New energy base power supply capacity calculation method based on high-frequency weather forecast

Abstract

The invention relates to the technical field of energy management and prediction, discloses a new energy base power supply capacity calculation method based on high-frequency weather forecast, and aims to solve the problem of insufficient prediction precision and robustness when an existing new energy power generation power prediction model processes high-frequency nonlinear meteorological data. The calculation method comprises the following steps: constructing a meteorological physical causal knowledge graph; carrying out physical constraint guided multi-scale modal decomposition based on the atlas; building a hierarchical predictor network to predict photovoltaic power; and establishing a closed-loop feedback loop of prediction error attribution and model self-calibration. By adopting the above technical scheme, the method can solve the defects of traditional decomposition, improve the effectiveness, prediction accuracy and interpretability of feature engineering, achieve the adaptive optimization of the model, and maintain the high-precision and high-robustness power supply capability calculation performance.

Description

Technical Field

[0001] This invention belongs to the field of energy management and forecasting technology, and specifically relates to a method for calculating the power supply capacity of new energy bases based on high-frequency weather forecasts. Background Technology

[0002] The field of new energy technology includes multiple branches such as solar photovoltaic power generation, wind power generation, and hydropower generation. The core of this technology lies in utilizing renewable natural resources for energy conversion to address the challenges of global climate change and increasing energy demand. As a key component, photovoltaic power generation directly converts solar energy into electrical energy through the photovoltaic effect. With the decline in the cost of photovoltaic modules and the expansion of deployment scale, new energy technologies have been widely applied in various scenarios such as centralized power plants and distributed power sources. The development of this field has not only promoted the transformation of the energy structure but also put forward higher requirements for ensuring the safe and stable operation of the power grid.

[0003] The method for calculating the power supply capacity of new energy bases refers to accurately predicting the future power generation of photovoltaic power plants by analyzing influencing factors such as meteorological conditions. This topic addresses the power fluctuations and uncertainties brought about by the grid connection of high-penetration new energy sources, proposing an optimized prediction method. This method establishes a prediction model that reflects the dynamic changes in power generation capacity by processing high-frequency weather forecast data and historical operating data. In this method, the model needs to effectively extract and characterize the nonlinear and non-stationary characteristics in the data to achieve accurate prediction of power supply capacity within a specific future time period, thereby providing support for power system dispatch decisions, reserve capacity allocation, and market transactions.

[0004] Existing technologies for processing photovoltaic power series mostly rely on a single prediction model, which struggles to effectively address the strong randomness and noise interference caused by sudden weather changes in the raw data. Signal decomposition techniques introduced to improve the signal-to-noise ratio, such as empirical mode decomposition (EMD), inherently suffer from mode aliasing, leading to mutual interference between different frequency features and preventing complete separation. While improved methods can partially alleviate this problem, they typically come at the cost of introducing significant reconstruction errors and remain ineffective in handling feature aliasing in high-frequency components, resulting in the loss of crucial details. Traditional deep learning models generally suffer from insufficient long-term dependency capture when processing long-term time series data, causing prediction accuracy to decrease with increasing time span. The deficiencies in data preprocessing and feature extraction, coupled with the inherent limitations of the prediction models themselves, make it difficult for existing methods to accurately characterize power variation patterns under complex meteorological conditions. This problem is particularly pronounced in power systems with a high proportion of renewable energy grid integration, directly impacting the grid's dispatch economy and operational reliability. Summary of the Invention

[0005] The purpose of this invention is to overcome the technical defects of existing new energy power generation prediction models when processing high-frequency, non-stationary, and strongly nonlinear meteorological data, which result in insufficient prediction accuracy and poor model robustness due to incomplete signal decomposition, unclear physical meaning of feature extraction, and a single multidimensional feature coupling mechanism. Therefore, this invention proposes a method for calculating the power supply capacity of new energy bases based on high-frequency weather forecasts.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting, comprising the following steps: S1: Construct a meteorological physical causal knowledge graph. The graph uses a directed acyclic graph data structure to represent the physical hierarchy and causal transmission relationship between meteorological variables in the numerical weather prediction (NWP) system. The graph is then solidified into a structured data body containing node identifiers, causal orientations, transfer function types, and associated weight parameters. S2: Based on the meteorological physical causal knowledge graph, perform a physical constraint-guided multi-scale mode decomposition on the high-frequency NWP time series data, deconstructing the original time series of each meteorological variable into a set of orthogonal mode component sequences with clear physical attribution; S3: Based on the topology of the meteorological physical causal knowledge graph, a hierarchical predictor network is built, in which each predictor node corresponds to a meteorological variable node in the graph, and the orthogonal modal component sequence set is used to drive the predictor network to perform sequential prediction calculations of intermediate physical quantities and final photovoltaic power from bottom to top. S4: Establish a closed-loop feedback loop for prediction error attribution and model self-calibration, calculate the deviation between the predicted and measured values ​​of the final photovoltaic power, and propagate the deviation in reverse along the hierarchical predictor network. Quantify the contribution of each predictor node and the upstream mode decomposition process to the total error, and automatically adjust the decomposition parameters in the physical constraint-guided multi-scale mode decomposition process and the node weights in the predictor network according to the contribution to form the iterative initial state for a new round of calculation.

[0007] As one embodiment of the present invention, the construction steps of the meteorological physical causal knowledge graph specifically include: S111: Initialize a basic physical model library, which stores deterministic physical equations related to photovoltaic power generation systems, including solar geometric position algorithms, atmospheric mass calculation models, Kerchev's radiation law, standard equations for PV and IV characteristic curves of photovoltaic cells, and a set of thermodynamic equations based on the Sandia National Laboratories (SNL) model. This set of equations defines the functional relationships between total solar radiation (GHI), direct normal radiation (DNI), diffuse radiation (DHI), ambient temperature, wind speed, and photovoltaic module backsheet temperature. S112: Obtain the historical NWP dataset of the geographical location of the new energy base. The dataset has a time resolution of 15 minutes and a time span of 36 consecutive calendar months. The data dimensions include, but are not limited to, GHI, DNI, DHI, ambient temperature, relative humidity, wind speed, wind direction, and cloud coverage. S113: A constraint-based structure learning algorithm, specifically the PC (Peter-Clark) algorithm, is adopted. The deterministic physical equations in the basic physical model library are used as hard constraints to perform causal structure mining on the historical NWP dataset and generate an initial directed acyclic graph. S114: Perform parameter quantization on the initial directed acyclic graph, calculate the conditional probability distribution or transfer function between any two nodes connected by an edge in the graph. For linear relationships, use the Pearson correlation coefficient for quantization, and for nonlinear relationships, use the maximum information coefficient (MIC) for quantization. Use the quantization result as the weight of the edge. The resulting meteorological physical causal knowledge graph is serialized into an XML file, where each node element contains a variable name, physical unit, and time series characteristic label (periodicity, randomness), and each edge element contains a source node, a target node, causal relationship strength (weight value), and transfer function type identifier.

