Method and System for Marine Weather Forecasting Using Bivariate Multi-Model Ensemble

Abstract

A marine weather forecasting method utilizing a multi-model ensemble technique according to an exemplary embodiment of the present invention performs a forecast of either sea wind or seawater flow using a probability density function of a bivariate normal distribution, wherein the sea wind and seawater flow are expressed as combination vectors of an east-west component vector u and a north-south component vector v, and the probability density function is a function with u and v as variables, and the correlation coefficient between u and v may be included in the probability density function. Furthermore, the weights of the correlation coefficients are calculated in real time from historical data updated based on the calculation time and may be updated according to the calculation time. Additionally, spatial interpolation reflecting the physical environmental characteristics of the observation area is performed for either sea wind or seawater flow, and a spatial distribution map can be generated by distributing ensemble weights corresponding to the space. Thus, in predicting sea wind or seawater flow, the present invention can not only reflect the speed and direction of vector variables but also mutually compensate for biases between prediction models, thereby improving the prediction reliability and accuracy of weather variables. In addition, by reflecting changes in model performance and environmental conditions over time, it is possible to provide a prediction system capable of stable operation over a long period, and to provide physically consistent and highly reliable spatial prediction information for observation points.

Description

Technology Field

[0001] The present invention relates to a marine weather forecasting technology, and more specifically, to a marine weather forecasting method and system that reduces forecasting errors, improves physical consistency and reliability, and enables stable and accurate marine weather forecasting during long-term operation by mutually complementing the uncertainty of a numerical model regarding a marine weather variable having a vector component to be predicted through a multi-model ensemble and reflecting ensemble weights that are dynamically updated over time. Background Technology

[0002] Ocean weather is a nonlinear phenomenon that changes complexly in space and time due to the interaction between the atmosphere and the ocean, and predictions regarding it inherently involve high uncertainty.

[0003] Conventional marine weather forecasting has been performed based on Numerical Weather Prediction (NWP) or Forecast Models, which is a method of predicting future conditions by establishing initial conditions using observational data and numerically interpreting physical equations.

[0004] However, such deterministic single prediction models have limitations in the reliability of predictions due to factors such as the fact that results can vary significantly depending on minute differences in initial conditions, the incompleteness of observational data, instrumentation and representativeness errors, and the accumulation of errors resulting from numerical approximations within the model. As suggested in the study by Lorenz (1963), the chaotic nature of atmospheric and oceanic flows makes long-term forecasting inherently difficult, and this uncertainty can further intensify as the forecast period lengthens.

[0005] Furthermore, because numerical weather prediction models approximate various physical processes, parameter uncertainty resulting from simplification during the parameterization process, computational errors when setting boundary conditions, and structural uncertainty stemming from subjective choices during model design are inevitably present. These complex factors degrade the accuracy of predictions for specific regions, seasons, or extreme marine weather phenomena. In particular, single models tend to exhibit systematic bias toward specific phenomena due to their unique representation of physical processes and numerical solutions; this leads to problems where prediction results differ from actual observations or variability is underestimated.

[0006] To address the limitations of this single-model approach, the Single-Model Ensemble technique was introduced to statistically quantify prediction uncertainty. This method utilizes the same numerical weather forecasting model but generates multiple prediction results by slightly varying initial conditions or parameters, and then statistically combines them to evaluate uncertainty. However, since the Single-Model Ensemble technique also fundamentally shares the same model structure and physical processes, it is difficult to overcome the systematic bias or physical limitations inherent in the model. In other words, while it can capture prediction variance based on initial conditions, structural uncertainty or bias stemming from the model structure itself is still reflected. Consequently, while the Single-Model Ensemble can quantify uncertainty to some extent, it has limitations in overcoming the inherent limitations of the model or fundamentally improving prediction reliability.

[0007] Therefore, in order to substantially reduce the uncertainty of marine weather forecasting and ensure the reliability of predictions, the introduction of new types of forecasting technology is required to complement the structural limitations of existing deterministic numerical weather prediction models and single-model ensemble techniques. This new approach must be able to minimize physical uncertainty and improve the ability to forecast extreme marine weather phenomena by utilizing the complementary characteristics of various models. The problem to be solved

[0008] To address this, one aspect of the present invention can provide a technology that can improve the prediction accuracy of weather variables by reflecting the correlation of magnitude and direction in the prediction of vector-based weather variables such as sea wind and seawater flow, while mutually complementing the uncertainties of different prediction models through a multi-model ensemble.

[0009] One aspect of the present invention can provide a marine weather forecasting technology capable of stable operation over a long period by considering the predicted change of weather variables over time.

[0010] One aspect of the present invention can provide a marine weather spatial prediction technology that utilizes a more continuous and reliable spatial distribution based on a spatial structure that reflects actual topography and water depth. means of solving the problem

[0011] A marine weather forecasting method based on a bivariate multi-model ensemble according to an embodiment of the present invention, devised to solve the above problem, performs a forecast of either sea surface wind or seawater flow using a probability density function of a bivariate normal distribution, wherein the sea surface wind and the seawater flow are expressed as combination vectors of an east-west component vector u and a north-south component vector v, and the probability density function is a function with u and v as variables, and the correlation coefficient between u and v may be included in the probability density function. At this time, the probability density function is expressed by Equation 1,

[0012] (Mathematical Formula 1)

[0013]

[0014] In the above mathematical formula 1, the above ρ uv is the correlation coefficient between the above u and the above v, and the above ρ uv can be expressed by mathematical equation 2, and

[0015] (Mathematical Formula 2)

[0016] The above r is the amplitude of the correlation coefficient, the above s is the offset, the above k is the number of periods, and the above φ is the phase, and the above r, the above s, the above k, and the above φ are the correlation coefficient ρ uv It is the weight of, and the above μ u and the above μ u is the bias-corrected ensemble average of the above u and the above v, respectively, and the μ u and the above μ u can be derived by mathematical formula 3, and

[0017] (Mathematical Formula 3)

[0018] The above is the u mean of the ensemble members, and the above is the v-average of the ensemble members, and the above a u , above b u , above a v and the above b v is a bias correction weight, and the above σ 2 u and the above σ 2 u are the corrected ensemble variances of the above u and the above v, respectively, and the above σ 2 u and the above σ 2 v can be derived by mathematical formula 4, and

[0019] (Mathematical Formula 4)

[0020] The above s2 u and the above s 2 v are the ensemble variances of the above u and the above v, respectively, and the above c u , above d u , above c v and the above d v is the variance-corrected weight.

[0021] The above correlation coefficient ρ uv The weights are calculated in real time from historical data updated based on the calculation time and are updated according to the calculation time, the bias correction weights are calculated in real time from historical data updated based on the calculation time and are updated according to the calculation time, and the variance correction weights are calculated in real time from historical data updated based on the calculation time and can be updated according to the calculation time.

