The current grid code in China in regard to solar forecasting is, in my opinion, underdeveloped, especially in contrast to the rate at which photovoltaics are being installed. As such, explaining the limitations of the grid code and resetting pathways to improve it are thought utilitarian for those scientists and policymakers who are responsible for or aware of the grid code but have not themselves worked on the problem of forecasting. In this perspective article, I should first explain with respect to China's grid code the perceived deficiencies in the current forecasting research and practices, and then outline a five-stage workflow that could completely mitigate the situation. Among other things, the over-reliance on accuracy as the basis for gauging the goodness of forecasts is identified as a root cause for the status quo, and thus, I advocate a holistic forecast verification procedure that encompasses consistency, quality, and value. With that in mind, the proposed workflow for better solar forecasting for grid integration purposes relies on the effective information flow among the weather department, grid operators, and individual plant owners, which is inline with the current grid code. What goes beyond this is that the proposal further introduces a couple of concepts called “hierarchical reconciliation” and “firm forecasting,” which are new but are able to eliminate forecast errors wholly, thus making solar power dispatchable on the system level. With a slight premium incurred, it is now possible to manage solar plants, or variable renewables in general, in the same style as managing conventional fire-powered generators.
INTRODUCTION
Solar forecasting, as a key constituent of modern energy meteorology, started to thrive in the early 2010s. As is the case of almost all other forecasting domains, much of the virtue of solar forecasting has hitherto been thought to be embedded in accuracy. Indeed, a large fraction of research initiatives on solar forecasting places accuracy as the foremost and often only criterion for gauging the goodness of solar forecasts. However, this view has been formally challenged via the collective effort of 33 experts,1 who advocate the joint use of consistency, quality, and value to verify the goodness of forecasts. Very briefly, consistency refers to the correspondence between forecasts and judgments; quality refers to the correspondence between forecasts and observations; value denotes the incremental benefits of forecasts to users. If the holistic forecast verification procedure as recommended by Yang et al.1 can be agreed upon at any rate, one should logically acknowledge the limitations of using just accuracy, which is but one among the many aspects of quality, to justify any progress in solar forecasting research.
In the power and energy industrial sector, accuracy measures, such as the root mean square error (RMSE) or mean absolute error (MAE), are commonly used as a basis to set penalty triggers and schemes, in order to penalize “bad” forecasts submitted to the grid operators by individual photovoltaic (PV) power plant owners. This is in fact the case for the State Grid Corporation of China (SGCC), which, together with the China Southern Power Grid (CSG), are the two major system operators of China. As reviewed by Yang et al.,2 the area managed by SGCC can be sub-divided into several zones (see the top panel of Fig. 1), each covering several provinces of China and running on a slightly different set of grid code. There are three conspicuous issues with the current grid code of SGCC pertaining to forecast submission. First is that the diverse penalty schemes across zones are not conducive to centralized forecast dissemination by the weather department, because each penalty scheme corresponds to a different definition of optimality—this implies a lack of consistency.3 Second is that the forecast penalties are set homogeneously within each zone, disrespecting the intrinsic diversity in climate and weather regimes (see the bottom panel of Fig. 1), which impacts the predictability at large—this indicates an incomplete process of quality assessment.4 Third is that the mapping from accuracy measure to monetary penalty is entirely ad hoc, with penalty multipliers allocated by empiricism, but without proper considerations of the economics of power system operations—this signifies a divorce between quality and value of forecasts.5 All three deficiencies are to be further elaborated below, before I also propose some means for improvement from a researcher's perspective. However, for now, I should first explain the kind of forecasting workflow that is currently mandated by the SGCC.
(Top) The zoning of the power grid in China. Whereas CSG and the Tibet grid are managed separately, all other regional grids are under the SGCC. (Bottom) The Köppen–Geiger climate classification over China. Abbreviations: CCG, Central China Grid; CSG, China Southern Power Grid; ECG, East China Grid; NCG, North China Grid; NEG, Northeast China Grid; NWG, Northwest China Grid; and SGCC, State Grid Corporation of China.
(Top) The zoning of the power grid in China. Whereas CSG and the Tibet grid are managed separately, all other regional grids are under the SGCC. (Bottom) The Köppen–Geiger climate classification over China. Abbreviations: CCG, Central China Grid; CSG, China Southern Power Grid; ECG, East China Grid; NCG, North China Grid; NEG, Northeast China Grid; NWG, Northwest China Grid; and SGCC, State Grid Corporation of China.