[0008] As one embodiment of the present invention, the physical constraint-guided multi-scale mode decomposition step specifically includes: S211: For each NWP meteorological variable time series to be decomposed, query all its parent node variables from the meteorological physical causal knowledge graph; S212: Extract the historical time series of the parent node variable and calculate its power spectral density (PSD) and autocorrelation function to determine its inherent periodic frequency, amplitude modulation law and statistical characteristics of random fluctuation components, and generate a set of physical prior constraint parameters, which includes the main periodic frequency, harmonic frequency, random component autocorrelation attenuation coefficient and probability distribution type. S213: Execute a fully adaptive noise set empirical mode decomposition (CEEMDAN) algorithm modified by a set of physical prior constraint parameters; in each iteration of the algorithm, the Gaussian white noise added to the signal to be decomposed is no longer fixed in amplitude, but dynamically adjusted according to the statistical characteristics of random components defined in the set of physical prior constraint parameters. S214: During the screening process of Empirical Mode Decomposition (EMD), the boundary conditions of the spline interpolation function used to construct the upper and lower envelopes of the signal are no longer free boundaries, but are constrained by the periodic frequency and amplitude modulation defined in the physical prior constraint parameter set, forcing the frequency components of the first intrinsic mode function (IMF) decomposed to be highly consistent with the principal periodic frequency of the parent node. S215: The matching degree of each decomposed IMF component is calculated with the physical prior constraint parameter set. By calculating mutual information and correlation coefficient, a physical attribution label is assigned to each IMF component, such as "GHI_IMF_Daily Cycle Component", "GHI_IMF_Cloud Covering High-Frequency Random Component", and "Temperature_IMF_Seasonal Trend Component", and finally the orthogonal modal component sequence set is formed.

[0009] As one embodiment of the present invention, the construction and computation steps of the hierarchical predictor network specifically include: S311: Based on the topology of the meteorological physics causal knowledge graph, instantiate a hierarchical network. Each computing node in the network uses a Kolmogorov-Arnold Network (KAN) as the basic predictor unit. The internal activation function of KAN is a B-spline function with a spline order of 3 and 10 grid points. S312: The hierarchical relationship of the network structure strictly corresponds to the causal chain of the graph. The KAN corresponding to the leaf node (such as the solar zenith angle and azimuth angle) in the graph is located at the bottom layer of the network, and its input is a deterministic calculated value; the KAN of other non-leaf nodes has the output of the KAN corresponding to all parent nodes in the graph as its input. S313: For example, the input of a "component backsheet temperature" KAN node receives the predicted temperature mode component sequence from the output of the "ambient temperature" KAN, the predicted wind speed mode component sequence from the output of the "wind speed" KAN, and the predicted radiation mode component sequence from the output of the "total radiation" KAN. S314: The calculation process proceeds from the bottom layer of the network to the top layer. First, it drives the bottommost KAN node to perform calculations using the corresponding deterministic or decomposed modal component sequences to generate the first-level intermediate physical quantity prediction results. Then, these prediction results are used as inputs to drive the KAN node of the next higher level to perform calculations. This process is repeated layer by layer until the topmost "photovoltaic output power" KAN node, which receives the prediction results of all direct physical precursors (such as the predicted module backsheet temperature and the predicted total radiation) and combines them with the electrical parameters of the photovoltaic array (such as nominal power and attenuation coefficient) to finally output the power supply capacity of the new energy base, i.e., the high-frequency photovoltaic power prediction time series.

[0010] As one embodiment of the present invention, the closed-loop feedback loop steps of prediction error attribution and model self-calibration specifically include: S411: Obtain the final photovoltaic power prediction time series and the actual power measurement time series collected from the power station side in the same period, and calculate the root mean square error (RMSE) and mean absolute error (MAE) between the two as the total system error; S412: Using a backpropagation mechanism based on the chain rule, the total system error is decomposed layer by layer in reverse along the hierarchical predictor network. The partial derivative of the prediction error of each KAN node to its direct successor node is calculated, thereby quantifying the contribution of each intermediate physical quantity prediction link to the total system error and generating an error contribution attribution vector. S413: For KAN nodes whose error contribution attribution vector values ​​exceed a preset error attribution threshold (e.g., 10% of the total error), lock their corresponding upstream input modal component sequence. S414: Start a Bayesian optimizer and use the decomposition parameters (such as the amplitude range of added noise and the boundary constraint strength of the spline interpolation function) used to generate these locked modal component sequences during the multi-scale modal decomposition process guided by the physical constraints as optimization variables, and perform parameter optimization with the goal of minimizing the total system error; S415: Simultaneously, for the KAN nodes whose error contribution exceeds the threshold, fine-tune the grid point positions and coefficients of their internal B-spline functions, as well as the connection weights, to optimize the local model parameters; after completing the parameter update, the entire calculation process returns to step S2, forming a complete iterative self-calibration.

[0011] This invention also provides a power supply capacity calculation system for new energy bases based on high-frequency weather forecasts. The system is used to execute the above method and includes: A meteorological physical causal knowledge graph construction module is configured with a basic physical model library and a structured learning algorithm engine. It is used to receive historical NWP data, generate and store a structured graph data body that represents the physical causal relationship between meteorological variables through constrained causal mining and parameter quantization. A physical constraint-guided multi-scale mode decomposition module integrates a CEEMDAN algorithm core that has been corrected by physical priors. This module queries the physical prior constraint parameter set of a specific meteorological variable from the knowledge graph construction module and decomposes the input real-time high-frequency NWP time series based on the parameter set, producing a set of orthogonal mode component sequences with clear physical attribution. A hierarchical predictor network module dynamically instantiates a hierarchical computing network composed of multiple Kolmogorov-Arnold network (KAN) units according to the topology of the knowledge graph, and receives the component sequence produced by the mode decomposition module, performs the calculation step by step from the physical bottom layer to the power top layer, and outputs the final photovoltaic power prediction sequence. A model self-calibration closed-loop control module integrates an error calculation unit, a partial derivative-based error backpropagation attribution unit, and a Bayesian optimizer. It is used to compare the predicted power with the measured power, decompose the error inversely to each predictor node and the mode decomposition process, and automatically adjust the decomposition parameters and network weights to achieve iterative optimization of the system and adaptation to dynamic changes.

