Probabilistic Forecasts Research Articles

Statistical Postprocessing of Week-1 and Week-2 Precipitation Forecasts over Taiwan

Abstract The predictability of precipitation is hindered by finer-scale processes not captured explicitly in global numerical models, such as convective interactions, cloud microphysics, and boundary layer dynamics. However, there is growing demand across various sectors for medium- (3–10-day) and extended-range (10–30-day) quantitative precipitation forecasts (QPFs) and probabilistic QPFs (PQPFs). This study uses a novel statistical postprocessing technique, APPM, that combines analog postprocessing (AP) with probability matching (PM) to produce week-1 and week-2 accumulated precipitation forecasts over Taiwan. AP searches for historical predictions that closely resemble the current forecast and create an AP ensemble using the observed high-resolution precipitation patterns corresponding to these forecast analogs. Frequency counting and PM are then separately applied to the AP ensemble to produce calibrated and downscaled PQPFs and bias-reduced QPFs, respectively. Evaluation over a 22-yr (1999–2020) period shows that raw ensemble forecasts from the GEFS of NOAA/NWS/Environmental Modeling Center, collected for the subseasonal experiment, are underdispersive with a wet bias. In contrast, the AP ensemble spread well represents forecast uncertainty, leading to substantially more reliable and skillful probabilistic forecasts. Furthermore, the AP-based PQPF demonstrates superior discrimination ability and yields notably greater economic benefits for a wider range of users, with the maximum economic value increasing by 30%–50% for the week-2 forecast. Compared to the raw ensemble mean forecast, the calibrated QPF exhibits lower mean absolute error and explains 3–8 times more variance in observations. Overall, the APPM technique significantly improves week-1 and week-2 QPFs and PQPFs over Taiwan. Significance Statement There are two significant challenges in improving precipitation forecasts beyond a few days in Taiwan. First, large-scale numerical models often struggle with accurately predicting precipitation locations, magnitudes, and providing sufficient detail. Second, probabilistic precipitation forecasts have been unreliable, failing to convey accurate uncertainty information to users. In response to these challenges, this study has developed a relatively simple yet effective technique that corrects the spatiotemporal distribution of predicted precipitation and downscales the forecasts from a 1° to 1-km spatial resolution. Our results demonstrate that this technique significantly alleviates these two issues, resulting in more accurate precipitation forecasts and more reliable probabilistic precipitation forecasts within a 2-week timeframe.

Evaluating Machine Learning–Based Probabilistic Convective Hazard Forecasts Using The HRRR: Quantifying Hazard Predictability and Sensitivity to Training Choices

Abstract The High-Resolution Rapid Refresh (HRRR) model provides hourly updating forecasts of convective-scale phenomena, which can be used to infer the potential for convective hazards (e.g., tornadoes, hail, and wind gusts), across the United States. We used deterministic 2019–20 HRRR, version 4 (HRRRv4), forecasts to train neural networks (NNs) to generate 4-hourly probabilistic convective hazard forecasts [neural network probability forecasts (NNPFs)] for HRRRv4 initializations in 2021, using storm reports as ground truth. The NNPFs were compared to the skill of a smoothed updraft helicity (UH) baseline to quantify the benefit of the NNs. NNPF skill varied by initialization time and time of day but was all superior to the UH forecast. NNPFs valid at hours between 1800 and 0000 UTC were most skillful in aggregate, significantly exceeding the baseline forecast skill. Overnight NNPFs (i.e., valid 0600–1200 UTC) were least skillful, indicating a diurnal cycle in hazard predictability that was present across all HRRRv4 initializations. We explored the sensitivity of HRRRv4 NNPF skill to NN training choices. Including an additional year of 2021 HRRRv4 forecasts for training slightly improved skill for 2022 HRRRv4 NNPFs, while reducing the training dataset size by 40% using only forecasts with storm reports was not detrimental to forecast skill. Finally, NNs trained with 2018–20 HRRRv3 forecasts led to a reduction in NNPF skill when applied to 2021 HRRRv4 forecasts. In addition to documenting practical predictability challenges with convective hazard prediction, these findings reinforce the need for a consistent model configuration for optimal results when training NNs and provide best practices when constructing a training dataset with operational convection-allowing model forecasts. Significance Statement Convective hazards, such as hail and tornadoes, are often challenging to predict. To improve hazard predictions, we used machine learning (ML) to generate forecasts of convective hazards across the United States leveraging forecasts of prior events. The ML hazard forecasts were consistently better than a non-ML approach and varied in skill based on the time of day, with nighttime forecasts being particularly challenging. ML forecasts of wind gusts and hail were more skillful than tornadoes. Different strategies of constructing the dataset of prior events led to differences in forecast performance; thus, this work provides recommendations for how to assemble these datasets and train ML models to generate improved forecasts of severe weather events.

