Correcting Coarse-Resolution Weather and Climate Models by Machine Learning from Global Storm-Resolving Simulations

Running a fine grid model with 107 - 109 of cells is possible using a supercomputer with 103 - 106 of CPUs but may not be always cost-effective. The most cost-effective way is to use a coarse grid model that is much smaller but with static/dynamic profiles very close to the fine grid model. This paper proposes a new layer optimization and upscaling method with the aim for creating a consistent coarse grid model. Unlike the industry's existing layer optimization and upscaling methods, the proposed method performs layer optimization and upscaling fully integrated with the Lorenz coefficient and curves (LCC). Coarse grid layers and their permeabilities are created by minimizing the difference between fine and coarse grid LCCs. The process consists of static and dynamic optimizations. The former is measured by LCC while the latter by pressure, GOR, and water-cut. A new LCC-based permeability upscaling method is developed to preserve the fine grid multiphase flow behaviors. A satisfactory coarse grid model is achieved when both static and dynamic criteria are met. The proposed method has been successfully applied to a giant carbonate oil field in the Caspian Sea that consists of a matrix dominated platform and a fracture/karst dominated rim. Due to the field's complex geology and high H2S content (15%), a dual porosity, dual permeability compositional model has been created to model compositional sour crude flow within and between the matrix and fracture/karst features. The reservoir drive mechanisms are fluid expansion, miscible gas injection and aquifer drive. The reservoir is undersaturated and has an abnormally high initial reservoir pressure. The fine-grid static model contains 104 million cells (370×225×625×2) and the optimized upscaled coarse-grid dynamic model has 8.3 million cells (370×225×50×2). The upscaled model can be run efficiently on the company's existing HPC infrastructure with a maximum of 64 CPUs. Excellent matches of the Lorenz coefficient maps for reservoir total/zones and Lorenz curves at all wells between the fine and coarse grid models have been achieved. Matches on the dynamic variables, e.g., pressure, gas breakthrough time, and GOR growth, in all producers are within the defined acceptable tolerances. The high quality of the static and dynamic matches between the coarse- and fine-grid models confirms that the reservoir properties of the coarse-grid model is very close to the fine-grid model and can be used a base model for history matching and uncertainty analysis.

Two modeling approaches are commonly utilized for simulating flow in fractured formations: the discrete fracture network (DFN) approach and the stochastic continuum (SC) approach. Although the DFN approach is the most accurate, it has computational and memory constraints. The SC approach ensures fast processing but results in system over-homogenization. The fracture continuum (FC) approach arises as an integrated technique that incorporates the merits of both approaches. The main objective of this research is to develop a computationally efficient technique based on the FC approach to simulate the radial groundwater flow towards wells through two-dimensional fractured media under both steady and transient conditions. A stochastic generation of the DFN is performed in a Monte Carlo framework taking into account wells positioning. The DFN flow system is solved by applying the mass balance equation at fracture intersections. Fracture segments are mapped onto grids of 1 × 1 m and 5 × 5 m resolution as conductivity and specific storage cells. The grid flow problem is solved via MODFLOW. Flow and head discrepancies between the proposed technique and the DFN approach (reference solution) are assessed in steady and transient conditions. A grid-conductivity correction is needed to preserve the DFN flow in the presence of wells. A porosity estimation is proposed to identify the grid-pressure transient response. Promising flow and head results are observed for fine and coarse grid models. Some of the studied cases show large discrepancies in the maximum drawdown obtained on the coarse grid model. Accordingly, a new technique is proposed to handle such discrepancies and is found efficient in transient simulations (e.g., 11% and 26.12% discrepancies are minimized to − 0.93% and − 1.03% for two studied cases). The adopted mapping technique is found efficient when the interest is to estimate the average drawdown over an aquifer as correlation coefficients of 0.89 and 0.97 are found for the coarse and fine grid models, respectively when compared to the DFN model. However, the technique has limitations in estimating the drawdown at locations of wells.