[0012] Compared with the prior art, the beneficial effects of the present invention are reflected in: This invention first constructs a meteorological physics causal knowledge graph, solidifying the physics knowledge of domain experts into a constraint framework guiding subsequent data processing. This fundamentally solves the problems of blind decomposition, mode aliasing, and unclear physical meaning of decomposition results in traditional signal decomposition methods. Secondly, the proposed physical constraint-guided multi-scale mode decomposition method utilizes prior knowledge from the causal graph to force the decomposition process towards physical interpretability, ensuring that each decomposed modal component corresponds to a specific physical driving factor (such as Earth's rotation or cloud disturbance), greatly improving the effectiveness and depth of feature engineering. Furthermore, the hierarchical predictor network created in this invention strictly replicates the causal transmission chain of the physical world, making the prediction process no longer a black-box mapping of a single model, but a transparent, traceable white-box process gradually synthesized from multiple intermediate physical quantities. This not only improves the accuracy of predictions but also enhances the interpretability of the model. Finally, by establishing a reverse attribution and self-calibration closed-loop feedback mechanism for prediction errors, the system can autonomously diagnose the source of errors and accurately target and optimize the key links that cause errors (whether it is mode decomposition parameters or specific predictor nodes), thus realizing the continuous adaptive evolution of the model. In this way, the system can always maintain high accuracy and strong robustness in power supply capacity calculation performance when facing constantly changing meteorological conditions and power plant operating status. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall process of a method for calculating the power supply capacity of a new energy base, provided in an embodiment of the present invention. Figure 2 A schematic diagram of the functional modules of a power supply capacity calculation system for a new energy base provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hierarchical predictor network structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the closed-loop feedback process for prediction error attribution and model self-calibration in an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the description of this invention, it should be understood that the terms "upper," "lower," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0015] Please see Figure 1 This invention provides a method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasts. A flowchart illustrating the overall computational framework is provided. Specifically, the method includes the following steps: Step S1, constructing a meteorological physical causal knowledge graph; Step S2, performing a physically constrained guided multi-scale modal decomposition on high-frequency numerical weather prediction (NWP) time series data based on the knowledge graph; Step S3, building and calculating a hierarchical predictor network according to the topology of the knowledge graph; and Step S4, establishing a closed-loop feedback loop for prediction error attribution and model self-calibration.

[0016] Step S1: Construct a meteorological physics causal knowledge graph. The core of this step lies in combining prior knowledge from meteorology and photovoltaic physics with data-driven causal discovery to generate a structured data body that accurately represents the physical hierarchy and causal transmission relationships between various meteorological variables. This knowledge graph is organized using a Directed Acyclic Graph (DAG) data structure.

[0017] In a specific embodiment, step S1 can be broken down into the following sub-steps: Step S111: Initialize a basic physical model library. This library is the source of physical constraints and stores the deterministic physical equations and empirical models for each link in the energy conversion chain of a photovoltaic power generation system. Specifically, the library includes: 1. Solar geometric position algorithms, such as the SPA (SolarPosition Algorithm) released by the National Renewable Energy Laboratory (NREL) in the United States. This algorithm can accurately calculate parameters such as solar zenith angle, azimuth angle, and Earth-Sun distance based on geographical location (latitude and longitude), altitude, date, and time.

[0018] II. Air Mass (AM) calculation models, such as the Katten and Young model, which calculates the effective path length of solar radiation through the atmosphere based on the solar zenith angle.

[0019] III. A derivative model of Kerchev's radiation law, used to describe the geometric relationship between total solar radiation (GHI), direct normal radiation (DNI), and diffuse radiation (DHI), namely GHI = DNI * cos(zenith angle) + DHI.

[0020] IV. The standard equations for the PV (power-voltage) and IV (current-voltage) characteristic curves of photovoltaic cells are usually based on single-diode or dual-diode models. These equations describe the electrical output characteristics of photovoltaic modules under specific irradiance and cell temperature.

[0021] V. Based on the photovoltaic module thermodynamic model released by Sandia National Laboratories (SNL). This model is a comprehensive set of equations, the core of which lies in establishing the functional relationship between the photovoltaic module backsheet temperature and multiple environmental input variables. Specifically, the equations define the module backsheet temperature as T_module=f(GHI,T_ambient,V_wind,...), where T_ambient is the ambient temperature and V_wind is the wind speed. The model quantifies thermodynamic processes such as radiative gain, convective heat dissipation, and radiative heat dissipation through a series of empirical coefficients (determined by module materials, installation methods, etc.).

[0022] Step S112: Obtain the historical NWP dataset of the geographical location of the new energy base. The quality and dimensionality of the data are fundamental to causal structure mining. In one embodiment, the data collection target is a large photovoltaic base located in Northwest China, with geographical coordinates of 38.5°N, 103.2°E. Historical meteorological data corresponding to this coordinate point is obtained from the ERA5 reanalysis dataset of the European Centre for Medium-Range Weather Forecasts (ECMWF). The temporal resolution of the dataset is set to 15 minutes to capture short-term drastic changes in weather conditions, which is crucial for high-frequency power supply capacity calculations. The time span is selected as 36 consecutive calendar months, for example, from 00:00 on January 1, 2019 to 23:45 on December 31, 2021, to ensure that the data covers multiple seasonal patterns, extreme weather events, and long-term climate trends. Data dimensions include, but are not limited to: Total Horizontal Radiation (GHI), in W / m²; Direct Normal Radiation (DNI), in W / m²; Dissipated Horizontal Radiation (DHI), in W / m²; Ambient temperature at 2 meters, in degrees Celsius (°C); Wind speed at 10 meters, in meters per second (m / s); Wind direction at 10 meters, in degrees (°); Relative humidity, in percentages (%); and Total Cloud Cover, in percentages (%). The acquired data underwent preprocessing, including imputation of missing values ​​and removal of outliers.

[0023] Step S113 employs a constraint-based structure learning algorithm for causal structure mining. Specifically, the Peter-Clark (PC) algorithm is selected. The PC algorithm determines causal relationships between variables by iteratively performing conditional independence tests. In this step, the deterministic physical equations defined in the basic physical model library from step S111 are integrated into the PC algorithm as hard constraints. For example, the relationship between the solar geometric position (zenith angle, azimuth angle) and GHI and DNI is a deterministic causal relationship in physics. Therefore, in the initial stage of the PC algorithm, these known causal edges (such as zenith angle -> GHI) are pre-placed in the adjacency matrix of the graph and marked as non-removable. This avoids the algorithm generating erroneous causal links that violate physical laws when the data is sparse or noisy. The conditional independence test of the PC algorithm uses the kernel conditional independence test (KCI-test), with a significance level α set to 0.01. The α value was determined based on multiple cross-validation experiments on a validation subset representing 10% of the total dataset, striking a balance between preventing overfitting (i.e., ensuring the model's expressive power). This step generates a preliminary directed acyclic graph containing nodes and directed edges.