Quantifying the value of probabilistic forecasts when trading renewable hybrid power parks in day-ahead markets: A Nordic case study

Renewable hybrid power parks (HPPs) that combine wind power, solar photovoltaic (PV) power and storage have emerged as promising electricity generation resources. However, HPPs face operational challenges due to the uncertainty in power production and electricity prices, which is why probabilistic forecasts that capture the uncertainty associated with forecast errors have gained attention. While the research community has proposed several methods to improve the accuracy of probabilistic forecasts, the question on how these forecasts can improve decision-making over deterministic forecasts is rarely quantified. This study assesses the value of probabilistic forecasts and analyze the improvement compared to deterministic forecasts in day-ahead markets. The value is quantified using almost two years of data from an operational HPP in Sweden. Results show that: (i) high grid connection capacities leverage the value of probabilistic models, (ii) a deterministic model is preferable for parks with a ratio of battery energy capacity to installed nominal power of the renewable power park equal to 0.6 MWh/MW, (iii) a probabilistic model allows utilizing the energy storage more effectively by reducing the energy throughput of the battery with 61%–87%, and (iv) a probabilistic model increases the unit profit when the forecast errors of the regulating price are higher than the spot price, (v) a simple probabilistic benchmark model, which is worse in terms of forecast accuracy, increases the unit profit compared to the analyzed deterministic models, and (vi) the more advanced probabilistic model analyzed in this study does not provide a significant improvement over a simple probabilistic benchmark model.

New Features in the pyCSEP Toolkit for Earthquake Forecast Development and Evaluation

Abstract The Collaboratory for the Study of Earthquake Predictability (CSEP) is a global community dedicated to advancing earthquake predictability research by rigorously testing probabilistic earthquake forecast models and prediction algorithms. At the heart of this mission is the recent introduction of pyCSEP, an open-source software tool designed to evaluate earthquake forecasts. pyCSEP integrates modules to access earthquake catalogs, visualize forecast models, and perform statistical tests. Contributions from the CSEP community have reinforced the role of pyCSEP in offering a comprehensive suite of tools to test earthquake forecast models. This article builds on Savran, Bayona, et al. (2022), in which pyCSEP was originally introduced, by describing new tests and recent updates that have significantly enhanced the functionality and user experience of pyCSEP. It showcases the integration of new features, including access to authoritative earthquake catalogs from Italy (Bolletino Seismico Italiano), New Zealand (GeoNet), and the world (Global Centroid Moment Tensor), the creation of multiresolution spatial forecast grids, the adoption of non-Poissonian testing methods, applying a global seismicity model to specific regions for benchmarking regional models and evaluating alarm-based models. We highlight the application of these recent advances in regional studies, specifically through the New Zealand case study, which showcases the ability of pyCSEP to evaluate detailed, region-specific seismic forecasts using statistical functions. The enhancements in pyCSEP also facilitate the standardization of how the CSEP forecast experiments are conducted, improving the reliability, and comparability of the earthquake forecasting models. As such, pyCSEP exemplifies collaborative research and innovation in earthquake predictability, supporting transparent scientific practices, and community-driven development approaches.

Can Expert Prediction Markets Forecast Climate-Related Risks?