[0022] In addition, a bivariate multi-model ensemble-based ocean weather forecasting method according to one embodiment of the present invention calculates ensemble weights for sea wind using relevant data from the past 30 days, and said sea wind-related data is collected at 12h intervals, but can be collected for a minimum of 3 days or more and a maximum of 30 days or less. In addition, ensemble weights for seawater flow are calculated using relevant data from the past 3 days, and said seawater flow-related data is collected at 12h intervals, but can be collected for a minimum of 1 day or more and a maximum of 3 days or less.

[0023] In addition, a bivariate multi-model ensemble-based ocean weather forecasting method according to one embodiment of the present invention performs spatial interpolation reflecting the physical environmental characteristics of the observation area for either sea wind or seawater flow, wherein the spatial interpolation may utilize an n-dimensional variational interpolation (DIVAnd: Data-Interpolating Variational Analysis in n dimension) technique. At this time, the n-dimensional variational interpolation technique utilizes Equation 5,

[0024] (Mathematical Formula 5)

[0025]

[0026] J(φ) is the cost function to be minimized, the said φ is the spatial field to be estimated, and the said Y i is the i-th observation, the above H i (φ) is the predicted value of the i-th observation position in the spatial field φ, the above R i is the error variance of the observation error, and the above φ b is the background field, and the above B is the background field covariance matrix It can represent spatial correlation structures.

[0027] Here, the physical environment characteristics may include physical constraints, boundary conditions, elevation, and water depth.

[0028] In addition, a bivariate multi-model ensemble-based ocean weather prediction method according to one embodiment of the present invention performs spatial prediction for an observation area of ​​either sea wind or seawater flow, wherein the space is a 3-dimensional space interpolated using an n-dimensional variational interpolation technique, and can generate a spatial distribution map by distributing ensemble weights corresponding to the space. At this time, the n-dimensional variational interpolation technique may utilize the aforementioned Equation 5.

[0030] Meanwhile, regarding a prediction system for either sea wind or seawater flow, the bivariate multi-model ensemble-based ocean weather prediction system comprises: a data collection unit that collects observational data and output data based on a prediction model; an ensemble weight calculation unit that applies at least a portion of the collected data to a prediction model based on a bivariate normal distribution probability density function and an ensemble weight calculation model, and calculates ensemble weights in real time from historical data updated based on the weight calculation time, wherein the ensemble weights include common weights for all observation points and individual weights for each observation point; an ensemble weight spatial distribution map generation unit that generates an ensemble weight spatial distribution map through spatial interpolation using a 3D spatial interpolation method based on the common weights and individual weights; and an ensemble prediction unit that calculates ensemble weights corresponding to unobserved sections using the ensemble weight spatial distribution map and calculates ensemble prediction results for the observation points and the unobserved sections. and a spatial prediction unit that performs spatial prediction for the observation point and the unobserved section using the generated ensemble weight spatial distribution map and calculates final spatial prediction data based thereon; wherein the sea wind and the seawater flow can be processed as the sum of the east-west component vector u and the north-south component vector v. Effects of the invention

[0031] The present invention, in predicting meteorological variables such as sea wind and seawater flow, expresses them using an east-west component vector (u) and a north-south component vector (v) while reflecting their correlations, and utilizes a bivariate multi-model ensemble technique that combines results calculated from multiple prediction models, thereby not only reflecting correlations in magnitude and direction but also mutually compensating for biases between prediction models, which can improve the reliability and accuracy of the prediction of meteorological variables.

[0032] In addition, by calculating weights for bivariate multi-model ensemble prediction in real time, changes in model performance and environmental conditions over time can be reflected, and the quality of the overall prediction can be prevented due to the performance degradation of a specific model, allowing for stable operation over a long period.

[0033] In addition, for the spatial prediction of sea wind and seawater flow, by utilizing an n-dimensional variational interpolation technique that applies elevation and water depth to spatially interpolate the gaps between observation points using elevation or water depth, and by generating a weighted spatial distribution map and performing spatial prediction, physically consistent and highly reliable spatial prediction information for observation points can be provided.

[0035] The technology proposed by the present invention Offshore wind power industry, shipping and shipbuilding industry, and offshore plant and coastal infrastructure industry It is applicable to various fields, and in particular, it possesses high industrial value as it is utilized in power generation forecasting, ensuring ship operation safety, calculating structural design loads, and marine disaster response systems.

[0036] 1. Offshore Wind Power and Renewable Energy Industry

[0037] The offshore wind power industry is an industry sector where the accuracy of power generation forecasting is directly linked to profitability, and the bivariate multi-model ensemble-based offshore wind forecasting technology of the present invention can be utilized as follows.

[0038] · During the turbine installation location selection stage Long-term wind resource assessment to use

[0039] · Through real-time and short-term forecasts Improvement in power generation prediction accuracy

[0040] · Extreme wind speeds, gusts, sudden changes in wind direction, etc. Predicting structural loads and ensuring operational stability

[0041] · Used for optimizing generator maintenance schedules and determining emergency stop criteria

[0042] In particular, the vector-based (u, v variables) prediction structure of the present invention is not based on simple wind speed. Precisely reflects even wind direction variabilityIt is possible, and has high practical utility compared to existing prediction systems.

[0044] 2. Shipping, Shipbuilding, and Smart Ship Operation Industry

[0045] The shipping industry is an industry where operational safety and fuel efficiency are key competitive advantages, and the present invention can be directly applied to the following fields.

[0046] · Route design incorporating headwind / tailwind conditions in conjunction with an automated ship route optimization system, and establishment of fuel-efficient navigation strategies considering ocean current directions

[0047] · For avoiding dangerous waters Path change based on high-precision wave and sea wind predictions

[0048] · Applied to weather judgment algorithms for autonomous ships

[0049] In particular, using a multi-model ensemble Prediction uncertainty quantification function It has high potential for direct application in ship operation risk management systems.

[0051] 3. Offshore Plant, Coastal Infrastructure, and Disaster Response Industries

[0052] In the field of offshore plants and coastal structures, predicting environmental loads directly related to structural safety is essential. The present invention can be utilized as follows.

[0053] · Offshore platforms, floating structures, breakwaters, etc. Calculation of design standard wind speed and flow velocity data and establishment of a real-time environmental load monitoring system

[0054] · Targeting coastal cities Prediction of storm surges, abnormal waves, and sudden changes in current velocity , development of disaster risk assessment and early warning systems

[0055] · Highly accurate prediction of the location and coordinates of drifters / floating objects using sea wind and ocean current prediction data in the event of a marine accident, and Development of a rapid search and rescue system

[0056] · For the protection of submarine cables and submarine pipeline facilities Long-term prediction of ocean flow and stability assessment

[0057] In particular, included in the present invention 3D spatial interpolation and weighted spatial distribution techniquesIt has the advantage of being able to provide reliable spatial prediction information even in sea areas where observation stations are scarce, and can be applied to various factors with movement, speed, and direction, such as weather forecasting or hydrology, in addition to the field of marine meteorology.

[0059] In addition, the marketability of the present invention can be highly evaluated in the following respects.