To examine spatial and temporal scales of katabatic flow, a distributed temperature sensing (DTS) optical fiber was deployed 2 km down a mild slope irregularly interrupted by small-scale drainage features as part of the Mountain Terrain Atmospheric Modeling and Observation (MATERHORN) experiment conducted at the U.S. Army Dugway Proving Ground, Utah. The fiber was suspended at two heights near the surface, enabling measurement of variations in lapse rate near the surface at meter-scale spatial resolution with 1-min temporal resolution. Experimental results derived from the DTS and tower-mounted instrumentation indicate that airflow through small-scale drainage features regulated the local cooling rate whereas topographic slope and distance along the drainage strongly influenced the larger-scale cooling rate. Empirical results indicate that local cooling rate decays exponentially after local sunset and basin-wide cooling rate decreases linearly with time. The difference in the functional form for cooling rate between local and basin-wide scales suggests that small-scale features have faster timescales that manifests most strongly shortly after local sunset. More generally, partitioning drainage flow by scale provides insight and a methodology for improved understanding of drainage flow in complex terrain.

A new wake model and comparison of eight algorithms for layout optimization of wind farms in complex terrain

This chapter describes the application of a variety of remote sensors to complex terrain studies, including sodars, lidars (aerosol mapping and Doppler), clear-air radars and wind profilers, and optical crosswind sensors. It outlines sensor characteristics relevant to operation in complex terrain settings and examines the origin of turbulence microstructure in complex terrain flows, a significant factor in the use of clear-air radars and sodars. It also summarizes the major programs that have made extensive use of remote sensors and promoted their integration with conventional in-situ instruments. Examples and case studies from these programs are then provided including observations of the initiation and destruction of drainage winds. The behavior of quasi-steady drainage winds and their irregularities during marine layer intrusions, seiches, waves, and strong ambient winds are also described. Recent results are presented that extend the use of Doppler lidar in channeled and merging valley flows as well as in the observation of large-scale mountain-plain circulations. The integration of remote and in-situ sensors in a major transport and dispersion study yields further examples of how different measurement methods can complement each other. Future directions in the study of complex terrain processes are suggested that can take advantage of newly developing remote sensors especially in application to larger-scale mountain-valley circulations.

Thermo-mechanical modelling of solar central receivers: Effect of incident solar flux resolution

The relationship between the surface and synoptic wind direction is examined climatologically in a complex terrain region. Surface winds were observed over a 1-yr period during the MISTRAL project in the Basel, Switzerland, area. The measurement sites were located in various topographical settings, namely, in broad and narrow valleys, on slopes, at hilltops, on passes, and at mountaintops. The synoptic winds above the MISTRAL area are approximated by upper-level winds from routine rawinsonde observations. The relationship between this synoptic wind and the surface wind at each site was compared to a conceptual model. According to the conceptual model used, there are four mechanisms for the forcing of near-surface winds by synoptic winds. Three of the four forcing mechanisms leading to channeled flow are identified in the MISTRAL area. In this region with its complex terrain, different channeling mechanisms act at different locations as well as different mechanisms may occur at the same location. The study shows that the type of channeling depends on the topography surrounding the observation site. The combination of several channeling mechanisms with the multitude of valley orientations in a complex terrain setting explains the variety of observed flow patterns. One mechanism, the thermal forcing of valley winds. is examined in more detail. Two minima in the averaged diurnal cycle of the wind speed are found. Both minima occur at the time when the direction of the thermally forced wind reverses - one in the morning and one in the evening. The daytime upvalley flow is. on average, stronger than the nighttime downvalley flow. In the MISTRAL region, the frequency of days with thermally driven flow does not have a significant annual cycle.

Complex terrain covers the vast majority of the Earth surface and affects local thermal wind flows by adding a gravitational component to the dynamics responding to local pressure gradients. The Leman Lake region, in Switzerland, thanks to its mountaineous topography and the presence of a large alpine lake, represents an ideal testbed for studying the interaction between regional and local scale flows in complex terrain. However, most of the research investigating local fluid dynamics in the area predominantly focuses on the impact of wind flow on internal lake circulations, and limited attention has so far been given to the role of the local lake breeze circulation in modulating wind flow in the area. Here, we use a set of high-resolution Weather Research and Forecast (WRF) model simulation experiments to investigate the diurnal and seasonal evolution of boundary layer dynamics in the Leman Lake region, with a focus on the environmental impacts associated with the city of Lausanne. In order to isolate urban impacts and explore the role of different land cover types, we compare simulation results from an “Urban” scenario featuring a realistic landscape representation, with simulation results from an hypothetical “Rural” scenario where urban areas are replaced by croplands. Analysis of results shows that the Lausanne urban area, although of limited extent, is able to modify wind flows locally, e.g., by anticipating the diurnal onset of the lake breeze circulation. In turn, regional wind flows interact with the local UHI by advecting cold air from the Leman Lake during the winter, in accordance with what has been observed by similar studies in world regions charaterized by different topographical and climatological conditions. In addition, we compare simulation results with Doppler Lidar vertical wind profiles as part of a recently initiated measurement campaign on site.