Abstract Prediction markets use a betting-like mechanism to aggregate information from disparate sources to produce probability forecasts that represent the collective view of participants. They have similarities with financial markets but are designed specifically to discover information rather than transfer assets or risks. Contracts that pay out a fixed amount if a well-defined event occurs are traded in these markets and, under certain conditions, the price at which these event-contracts trade can be interpreted as a probability estimate that the event will occur. This article examines the performance of 24 prediction markets for climate-related variables that have been run over the past 5 years. The predicted variables included the monthly temperature and rainfall for the United Kingdom, an index of El Niño–Southern Oscillation, Atlantic hurricane activity, and U.K. wheat yield. The markets had horizons of 2–12 months. Invitations to participate in the markets were extended to individuals and teams with relevant expertise. Participants did not have to pay to take part but did receive cash rewards based on their performance. Trades were made through an automated market maker overcoming the problems of low activity that have affected previous prediction markets for specialized topics. The predictions of the markets were consistent with good reliability, given the resolving power afforded by the sample size. Whether this level of reliability would persist for longer, multiyear, horizons relevant to climate change cannot be known without running markets on such time scales, something that we strongly advocate.

Time-Dependent Aftershock Probabilistic Forecasting and Ensemble Model Implementation: Insights from the 2021 Ms 7.4 Maduo Earthquake

Abstract On 22 May 2021, an Ms 7.4 earthquake with a focal depth of 17 km struck the Maduo region of Qinghai province, breaking a 3.8-year quiescence of strong earthquakes (magnitude >7.0) in mainland China. This event has increased stress on the Maqin–Maqu segment of the Kunlunshan fault, heightening the potential for a large earthquake in the region. In this study, we employed the epidemic-type aftershock sequence (ETAS) model and Reasenberg–Jones (R–J) model to fit the aftershock sequence following the mainshock, analyzing the temporal response and its ability to influence future generations. Concurrently, ensemble models were explored to leverage the strengths of both the ETAS and R–J models. Short-term forecasts of the probability and occurrence rate of aftershock events with varying magnitudes were conducted for the next three days. Then statistical methods, that is, receiver-operating characteristic diagrams and information gain, were used to evaluate the unconditional and relative performance. Our findings include that the ETAS model indicates a high decay rate with many aftershocks triggered by previous events, whereas the R–J model shows a normal decay rate with a higher proportion of strong aftershocks. The ETAS and R–J models perform better than random guesses for different aftershock magnitudes, but the ETAS model is somewhat affected by the problem of missing small earthquakes for a short period of time after the mainshock. The ensemble model that chooses the minimum strategy shows promise, especially in the early period, suggesting a reasonable and cautious decision approach should be chosen during the unstable stage. This study highlights the importance of short-term aftershock probability estimation for seismic research and decision-making. In the absence of more accurate models, current analytical approaches within the operational earthquake forecasting framework remain valuable. Continuous testing, feedback, and refinement of forecasting models, along with the development of ensemble models, are essential for enhancing seismic risk assessment and mitigation strategies.

Supervised Learning-Based Prediction of Lightning Probability in the Warm Season

The accurate prediction of lightning is crucial for forecasters to respond effectively to its related hazards. The rapid development and confined spatial extent of convective storms, in which lightning frequently occurs, pose considerable challenges for accurately predicting their locations using numerical weather prediction (NWP) models. Lightning occurrence is often prognosed using thermodynamic parameters, convective available potential energy (CAPE), the severe weather threat index (SWEAT), the lifted index (LI), etc. A high-resolution NWP model provides a prediction of these thermodynamic parameters at high spatiotemporal resolution with high accuracy for a few hours. However, a complicated algorithm is required to handle all the useful high-resolution variables from the NWP model. The recently emerging machine learning technique can solve this issue by properly handling these “big data” without any model distributional assumption. In this study, we developed a random forest algorithm for nowcasting and very short-range forecasting (useful for ~6 h), named LightningRF. LightningRF was trained by using lightning occurrence as a response variable and characteristic parameters from the NWP as predictors. It was also applied to analysis and forecast fields, showing a high probability of lightning within the observed lightning regions. This highlights the potential of helping forecasters improve their lightning forecasting skills using real-time probabilistic forecasts from a trained model.