[0060] 1. As maritime logistics, offshore wind power, and seabed resource development expand globally Demand for precise marine weather forecasting technology It is continuously increasing. In particular, vector-based predictions such as wind speed and current are directly linked to offshore wind power generation forecasting, ship fuel optimization, and structural safety assessment.

[0061] 2. Existing commercial prediction services often rely primarily on a single prediction model or simple ensemble statistics, such as the present invention Integrated bivariate-based, multi-model-based, and spatial prediction technology is positioned as a differentiated high-value-added service. possible.

[0062] 3. The technology in question Targeting both B2G and B2B markets simultaneously, including domestic and international meteorological agencies, oceanographic research institutions, port authorities, shipping companies, power generation companies, and defense contractors. You can formulate a stable market entry strategy.

[0063] As described above, the technology to which the present invention applies is directly applicable to various industrial fields and possesses high industrial applicability that aligns with the increasing trend of industrial demand for precise forecasting. Furthermore, by improving forecasting accuracy and stability compared to existing single-model-based forecasting technologies, it is possible to secure a competitive business advantage in the high-value-added forecasting service market. Brief explanation of the drawing

[0064] FIG. 1 is a schematic diagram for schematically explaining a bivariate multi-model ensemble-based ocean weather forecasting method according to one embodiment of the present invention, and FIG. 2 is a flowchart for schematically explaining a bivariate multi-model ensemble-based ocean weather forecasting method according to an embodiment of the present invention, and FIG. 3 is a diagram showing the distribution of sea wind (and / or seawater flow) by classifying it into sectors according to its orientation direction in one embodiment of the present invention, and displaying the sea wind (and / or seawater flow) in a u,v coordinate system. FIG. 4 is a diagram showing the correlation coefficients for major sectors of offshore wind according to an embodiment of the present invention, and FIG. 5 is a diagram illustrating a dynamic ensemble weight calculation process according to an embodiment of the present invention, and FIG. 6 is a diagram illustrating a predicted scatter plot using static and dynamic ensemble techniques according to an embodiment of the present invention, and FIG. 7 is a diagram illustrating a predicted time series using static and dynamic ensemble techniques according to an embodiment of the present invention, and FIG. 8 is a diagram illustrating ensemble prediction results for an observation point according to an embodiment of the present invention, and FIG. 9 is a diagram illustrating a weighted spatial distribution map using a conventional Kriging spatial interpolation technique, and FIG. 10 is a diagram illustrating a weighted spatial distribution map using the DIVAnd (Data-Interpolating Variational Analysis in n dimension) technique according to an embodiment of the present invention, and FIG. 11 is a diagram illustrating a spatial prediction result using DIVAnd according to an embodiment of the present invention, and FIG. 12 is a block diagram schematically illustrating a bivariate multi-model ensemble-based ocean weather forecasting system according to one embodiment of the present invention. Specific details for implementing the invention

[0065] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but can be implemented in various different forms. These embodiments may be provided to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention. Throughout the entire specification, the same reference numerals refer to the same components.

[0066] The terms used herein are for describing the embodiments and are not intended to limit the invention. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used herein, 'comprise' and / or 'comprising' do not exclude the presence or addition of one or more other components, steps, actions, and / or elements to the mentioned components, steps, actions, and / or elements.

[0067] Unless otherwise defined, all terms used herein have the same meaning as generally understood by those skilled in the art to which this invention pertains. Terms defined in commonly used dictionaries should be interpreted as having meanings consistent with the context of the relevant technology and should not be interpreted as having an ideal or overly formal meaning unless explicitly defined in this application.

[0069] The structure, operating principle, and effects of the present invention will be explained in more detail below with reference to the attached drawings.

[0071] FIG. 1 is a schematic diagram for schematically explaining a marine weather forecasting method based on a bivariate multi-model ensemble according to an embodiment of the present invention; FIG. 2 is a flowchart for schematically explaining a marine weather forecasting method based on a bivariate multi-model ensemble according to an embodiment of the present invention; FIG. 3 is a diagram showing the distribution of sea wind (and / or seawater flow) by classifying it by sector according to its orientation in an embodiment of the present invention and displaying the sea wind (and / or seawater flow) in a u,v coordinate system; FIG. 4 is a diagram showing the correlation coefficients of sea wind by major sectors according to an embodiment of the present invention; FIG. 5 is a diagram illustrating the process of calculating dynamic ensemble weights according to an embodiment of the present invention; FIG. 6 is a diagram illustrating a forecast scatter plot using static and dynamic ensemble techniques according to an embodiment of the present invention; FIG. 7 is a diagram illustrating a forecast time series using static and dynamic ensemble techniques according to an embodiment of the present invention; and FIG. 8 is an ensemble for an observation point according to an embodiment of the present invention Figure 9 is a diagram illustrating a prediction result, Figure 10 is a diagram illustrating a weighted spatial distribution using a conventional Kriging spatial interpolation technique, Figure 11 is a diagram illustrating a spatial prediction result using DIVAnd (Data-Interpolating Variational Analysis in n dimension) according to an embodiment of the present invention, Figure 11 is a diagram illustrating a spatial prediction result using DIVAnd according to an embodiment of the present invention, and Figure 12 is a block diagram for schematically explaining a bivariate multi-model ensemble-based marine weather prediction system according to an embodiment of the present invention.

[0073] Referring to FIGS. 1 and 2, a marine weather forecasting method based on a bivariate multi-model ensemble according to one embodiment of the present invention may include: a step of collecting observation data (S100); a step of calculating bivariate multi-model ensemble weights (S200); a step of ensemble forecasting for observation points (S300); a step of generating an ensemble weight spatial distribution map using a 3D spatial interpolation method (S400); and a step of analyzing spatial forecasting performance (S500).

[0074] A bivariate multi-model ensemble-based ocean weather prediction method according to one embodiment of the present invention can be performed by predicting either sea wind or seawater flow using a probability density function of a bivariate normal distribution, wherein sea wind and seawater flow are expressed as combination vectors of an east-west component vector u and a north-south component vector v, and the probability density function is a function with u and v as variables, and the correlation coefficient between u and v is included in the probability density function.

[0075] In one embodiment, the probability density function can be expressed by Equation 1.

[0076]

[0078] In mathematical equation 1, ρ uv is the correlation coefficient between u and the above v, and can be expressed by mathematical formula 2.

[0079]

[0080] In Equation 2, r is the amplitude of the correlation coefficient, s is the offset, k is the number of periods, and φ is the phase, and r, s, k, and φ are the correlation coefficients ρ uv It is the weight of.

[0082] μ in mathematical formula 2 u and μ u is the bias-corrected ensemble average of u and v, respectively, and can be derived by Equation 3.

[0083]

[0084] In mathematical formula 3 is the u mean of the ensemble members, and is the average of the ensemble members' v, and a u , b u , a v , b v is the bias correction weight.

[0086] σ in mathematical equation 1 2 u and σ 2 u is the corrected ensemble variance of u and v, respectively, and can be derived by Equation 4.