THE PRESENT
Solar forecasting for grid integration in China adopts a top-down–bottom-up workflow. In that, the Public Service Center of the China Meteorological Administration (CMA) disseminates numerical weather prediction (NWP) and satellite-based irradiance forecasts to provincial meteorological bureaus, which are tasked to dynamically or statistically downscale forecasts to PV-plant-level forecasts, and then further disseminate the downscaled forecasts to end users; this is the top-down part of the forecasting process. With the disseminated forecasts, solar power forecast providers or individual plant owners need to convert the irradiance forecasts to PV power forecasts, and then submit those to the grid operators for centralized management; this constitutes the bottom-up phase of the forecasting process. At the moment, the mesoscale NWP forecasts of CMA are issued on a 9-km lattice, with grid cells being larger than most PV plants, so the need for spatial downscaling is apparent; the satellite-based forecasts have a higher resolution of 4 km, and downscaling is less relevant, because in theory the Advanced Geostationary Radiation Imager onboard Fengyun-4A (FY-4A) allows irradiance retrieval at 0.5 km from its visible channel,6 which would be sufficiently refined for solar forecasting. As for the temporal resolution, both the raw NWP- and satellite-based forecasts have a 15-min temporal resolution, which aligns well with the grid code, which mandates day-ahead (2–7-day horizons, depending on the SGCC zones) and intra-day (4-h horizon) forecasts to be submitted in a rolling fashion at a 15-min resolution. Table I lists the existing forecast products provided by CMA to facilitate grid integration, alongside the variables disseminated and forecasting methods used.
Forecast products provided by CMA to facilitate grid integration. Notation: , forecast span; , forecast resolution; and , the refresh rate. Abbreviations: BNI, beam normal irradiance; CMA, China Meteorological Administration; DHI, diffuse horizontal irradiance; FY-4A, Fengyun-4A; GHI, global horizontal irradiance; NWP, numerical weather prediction; and WSP, wind and solar power.
Horizon . | . | . | . | Variables . | Method . |
---|---|---|---|---|---|
Long term | 12 mon | 1 mon | 1 mon | Monthly average GHI, temperature, and wind speed | Ensemble machine learning on historical and reanalysis data |
Medium term | 240 h | 15 min | 12 h | GHI, temperature, humidity, pressure, cloud cover, and wind speed and direction on several vertical layers | CMA-WSP NWP model |
Short term | 72 h | 15 min | 12 h | GHI, temperature, humidity, pressure, cloud cover, and wind speed and direction on several vertical layers | CMA-WSP NWP model |
Very short term | 4 h | 15 min | 15 min | GHI, BNI, and DHI | Statistical extrapolation of the FY-4A satellite-derived irradiance |
Horizon . | . | . | . | Variables . | Method . |
---|---|---|---|---|---|
Long term | 12 mon | 1 mon | 1 mon | Monthly average GHI, temperature, and wind speed | Ensemble machine learning on historical and reanalysis data |
Medium term | 240 h | 15 min | 12 h | GHI, temperature, humidity, pressure, cloud cover, and wind speed and direction on several vertical layers | CMA-WSP NWP model |
Short term | 72 h | 15 min | 12 h | GHI, temperature, humidity, pressure, cloud cover, and wind speed and direction on several vertical layers | CMA-WSP NWP model |
Very short term | 4 h | 15 min | 15 min | GHI, BNI, and DHI | Statistical extrapolation of the FY-4A satellite-derived irradiance |
Several observations and conclusions can be made based on just the short description above. First and foremost, it is evident that physics-based solar forecasting (i.e., NWP and satellite) has already become the recognized and preferred approach of CMA, which is inline with the latest recommendations from the academia.7,8 However, insofar as the forecasting methodologies are concerned, they are still subject to further refinements, e.g., replacing the time series extrapolation with optical flow or cloud motion vector for satellite-based forecasting,9,10 which can better capture the cloud dynamics, or promoting the deterministic-style NWP forecasts to dynamical ensemble forecasts with perturbed initial conditions, which brings forward accuracy improvement and the capability of uncertainty quantification.11,12 Next, one can see that the kind of solar forecasts useful to grid integration is always multi-step-ahead in nature, which contrasts the one-step-ahead forecasts that often appeared in research. Since extending one-step-ahead forecasts to multi-step-ahead ones is not always straightforward,13,14 especially for data-driven models, the operational aspects of solar forecasting ought to be factored in while designing forecasting models.15,16 Third, it is also evident that the roles of solar forecast providers and individual plant owners have now completely changed: rather than issuing PV power forecasts from scratch, one only needs to handle the irradiance-to-power conversion process (i.e., solar power curve modeling), e.g., via a physical model chain, which takes as input the system design parameters and the (forecast) weather information, and outputs the (forecast) PV power, see Mayer and coworkers17–21 for the latest advances in model chain. In the fourth place, the importance of downscaling is now confirmed, which seeks to reduce the spatial scale mismatch between gridded forecasts and plant-level power output. In fact, this topic is scantly studied by solar engineers,22,23 but has always been a research hotspot of the atmospheric science community; more specifically, dynamical downscaling can be handled by running limited-area NWP models such as the Weather Research and Forecasting (WRF) model or Application of Research to Operations at Mesoscale (AROME) model.24,25 Last but not least, there is a high resemblance in the time specifications (i.e., horizon, resolution, and refresh rate) of CMA's forecasts and those requested by the grid operators,2 which suggests that the serviceability (or value) of irradiance forecasts is devised to be materialized in terms of the capacity to influence grid-side decision-making. With this in mind, we shall examine next the three aforementioned deficiencies in the current grid code of SGCC.