Prediction of geomagnetic storms associated with interplanetary coronal mass ejections

Geomagnetic storms have a significant impact on the performance of technical systems both in space and on Earth. The sources of strong geomagnetic storms are most often interplanetary coronal mass ejections (ICMEs), generated by coronal mass ejections (CMEs) in the solar corona. The ICME forecast is based on regular optical observations of the Sun, which make it possible to detect CMEs at the formation stage. It is known that the intensity of geomagnetic storms correlates with the magnitude of the southern component of the magnetic field (Bz) of the ICME. However, it is not possible yet to predict the sign and magnitude ofBzfrom solar observations for the operational forecast of an arbitrary CME. Under these conditions, a preliminary forecast of the magnetic storm probability can be obtained under the assumption that the strength of the storm is related to the magnitude of the magnetic flux from the eruption region, observed as dimming. In this paper we examine the relationship between the integral magnetic flux from the dimming region and the probability that CMEs associated with them will cause geomagnetic storms, using a series of 37 eruptive events in 2010–2012. It is shown that there is a general trend toward an increase in the ICMEs geoefficiency with an increase in the magnitude of the magnetic flux from the dimming region. It has been demonstrated that the frequency of moderate and severe storms observation increases in cases of complex events associated with the interaction of CMEs with other solar wind streams in the heliosphere.

Insights to key operational questions in forecast-informed dam release operation: case of Hume Dam

ABSTRACT Dam operators decide on dam releases given downstream demands and other factors. Downstream tributaries complicate the problem due to significant uncertainties in their flows. Demands can also be highly uncertain. Incorporating probabilistic ensemble forecasts into the decision-making process can mitigate these uncertainties. We investigate how ensemble flow and demand forecasts can inform Hume Dam releases, in the Murray-Darling Basin, for delivery to Lake Mulwala, about four days downstream. We develop and simulate three forecast-based decision rules governing the releases. The first and second rules are based on the expectations and medians respectively of ensemble forecasts of water excess and shortage, forced by ensemble tributary inflow and demand forecasts. The third rule is based on the risks of an excess or shortage per the ensemble forecasts of water excess and shortage. We analyse the results to answer the following key operational questions: (1) Which decision rule is best for managing Hume Dam releases for water supply and under what conditions? (2) How far ahead of flow and demand forecasts should Hume Dam operators adopt? (3) How does limiting the day-to-day change in Hume Dam releases influence the value of forecasts? (4) What is the value of improved demand forecasts?

Process-based forecasts of lake water temperature and dissolved oxygen outperform null models, with variability over time and depth

Near-term iterative ecological forecasting has great potential for providing new insights into our ability to predict multiple ecological variables. However, true, out-of-sample probabilistic forecasts remain rare, and variability in forecast performance has largely been unexamined in process-based forecasts which predict multiple ecosystem variables. To explore how forecast performance varies for water temperature and dissolved oxygen, two freshwater variables important for lake ecosystem functioning, we produced probabilistic forecasts at multiple depths over two open-water seasons in Lake Sunapee, NH, USA. Our forecasting system, FLARE (Forecasting Lake And Reservoir Ecosystems), uses a 1-D coupled hydrodynamic-biogeochemical process model, which we assessed relative to both climatology and persistence null models to quantify how much information process-based FLARE forecasts provide over null models across varying environmental conditions. We found that FLARE water temperature forecasts were always more skillful than FLARE oxygen forecasts. Specifically, temperature forecasts outperformed both null models up to 11 days into the future, as compared to only two days for oxygen. Across different years, we observed variable forecast skill, with performance generally decreasing with depth for both variables. Overall, all temperature forecasts and surface oxygen, but not deep oxygen, forecasts were more skillful than at least one null model >80 % of the forecasted period, indicating that our process-based model was able to reproduce the dynamics of these two variables with greater reliability than the null models. However, process-based oxygen forecasts from deeper waters were less skillful than both null models during a majority of the forecasted period, which suggests that deep-water oxygen dynamics are dominated by autocorrelation and seasonal change, which are inherently captured by the null forecasts. Our results highlight that forecast performance varies among lake water quality metrics and that process-based forecasts can provide important information in conjunction with null models in varying environmental conditions. Altogether, these process-based forecasts can be used to develop quantitative tools which inform our understanding of future ecosystem change.