[0087]

[0088] In mathematical equation 4, s 2 u and s 2 v are the ensemble variances of u and v, respectively, and c u , d u , c v , d v is the variance-corrected weight.

[0090] Observation data collection stage (S100)

[0091] In one embodiment, the observation data collection step (S100) may include a process of collecting marine weather information, such as waves, wind, and water temperature, from marine weather observation equipment installed at an observation point within a marine area to be observed (hereinafter referred to as the observation area). At this time, the marine weather observation equipment may include a marine weather buoy under the jurisdiction of the Korea Meteorological Administration located above the sea surface to provide weather forecasts and marine information services based on real-time observation data, and if a marine weather buoy is not located at the observation point, a marine observation buoy of the Korea Hydrographic and Oceanographic Agency may be additionally used. Weather information, i.e., observation data, at the observation point can be collected through such equipment.

[0092] In one embodiment, marine meteorological observation equipment that has missing data for more than 3 months or has a missing data rate of 40 to 70% during the observation data collection period may be excluded from the observation point.

[0093] In one embodiment, observation data collection may be performed twice daily at 00 UTC and 12 UTC.

[0094] In one embodiment, the observation data may include data on vector-based meteorological variables such as sea wind and ocean current. Since sea wind and ocean current are vector variables that possess both magnitude and direction, the observation data needs to be distinguished by magnitude and direction. For example, in the case of "wind speed (or current velocity) 15 m / s, wind direction (or current direction) 315° northwest," the sea wind means that a wind of 15 m / s occurred in the northwest to southeast direction, and the ocean current means that ocean current occurred in the northwest direction. However, in the case of direction, since the value circulates in a circular pattern after 359°, there are difficulties in mathematically expressing the magnitude or performing calculations.

[0095] In one embodiment of the present invention, vector-based meteorological variables such as sea wind and seawater flow are represented by wind speed (or flow velocity) and wind direction (or flow direction) as an east-west component vector (u) and a north-south component vector (v). Accordingly, mathematical and physical operations using u and v can be performed, and sea wind and seawater flow can be predicted more accurately.

[0096] In the situation given above as an example, "wind speed (or current speed) 15 m / s, wind direction (or current direction) 315° northwest," if the following formula is used, u can be converted to 10.6 m / s and v to -10.6 m / s in the case of seawater flow, unlike seawater wind, because the directional component is to, u can be converted to -10.6 m / s and v to 10.6 m / s.

[0097] Speed ​​= , Direction =

[0099] Therefore, sea surface wind and ocean current are expressed as vector variables, simultaneously representing both magnitude and direction, while the collected observational data can be stored in a converted form of u and v. This enables bivariate prediction that considers both magnitude and direction. Furthermore, compared to conventional techniques that consider only a single variable, prediction performance can be further improved.

[0101] Bivariate multi-model ensemble weight calculation step (S200)

[0102] In one embodiment, the bivariate multi-model ensemble weight calculation step (S200) may perform a process of generating prediction data by inputting collected observation data into a plurality of prediction models, and calculating bivariate multi-model ensemble weights for an observation point using observation data within a set specific past period (e.g., 3, 5, 7 days, etc.) and the generated prediction data. To perform this, the bivariate multi-model ensemble weight calculation step (S200) may include a multi-model ensemble member selection step (S210), a multi-model ensemble combination step (S220), and an ensemble weight calculation step (S230).

[0104] Multi-model ensemble member selection step (S210)

[0105] In one embodiment, the multi-model ensemble member selection step (S210) may be performed by selecting multi-model ensemble candidates to ensemble combine different prediction models when predicting sea wind and seawater flow, and may include a process of selecting ensemble members by analyzing the prediction performance of a plurality of candidate prediction models.

[0106] In one embodiment, the candidate prediction model may be a physics-based prediction model or an artificial intelligence-based prediction model. In this case, at least one of GDAPS(UM), GDAPS(KIM), GFS(NCEP), and ECMWF(HRES) may be selected as the physics-based prediction model, and at least one of FourCast-Net, Pangu-Weather, and GraphCast may be selected as the artificial intelligence-based prediction model, and such prediction models may be used for sea wind forecasting.

[0107] In one embodiment, since the area and grid resolution of the model are unified based on the model with the smallest area among the grid areas of the input numerical models, it can be selected among the numerical prediction models that perform the prediction, and such a numerical prediction model can be used for seawater flow prediction.

[0108] In one embodiment, error indicators such as RMSE (Root Mean Square Error) and BIAS (Bias) may be used as criteria for selecting members of a multi-model ensemble. Through such selection, the prediction error of the ensemble model can be reduced.

[0110] Multi-model ensemble combination step (S220)

[0111] In one embodiment, the multi-model ensemble combining step (S220) may be performed by combining prediction models selected as members of the multi-model ensemble to compensate for the bias of a single prediction model. In one embodiment, the multi-model ensemble combining step (S232) is based on a super-ensemble technique. The initial super-ensemble technique first proposed by Krishnamurti et al. (1999) was a method of assigning equal weights to all models; however, in one embodiment of the present invention, a Bivariate EMOS (BiEMOS) technique, which is an extension of Ensemble Model Output Statistics (EMOS), may be applied to assign weights that consider the performance and uncertainty of each prediction model. More specifically, probability distribution parameters may be estimated for each prediction model selected as a member of the multi-model ensemble, and the prediction models may be combined using these parameters. By performing a multi-model ensemble through this method, the overall error of the prediction can be reduced and the prediction results improved. In addition, the complementary characteristics of the prediction models can be reflected in the prediction results, and the reliability of the prediction results can be ensured.

[0113] Ensemble weight calculation step (S230)

[0114] In one embodiment, the ensemble weight calculation step (S231) may include an ensemble weight learning step (S231) and a dynamic ensemble weight calculation step (S232). Additionally, the ensemble weight calculation step (S231) may include a process of performing multi-model ensemble prediction using an RMSE metric within a set learning period using a Bivariate Ensemble Model Output Statistics (BiEMOS), selecting an optimal weight learning period, and then calculating ensemble weights. The ensemble weight according to one embodiment of the present invention is an average corrected weight (coefficient; a u , b u , a v, b v ), variance-corrected weights (coefficients; c u , d u , c v , d v It may mean the weights (r, s, k, φ) in the correlation coefficient, etc., and can be calculated through the aforementioned mathematical formulas 2 to 4, etc.

[0115] 1) Ensemble weight learning step (S231)

[0116] In one embodiment, the ensemble weight learning step (S231) can be performed by using the prediction results (including u and v) of each of the numerical models to derive a probabilistic prediction in the form of a bivariate normal distribution through BiEMOS.

[0117] (1) Weights in the correlation coefficient

[0118] The ensemble weight learning step may include a step of modeling the correlation coefficient between u and v as a trigonometric function in the direction of the ensemble average. For example, since u and v of the sea wind have the characteristic of changing systematically depending on the wind direction, the correlation coefficient can be calculated using trigonometric function-based parameters to model this direction-dependent correlation structure. In one embodiment, the correlation coefficient ρ between u and v can be expressed by the aforementioned Equation 2.