[0024] Step S114 involves parameter quantization of the initial directed acyclic graph and solidifying the graph structure into a structured data volume. This step aims to assign a quantized weight to each causal edge in the graph, representing the strength and type of the causal relationship. For any two nodes (variables) connected by an edge in the graph, the transfer function or conditional probability distribution between them is calculated. Specifically, for variable pairs judged to have a linear relationship, such as 'ambient temperature' pointing to 'component backsheet temperature', the Pearson correlation coefficient is used for quantization. Based on the historical dataset in step S112, the Pearson correlation coefficient for this edge is calculated to be 0.89, indicating a strong positive correlation. For nonlinear relationships, such as 'cloud cover' pointing to 'total radiation (GHI)', the maximum information coefficient (MIC) is used for quantization. MIC can capture a wide range of functional relationships, not limited to linear ones. The calculated MIC value for this edge is 0.72, indicating a significant nonlinear negative correlation.

[0025] Finally, the resulting meteorological physics causal knowledge graph is serialized into an XML (eXtensible Markup Language) file for storage and retrieval. An example of the structure of this XML file is as follows: XML 1 <knowledgegraph name="PV_Causal_Graph_SiteA"> 2 <Node id="var_ghi" name="Global Radiation" unit="W / m^2"> 3 <timeseries a_label="periodic,stochastic" / > 4 5 <node id="var_temp_amb" name="环境温度" unit="C"> 6 <timeseries a_label="periodic,trend" / > 7< / node> 8... 9 <edge source="var_zenith" target="var_ghi" type="deterministic_nonlinear"> 10 <weight value="1.0" method="PhysicsEquation" / > 11 <function type="Cosine" / > 12< / edge> 13 <edge source="var_cloud_cover" target="var_ghi" type="stochastic_nonlinear"> 14 <weight value="0.72" method="MIC" / > 15 <function type="Unknown" / > 16< / edge> 17 <edge source="var_temp_amb" target="var_module_temp" type="stochastic_linear"> 18 <weight value="0.89" method="Pearson" / > 19 <function type="Linear" / > 20< / edge> 21... 22< / knowledgegraph> This structured data volume clearly defines the attributes (name, unit, time series characteristics) of each meteorological variable node and the attributes (source node, target node, causal relationship strength i.e. weight, transfer function type identifier) ​​of each causal edge.

[0026] As an alternative, causal structure learning can employ the Fast Greedy Equivalence Search (FGES) algorithm. FGES is computationally more efficient than PC algorithms, especially when dealing with high-dimensional datasets. It finds the optimal graph structure by greedily searching a scoring function (such as the BIC score). However, FGES makes strict assumptions about the data distribution and may get stuck in local optima when the data volume is insufficient. Another alternative is the Linear Non-Gaussian Acyclic Model (LiNGAM) algorithm, which can directly identify causal directions from the data, provided that the data follows a non-Gaussian distribution and the relationships are linear. The applicability of LiNGAM is limited for non-linear and partially Gaussian distributed variables commonly found in meteorological data. In contrast, the constrained PC algorithm used in this embodiment, by introducing physical priors, has better universality for data distribution and relationship types while ensuring physical consistency.

[0027] Step S2 involves performing a physically constrained multi-scale modal decomposition on the high-frequency NWP time series data based on the aforementioned meteorological physical causal knowledge graph. The purpose of this step is to deconstruct the original meteorological variable time series, which is influenced by multiple physical processes, into a set of mutually orthogonal modal component sequences, each with a clearly defined physical attribution. This provides cleaner and more interpretable input features for subsequent prediction models.

[0028] In a specific embodiment, step S2 can be broken down into the following sub-steps: Step S211: For each NWP meteorological variable time series to be decomposed, query all its parent node variables from the meteorological physical causal knowledge graph. Parent node variables are the direct physical causes of changes in the current variable. For example, when decomposing the 'Total Radiation (GHI)' time series, the system parses the XML file generated in step S114, queries all edges with 'var_ghi' as the target node, and thus identifies its parent nodes, including 'solar zenith angle', 'cloud coverage', and 'atmospheric mass', etc.

[0029] Step S212 involves extracting the historical time series of the parent node variables and analyzing their statistical properties to generate a set of physical prior constraint parameters. Specifically, the power spectral density (PSD) and autocorrelation function (ACF) are calculated for the time series of each parent node variable. For example, PSD analysis of the 'solar zenith angle' sequence reveals that at a frequency f_day = 1.157 x 10⁻ 5 Extremely sharp peaks exist at Hz (corresponding to a 24-hour cycle) and its harmonic frequencies (such as 12-hour and 8-hour cycles), revealing its strong diurnal periodicity. ACF analysis of the 'cloud cover' sequence reveals a rapid exponential decay in its autocorrelation function, with a decay coefficient λ_cloud = 0.95, indicating that cloud disturbance is a high-frequency, short-duration random fluctuation process. Based on these analytical results, a set of physical prior constraint parameters is generated for the GHI sequence to be decomposed. This parameter set is constructed as a structured object, for example: ConstraintSet_GHI={ main_freq: 1.157e-5Hz, harmonic_freqs:[2.314e-5Hz,3.471e-5Hz], stochastic_acf_decay:0.95, stochastic_dist_type:'Laplace', stochastic_dist_params:{mu:0,b:5.2W / m^2}} The probability distribution type of the random component (e.g., Laplace distribution) and its parameters (e.g., scale parameter b) are obtained by fitting the non-periodic residuals of the parent node's 'cloud coverage' sequence.

[0030] Step S213 executes a fully adaptive ensemble empirical mode decomposition (CEEMDAN) algorithm modified with a set of physical prior constraint parameters. CEEMDAN is an improved empirical mode decomposition (EMD) method that suppresses mode aliasing by adding paired, opposite-signed auxiliary white noise. In this embodiment, a key modification is made to the standard CEEMDAN: in each iteration of the algorithm, the amplitude of the auxiliary noise sequence added to the signal to be decomposed is no longer fixed or only related to the signal standard deviation, but is dynamically adjusted according to the statistical characteristics of the random components defined in the set of physical prior constraint parameters generated in step S212. Specifically, when decomposing the GHI sequence, the added auxiliary noise sequence is sampled from a Laplace distribution (μ=0, b=5.2W / m²), instead of traditional standard Gaussian white noise. This modification makes the added noise statistically closer to the real physical disturbance source (such as clouds), thus more effectively separating high-frequency random modes.