Improving sub-seasonal extreme precipitation forecasts over China through a hybrid statistical-dynamical framework

Skillful and reliable sub-seasonal extreme precipitation forecasts are crucial for disaster prevention and mitigation. In this study, we introduce a hybrid statistical-dynamical framework to predict monthly maximum one-day precipitation (Rx1D) and monthly maximum five-day precipitation (Rx5D) over China from May to October. In the hybrid statistical-dynamical framework, the ECMWF forecasts of precipitation and boreal summer intraseasonal oscillation (BSISO) indices are used as predictors to establish calibration model and bridging models, separately. The calibration model and bridging models are then merged to generate probabilistic forecasts of Rx1D and Rx5D. Our results suggest that the bridging models show better performance in predicting Rx1D and Rx5D than calibration model in May, June, and July when the BSISO indices are used as predictors. The forecast skill of calibration model is higher compared to bridging models in August, September, and October. The BMA merged forecasts take advantage of both calibration model and bridging models, and can provide skilful and reliable forecasts for both Rx1D and Rx5D prediction. To have a more comprehensive assessment, we also evaluate the prediction skill of the occurrence of extreme precipitation events with exceedance probabilities of 50%, 20%, and 5% for both Rx1D and Rx5D. The Brier skill score of merged forecasts indicates that the hybrid statistical-dynamical framework can also provide skilful forecasts for the occurrence of extreme precipitation events greater than one-in-5-year return value of Rx1D (5Rx1D) and one-in-5-year return value of Rx5D (5Rx5D) in comparison to long-term climatology. These findings demonstrate the great potential of combining dynamical models and statistical models in improving sub-seasonal extreme precipitation forecasts.

Intelligent crude oil price probability forecasting: Deep learning models and industry applications

The crude oil price has been subject to periodic fluctuations because of seasonal changes in industrial demand and supply, weather, natural disasters and global political unrest. An accurate forecast of crude oil prices is of utmost importance for decision makers and industry players in the energy sector. Despite this, the volatility of crude oil prices contributes to the uncertainty of the energy industry, which was particularly challenging following the recent global spread of the COVID-19 epidemic and the Russia–Ukraine conflict. This paper proposes a hybrid deep learning (DL) modelling framework to deal with the volatility of crude oil prices, applying ensemble empirical mode decomposition (EEMD), convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) integrated with quantile regression (QR); named EEMD-CNN-BiLSTM-QR. Two real-world datasets on crude oil prices from the West Texas Intermediate and Brent Crude Oil markets were employed to validate the EEMD-CNN-BiLSTM-QR hybrid modelling framework. Given that the probability density forecast can capture uncertainty, an in-depth analysis was carried out and prediction accuracy calculated. The findings of this study demonstrate that the proposed EEMD-CNN-BiLSTM-QR DL modelling framework, which uses the probability density forecast method, is superior to other tested models in terms of its ability to forecast crude oil prices. The novelty of this study stems mainly from its use of QR, which allows for the description of the conditional distribution of predicted variables and the extraction of more uncertain information for probability density forecasts.

Statistical analysis of the ground deformation of Vulcanian explosions at Sakurajima volcano, Japan

The forecast of pulsatory explosions during volcanic unrest periods is an essential issue for the assessment and mitigation of volcanic hazards. Although various precursors are detectable through geophysical and geochemical monitoring, difficulties remain in precisely constraining possible scenarios. A probabilistic approach is effective in assessing risk while considering various uncertainties. Sakurajima volcano characterized by frequent Vulcanian activity is one of the suitable fields for the probabilistic forecast of pulsatory explosions. Their inflation-deflation patterns of ground deformation related to Vulcanian explosions are useful for evaluating the imminence and size of the next event. The large database obtained from its vigorous activity can contribute to statistical analysis. In this study, aiming the probabilistic forecast of the timing and size of explosions, we investigated the duration of inflation and volume changes at the pressure source using strain records of over 5000 events of Sakurajima volcano. Then, a stochastic model was estimated to explain the distribution of these events. The log-logistic distribution was found to be an appropriate model for data distribution, indicating the presence of competing processes, such as pressurization and depressurization, in the conduit. The model parameters of the log-logistic distribution temporally fluctuated reflecting the volcanic activity, especially increasing the magma supply from a deep region. We also suggested a methodology to constrain the probabilities of the likely timing and size of an imminent explosion using real-time strain monitoring and an estimated model distribution. Although some improvements would be needed for practical forecasting, our approach could be useful in predicting possible ash hazards.