The definition of consistency can take several slightly differing forms, but all of which narrate the correspondence between forecasts and judgments. (Interested readers may go beyond the current discussion and learn about “hedging” and “propriety,” which are two important concepts that relate judgment to forecasts.26–29) If the forecasts are deterministic (i.e., single-valued), consistency is related to how a statistical functional can be optimally elicited from a distributional forecast according to a scoring function specified ex ante.30 Stated differently, consistency is concerned with drawing out a single value from a predictive distribution, such that the pre-specified error metric can be minimized. For instance, it has been demonstrated in a solar forecasting context that the RMSE is minimized by eliciting the mean of the distributional forecast, MAE is minimized by eliciting the median of the distributional forecast, whereas the squared percentage error (SPE) is minimized by eliciting the β-median (with ) of the distributional forecast.3,31 Currently, different SGCC zones use different error metrics to penalize forecast submissions, e.g., the Central China Grid (CCG) adopts an MAE-based penalty scheme, whereas the East China Grid (ECG) adopts an SPE-based penalty scheme. On this point, if consistency is to be respected at any rate, the current grid code of specifying different error metrics for different zones is deficient, because it implies a need for CMA to issue several versions of deterministic weather forecasts.
The concept of quality is quite straightforward to comprehend, as it can be broken down into several aspects, such as accuracy, association, calibration, refinement, discrimination, or skill, each being quantifiable via one or more statistics.32 Take for instance accuracy, it can be gauged by RMSE, MAE, SPE, among others. As for association, it is quantified by the correlation coefficient or coefficient of determination. It should be well known that the quality of solar forecasts, besides being dependent on the forecasting model itself, is influenced by a wide range of factors, such as the climate and weather regime, spatial and time scales, or the sky condition. Clearly then, it is not fair to penalize forecasts without considering the intrinsic difficulty of the forecasting problem—the issue is one of predictability,4,33–35 which is a grossly under-represented research direction of solar forecasting. As shown in Fig. 1, other than the ECG, which is quite homogeneous in terms of the Köppen–Geiger climate class, all zones exhibit diverse climate regimes. Hence, the current grid code is flawed in the sense that it confuses accuracy with the overall quality of forecasts. Yang36,37 advocated using the RMSE skill score as the main aspect of forecast quality during solar forecast evaluation, because RMSE can be decomposed via the Murphy–Winkler factorizations38 into six aspects of quality, namely, unconditional bias, association, type-1 conditional bias, resolution, type-2 conditional bias, and discrimination, allowing comprehensive assessment of forecast quality.39,40 Better still, skill is further related to the predictability of irradiance, which has the potential to promote a fair forecast penalty scheme.4,33
The third type of goodness of forecasts, that is, value, is quite a versatile idea, owing to the eclectic mix of ways to quantify it. Yet/However in a grid integration context, it refers mostly to the incremental monetary benefits gained by using one set of forecasts instead of another. Broadly speaking, there are two groups of users of solar forecasts: the energy market participants and the grid operators.8 As such the value of solar forecasts must be quantified from both perspectives. From the energy market participants' perspective, the PV plant owners, for instance, seek to avoid as far as possible the penalty incurred on them by submitting forecasts that minimize the pre-specified scoring function. From the power system operators' perspective, it is the cost of managing the variable solar power injection into the grid, e.g., setting the spinning reserves, that matter. If one proceeds from these two perspectives separately, it is thought possible to conclude one set of forecasts has a higher value than another. However, when these two perspectives are considered together, the current grid code, which imposes the penalty multiplier in an ad hoc fashion, does not seem to be able to facilitate joint decision-making, which is evidently deficient. Furthermore, because the value of forecasts is also tied to consistency and quality,1 the aforementioned drawbacks of the current grid code also prevent proper valorization of solar forecasts.
