) are tested. The results indicate that: i) the streamflow observations can be well tracked by 95% prediction intervals. ii) reducing streamflow observation error is the most efficient way to improve both deterministic and probabilistic predictions. iii) data assimilation scheme plays a vital role in hydrological predictions, especially for probabilistic ones (i.e., contributes 22% for the ensemble prediction uncertainty). Therefore, to improve the prediction accuracy, it is necessary to optimize the parameter assimilation schemes in hydrological data assimilation model.

A short-term wind speed prediction method based on interval type 2 fuzzy model considering the selection of important input variables

Wind speed prediction is a challenging task in wind energy resource prediction. Wind speed time series, which are a highly variable data source, require highly nonlinear time characteristics for prediction. Most prediction models use hybrid models to improve prediction accuracy, which increases the complexity of the model. This paper proposes an interval type-2 (IT2) wind speed prediction model based on the selection of important input variables. First, the two-stage fuzzy curve are used to determine the model input so as to reduce redundant input and further reduce the complexity of the model. Second, a Gaussian membership function with a variable centre is used to further improve the prediction accuracy of the model. Finally, the recursive least squares method is used to complete the identification of the conclusion parameters. The proposed model has high generalisability and IT2 robustness, and can thus realise deterministic wind speed prediction. The wind speed prediction results for a wind farm in Colorado, USA, in two time periods show that compared with the type-1 fuzzy model and other type-2 fuzzy models, the proposed model provides more accurate short-term wind speed forecasting.

An integrated power load point-interval forecasting system based on information entropy and multi-objective optimization

During an era of rapid growth in electricity demand throughout society, accurate forecasting of electricity loads has become increasingly important to guarantee a stable power supply. Nevertheless, historical models do not address the structure of the data itself, and a single model cannot accurately determine the nonlinear characteristics of the data. This would not allow for accurate and stable predictions. With the aim of filling this gap, this paper proposes an innovative intelligent power load point-interval forecasting system. The system discretizes the time series, then performs efficient dimensionality reduction by fuzzification, and multi-level optimization of five benchmark deep learning models by the proposed multi-objective optimization algorithm, and finally analyzes the uncertainty of the prediction results. Experiments comparing the developed prediction system with other models were conducted on three datasets, and the prediction results were discussed for validation from multiple perspectives. The simulation results show that the proposed model has superior prediction accuracy, robustness and uncertainty analysis capability, and can provide accurate deterministic prediction information and fluctuation interval analysis to ensure the long-term safety and stability and operation of the grid.

A Posteriori Random Forests for Stochastic Downscaling of Precipitation by Predicting Probability Distributions

AbstractThis work presents a comprehensive assessment of the suitability of random forests, a well‐known machine learning technique, for the statistical downscaling of precipitation. Building on the experimental and validation framework proposed in the Experiment 1 of the COST action VALUE—the largest, most exhaustive intercomparison study of statistical downscaling methods to date—we introduce and thoroughly analyze a posteriori random forests (AP‐RFs), which use all the information contained in the leaves to reliably predict the shape and scale parameters of the gamma probability distribution of precipitation on wet days. Therefore, as opposed to traditional random forests, which typically provide deterministic predictions, our AP‐RFs allow realistic stochastic precipitation samples to be generated for wet days. Indeed, as compared to one particular implementation of a generalized linear model that exhibited an overall good performance in VALUE, our AP‐RFs yield better distributional similarity with observations without loss of predictive power. Noteworthy, the new methodology proposed in this paper has substantial potential for hydrologists and other impact communities which are in need of local‐scale, reliable stochastic climate information.

Uncertainty in Collapse Strength Prediction of Sandwich Pipelines

Abstract This paper assesses the uncertainty in the collapse strength of sandwich pipelines (SPs) under external pressure predicted by various strength models in three categories based on interlayer adhesion conditions. First, the validity of the strength models is verified by comparing their predictions with sandwich pipeline collapse test data and the corresponding model uncertainty factors are derived. Then, a parametric analysis of deterministic collapse strength predictions by models is conducted, illustrating insights of models’ behavior for a wide range of design configurations. Furthermore, the uncertainty among different model predictions is perceived at different configurations of outer and inner pipes and core thicknesses. A case study of a realistic sandwich pipeline is developed, and probabilistic models are defined to basic design parameters. Uncertainty propagation of models’ predictions is assessed by the Monte Carlo simulation method. Finally, the strength model predictions of sandwich pipelines are compared with that of an equivalent single-walled pipe.

Probability prediction of short-term user-level load based on random forest and kernel density estimation

With the development of smart grids and the popularization of smart meters, grid companies have obtained a large amount of fine-grained user electricity consumption data, making it possible to forecast individual users’ electricity load. Traditional deterministic forecasting cannot measure the uncertainty of users’ future electricity consumption behavior. This paper aims to establish a short-term load probability forecasting model for individual users, which can output the probability prediction interval and the probability density curve. Firstly, historical data is applied to train a deterministic prediction model based on random forest. Then, when predicting the user load at a future time, the output of each tree in the forest is formed into a prediction set, and the kernel density estimation method is used to obtain the probability prediction result of the user load. The proposed RF-KDE method obtains promising results on a public dataset. It also shows the advantages of easy parameters adjustment and fast training speed. Besides, tests conducted on the public dataset confirm that the proposed method can be applied to users with different electricity consumption behaviors.

Mixed-Effects Modeling Framework for Amsterdam and Copenhagen for Outdoor NO2 Concentrations Using Measurements Sampled with Google Street View Cars.