[0119] In mathematical equation 2, θ is the ensemble mean direction, r class s is the amplitude and offset of the correlation coefficient, k is the period number, φε is the phase, and kθ can be the period adjustment term of θ. Here, r, s, k, and φ represent the ensemble weights in the correlation coefficients, and these parameters can be estimated using a weighted nonlinear least squares method based on past observations under the condition |r| + |s| ≤ 1. In this way, by reflecting the correlation coefficient between u and v in the ensemble prediction, correlation can be secured not only in magnitude (e.g., wind speed, flow velocity) but also in direction (wind direction, flow direction).

[0121] To support the above, the observed distribution of sea wind (and / or seawater flow) according to the experimental example of the present invention was analyzed. Referring to FIG. 3, each sector represents the average wind direction (and / or ocean current direction) of the ensemble members, and the scatter plot within the sector diagram represents the distribution of the observed data. Sectors are classified into Sectors 2 through 9 according to the direction of the sea wind (and / or seawater flow). For example, regarding the sea wind, if the observed sea wind is a northerly wind, it is indicated in Sector 6, and if it is an easterly wind, it is indicated in Sector 8. Meanwhile, if the wind speed is less than a predetermined standard, it is indicated in Sector 1.

[0122] As such, Figure 3 allows us to identify the distribution characteristics of u and v of the observation data by sector. Furthermore, meteorological variables such as sea wind (and / or ocean current) exhibit direction-dependent characteristics, as the trends in the distribution and correlation of u and v vary by sector, i.e., by direction. Through this, it can be seen that a vector-based bivariate (u, v) probability model is required for the prediction of sea wind.

[0123] Referring to Figure 4, the bold dots represent the empirical correlation coefficients calculated from the central wind direction for each sector, and the dotted lines represent the optimized results obtained by adjusting the correlation coefficient values ​​using Equation 2. As illustrated in Figure 4, the bold dots show a pattern of repetitively cyclic positive (+) and negative (-) correlations depending on the wind direction, while the dotted lines approximate the bold dots and exhibit a periodic sine curve shape. Therefore, it can be confirmed that defining the correlation coefficients of u and v using Equation 2 is empirically valid. Through this, it can be understood that the correlation coefficient between the bivariates (u, v) of the sea wind is not a single fixed correlation coefficient, but a correlation coefficient that changes continuously according to changes in wind direction. Thus, it can be seen that incorporating a wind direction-dependent dynamic correlation model can further improve the accuracy of wind direction prediction. Meanwhile, the numbers next to the dots in Figure 4 (numerical values ​​such as 94, 52, and 74) represent the number of sample data points.

[0124] (2) Average Correction Weight

[0125] The ensemble weight learning step (S231) may include a step of performing bias correction on the ensemble mean. The ensemble mean may have seasonal or regional bias. According to one embodiment of the present invention, to correct such bias, linear regression-based parameters may be used, and bias correction may be performed through the estimation of these parameters. In one embodiment, the bias-corrected ensemble mean of u and v μ u and μ u It can be expressed by the aforementioned mathematical formula 3.

[0126] In mathematical formula 3, , is the mean of u and v of the ensemble members, and a u , b u , a v , b v represents the ensemble weight as a bias correction parameter. Here, parameter a u , b u , a v , b v can be estimated using the linear least squares method between the observed and predicted data produced during the ensemble weight training period. μ u and μ u is the bias correction of the raw ensemble mean, representing the center point of the final prediction distribution, and the probability density function f(u, v) is (μ u , μ u It can have a maximum value in ).

[0127] (3) Variance-corrected weights

[0128] The ensemble weight learning step (S231) may include a step of performing correction for ensemble variance. In the case of ensemble variance, since predictions using conventional ensemble techniques have a variance smaller than the actual uncertainty, ensemble variance correction can be performed by calculating a variance correction parameter to solve this underdispersion problem. In one embodiment, the corrected ensemble variance of u and v σ 2 u and σ 2 v It can be expressed by the aforementioned mathematical formula 4.

[0129] In mathematical equation 4, s 2 u, s 2 v is the u, v variance of the ensemble members, and c u , d u , c v , d v represents the ensemble weight as a variance correction parameter. Here, the parameter c u , d u , c v , d vcan be determined by the maximum likelihood estimation method under non-negative constraints, and the contrapositive function (e.g. It can be calculated using ).

[0130] In this way, unlike conventional single EMOS techniques, the bivariate prediction uncertainty can be quantified by utilizing correlation coefficients between bivariates and correction parameters, namely ensemble weights, and thereby the problem of ensemble mean bias and variance deficiency can be resolved, which can improve the accuracy of ensemble prediction.

[0131] 2) Dynamic ensemble weight calculation step (S232)

[0132] In one embodiment, the dynamic ensemble weight calculation step (S232) may proceed in a manner that performs the process of calculating dynamic ensemble weights to reflect ocean weather changes and short-term variability of observation data.

[0133] Weather forecasting must be able to respond adaptively to seasonal changes, changes in weather patterns, or rapidly changing marine weather conditions. Conventionally, fixed ensemble weights have been calculated and used based on data from past periods; however, such fixed weights not only make it difficult to respond to rapid weather changes caused by recent climate change but also present the problem of increasing prediction errors over time. Therefore, in calculating ensemble weights according to one embodiment of the present invention, a dynamic ensemble technique may be used instead of the static ensemble technique that uses conventional fixed weights. By using a dynamic ensemble technique, weights can be calculated in real time from the most recent historical data based on the time of weight calculation. That is, depending on the time of calculation, the historical data used for weight calculation is updated, and the weights can also be updated.

[0134] In one embodiment, as shown in FIG. 5, dynamic ensemble weights can be calculated by selecting a specific period based on the time of ensemble weight calculation and utilizing the observation and prediction results of that period.

[0135] In the experimental example of the present invention, the ensemble weights for sea wind and seawater flow were calculated based on the prediction data of a multi-model ensemble.

[0136] In this case, for sea wind, since only the prediction result corresponding to a single prediction time is utilized for the same reference time, a certain amount of training data is required to derive statistically significant weights. Accordingly, the minimum training period required for calculating ensemble weights was set to approximately 3 days. Meanwhile, if the training period exceeds 30 days, excessive computational time is required to calculate weights corresponding to each space in the spatial prediction step described later; therefore, it is desirable to limit the maximum training period to within 30 days.

[0137] Meanwhile, since the prediction time for ocean current is fixed at 24 hours, the entire time interval included in a single prediction result can be integrated and utilized as training data, rather than calculating weights separately for each prediction time period. Accordingly, ensemble weights for ocean current can be calculated using only one-day data. On the other hand, unlike sea wind, ocean current prediction models exhibit different time-series trends depending on the model; therefore, if the training period is excessively increased, the characteristics between models may cancel each other out, causing the weights to converge to zero. Accordingly, it is desirable to limit the maximum training period for calculating ensemble weights for ocean current to three days.