Having understood the pitfalls of the current grid code of SGCC, remedies must be sought. In dealing with inconsistency, an easy action is to simply standardize the penalty schemes, such that there is only one scoring function that is shared across the entire power system, and the elicitable functional for which that scoring function is strictly consistent can be issued by CMA. As for quality, the involvement of skill score and predictability is vital, in that, the grid operators may moderate the penalty in an objective and justified way. Although both recommendations here are logical, which motivated the work of Yang et al.,1 they do not solve the fundamental problems that motivate solar forecast submission—forecasts are needed because (1) the grid needs to be managed as a whole, and (2) solar power is inherently variable. In view of that, I propose next a five-stage solar forecasting workflow for grid integration that can completely revolutionize the way in which solar power is used, that is, making it dispatchable.
THE FUTURE
Currently, the solar forecasting workflow in China terminates after the forecasts are received by the grid operators; this, as mentioned earlier, corresponds to a top-down–bottom-up information flow. More specifically, this workflow has three stages: (1) irradiance and auxiliary parameter forecasting, which is handled by CMA, (2) forecast post-processing,41 which is handled by provincial meteorological bureaus, and (3) irradiance-to-power conversion, which is handled by the energy market participants. Arguably, another round of post-processing could be applied to the PV power forecasts, but the concept and techniques do not differ much from those of irradiance post-processing. However, inasmuch as the power system is to be managed in its entirety, the top-down–bottom-up information flow is suboptimal. This is because the participants who submit forecasts to the grid operators only have access to their own plant-level information, whereas the grid operators have access to system-level information that is also helpful to managing the solar power injection into the grid. In this regard, a potential fourth stage of the workflow, namely, the hierarchical forecast reconciliation is thought to be profoundly beneficial.
The power system is of a hierarchical nature, in that, both the load and solar power generation at individual household and plant levels sum up to the load and solar power generation at the distribution-feeder level, which further sum up to those at the transmission level. Whereas the actual load and generation are aggregate consistent, their forecasts are not.42,43 On this point, hierarchical forecast reconciliation, which takes as input the base forecasts generated on various levels of the hierarchy and outputs a set of reconciled forecasts (i.e., aggregate-consistent forecasts), has flourished over the past decade or so, especially in the energy forecasting domain.44–47 Hierarchical solar power forecasting has also been studied, in both deterministic and probabilistic settings,43,48–51 and the techniques are mature enough to support demonstrations in an actual power system context.
Hierarchical solar forecasts have over the conventional solar forecasts several advantages. First, they are aggregate consistent, which means the forecasts at the lower levels of the hierarchy sum up exactly to the forecasts at the higher levels; this facilitates joint decision-making. The second benefit is that the reconciled forecasts are often of higher quality than the base forecasts owing to their geometric advantage52 and ability to exploit the shared information across the entire hierarchy;53,54 if we are to argue that higher quality corresponds to higher value, this property of the reconciled forecasts is commendable. Third, hierarchical forecasting is computationally feasible even for very big data with millions of time series, owing to the natural construct of the summing matrix that allows recursive calculation of the matrix inverse needed by least squares;55,56 these data scale is essentially the kind concerning power systems. The last and possibly the most attractive feature of hierarchical forecasting is that most of its benefits are general, and are not impacted by how the base forecasts are created; this is a very appealing property, since different market participants may employ different approaches to generate, or covert from irradiance forecasts, base forecasts. All these benefits have been theoretically and empirically demonstrated before, and the reader is referred to the above-listed references for details.