[0031] Step S214: During the sifting process of Empirical Mode Decomposition (EMD), constraints are imposed on the boundary conditions of the spline interpolation function. In the standard EMD sifting process, the cubic spline interpolation function used to construct the upper and lower envelopes of the signal typically employs natural boundary conditions (i.e., the second derivative at the endpoints is zero). In this embodiment, the boundary conditions are modified to be constrained by periodic frequency and amplitude modulation defined in the set of physical prior constraint parameters. Specifically, when sifting the first intrinsic mode function (IMF) of the GHI sequence, the B-spline interpolation function used to construct the upper and lower envelopes has its second derivative at both ends of the signal forcibly set to a specific non-zero value related to the principal periodic frequency f_day. This specific value is obtained by spline interpolating a pure cosine function with the same frequency as f_day and calculating its endpoint second derivative. This forces the frequency components of the first IMF to closely match the principal periodic frequency of the parent node 'solar zenith angle', ensuring that the physically dominant periodic components are preferentially and accurately separated.

[0032] Step S225 involves calculating the matching degree between each IMF component obtained after decomposition and the set of physical prior constraint parameters, and assigning a physical attribution label to each IMF component. The matching degree calculation is performed using various indicators, such as mutual information, correlation coefficient, and spectral similarity.

[0033] As a specific computational example, for the IMF set {IMF1, IMF2, ..., IMFn} obtained from GHI sequence decomposition: First, the mutual information between IMF1 and its parent node's 'solar zenith angle' time series was calculated, yielding a value of 1.25, significantly higher than the mutual information values ​​between other IMFs and this parent node (all less than 0.2). Simultaneously, the main peak of IMF1's power spectrum is located near f_day. Based on this, IMF1 was assigned the label: "GHI_IMF_daily periodic component".

[0034] Second, the Pearson correlation coefficient between IMF5 and the high-pass filtered sequence (representing high-frequency perturbations) of its parent node 'cloud coverage' was calculated, yielding a value of -0.68. Simultaneously, the decay pattern of the autocorrelation function of IMF5 is very close to the stochastic_acf_decay value (0.95) in the constraint parameter set. Therefore, IMF5 is labeled as: "GHI_IMF_Cloud-Obscured High-Frequency Random Components".

[0035] Third, the environmental temperature series was similarly decomposed, and one IMF with an extremely low frequency and a period close to one year had a correlation coefficient of 0.92 with the standard cosine function representing seasonal variation with a period of 365.25 days. Therefore, this IMF was labeled: "Temperature_IMF_Seasonal Trend Component".

[0036] Through this step, the original NWP time series is transformed into a set of orthogonal modal component sequences, where each sequence is associated with a clear physical driving factor.

[0037] As an alternative to mode decomposition, variational mode decomposition (VMD) can be used. VMD is mathematically more rigorous, effectively suppresses mode aliasing, and has better robustness to noise. It works by transforming the signal decomposition problem into a variational optimization problem. However, a major limitation of VMD is the need to pre-specify the number of modes K to be decomposed. In meteorological signals, the number of physical processes is unknown and may vary over time; an inappropriate choice of K value can lead to under- or over-decomposition. Another alternative is wavelet transform (WT). WT is a mature time-frequency analysis tool that can provide local information about the signal at different scales. However, WT's performance is highly dependent on the chosen mother wavelet function; for meteorological signals with complex and variable nonlinear and non-stationary characteristics, there is no universally optimal mother wavelet. In contrast, the modified CEEMDAN method proposed in this embodiment has an adaptive decomposition process that does not require a preset number of modes. Furthermore, by introducing physical prior constraints, its decomposition results maintain both data-driven flexibility and physical interpretability, resulting in a more prominent overall advantage.

[0038] Step S3 involves constructing a hierarchical predictor network based on the topological structure of the meteorological physical causal knowledge graph. Using the orthogonal modal component sequence set, the predictor network is driven from the bottom up to perform sequential prediction calculations of intermediate physical quantities and the final photovoltaic power. This step aims to construct a prediction model whose structure is isomorphic to the causal chain of physical reality, thereby transforming the black-box prediction problem into a transparent, traceable, white-box or gray-box computational process that is progressively synthesized from multiple intermediate physical quantities.

[0039] Please see Figure 3 The figure is a schematic diagram of the hierarchical predictor network structure in an embodiment of the present invention. In a specific embodiment, step S3 can be further decomposed into the following sub-steps: Step S311: Based on the topology of the meteorological physics causal knowledge graph, a hierarchical network is instantiated. Each computational node in the network uses a Kolmogorov-Arnold Network (KAN) as its basic predictor unit. KAN is chosen because of its superior interpretability and higher parameter efficiency compared to traditional multilayer perceptrons (MLPs). Specifically, KAN places the activation function on the network edges, rather than on the nodes, and uses a learnable spline function as the activation function. In this embodiment, the activation function within each KAN unit is a B-spline function, with its spline order set to 3 (i.e., cubic B-spline) to ensure the second-order continuity and smoothness of the function curve. The initial number of spline function grid points is set to 10, and these grid points are uniformly distributed within the domain of the B-spline. The overall structure of KAN is defined as a two-layer network, namely KAN([input_dim,hidden_dim,output_dim]), where the width of hidden_dim is dynamically set to twice the input dimension input_dim to provide sufficient function fitting capability.

[0040] In step S312, the hierarchical relationship of the network strictly corresponds to the causal chain of the meteorological physical causal knowledge graph. Leaf nodes in the graph, i.e., variables without parent nodes, such as the solar zenith angle and azimuth angle calculated using the Solar Geometric Position Algorithm (SPA), serve as deterministic inputs, forming the lowest-level input of the entire predictor network. All other non-leaf nodes in the graph are instantiated as corresponding KAN computation nodes in the predictor network. The input of each KAN node is set to the output of the KAN nodes corresponding to all its parent nodes in the knowledge graph. This structure ensures that the transmission of information in the network strictly follows the causal transmission path in the physical world. For example, a KAN node representing 'atmospheric mass' will receive the output sequence from the upstream 'solar zenith angle' KAN node.

[0041] Step S313 involves configuring the input and output of a specific KAN node. Taking a KAN node used to predict "module backsheet temperature" as an example, the construction and configuration process of this node is as follows: First, query the 'module backsheet temperature' (var_module_temp) node in the knowledge graph and identify all its parent nodes, such as 'total radiation' (var_ghi), 'ambient temperature' (var_temp_amb), and 'wind speed' (var_wind). Second, from the output of step S2, i.e., the set of orthogonal modal component sequences, obtain all the decomposed modal component sequences corresponding to these parent node variables. Specifically, the input vector of the "Module Backsheet Temperature" KAN node will be concatenated from the following sequences: the "GHI_IMF_Diurnal Periodic Component" prediction sequence, the "GHI_IMF_Cloud Covering High-Frequency Random Component" prediction sequence, the "Temperature_IMF_Seasonal Trend Component" prediction sequence, the "Temperature_IMF_Diurnal Periodic Component" prediction sequence, and the "Wind Speed_IMF_Average Wind Component" and "Wind Speed_IMF_Gust Disturbance Component" prediction sequences. The output of the KAN node is defined as a single time series, namely the predicted "Module Backsheet Temperature" time series.