Parametric model for post-processing visibility ensemble forecasts

Abstract. Although, by now, ensemble-based probabilistic forecasting is the most advanced approach to weather prediction, ensemble forecasts still suffer from a lack of calibration and/or display systematic bias, thus requiring some post-processing to improve their forecast skill. Here, we focus on visibility, a weather quantity that plays a crucial role in, for example, aviation and road safety or ship navigation, and we propose a parametric model where the predictive distribution is a mixture of a gamma and a truncated normal distribution, both right censored at the maximal reported visibility value. The new model is evaluated in two case studies based on visibility ensemble forecasts of the European Centre for Medium-Range Weather Forecasts covering two distinct domains in central and western Europe and two different time periods. The results of the case studies indicate that post-processed forecasts are substantially superior to raw ensembles; moreover, the proposed mixture model consistently outperforms the Bayesian model averaging approach used as a reference post-processing technique.

Physics-Based vs Data-Driven 24-Hour Probabilistic Forecasts of Precipitation for Northern Tropical Africa

Abstract Numerical weather prediction (NWP) models struggle to skillfully predict tropical precipitation occurrence and amount, calling for alternative approaches. For instance, it has been shown that fairly simple, purely data-driven logistic regression models for 24-h precipitation occurrence outperform both climatological and NWP forecasts for the West African summer monsoon. More complex neural network–based approaches, however, remain underdeveloped due to the non-Gaussian character of precipitation. In this study, we develop, apply, and evaluate a novel two-stage approach, where we train a U-Net convolutional neural network (CNN) model on gridded rainfall data to obtain a deterministic forecast and then apply the recently developed, nonparametric Easy Uncertainty Quantification (EasyUQ) approach to convert it into a probabilistic forecast. We evaluate CNN+EasyUQ for 1-day-ahead 24-h accumulated precipitation forecasts over northern tropical Africa for 2011–19, with the Integrated Multi-satellitE Retrievals for GPM (IMERG) data serving as ground truth. In the most comprehensive assessment to date, we compare CNN+EasyUQ to state-of-the-art physics-based and data-driven approaches such as monthly probabilistic climatology, raw and postprocessed ensemble forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), and traditional statistical approaches that use up to 25 predictor variables from IMERG and the ERA5 reanalysis. Generally, statistical approaches perform about on par with postprocessed ECMWF ensemble forecasts. The CNN+EasyUQ approach, however, clearly outperforms all competitors in terms of both occurrence and amount. Hybrid methods that merge CNN+EasyUQ and physics-based forecasts show slight further improvement. Thus, the CNN+EasyUQ approach can likely improve operational probabilistic forecasts of rainfall in the tropics and potentially even beyond. Significance Statement Precipitation forecasts in the tropics remain a great challenge despite their enormous potential to create socioeconomic benefits in sectors such as food and energy production. Here, we develop a purely data-driven, machine learning–based prediction model that outperforms traditional, physics-based approaches to 1-day-ahead forecasts of rainfall occurrence and rainfall amount over northern tropical Africa in terms of both forecast skill and computational costs. A combined data-driven and physics-based (hybrid) approach yields further (slight) improvement in terms of forecast skill. These results suggest new avenues to more accurate and more resource-efficient operational precipitation forecasts in the Global South.

Global Probabilistic Forecasting for Multiple Tight Gas Wells Using Deep Autoregressive Networks

Summary Traditional methods for forecasting production rate, such as Arps, analytical techniques, and recurrent neural network (RNN)–based deep learning, are mainly point prediction techniques developed within the framework of single-well forecasting. These methods often face limitations stemming from single-well historical production data and model assumptions, hindering their ability to consider the influence of development patterns of other production wells within the block on the target well. In addition, they struggle to predict the multiple production rate time series simultaneously and often fail to quantify uncertainty in predictions or adequately exploit extensive relevant historical production data. To tackle these challenges, we propose a model based on the deep autoregressive recurrent neural network (DeepAR), leveraging related multiwell production rate data to enable global modeling and probabilistic forecasting. This model incorporates dynamic covariate and static categorical variable data, integrating Bayesian inference and using gradient descent algorithms and maximum likelihood estimation methods to derive a comprehensive historical-future production probability evolution pattern shared across multiple wells. Leveraging data from 943 tight gas wells, a comprehensive evaluation of the DeepAR model’s performance was undertaken, encompassing the comparison of prediction accuracy with long short-term memory (LSTM), classification prediction, cold-start prediction, and single-well multitarget prediction scenarios, summarizing the applicability conditions for each. The research findings highlight that DeepAR integrates the acquired comprehensive production probability evolution pattern with specific production historical data of the target well to formulate a “comprehensive + specific” production probability prediction approach, resulting in improved stability and accuracy. On average, DeepAR demonstrates a 58.79% reduction in normalized deviation (ND) compared to the LSTM model, showcasing enhanced stability, particularly in scenarios involving frequent well shut-ins and openings. Moreover, DeepAR can learn static categorical features, with the classification model resulting in a 27.15% reduction in the ND compared to the unclassified model. Furthermore, DeepAR adeptly addresses the challenge of limited data availability, achieving cold-start prediction and facilitating multitarget single-well training and prediction while considering the interdependency among multiple variables over time and effectively mitigating the issue of missing auxiliary variables during the prediction phase. This study contributes to a broader understanding of production forecasting and analysis of production dynamics methods from a big data perspective.