Hierarchical solar forecast reconciliation should be, and in fact can only be, performed by grid operators, who possess not just the bottom-level submitted forecasts but also their own higher-level forecasts. However, the reconciliation changes/adjusts the forecasts, which is somewhat problematic, because the numerical values of the forecasts are tied to the penalty and, thus, the monetary gain of market participants, and any alteration could make the remuneration more or less to the participants' favor. What this implies is that the top-down–bottom-up information flow is no longer apt, and another round of top-down forecast dissemination is necessitated—this time, it is the reconciled forecasts that are shared by the grid operators to market participants. This new form of information flow is analogous to dispatching conventional generators, where the generation targets are released to the participants after unit commitment and bidding; the current grids have already the capability of supporting such information flow. However, one issue remains—solar power is not dispatchable.
Therefore, the last piece of the puzzle, as to reimagining a new solar forecasting workflow, is firm forecasting, which is able to remove forecast errors entirely, through the use of battery storage, PV overbuilding, and proactive curtailment, and other firm power enablers, such as solar–wind–hydro blending, demand response (i.e., load flexibility), or even a small amount of conventional spinning reserve.57 In a recent review article, Remund et al.58 has reported the latest advances in firm generation/forecasting made by the International Energy Agency Photovoltaic Power Systems Program (IEA PVPS) Task 16, and a possibly exhaustive collection of studies incorporating different combinations of firm power enablers over power grids in different countries has been presented. Common to all case studies depicted therein, battery storage and PV overbuilding and proactive curtailment are found to be essential prerequisites for firming up generation/forecasts. Therefore, the following discussion is confined to the two, to keep the firm-forecasting idea simple, but the reader should be aware that incorporating other firm power enablers usually implies further economic benefits, and the optimal combination would differ from one scenario to another.
In a nutshell, both the battery storage and PV overbuilding and proactive curtailment are able to modify the solar generation profile, such that it can align with the generation target (i.e., the reconciled forecasts) perfectly.59–61 Whereas most readers should already be acquainted with how battery storage works, PV overbuilding and proactive curtailment is an exceedingly recent conception. This conception exploits the fact that PV is getting increasingly affordable, and expanding the installed capacity is able to elevate the actual generation profile above the generation target and, thus, fulfills it at a substantially lower cost as compared to a storage-only solution—PV overbuilding and proactive curtailment is thus a strategy Remund et al.58 term applying implicit storage. This idea is illustrated in Fig. 2, in which the actual daily energy production of a PV cluster (solid line) is plotted against the generation target (dashed line), over the course of a year. One can see that the energy deficit, which has to be fulfilled by storage, as represented by the integral of the blue-shaded areas, is significantly reduced after overbuilding the PV by a factor of 1.5, which, despite an increase in curtailment, translates to a substantial cost saving on storage. The question is therefore how to optimally the sizes of the battery and overbuilt part of the PV. The scientific problem behind the lowest-cost firm forecasting is one of optimization, and the earliest attempt took an iterative approach,62 which was later improved and formalized through mathematical programming.63,64
An illustration of the firm forecasting concept. (Top) The daily energy production of a PV cluster (solid line) strives to satisfy the generation target (dashed line). The energy surplus and deficit are painted with orange and blue, respectively. (Bottom) The situation with a PV overbuilding factor of 1.5, which shows a significant reduction in energy deficits, and thus a much lower requirement for energy storage.
An illustration of the firm forecasting concept. (Top) The daily energy production of a PV cluster (solid line) strives to satisfy the generation target (dashed line). The energy surplus and deficit are painted with orange and blue, respectively. (Bottom) The situation with a PV overbuilding factor of 1.5, which shows a significant reduction in energy deficits, and thus a much lower requirement for energy storage.