[0042] Step S314: Execute the calculation process. This calculation process proceeds sequentially from the bottom layer (leaf nodes) of the network upwards, forming a forward computation propagation. First, the KAN node located at the bottom layer of the network is driven, for example, a KAN node predicting 'DHI'. This node uses the modal component sequences corresponding to its parent nodes (such as 'total radiation', 'solar zenith angle'), which have already been decomposed in step S2, as input. It performs calculations using its internal B-spline activation function and weights to generate the first-level intermediate physical quantity prediction results, i.e., the predicted time series of 'DHI'. Subsequently, these first-level prediction results are passed as input to higher-level KAN nodes in the network. For example, a KAN node predicting 'component effective irradiance' will receive the predicted 'DHI' and predicted 'DNI' sequences output from the bottom-level KAN nodes and perform calculations in conjunction with the deterministic input 'solar zenith angle' sequence. This process is progressive, with information flowing upwards along the causal chain. Ultimately, the top-level KAN node, representing "PV output power," receives predictions for all direct physical antecedent variables, such as the predicted sequence of "module effective irradiance" and the predicted sequence of "module backsheet temperature." Furthermore, this top-level KAN node receives a set of static power plant configuration parameters as constant inputs, including the nominal power of the PV array (in MW), the power temperature coefficient of the PV modules (e.g., -0.38% / °C), and the overall degradation coefficient considering the service life (e.g., 0.98). The final output of this top-level KAN node is the power supply capacity of the renewable energy base, represented as a high-frequency (e.g., 15-minute resolution) PV power prediction time series for one or more future scheduling cycles.

[0043] As an alternative to building a predictor network, a Graph Neural Network (GNN) can be used. GNNs are naturally well-suited for processing graph-structured data and can aggregate information between nodes through message passing mechanisms. Specifically, a Graph Convolutional Network (GCN) or a Graph Attention Network (GAT) can be constructed, with its nodes and edges directly mapped to the meteorological physics causal knowledge graph. The advantage of this approach is that it can learn the dynamic evolution of the entire physical system end-to-end. However, the standard GNN message passing mechanism usually occurs between all nodes, which may obscure specific, directed causal paths and reduce the interpretability of the model. In addition, GNN training is sensitive to the connectivity of the graph, and its advantages may not be fully realized for the strict, hierarchical directed acyclic graph structure of this invention. Another alternative is to use a set of independent, traditional machine learning models (such as Support Vector Regression (SVR) or Gradient Boosting Decision Tree (GBDT)) to train a prediction model separately for each non-leaf node variable. This approach is simple to implement, but it ignores the hierarchical dependencies between variables, and independent training of each model may lead to the accumulation and amplification of errors in the prediction chain, lacking an overall optimization framework. In contrast, the hierarchical network based on KAN used in this embodiment not only strictly follows the physical causal structure, but also utilizes the interpretability and function learning capabilities of KAN. At the same time, the entire network is treated as a unified whole for end-to-end computation and optimization, achieving a deep integration of structured knowledge and data-driven learning.

[0044] Step S4 establishes a closed-loop feedback loop for prediction error attribution and model self-calibration. This step aims to achieve the system's self-evolution and adaptability to dynamic changes. By comparing the final predicted value with the measured value, the system can quantify the total error and propagate this error backward along the prediction network, thereby accurately diagnosing the source of the error and targeting and optimizing the key links causing the error (whether it is the upstream mode decomposition parameters or specific predictor nodes).

[0045] Please see Figure 4 The figure illustrates the closed-loop feedback process of prediction error attribution and model self-calibration in an embodiment of the present invention. In a specific embodiment, step S4 can be further decomposed into the following sub-steps: Step S411: Obtain and calculate the total system error. Specifically, the system obtains the final photovoltaic power prediction time series output in step S314, and obtains the actual power metering time series collected by the power plant outlet metering meter within the same time period from the SCADA (Supervisory Control and Data Acquisition) system of the new energy base. The time resolution and length of the two time series are aligned, for example, both using a 15-minute resolution, covering the past 24 hours. Subsequently, the root mean square error (RMSE) and mean absolute error (MAE) between the two series are calculated as quantitative indicators of the total system error. In a specific embodiment, a weighted average absolute percentage error (wMAPE) is also introduced, where the weight is proportional to the actual power value, so that the prediction error during the high power output period of the daytime contributes more to the total error indicator, because the economic impact of the error during these periods is more significant.

[0046] Step S422 employs a chain rule-based error backpropagation mechanism to perform inverse attribution of the total systematic error. Mathematically, this process is equivalent to calculating the partial derivative of the total systematic error (e.g., RMSE) with respect to the output of each KAN node in the hierarchical predictor network. Starting from the top-level "photovoltaic output power" KAN node, the error gradient is propagated downstream (upstream in the network structure) layer by layer. For any KAN node i, its contribution C_i to the total error is quantified as C_i = |∂(RMSE) / ∂(Output_i)|, where Output_i is the output time series of that node. Using the chain rule, this partial derivative can be decomposed into (∂(RMSE) / ∂(Input_j))*(∂(Input_j) / ∂(Output_i)), where node j is the direct successor of node i. This process continues until the lowest-level KAN node in the network. Finally, an error contribution attribution vector or attribution graph is generated, with the same dimension as the number of KAN nodes in the network. The value of each element represents the contribution of the corresponding intermediate physical quantity prediction link to the total system error.

[0047] Step S413: Lock the critical nodes whose error contribution exceeds the threshold and their upstream inputs. Normalize the values ​​in the error contribution attribution vector generated in step S412 so that their sum is 1. Then, set a preset error attribution threshold, for example, any KAN node whose contribution exceeds 10% of the total error is identified as a critical error source. This threshold is dynamically adjustable; for example, when the overall system error is large, the threshold can be appropriately relaxed to capture the main contradiction; when the system enters the fine optimization stage, the threshold is tightened to locate small error sources. Once a KAN node (e.g., the "component backsheet temperature" prediction node) is locked, the system will further trace its inputs, that is, lock the upstream modal component sequences (e.g., "GHI_IMF_daily cycle component" and "wind speed_IMF_gust disturbance component") that are directly fed into the node from step S2.