A model for the simulation of wind flow in complex terrain is presented based on the Reynolds averaged Navier–Stokes (RANS) equations. For the description of turbulence, the standard k-ε, the renormalization group (RNG) k-ε, and a Reynolds stress turbulence model are applied. Additional terms are implemented in the momentum equations to describe stratification of the Earth’s atmosphere and to account for the Coriolis forces driven by the Earth’s rotation, as well as for the drag force due to forested canopy. Furthermore, turbulence production and dissipation terms are added to the turbulence equations for the two-equation, as well as for the Reynolds stress models, in order to capture different types of land use. The approaches for the turbulence models are verified by means of a homogeneous canopy test case with flat terrain and constant forest height. The validation of the models is performed by investigating the WindForS wind test site. The simulation results are compared with five-hole probe velocity measurements using multipurpose airborne sensor carrier (MASC) systems (unmanned small research aircraft)—UAV at different locations for the main wind regime. Additionally, Reynolds stresses measured with sonic anemometers at a meteorological wind mast at different heights are compared with simulation results using the Reynolds stress turbulence model.

The Alaiz Experiment (ALEX17), located in Navarre (Spain), is a full-scale experiment that aims to peer into the microscale flow of a large-scale and complex topography influenced by mesoscale forcing. The experimental site site consists of a mountain range of 1000m a.s.l with a well-known wind regime. To the North, aligned with the prevailing wind, measurement equipment are located at a valley 500m lower in altitude. The measurement campaign started in November/2017, with an intense observational period (IOP) of six months as well as 1-year of met mast measurements along with long-term means from a reference mast.<br> <br> The multi-lidar scanning patterns can provide 2D and 3D wind reconstruction over a 125m a.s.l. Z-shaped transect 10km long, being a unique feature from this experiment. The intersecting scans used for wind speed reconstruction can also be rendered into virtual met masts.<br> <br> Results show this multi-lidar layout is capable of capturing main flow patterns as speed-ups on ridge tops and valley flow between ridges, with valid data from 100m to 3500m. The presentation summarizes the main layout and results of the experiment, pointing the way to access the open-source dataset of all equipment and documentation of ALEX17.

A variety of programmes and field experiments were carried out in order to develop and evaluate models of transport and diffusion of pollutants in complex terrain areas. As part of this programme, in this study, we have focused our interest on analyzing the basic features of different flow fields and thermal structures developed in a complex area and their relation to air pollution problems. The area is located in the province of Barcelona (in the northeast of Spain) close to a wide industrial zone, thus a pollutant flux could affect this region. In order to carry out the main purpose of this study we have analysed data from a Doppler Sodar (FAS 64) and a network of near surface meteorological and air quality stations. In addition, different dynamical simulations given by a numerical mesoscale model (MM5) are also analyzed. The results show that the main flow fields and thermal structures generated in this area are: sea breeze, slope drainage winds, channelling winds created by terrain constrictions and cool-air accumulation in low-lying regions. This last structure, developed specially in winter time, gives rise to stagnant cold air masses and strong thermic inversions, with average lapse rate of −4 degrees on 100 m, which contribute to increase air pollution concentration, especially SO2. Hourly and daily averaged SO2 concentration can be higher than 350 and 138 µg m−3 respectively. In addition, as “La Plana” is located not far from the Mediterranean Sea, during summertime the sea breeze arrives into this zone via its southern entrance, thereby reaching the whole area. The arrival of the sea breeze in to “La Plana”, which advects pollutants from the nearby industrial area, is the main cause of some of these pollutants, especially ozone and its precursors, attaining high concentrations during afternoon hours. The contribution of the sea breeze is variable, but could represent between a 25% to a 30% of its total value.

Compositional upscaling in fractured reservoirs during gas recycling

We analyze trends in temperature from 18 temperature stations and one upper air sounding site at 30°–35° S in central Chile between 1979–2015, to explore geographical and season temperature trends and their controls, using regional ocean-atmosphere indices. Significant warming trends are widespread at inland stations, while trends are non-significant or negative at coastal sites, as found in previous studies. However, ubiquitous warming across the region in the past 8 years, suggests the recent period of coastal cooling has ended. Significant warming trends are largely restricted to austral spring, summer and autumn seasons, with very few significant positive or negative trends in winter identified. Autumn warming is notably strong in the Andes, which, together with significant warming in spring, could help to explain the negative mass balance of snow and glaciers in the region. A strong Pacific maritime influence on regional temperature trends is inferred through correlation with the Interdecadal Pacific Oscillation (IPO) index and coastal sea surface temperature, but the strength of this influence rapidly diminishes inland, and the majority of valley, and all Andes, sites are independent of the IPO index. Instead, valley and Andes sites, and mid-troposphere temperature in the coastal radiosonde profile, show correlation with the autumn Antarctic Oscillation which, in its current positive phase, promotes subsidence and warming at the latitude of central Chile.