[0138] In this way, by applying dynamic ensemble weights that are updated in real time using historical data from the calculation period to the prediction of weather variables, changes in the performance of each prediction model and changes in environmental conditions over time can be effectively reflected in the multi-model ensemble. Through this, the impact of performance degradation of a specific model on the overall prediction results can be minimized, thereby maintaining stable prediction performance over a long period.

[0140] Ensemble prediction step for observation points (S300)

[0141] In one embodiment, the prediction of a weather variable using a bivariate multi-model ensemble can be achieved through a process of performing a final probabilistic prediction using a probability density function of a bivariate normal distribution using parameters, i.e., ensemble weights, corrected through at least one of Equations 2 to 4, and then calculating an ensemble prediction result from this probability density function.

[0142] In addition, ensemble prediction for observation points can be performed through this method. In one embodiment, referring to FIG. 8, the ensemble prediction step (S300) for observation points can perform ensemble prediction by calculating dynamic ensemble weights according to the time of collection of observation data for each observation point and reflecting the results in a bivariate multi-model ensemble.

[0143] To perform this process, the ensemble prediction step (S300) for the observation point may include an ensemble prediction step (S310) using a probability density function and an ensemble prediction step (S320) using a dynamic ensemble technique.

[0145] 1) Ensemble prediction step using probability density function (S310)

[0146] First, the ensemble prediction step using a probability density function can perform probabilistic prediction of meteorological variables using a bivariate multi-model ensemble by using the probability density function f(u,v) expressed by the aforementioned mathematical equation 1 related to the bivariate normal distribution of u and v.

[0147] In mathematical formula 1, ρ 2 uv is (ρ uv ) 2 It refers to the square of the correlation coefficient between u and v. Correlation coefficient ρ uv -1≤ρ uv A value within the range ≤1, and its squared value ρ 2 uv 0≤ρ 2 uv It has a value within the range ≤1. Such ρ uv and ρ2 uv The strength of the correlation between u and v can be quantitatively determined through the magnitude of .

[0148] For example, ρ uv When = 1, ρ 2 uv Furthermore, it becomes 1, meaning that a very strong positive correlation exists between u and v, and conversely, ρ uv When = 0, ρ 2 uv It becomes 0, indicating that there is no correlation between the two variables.

[0149] Meanwhile, ρ uv When = 0.7, ρ 2 uv = 0.49, meaning that a relatively strong positive correlation exists between u and v, whereas ρ uv Even in cases with negative correlation coefficient values ​​such as = -0.7, ρ 2 uv It becomes 0.49, which means the same level of correlation in terms of the strength of correlation.

[0150] In this way, ρ 2 uv It is used as an indicator to quantify the strength of correlation, and ρ uv The sign of can be used to determine the direction of correlation, that is, the positive (+) or negative (-) direction.

[0151] In mathematical equation 1, ρ 2 uv is within the normalization constant term (1-ρ 2 uv It is located in the ) part, and ρ 2 uv As the value of increases, the correlation strength between u and v increases, whereas (1-ρ 2 uv Since ) decreases, the normalization constant increases relatively, and accordingly, it exhibits the effect of increasing the overall amplitude of the probability density function.

[0152] ρ 2 uv In the exponential term, -1 / 2(1-ρ 2 uv It is located in the form of ), and ρ 2 uv As the value of increases (1-ρ 2 uv Since ) decreases, -1 / 2(1-ρ 2 uv The absolute value of ) increases, -1 / 2(1-ρ 2 uv Since ) always has a negative (-) value, the exp(·) function decreases more rapidly and converges to 0. As a result, the probability density becomes more concentrated around the center (mean).

[0154] The ensemble prediction result can be calculated from the probability density function f(u,v) using the following mathematical formula.

[0155] ,

[0156] Here, P is a probability distribution predicted from the probability density function f(u, v), and X is a probability vector following the probability distribution P, and is a 2D real vector to be optimized, and is the Euclidean distance. Here, bmedP is the point that minimizes the average distance to all points belonging to the probability distribution P, corresponding to the median in a bivariate space, and this is the final ensemble prediction value It is used as.

[0157] In one embodiment, in the case of a bivariate normal distribution which is an elliptical symmetric distribution, the spatial median coincides with the mean vector, so the final predicted value bmedP can be calculated using the bias-corrected ensemble mean of u and v.

[0159] As such, by utilizing the probability density function of the bivariate normal distribution for the prediction of meteorological variables, it is possible to make predictions that reflect u and v of the meteorological variables, and by mutually compensating for biases between prediction models, the reliability and accuracy of the prediction can be improved.

[0161] 2) Ensemble prediction step using dynamic ensemble technique (S320)

[0162] In the ensemble prediction step (S320) using a dynamic ensemble technique, for ensemble prediction of observation points, dynamic ensemble weights are calculated according to the time of collection of observation data for each observation point, and the results are reflected in Equation 1, Equation 5, etc. to perform ensemble prediction.

[0163] In one embodiment, taking the prediction of seawater flow as an example, the prediction results using a static ensemble technique (hereinafter static prediction) and a dynamic ensemble technique (hereinafter dynamic prediction) can be compared through a scatter plot. As shown in Fig. 6, when examining the static prediction results, a problem occurred in which the predicted values ​​were concentrated near zero as prediction models with different time series fluctuation characteristics were combined with equal weights. This problem may occur because the ensemble weights are calculated using long-term (e.g., one year) data, and the weights are learned only in a direction that reduces error without considering the real-time variability of seawater flow. On the other hand, when examining the dynamic prediction results, unlike the static prediction results, it can be seen that the prediction results tend to correspond one-to-one with the observed data.

[0164] In one embodiment, with reference to FIG. 7, the time series results of static prediction and dynamic prediction in seawater flow prediction can be compared. As seen in FIG. 6, the time series prediction results also show that the dynamic prediction results effectively reproduce the changes in flow velocity of the observation data.

[0166] As described above, when predicting marine weather, by ensembling bivariate multi-models of u and v and applying dynamic ensemble techniques to update weights, it is possible to make predictions that reflect the variability of observations, and as the prediction error is reduced, the prediction results can be improved.

[0168] Step of generating an ensemble weight spatial distribution map using 3D spatial interpolation (S400)

[0169] In one embodiment, the step of generating an ensemble weight spatial distribution map using three-dimensional spatial interpolation (S400) may proceed by performing a process of calculating a predicted value for each space and spatially distributing ensemble weights by reflecting the physical environmental characteristics of the observation area. At this time, the physical environmental characteristics may include physical constraints, boundary conditions (such as coastlines and topography), elevation, and water depth.

[0170] To perform this, a 3D spatial interpolation step (S410) and an ensemble weight spatial distribution map generation step (S420) may be included.

[0172] 3D spatial interpolation execution step (S410)

[0173] In one embodiment, the 3D spatial interpolation step (S410) may be performed by introducing n-dimensional variational interpolation (DIVAnd: Data-Interpolating Variational Analysis in n dimension) to perform gap correction between observation points that reflect physical elements such as terrain.