Insofar as optimization is concerned, it involves an objective function that needs to be minimized, as well as various constraints narrating what rules have to be respected and what physical limits have to be imposed. In that, firm forecasting seeks to minimize a quantity known as the firm forecasting premium (or the perfect forecast metric, as per the terminology of Perez et al.60) which is the cost multiplier of delivering 1 kWh of firm PV power as compared to sending 1 kWh of unconstrained65 PV power, which can be further written in terms of the levelized cost of electricity. Of course, such a premium is mainly composed of: (1) the cost of battery storage, and (2) the cost to overbuild PV. As for constraints, the definition of firm forecasting implies that the generation target (i.e., the reconciled forecasts) has to be satisfied on a 24/365 basis. In addition to that, other constraints, such as the battery operation constraints or PV operation constraints, also apply. The reader is referred to Yang et al.63 for mathematical details. Studies have shown that, according to the current prices of battery storage and PV, the firm forecasting premium is typically below 2, indicating that less than doubling the current turnkey cost of PV can already firm up the forecasts.59,60 To put it more plainly, suppose one pays for $1000 kWac turnkey for PV, investing a few hundred dollars more per kW can make that PV system dispatchable. To summarize the discussion thus far, Fig. 3 shows the new five-stage solar forecasting workflow. The first four stages have been elaborated. As for the last stage, it commences when the market participants receive the reconciled forecasts, i.e., the generation targets, from the grid operators; the flow chart should be self-explanatory.
(Left) The new five-stage solar forecasting workflow. (Right) The flow chart of the firm forecasting strategy. When actual generation is greater than the generation target, the excess power is either curtailed or stored (the solid path); when actual PV generation is less than the generation target, the deficit amount of power is drawn from the batteries (the dashed path).
(Left) The new five-stage solar forecasting workflow. (Right) The flow chart of the firm forecasting strategy. When actual generation is greater than the generation target, the excess power is either curtailed or stored (the solid path); when actual PV generation is less than the generation target, the deficit amount of power is drawn from the batteries (the dashed path).
Dispatchable PV power is certainly attractive, as it can inherit most if not all virtues of those grid operations that the current grid operators have already become accustomed to, such as unit commitment and economic dispatch. However, there is still a list of unresolved issues limiting the technical readiness of the new solar forecasting workflow. The first issue is in regard to the cost-sharing of the firm forecasting premium, i.e., who should bear what proportion of the premium is the question; this topic saddles on energy policy and sharing economy.66–68 The second question, which follows from the first, is where the battery storage and the overbuilt part of PV should be situated, should they be installed near substations or at individual plants? The answer to this may seem apparent, as geographical smoothing61,69 promotes the centralized installation, such that a cluster of distributed PV can be firm-forecast and therefore dispatched as a whole, but this again would make complex the energy economics, where the individual and communal benefits may conflict. In China, the current policy mandates pairing storage with a capacity that is 10%–30% of the installed PV capacity for each newly commissioned plant, but how the storage should be controlled, and by whom, are still unclear at the moment. The third issue is technical, in that, all investigations on firm forecasting thus far are conducted with data on an hourly scale, whereas the actual battery management, such as the model predictive control70,71 or fault diagnosis and contingency actions,72,73 has often a timescale of a few seconds to a minute; this gap between theory and practice is to be addressed through more refined control simulations and experiments.
CONCLUSION
Whereas there are no doubt many other issues, which I shall not expand on too much here, the most challenging one, by which I myself have been confronted for a long while, is the clash between the goodness of firm forecasts and that of conventional forecasts, which seems insoluble at the moment. Conventional forecasts are said to be good if their superiority can be justified in joint terms of consistency, quality, and value. To that end, much effort has hitherto been devoted, so that the forecasting methods and models can excel in those aspects.74,75 However, as shown by Perez et al.,60 forecasts of higher quality do not necessarily correspond to a lower firm forecasting premium. Instead, it is often the persistence forecasts, which have a very well balance between over- and under-forecast situations, that can achieve an attractive firm forecasting premium. What this implies is that much of the solar forecasting knowledge acquired to date is to be abandoned. Disruptive technologies arrive as “black swan events,” and firm forecasting is the current example. On this point, this perspective article serves as a modest spur to induce the community to come forward with its valuable suggestions and solutions.
ACKNOWLEDGMENTS
This work is supported by the National Natural Science Foundation of China (Project No. 42375192), and China Meteorological Administration Climate Change Special Program (CMA-CCSP; Project No. QBZ202315).
AUTHOR DECLARATIONS
Conflict of Interest
The authors have no conflicts to disclose.
Author Contributions
Dazhi Yang: Conceptualization (lead); Visualization (lead); Writing – original draft (lead).
DATA AVAILABILITY
Data sharing is not applicable to this article as no new data were created or analyzed in this study.