Leveraging Deterministic Weather Forecasts for In Situ Probabilistic Temperature Predictions via Deep Learning

Abstract We propose a neural network approach to produce probabilistic weather forecasts from a deterministic numerical weather prediction. Our approach is applied to operational surface temperature outputs from the Global Deterministic Prediction System up to 10-day lead times, targeting METAR observations in Canada and the United States. We show how postprocessing performance is improved by training a single model for multiple lead times. Multiple strategies to condition the network for the lead time are studied, including a supplementary predictor and an embedding. The proposed model is evaluated for accuracy, spread, distribution calibration, and its behavior under extremes. The neural network approach decreases the continuous ranked probability score (CRPS) by 15% and has improved distribution calibration compared to a naive probabilistic model based on past forecast errors. Our approach increases the value of a deterministic forecast by adding information about the uncertainty, without incurring the cost of simulating multiple trajectories. It applies to any gridded forecast including the recent machine learning–based weather prediction models. It requires no information regarding forecast spread and can be trained to generate probabilistic predictions from any deterministic forecast. Significance Statement Weather is difficult to predict a long time in advance because we cannot measure the state of the atmosphere precisely enough. Consequently, it is common practice to run forecasts several times and look at the differences to evaluate how uncertain the prediction is. This process of running ensemble forecasts is expensive and consequently not always feasible. We propose a middle ground where we add uncertainty information to forecasts that were run only once, using artificial intelligence. Our method increases the value of these forecasts by adding information about the uncertainty without incurring the cost of multiple full simulations.

Predicting Oil Recovery Under Uncertainty for Huff ‘n’ Puff Gas Injection: A Field Case Study in the Permian Basin

Summary The extent and size of the fracture network connected to the wellbore are among the most uncertain parameters for multifractured horizontal wells. The uncertainty has a profound impact on production under primary depletion and has an even bigger impact on the recovery during huff ‘n’ puff (HnP) gas injection. In this paper, we quantify the uncertainty in reservoir properties and enhanced oil recovery (EOR) for a pair of wells in the Permian basin. In addition, we report on the lessons learned from modeling a field pilot HnP project involving four cycles using reservoir simulation. We create a base sector model for two wells completed in Wolfcamp B and C formations, where fractures are incorporated using an embedded discrete fracture model. For each well, we identify 13 uncertain parameters (describing fracture properties and compaction, initial conditions, and relative permeability). We use a Bayesian assisted-history-matching (AHM) method to match primary production data. Our study emphasizes the critical role of grid refinement for the accurate depiction of the gas-injection process by comparing three models with varying grid sizes. Additionally, we study the impact of molecular diffusion on oil and gas recovery using different diffusion coefficients. Another significant factor is the dynamic nature of the fracture network during cyclic injection, as indicated by evidence of communication between the wells. This makes the accessible fracture area during injection highly uncertain. To address interwell interference, we incorporate an additional fracture network activated during injection. To manage the uncertainty of active fracture area, we created four scenarios within the literature range, approximately 0.5, 1, 2, and 4 million ft² per stage, and tuned their properties to match historical HnP data. The production predictions from the AHM solutions were extended beyond the initial four HnP cycles to provide probabilistic forecasts. We demonstrate that grid refinement has a minimal effect on production under primary depletion but significantly influences HnP recovery. We attribute this disparity to numerical dispersion, which causes the artificial mixing of gas in the large gridblocks, leading to an overly optimistic recovery. Activating the molecular-diffusion mechanism results in a negligible impact on the recovery in our case. The contribution of molecular diffusion becomes noticeable only when diffusion coefficients exceed ten times the reported values. This leads us to conclude that molecular diffusion is of second-order importance compared with advection and can generally be neglected in our case. This finding should not be generalized to all shale formations. In reservoirs characterized by lower permeability and fluids with higher diffusion coefficients, molecular diffusion could still play a crucial role. We highlight that the contacted-fracture area, which is often underestimated in the reservoir simulation practice, is a critical factor for accurate modeling. Production for 37 filtered AHM solutions is extended under the four scenarios with variable fracture surface areas. Based on our analysis, we report that the HnP results in an average of 76% improvement in oil recovery compared with natural depletion.