[0174] Conventionally, the Kriging spatial interpolation technique, which is widely used in the field of geostatistics, has been used for three-dimensional spatial interpolation. Since this technique uses linear combinations based on covariance functions or semivariograms to estimate observed values ​​at observation points and values ​​at unobserved points, it has inherent limitations, such as causing physical discrepancies when complex physical boundary conditions are involved, such as in ocean-atmosphere environments.

[0175] Referring to Fig. 9, when spatial interpolation is performed using the kriging technique, it can be seen that the weights are biased toward the observation points, and the interpolation results for sections without observation points are irregular or unclear. Furthermore, regarding topographic features such as islands located between observation points, physical constraints are not reflected, resulting in only simple distance interpolation. As such, the spatial interpolation method using the kriging technique does not sufficiently reflect the spatial characteristics of the marine environment, and thus has limitations in performing spatial interpolation of the marine spatial field.

[0176] On the other hand, DIVAnd is a type of variational analysis that can generate a spatial field satisfying physical constraints while combining observational data and background information in an optimal manner. In one embodiment, in the case of sea wind and seawater flow as in the present invention, since prediction is performed considering both magnitude and direction, it is necessary to perform spatial interpolation that reflects vertical elements such as elevation and depth as well as horizontal elements. The basic principle of DIVAnd can be explained from the principle of constructing a spatial field φ(X) that minimizes the cost function, as shown in Equation 5 below.

[0177]

[0178] In mathematical formula 5, is the cost function to be minimized, and is an objective function that simultaneously minimizes the goodness of fit to the observed data and spatial smoothness. In the first term, φ is the spatial field to be estimated, Y i is the i-th observation, and H i (φ) is the predicted value of the i-th observation position in the spatial field φ, and the first term is Y i The value H predicted from and the space field φ iA data fitting term that minimizes the difference between (φ), where R i is the error variance of the observation error. Also, in the second term, φ b represents the background field (prior information), and the estimated space field φ is φ b It is a normalization term that constrains the system so as not to deviate abruptly from, where B represents the spatial correlation structure as the background field covariance matrix.

[0179] Referring to Fig. 10, when performing spatial interpolation using DIVAnd, water depth can be considered even in sections without observation points; therefore, it can be confirmed that spatial interpolation reflecting the characteristics of the ocean spatial field is possible, for example, even for the open sea section of the East Sea, through interpolation considering water depth. Furthermore, since spatial interpolation considering priority boundary conditions is possible, mutual influence can be minimized by considering the actual topography even when geographical features such as islands exist between observation points. In the case of sea wind, elevations such as islands and land (including mountains) can be reflected instead of water depth as priority boundary conditions. For reference, Figs. 9 and 10 respectively show the same ensemble weight a using the Kriging spatial interpolation technique and the DIVAnd technique. u The result of performing spatial interpolation is illustrated using [this] as an example. Here, Figures 9 and 10 illustrate the spatial distribution of weights for seawater flow prediction.

[0180] Comparing Figures 9 and 10, it can be seen that the spatial interpolation performed using DIVAnd reflects physical constraints, boundary conditions such as coastlines and topography, and characteristics of the marine space such as elevation and water depth.

[0181] In this way, through 3D spatial interpolation using DIVAnd, it is possible to reproduce a correlation structure similar to the actual environment of the observation area and maintain the continuity and consistency of the spatial field. In addition, spatial interpolation can be performed without false diffusion even in complex terrains such as islands, bays, and straits.

[0183] Ensemble weight spatial distribution generation step (S420)

[0184] In one embodiment, the step of generating an ensemble weight spatial distribution map (S420) may be performed in a manner that distributes ensemble weights in a three-dimensional space spatially interpolated using DIVAnd.

[0185] Conventionally, when calculating ensemble prediction results for observation points, ensemble weights were calculated for each point, and when calculating ensemble prediction results for space, the same weights were applied across the entire space using the weights of specific points. However, since the predicted values ​​may vary from space to space due to the physical characteristics of the ocean region, applying the same weights across the entire space can significantly degrade prediction performance.

[0186] Therefore, as illustrated in FIG. 11, in order to perform a prediction that more accurately reflects the physical environmental characteristics of the observation area, it is necessary to distribute ensemble weights in a three-dimensional space that is spatially interpolated to reflect physical elements such as terrain (e.g., elevation, water depth, etc.). In one embodiment according to the present invention, an ensemble weight spatial distribution map can be generated by distributing ensemble weights in a three-dimensional space that is spatially interpolated using the DIVAnd technique. In addition, multi-model ensemble spatial prediction data can be generated based on the generated spatial distribution map. Through this, the error in spatial prediction can be minimized and the prediction reliability can be improved.

[0188] Spatial prediction performance analysis step (S500)

[0189] In one embodiment, the spatial prediction performance step (S500) may perform a process of analyzing spatial prediction performance by utilizing high-resolution satellite analysis data provided by CMEMS (Copernicus Marine Environment Monitoring Service) (hereinafter CMES data) or NOAA NCEI Blended Seawinds data (hereinafter NOAA data) as a verification standard for the ensemble weight spatial distribution map generated in the previous step. In one embodiment, when CMEMS data is used as a verification standard, bias can be identified by using deviation distribution analysis in the spatial prediction performance analysis. Meanwhile, when NOAA data is used as a verification standard, error can be identified by using the RMSE metric.

[0191] Bivariate multi-model ensemble-based ocean weather prediction system (10)

[0192] As illustrated in FIG. 12, a bivariate multi-model ensemble-based marine weather prediction system (10) according to one embodiment of the present invention may include a data collection unit (1), an ensemble weight calculation unit (2), an ensemble weight spatial distribution map generation unit (3), an ensemble prediction unit (4), and a spatial prediction unit (5) to perform marine weather prediction.

[0193] In one embodiment, the data collection unit (1) can collect observational data and prediction model-based output data used in the bivariate multi-model ensemble-based marine weather prediction system (10) of the present invention. More specifically, the data collection unit (1) can collect marine weather information including sea wind and seawater flow data from marine weather observation equipment installed at observation points within the observation area. In one embodiment, the data collection unit (1) can collect prediction results of a plurality of numerical weather prediction models (NWP) produced at 00 UTC and 12 UTC every day, or collect execution results of artificial intelligence-based weather prediction models. The collected data may include east-west component variables u and north-south component variables v that have vector characteristics, such as sea wind and seawater flow. In one embodiment, the data collection unit (1) can collect data having vector characteristics among the observational data by converting them into east-west component variables u and north-south component variables v.

[0195] In one embodiment, the ensemble weight calculation unit (2) can calculate bivariate multi-model ensemble weights for each observation point based on observation data and prediction data provided by the data collection unit (1). In one embodiment, wind speed (or flow velocity) and wind direction (or flow direction) can be expressed as variables u and v and applied to the ensemble weight calculation model. In one embodiment, the ensemble weight calculation unit (2) can calculate weights by utilizing historical data updated based on the weight calculation time point, setting a learning period, and using multi-model ensemble prediction results based on prediction error statistics such as RMSE. At this time, since the calculated weights are updated at every prediction time point, they can be called dynamic ensemble weights. In one embodiment, the ensemble weight calculation unit (2) can calculate both common ensemble weights for all observation points and individual ensemble weights for each observation point.