High-resolution air quality (AQ) maps based on street-by-street measurements have become possible through large-scale mobile measurement campaigns. Such campaigns have produced data-only maps and have been used to produce empirical models [i.e., land use regression (LUR) models]. Assuming that all road segments are measured, we developed a mixed model framework that predicts concentrations by an LUR model, while allowing road segments to deviate from the LUR prediction based on between-segment variation as a random effect. We used Google Street View cars, equipped with high-quality AQ instruments, and measured the concentration of NO2 on every street in Amsterdam (n = 46.664) and Copenhagen (n = 28.499) on average seven times over the course of 9 and 16 months, respectively. We compared the data-only mapping, LUR, and mixed model estimates with measurements from passive samplers (n = 82) and predictions from dispersion models in the same time window as mobile monitoring. In Amsterdam, mixed model estimates correlated rs (Spearman correlation) = 0.85 with external measurements, whereas the data-only approach and LUR model estimates correlated rs = 0.74 and 0.75, respectively. Mixed model estimates also correlated higher rs = 0.65 with the deterministic model predictions compared to the data-only (rs = 0.50) and LUR model (rs = 0.61). In Copenhagen, mixed model estimates correlated rs = 0.51 with external model predictions compared to rs = 0.45 and rs = 0.50 for data-only and LUR model, respectively. Correlation increased for 97 locations (rs = 0.65) with more detailed traffic information. This means that the mixed model approach is able to combine the strength of data-only mapping (to show hyperlocal variation) and LUR models by shrinking uncertain concentrations toward the model output.

An approach to investigate the supplementary inconsistency between time series data for predicting road pavement performance models

ABSTRACT Road agencies commonly face high level complexity of comprehensive pavement maintenance scheduling complications. This could be due to uncertainty of some parameters that substance in the predicting models. Disregarding this uncertainty of such parameters in the road maintenance scheduling problems may lead to imprecise solutions and unreliable road conditions. Frequently time series data that have a hierarchical structure are used for predicting road surface or pavement structure distresses models. However, further realistic and accurate predicted models are those consider supplementary inconsistency between time series data with utilising a predictive approach that imposes knowledge of the input parameters. The purpose of this study is to test this theory by comparing the predictions of models developed using two approaches and their accuracies in utilising the same data. These approaches are the traditional regression model (TRM) with only one level of heterogeneity approach and the hierarchical regression model (HRM) with many levels of heterogeneities approach. A Metro North Western (MNW) urban road network in Victoria (Australia) was considered as the case study for this study investigation where covering the condition of approximately 305 Km of asphalt surfaced pavements. By considering the structure of time series data, the study revealed that more accurate models could predict the pavement behaviour, the deterioration rate, and significantly influenced parameters. In addition, the study highlighted that it is essential to deliberate and process the hierarchical structure of datasets to consider the observed and unobserved heterogeneities in time series data for developing any pavement performance model using either deterministic or probabilistic prediction approaches. Multiple simulation scenarios have been performed for both approaches and the HRM approach shows more reasonable outputs. For predicted deterministic SIR models, road intervention (SIR ≥ 20) are triggered at age 12.5 when using TRM approach and at age 10.5 when using HRM approach. For predicted probabilistic SIR models, the surface condition of a section with SIR weighted average condition values of 2 is triggered at age 17.5 when using TRM approach and at age 10.5 when using HRM approach. This also refers to earlier maintenance scheduling is required when using the later approach because it deals confidentially with each road sections time series data. In that regard, ignoring the uncertainty in time series data of such factors in the road surfacing maintenance scheduling problems may lead to suboptimal solutions and unstable pavement conditions.

HydroFlow: Towards probabilistic electricity demand prediction using variational autoregressive models and normalizing flows

We present HydroFlow, a novel deep generative model for predicting the electricity generation demand of large-scale hydropower stations. HydroFlow uses a latent stochastic recurrent neural network to capture the dependencies in the multivariate time series. It not only utilizes the hidden state of the neural network, but also considers the uncertainty of variables related to natural and social factors. We also introduce an end-to-end approach based on generative flows to approximate the posterior distribution of time series with exact likelihoods. Our model is powerful as adding stochasticity to different factors (e.g., reservoir capacity and water-flow measurements) and thus overcomes the expressiveness limitations of deterministic prediction methods. It also enables trainable latent transformations that can improve the model interpretability. We evaluate HydroFlow on the data collected from the hydropower stations of a large-scale hydropower development company. Experimental results show that our model significantly outperforms the state-of-the-art baseline methods while providing explainable results.

Analytical solution of impulse response function of finite-depth water waves

Short-term deterministic water wave prediction is of great importance to the operations of ships and offshore structures. The use of the convolution integral of input wave profiles and an impulse response function is one of the candidates to predict water waves on the basis of the linear time-invariant system. However, this approach is not often used today because of the non-causality and the absence of the analytical solution of the impulse response function of finite-depth water; only the deep water impulse response function is available. In the paper, we present the analytical solution of the impulse response function using the dispersion relation of finite-depth water waves. In addition, the prediction abilities in space and time are investigated. It is shown that non-causal influence decreases exponentially and the predictable time increases linearly as the distance between measuring and prediction points. We conducted a tank experiment to predict irregular waves based on the JONSWAP spectrum. The prediction accuracy by the impulse response function of finite-depth water is higher than that by the deep water impulse response function. Besides, we find that the errors of the use of the analytical solution and that of the numerical solution obtained by the inverse discrete Fourier transform are comparable.