How dependent are quantitative volcanic ash concentration and along‐flight dosage forecasts to model structural choices?

AbstractProducing quantitative volcanic ash forecasts is challenging due to multiple sources of uncertainty. Careful consideration of this uncertainty is required to produce timely and robust hazard warnings. Structural uncertainty occurs when a model fails to produce accurate forecasts, despite good knowledge of the eruption source parameters, meteorological conditions and suitable parameterizations of transport and deposition processes. This uncertainty is frequently overlooked in forecasting practices. Using a Lagrangian particle dispersion model, simulations with varied output spatial resolution, temporal averaging period and particle release rate are performed to quantify the impact of these structural choices. This experiment reveals that, for the 2019 Raikoke eruption, structural choices give measurements of peak ash concentration spanning an order of magnitude, significantly impacting decision‐relevant thresholds used in aviation flight planning. Conversely, along‐flight dosage estimates exhibit less sensitivity to structural choices, suggesting it is a more robust metric to use in flight planning. Uncertainty can be reduced by eliminating structural choices that do not result in a favourable level of agreement with a high‐resolution reference simulation. Reliable forecasts require output spatial resolution 80 km, temporal averaging periods 3 h and particle release rates 5000 particles/h. This suggests that simulations with relatively small numbers of particles could be used to produce a large ensemble of simulations without significant loss of accuracy. Comparison with previous Raikoke simulations indicates that the uncertainty associated with these constrained structural choices is smaller than those associated with satellite constrained eruption source parameter and internal model parameter uncertainties. Thus, given suitable structural choices, other epistemic sources of uncertainty are likely to dominate. This insight is useful for the design of ensemble methodologies which are required to enable a shift from deterministic to probabilistic forecasting. The results are applicable to other long‐range dispersion problems and to Eulerian dispersion models.

MEMPSEP‐I. Forecasting the Probability of Solar Energetic Particle Event Occurrence Using a Multivariate Ensemble of Convolutional Neural Networks

AbstractThe Sun continuously affects the interplanetary environment through a host of interconnected and dynamic physical processes. Solar flares, Coronal Mass Ejections (CMEs), and Solar Energetic Particles (SEPs) are among the key drivers of space weather in the near‐Earth environment and beyond. While some CMEs and flares are associated with intense SEPs, some show little to no SEP association. To date, robust long‐term (hours‐days) forecasting of SEP occurrence and associated properties (e.g., onset, peak intensities) does not effectively exist and the search for such development continues. Through an Operations‐2‐Research support, we developed a self‐contained model that utilizes a comprehensive data set and provides a probabilistic forecast for SEP event occurrence and its properties. The model is named Multivariate Ensemble of Models for Probabilistic Forecast of Solar Energetic Particles (MEMPSEP). MEMPSEP workhorse is an ensemble of Convolutional Neural Networks that ingests a comprehensive data set (MEMPSEP‐III by Moreland et al. (2024, https://doi.org/10.1029/2023SW003765)) of full‐disc magnetogram‐sequences and in situ data from different sources to forecast the occurrence (MEMPSEP‐I—this work) and properties (MEMPSEP‐II by Dayeh et al. (2024, https://doi.org/10.1029/2023SW003697)) of a SEP event. This work focuses on estimating true SEP occurrence probabilities achieving a 2.5% improvement in reliability and a Brier score of 0.14. The outcome provides flexibility for the end‐users to determine their own acceptable level of risk, rather than imposing a detection threshold that optimizes an arbitrary binary classification metric. Furthermore, the model‐ensemble, trained to utilize the large class‐imbalance between events and non‐events, provides a clear measure of uncertainty in our forecast.