[0196] In one embodiment, the ensemble weight spatial distribution map generation unit (3) can generate a weight spatial distribution map through three-dimensional spatial interpolation using common ensemble weights and individual ensemble weights calculated by the ensemble weight calculation unit (2). More specifically, the ensemble weight spatial distribution map generation unit (3) can use an n-dimensional variational interpolation (DIVAnd: Data-Interpolating Variational Analysis in n dimension) technique, and in this case, physical environmental characteristics such as elevation and water depth can be reflected in the spatial interpolation process. In one embodiment, the ensemble weight spatial distribution map generation unit (3) can generate a spatial distribution map by distributing ensemble weights corresponding to the spatially interpolated three-dimensional space. Through this, ensemble weights can be distributed not only at observation points but also at unobserved sections to generate a spatially continuous ensemble weight distribution map.

[0197] In one embodiment, the ensemble prediction unit (4) can calculate an ensemble prediction result using the observation point ensemble weight generated by the ensemble weight calculation unit (2) and the ensemble weight corresponding to the unobserved interval provided by the ensemble weight spatial distribution generation unit (3).

[0198] In one embodiment, the spatial prediction unit (5) can perform spatial prediction for observed points and unobserved sections using the ensemble weight spatial distribution map provided by the ensemble weight spatial distribution map generation unit (3), and can produce final spatial prediction data based thereon. The generated prediction data can be saved in the form of a .nc file.

[0200] Although representative embodiments of the present invention have been described in detail above, those skilled in the art will understand that various modifications can be made to the above-described embodiments without departing from the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined by the claims set forth below as well as equivalents thereof. Explanation of the symbols

[0201] 10: Bivariate Multi-Model Ensemble-Based Ocean Weather Forecasting System 1 : Data Collection Department 2 : Ensemble Weight Calculation Unit 3 : Ensemble Weight Space Distribution Generation Unit 4 : Ensemble Prediction Unit 5 : Spatial Prediction Unit

Claims

Claim 1 A bivariate multi-model ensemble-based ocean weather forecasting method characterized by performing a prediction of either sea wind or seawater flow using a probability density function of a bivariate normal distribution, wherein the sea wind and seawater flow are expressed as combination vectors of an east-west component vector u and a north-south component vector v, wherein the probability density function is a function with u and v as variables, and the correlation coefficient between u and v is included in the probability density function, and an ensemble weight for the sea wind is calculated using relevant data from the past 30 days, wherein the sea wind-related data is collected at 12h intervals, but for a minimum of 3 days and a maximum of 30 days. Claim 2 delete Claim 3 In claim 1, the probability density function is expressed by Equation 1, (Equation 1) In the above mathematical formula 1, the above ρ uv is the correlation coefficient between the above u and the above v, and the above ρ uv is expressed by mathematical formula 2, and (mathematical formula 2) The above r is the amplitude of the correlation coefficient, the above s is the offset, the above k is the number of periods, and the above φ is the phase, and the above r, the above s, the above k, and the above φ are the correlation coefficient ρ uv It is the weight of, and the above μ u and the above μ u is the bias-corrected ensemble average of the above u and the above v, respectively, and the μ u and the above μ u is derived by mathematical formula 3, and (mathematical formula 3) The above is the u mean of the ensemble members, and the above is the v-average of the ensemble members, and the above a u , above b u , above a v and the above b v is a bias correction weight, and the above σ 2 u and the above σ 2 u are the corrected ensemble variances of the above u and the above v, respectively, and the above σ 2 u and the above σ 2 v is derived by mathematical formula 4, (mathematical formula 4) The above s 2 u and the above s 2 v are the ensemble variances of the above u and the above v, respectively, and the above c u , above d u , above c v and the above d v is the variance-corrected weight, and the above correlation coefficient ρ uv A bivariate multi-model ensemble-based ocean weather forecasting method characterized in that the weights are calculated in real time from historical data updated based on the calculation time and are updated according to the calculation time, the bias correction weights are calculated in real time from historical data updated based on the calculation time and are updated according to the calculation time, and the variance correction weights are calculated in real time from historical data updated based on the calculation time and are updated according to the calculation time. Claim 4 delete Claim 5 A bivariate multi-model ensemble-based ocean weather forecasting method characterized by performing a prediction of either sea wind or seawater flow using a probability density function of a bivariate normal distribution, wherein the sea wind and the seawater flow are expressed as combination vectors of an east-west component vector u and a north-south component vector v, the probability density function is a function with u and v as variables, and the correlation coefficient between u and v is included in the probability density function, and an ensemble weight for the seawater flow is calculated using relevant data from the past 3 days, and the data related to the seawater flow is collected at 12h intervals, but is collected for a minimum of 1 day or more and a maximum of 3 days or less. Claim 6 delete Claim 7 delete Claim 8 delete Claim 9 A marine weather forecasting method based on a multi-model ensemble technique according to claim 1 or 5, wherein spatial prediction is performed for an observation area of ​​either sea wind or seawater flow, wherein the space is a 3-dimensional space spatially interpolated using an n-dimensional variational interpolation technique, and when outputting the observed value and the predicted value through spatial interpolation into the 3-dimensional space, the method includes a process of generating a spatial distribution map in which dynamic ensemble weights corresponding to the space are distributed, and wherein the dynamic ensemble weights are updated according to the time of calculation. Claim 10 delete Claim 11 In a prediction system for either sea wind or seawater flow, the bivariate multi-model ensemble-based ocean weather prediction system comprises: a data collection unit that collects observational data and output data based on a prediction model; an ensemble weight calculation unit that applies at least a portion of the collected data to a prediction model based on a bivariate normal distribution probability density function and an ensemble weight calculation model, and calculates ensemble weights in real time from historical data updated based on the weight calculation time, wherein the ensemble weights include common weights for all observation points and individual weights for each observation point; an ensemble weight spatial distribution map generation unit that generates an ensemble weight spatial distribution map through spatial interpolation using a 3D spatial interpolation method based on the common weights and individual weights; and an ensemble prediction unit that calculates ensemble weights corresponding to unobserved sections using the ensemble weight spatial distribution map and calculates ensemble prediction results for the observation points and the unobserved sections. A bivariate multi-model ensemble-based ocean weather forecasting system comprising: a spatial prediction unit that performs spatial prediction for the observation point and the unobserved section using the generated ensemble weight spatial distribution map and calculates final spatial prediction data based thereon, wherein the sea wind and the seawater flow are expressed as the sum of an east-west component vector u and a north-south component vector v, the ensemble weight for the sea wind is calculated using relevant data from the past 30 days, and the sea wind-related data is collected at 12h intervals, but for a minimum of 3 days and a maximum of 30 days.

Images