In this study, various upscaling techniques and their effects on Barik tight gas Formation simulation modelling results were investigated. The intent of this upscaling study is to recommend coarse models that provide approximately the same flow behavior or well performance as the fine grid model. The study will help to develop an awareness of the range of applicability for the upscaled coarse models. It will also allow coarse grid models to be used appropriately and with greater confidence in production optimization. These techniques comprised several alternative ways of grouping reservoir data, with respect to petrophysical rock types (herein referred to as RT). This scheme defines 5 individual rock types, with RTs 1 to 4 broadly defining pay and RT 5 defining non-pay. The layers in the simulation models are made up of single or multiple grouped RTs from the same zone. Keeping each layer in the model, ordered as it was in the well log, resulted in 85 simulation layers for the fine grid model. The upscaled or coarse models have 16 to 35 simulation layers, with the smallest being the model where RTs 1 to 4 are grouped together. Different RTs sorting were tested in each of the upscaled models. The results of this study suggest that re-ordering the log information so that all of the rock for each rock type was grouped together inside each stratigraphic unit appears to be an acceptable upscaling technique that gives reasonable efficiency and accuracy. This type of upscaling was required since the study showed that any upscaling method that averaged the properties of RT 1 with the properties of other rock types within the same simulation layer could result in optimistic EUR estimates of up to 30% relative absolute error when compared to a fine grid model. The choice of upscaling method is particularly significant in a low Kv/Kh ratio environment, but less so within environments with a high Kv/Kh ratio. The upscaling technique has only been tested for the Barik formation and should not be used elsewhere without proper testing. The upscaling technique works for the Barik formation because of several distinctive features of the reservoir and the wells. In applying this upscaling technique we assume that: The wells are all hydraulically fractured and thus the flow is generally horizontal into the fractures and from the fractures into the well.The fracture extends from the top to the bottom of each stratigraphic unit that the fracture encounters. If a fracture only partially penetrates a stratigraphic unit, this method may not work.The gas is relatively dry and thus the flow is less affected by gravity than for reservoirs containg flowing liquids.The reservoir is made up of rock with a wide range of permeability, but the flow is dominated by the well connected, higher permeability rocks (RT 1).The reservoir is very stratified, and the correlation length of the rock types has to be very much greater than the well spacing. Re-ordering of the RTs in a given upscaling method will result in acceptable accurate estimates of hydrocarbon recovery, compared to the fine grid model. It showed that the order of layering did not matter for the area studied, because a conductive fracture connects all the layers. This method probably really is only applicable for dry gas reservoirs (where gravity is not important) and in fractured wells (where horizontal flow into the fractures and into wells dominates). In oil reservoirs and rich condensate reservoirs where gravity is an important factor, the ordering of the rock types within the simulation layer may matter. It will also be shown that an approximately 90% reduction in the simulation modelling computing time could be achieved if the appropriate upscaling technique is used. To achieve this reduction in computing time, some compromises were made, including assuming RT 4 is non-pay in upscaled models where RTs 4 &amp; 5 are grouped together in the same simulation layers, resulting in reduction of HCPV. It is important to mention that some RT 5 zones in the log have thin instances of other rock types, which are not accounted for in the upscaled models and could result in an error in average pore volume preservation.

The paper investigates the applicability of machine learning (ML) to weather prediction by building a reservoir computing‐based, low‐resolution, global prediction model. The model is designed to take advantage of the massively parallel architecture of a modern supercomputer. The forecast performance of the model is assessed by comparing it to that of daily climatology, persistence, and a numerical (physics‐based) model of identical prognostic state variables and resolution. Hourly resolution 20‐day forecasts with the model predict realistic values of the atmospheric state variables at all forecast times for the entire globe. The ML model outperforms both climatology and persistence for the first three forecast days in the midlatitudes, but not in the tropics. Compared to the numerical model, the ML model performs best for the state variables most affected by parameterized processes in the numerical model.