[0048] Step S414: A Bayesian optimizer is initiated to optimize the parameters of the upstream mode decomposition process. For each modal component sequence locked in step S413, the system uses the decomposition parameters used in its generation process, located in steps S213 and S214, as optimization variables. These parameters constitute a multidimensional search space, with specific variables including: the range of values ​​for the scaling parameter b of the Laplace distribution used to correct auxiliary noise in the CEEMDAN algorithm, the ratio coefficient of noise amplitude to signal standard deviation, and the boundary constraint strength factor applied in spline interpolation. The Bayesian optimizer aims to minimize the total systematic error (e.g., wMAPE) calculated in step S411. The optimizer first constructs a Gaussian process surrogate model based on the current parameters to fit the unknown functional relationship between "decomposition parameters - systematic error," and then uses the "Expected Improvement" acquisition function to select the next set of parameter combinations that are most promising for improving performance. This process is iterated, for example, 50 iterations, to find a set of near-optimal decomposition parameters within an acceptable computational time.

[0049] Step S415 involves simultaneously fine-tuning the local model parameters of KAN nodes whose error contribution exceeds the threshold. For the KAN nodes locked in step S413, the system performs one or more additional gradient descent updates. Unlike the initial training, this fine-tuning uses a very small learning rate, such as 1 / 100 of the initial learning rate, to avoid destructive interference with the already learned good patterns. The optimization goal is to fine-tune the grid point positions and corresponding coefficients of the B-spline functions within the KAN node, as well as the weights connecting different spline basis functions. In a specific embodiment, for the spline basis function with the highest error contribution, its grid points are allowed to move within a small range to better capture local nonlinear features in the input-output relationship. After updating the decomposition parameters (through Bayesian optimization) and predictor node parameters (through gradient fine-tuning), the entire computation process returns to step S2, using the new parameters to decompose and predict the latest NWP data, forming a complete iterative self-calibration loop. This loop can be configured to execute periodically (e.g., every 24 hours) or triggered when the total system error continuously exceeds the warning line.

[0050] As an alternative to error attribution and self-calibration, sensitivity analysis-based methods can be employed. For example, by introducing small perturbations into the parameters of each input modal component sequence and each KAN node, the magnitude of change in the final output power prediction can be observed to assess the sensitivity of each stage and use it as the basis for error attribution. The advantage of this method is its intuitiveness and independence from gradient calculation, but it is computationally expensive because it requires multiple perturbation simulations for each parameter. Another alternative is to use a reinforcement learning (RL) framework, treating the entire prediction system as an agent, with decomposed parameters and network weights as its action space, and negative prediction errors as a reward. Through interaction with the environment (real weather and power data), the agent learns an optimal policy to dynamically adjust the model parameters. This approach has strong adaptive learning capabilities, but its training process is often unstable, requiring extensive trial and error and complex reward function design, posing significant challenges in practical deployment. In contrast, this embodiment combines a closed-loop feedback mechanism of gradient backpropagation and Bayesian optimization. It utilizes the internal structural information of the model (gradient) to achieve accurate attribution, and solves the parameter optimization problem of the non-differentiable part through an efficient global optimization algorithm (Bayesian optimization), achieving a better balance in efficiency, stability and optimization effect.

[0051] This invention also provides a power supply capacity calculation system for new energy bases based on high-frequency weather forecasts. Please refer to [link / reference]. Figure 2This figure is a functional module diagram of a power supply capacity calculation system for a new energy base provided in an embodiment of the present invention. The system is used to execute the above method, and the system includes: A meteorological physics causal knowledge graph construction module is configured to execute step S1. Internally, it integrates a basic physical model library and a constraint-based structure learning algorithm engine, such as a PC algorithm engine. This module receives historical NWP datasets and power plant geographic information as input. Through constraint-based causal structure mining and parameter quantization, it generates and persistently stores a structured graph data body representing the physical hierarchy and causal transmission relationships among various meteorological variables in XML file format.

[0052] A physically constrained multi-scale mode decomposition module is configured to execute step S2. Its core is a physically prior-corrected CEEMDAN algorithm. The module's function is to, upon receiving a real-time high-frequency NWP time series, first query the physical prior constraint parameter set of the variable to be decomposed from the knowledge graph construction module, then adaptively decompose the input NWP time series based on this parameter set, ultimately producing a set of orthogonal mode component sequences with clear physical attributions (such as daily cycles, seasonal trends, and random disturbances), and then pass these sequences to downstream modules.

[0053] A hierarchical predictor network module, configured to perform step S3, has the following structure: Figure 3 As shown, this module dynamically instantiates a hierarchical computational network composed of multiple Kolmogorov-Arnold Network (KAN) units based on the graph topology structure output by the meteorological physics causal knowledge graph construction module. This module receives the set of orthogonal modal component sequences produced by the mode decomposition module as input and drives the KAN nodes in the network to perform calculations step-by-step according to the causal order from the physical bottom layer (e.g., solar geometric position) to the power top layer (photovoltaic output power), ultimately outputting a time series prediction of the high-frequency power supply capacity of the new energy base.

[0054] A model self-calibration closed-loop control module is configured to execute step S4, the process of which is... Figure 4As shown, this module integrates an error calculation unit, an error backpropagation attribution unit based on partial derivatives and the chain rule, and a Bayesian optimizer. Its function is to continuously compare the predicted power output by the predictor network module with the measured power obtained from an external SCADA system to calculate the total system error. Then, through error backpropagation, the total error is decomposed inversely to each KAN node in the hierarchical predictor network, and further traced back to the decomposition parameters used by the physically constrained multi-scale mode decomposition module. Finally, through the Bayesian optimizer and gradient fine-tuning mechanism, the decomposition parameters and network node weights that contribute the most to the total error are automatically adjusted to achieve iterative optimization of the entire computing system and continuous adaptation to dynamic environmental changes.

[0055] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting, characterized in that, Includes the following steps: S1: Construct a meteorological physical causal knowledge graph, which is organized in a directed acyclic graph data structure containing nodes and directed edges, to structurally represent the preset physical hierarchy and causal transmission relationship between meteorological variables. S2: Based on the meteorological physical causal knowledge graph, a physical constraint-guided multi-scale modal decomposition is performed on the meteorological variable time series data obtained from high-frequency numerical weather forecasts, deconstructing the meteorological variable time series data into a set of mutually orthogonal modal component sequences with physical attribution labels; S3: Based on the topological structure of the meteorological physical causal knowledge graph, a hierarchical predictor network is built and calculated. Using the modal component sequence set, the predictor network is driven from bottom to top to predict the intermediate physical quantities and the final photovoltaic power in sequence, generating the final photovoltaic power prediction time series. S4: Establish a closed-loop feedback loop for prediction error attribution and model self-calibration, compare the final photovoltaic power prediction time series with the measured power measurement time series to calculate the total system error, and propagate the total system error backward along the predictor network to attribute it. Based on the attribution results, perform targeted parameter optimization on the multi-scale mode decomposition process and the predictor network.

2. The method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting according to claim 1, characterized in that, The meteorological physical causal knowledge graph includes the attributes of each meteorological variable as a node and the attributes of each causal relationship as a directed edge. The modal component sequence set includes multiple intrinsic mode function sequences decomposed from each meteorological variable and their corresponding physical attribution labels. The hierarchical predictor network includes multiple computational nodes corresponding to the non-leaf nodes in the knowledge graph and their hierarchical connection relationships. The closed-loop feedback loop includes an error contribution attribution vector, a Bayesian optimizer for optimizing mode decomposition parameters, and a gradient descent update mechanism for fine-tuning predictor node parameters.

3. The method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting according to claim 1, characterized in that, The specific steps for constructing the meteorological physical causal knowledge graph are as follows: S111: Initialize a basic physical model library, which stores deterministic physical equations and empirical models characterizing each link in the energy conversion chain of a photovoltaic power generation system; S112: Obtain the historical numerical weather forecast dataset of the geographical location of the new energy base, wherein the dataset has a preset high-frequency time resolution and a continuous time span; S113: Apply a constraint-based structure learning algorithm, using the historical numerical weather forecast dataset as input and the deterministic physical equations defined in the basic physical model library as hard constraints, to perform causal structure mining and generate an initial directed acyclic graph; S114: Perform parameter quantization on the initial directed acyclic graph, calculate and assign a quantized weight representing the strength and type of causal relationship to each directed edge in the graph, and serialize the solidified graph structure and parameters into a structured data body to form the meteorological physical causal knowledge graph.

4. The method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting according to claim 3, characterized in that, The constraint-based structural learning algorithm is the PC algorithm, which determines the causal relationship between variables by iteratively performing conditional independence tests. The parameter quantization step is specifically as follows: for any two variable nodes connected by an edge in the graph, calculate the transfer function or conditional probability distribution between them, and use the key parameters of the function as the quantization weights. The structured data volume defines the attributes of each meteorological variable node, including name, unit, time series characteristics, and the attributes of each causal edge, including source node, target node, causal relationship strength weight, and transfer function type identifier.

5. The method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting according to claim 1, characterized in that, The physical constraint-guided multi-scale mode decomposition steps are as follows: S211: For each meteorological variable time series to be decomposed, query and determine all its parent node variables from the meteorological physical causal knowledge graph; S212: Based on the physical characteristics of the parent node variable, generate a set of physical prior constraint parameters, which defines the statistical characteristics of the periodic component, random component and trend component in the sequence to be decomposed; S213: Execute a fully adaptive noise set empirical mode decomposition algorithm modified by the physical prior constraint parameter set to decompose the meteorological variable time series; S214: The matching degree of the set of intrinsic mode function components obtained after decomposition is calculated with the set of physical prior constraint parameters, and a unique physical attribution label is assigned to each intrinsic mode function component to form the set of mode component sequences.

6. The method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting according to claim 5, characterized in that, The modified fully adaptive noise set empirical mode decomposition algorithm includes the following modifications: In each iteration of the algorithm, the amplitude of the auxiliary noise sequence added to the signal to be decomposed is dynamically adjusted according to the statistical characteristics of the random components defined in the set of physical prior constraint parameters, and the auxiliary noise sequence is sampled from a preset non-Gaussian distribution. Furthermore, during the screening process of performing empirical mode decomposition, the spline interpolation function used to construct the upper and lower envelopes of the signal is subject to the boundary conditions of the periodic frequency and amplitude modulation defined in the physical prior constraint parameter set, so as to force the frequency components of the first intrinsic mode function decomposed to match the principal periodic frequency of the parent node.

7. The method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting according to claim 1, characterized in that, The specific steps for constructing and calculating the hierarchical predictor network are as follows: S311: Based on the topology of the meteorological physics causal knowledge graph, instantiate a hierarchical network, wherein each non-leaf node in the graph is instantiated as a corresponding computing node in the network, and each computing node uses a Kolmogorov-Arnold network as the basic predictor unit. S312: Set the connection relationship of the hierarchical network so that the input of each computing node is the output of the computing nodes corresponding to all its parent nodes in the knowledge graph, ensuring that the transmission of information in the network follows the causal transmission path defined by the knowledge graph; S313: Configure the input and output of each computing node, and use the modal component sequence corresponding to the parent node variable in the modal component sequence set as the input vector of the corresponding computing node; S314: The forward computation is performed sequentially from the bottom leaf nodes of the network to the top non-leaf nodes, generating the prediction time series of intermediate physical quantities at each level, until the final photovoltaic power prediction time series of the top node is output.

8. The method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting according to claim 7, characterized in that, The Kolmogorov-Arnold network unit places a learnable activation function on the network edge. The activation function is a B-spline function, and the spline order and the initial number of grid points of the B-spline function are preset. Taking a computing node for predicting "component backsheet temperature" as an example, the input vector of the node is formed by splicing all the modal component sequences extracted from the modal component sequence set corresponding to its parent node "total horizontal irradiance", "ambient temperature" and "wind speed".

9. The method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting according to claim 1, characterized in that, The closed-loop feedback loop steps for prediction error attribution and model self-calibration are as follows: S411: Obtain the final photovoltaic power prediction time series and the measured power measurement time series collected from the power plant metering system, and calculate the weighted average absolute percentage error between the two as the total system error; S412: Apply the error backpropagation mechanism based on the chain rule to propagate the total system error from the top-level computing node to each layer computing node along the hierarchical predictor network, and generate an error contribution attribution vector that quantifies the contribution of each node to the total system error. S413: Normalize the error contribution attribution vector and, based on the preset error attribution threshold, lock the key computing nodes whose error contribution exceeds the threshold and their upstream modal component sequence inputs. S414: For the locked key nodes and upstream input, a dual-path optimization process is started simultaneously to update the decomposition parameters of the multi-scale mode decomposition process and the model parameters of the key computing nodes.

10. The method for calculating the power supply capacity of a new energy base based on high-frequency weather forecasting according to claim 9, characterized in that, The dual-path optimization process is as follows: For the locked upstream modal component sequence, a Bayesian optimizer is started, using the modal decomposition parameters that generated the sequence as optimization variables, with the goal of minimizing the total system error. By constructing a Gaussian process surrogate model and applying the acquisition function for iterative optimization, a set of updated decomposition parameters is output. Furthermore, for the locked key computing nodes, perform one or more gradient descent updates using a preset small learning rate to fine-tune the grid point positions and coefficients of the B-spline function inside the node and the connection weights, and output the updated model parameters. Using the updated decomposition parameters and model parameters, return to step S2 to form a complete iterative self-calibration loop.

Images