A new energy power prediction and uncertainty quantification method and system

Abstract

The application relates to the technical field of new energy power prediction, and particularly discloses a new energy power prediction and uncertainty quantification method and system, which is realized based on a three-level distributed architecture composed of an edge layer, a regional layer and a cloud layer, and comprises the following steps: obtaining multi-source layered operation data; the multi-source layered operation data comprises local unit dynamic characteristic data of the edge layer, regional power grid topology data and regional weather forecast data of the regional layer, and cross-regional power transmission channel state data of the cloud layer; and in actual power grid dispatching operation, the multi-source layered operation data is deeply combined with a layered power grid business knowledge graph, so that the defect that a traditional pure data-driven model is prone to failure when encountering an unseen working condition is effectively overcome.

Description

Technical Field

[0001] This invention relates to the field of new energy power prediction technology, and in particular to a method and system for new energy power prediction and uncertainty quantification. Background Technology

[0002] With the large-scale grid connection of new energy sources, the requirements for power forecasting and uncertainty quantification accuracy of new energy sources such as wind power and photovoltaics are increasing. Existing power forecasting systems typically adopt a centralized cloud processing architecture or an isolated local forecasting mode at power plants. Their forecasting logic mostly relies on purely data-driven statistical models or conventional machine learning models, achieving forecasts through simple numerical fitting of historical power and meteorological data.

[0003] However, in the actual complex power grid dispatching operation scenario, if a completely centralized cloud architecture is adopted, the centralized uploading of massive high-frequency operation data of the underlying power stations will not only cause great communication delays and data privacy and security risks, but also cannot meet the rapid control needs of the underlying units in response to emergencies.

[0004] If a completely isolated local forecasting model is adopted, the forecasting model will lose its output uncertainty quantification results due to the lack of macro-business knowledge such as regional power grid topology, large-scale weather evolution trends and cross-regional transmission channel status. When encountering meteorological transients or power grid topology switching, the model is prone to outputting serious forecast bias due to the lack of global causal logic support, which in turn endangers the overall dispatch security of the power grid. Summary of the Invention

[0005] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a method and system for predicting renewable energy power and quantifying uncertainty, thereby improving the safety margin of renewable energy power prediction across the entire power grid.

[0006] To achieve the above objectives, a first aspect of this invention proposes a method for predicting new energy power and quantifying uncertainty, implemented based on a three-level distributed architecture consisting of an edge layer, a regional layer, and a cloud layer, comprising:

[0007] Acquire multi-source hierarchical operation data; the multi-source hierarchical operation data includes local unit dynamic characteristic data at the edge layer, regional power grid topology data and regional weather forecast data at the regional layer, and cross-regional power transmission channel status data at the cloud layer;

[0008] A hierarchical power grid business knowledge graph is constructed based on the multi-source hierarchical operation data; the hierarchical power grid business knowledge graph consists of a local knowledge graph at the edge layer, a regional knowledge graph at the regional layer, and a global knowledge graph at the cloud layer.

[0009] The multi-source hierarchical operation data and the hierarchical power grid business knowledge graph are input into the multi-scale causal reasoning model to generate initial prediction sequences of new energy power at each level.

[0010] Based on the initial prediction sequences of new energy power at each level, three-level uncertainty quantification collaborative processing is performed sequentially at the edge layer, regional layer, and cloud layer to generate a target power prediction sequence containing uncertainty intervals.

[0011] The actual power data is acquired, and the real-time prediction error of each level is calculated through a hierarchical dynamic ensemble learning mechanism. Based on the real-time prediction error, the multi-scale causal inference model is updated in a closed loop.

[0012] To achieve the above objectives, a second aspect of the present invention proposes a new energy power prediction and uncertainty quantification system, implemented based on a three-level distributed architecture consisting of an edge layer, a regional layer, and a cloud layer, comprising:

[0013] The multi-source data acquisition module is used to acquire multi-source hierarchical operation data; the multi-source hierarchical operation data includes local unit dynamic characteristic data at the edge layer, regional power grid topology data and regional weather forecast data at the regional layer, and cross-regional power transmission channel status data at the cloud layer.

[0014] The knowledge graph management module is used to construct a hierarchical power grid business knowledge graph based on the multi-source hierarchical operation data; the hierarchical power grid business knowledge graph consists of a local knowledge graph at the edge layer, a regional knowledge graph at the regional layer, and a global knowledge graph at the cloud layer.

[0015] The multi-scale causal prediction module is used to input the multi-source hierarchical operation data and the hierarchical power grid business knowledge graph into the multi-scale causal inference model to generate the initial prediction sequence of new energy power at each level.

[0016] The uncertainty quantification and coordination module is used to perform three-level uncertainty quantification and coordination processing sequentially in the edge layer, regional layer and cloud layer based on the initial prediction sequence of new energy power at each level, to generate a target power prediction sequence containing uncertainty intervals;

[0017] The closed-loop adaptive update module is used to acquire actual power data, calculate the real-time prediction error of each level through a hierarchical dynamic ensemble learning mechanism, and perform closed-loop adaptive update of the multi-scale causal inference model based on the real-time prediction error.

[0018] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described method for predicting new energy power and quantifying uncertainty.

[0019] This invention constructs a three-level distributed architecture consisting of an edge layer, a regional layer, and a cloud layer. In the actual power grid dispatching and operation process, it deeply integrates multi-source hierarchical operation data with hierarchical power grid business knowledge graphs, effectively overcoming the defect of traditional pure data-driven models that are prone to failure when encountering unseen operating conditions.

[0020] In specific business processes, the multi-scale causal inference model can mine an initial prediction sequence that conforms to the actual operation logic of the power grid based on physical topology and operating characteristics. Then, after bottom-up three-level uncertainty quantification collaborative processing, it converges and corrects the statistical error boundary of causal prediction layer by layer, and finally outputs a target prediction sequence with a high reliability confidence interval.

[0021] In addition, the system relies on the actual power data acquired to drive a hierarchical dynamic integrated learning mechanism, enabling the prediction models at each level to perform closed-loop adaptive updates in response to equipment dynamic aging or local climate evolution during long-term actual operation. This significantly improves the safety margin and refined scheduling level of the entire network's new energy power prediction while ensuring the privacy of underlying data and the timeliness of business response. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the new energy power prediction and uncertainty quantification method provided by the present invention;

[0023] Figure 2 This is a waveform diagram of the multi-scale eigenvalue decomposition signal of new energy power based on discrete wavelet transform in the new energy power prediction and uncertainty quantification method provided by this invention.

[0024] Figure 3 This is a heatmap of the regional node correlation matrix and graph attention network weight distribution characteristics in the new energy power prediction and uncertainty quantification method provided by this invention;

[0025] Figure 4 This is the optimal path matching data diagram of dynamic time warping (DTW) of the target knowledge increment and historical attribute evolution curve in the new energy power prediction and uncertainty quantification method provided by this invention.

[0026] Figure 5 This is a three-dimensional response surface plot of the knowledge credibility comprehensive score in the new energy power prediction and uncertainty quantification method provided by the present invention, which varies with the dynamic topology deviation index and the difference in temporal mutation degree.

[0027] Figure 6 This is a two-dimensional curve of the nonlinear evolution of the dynamic kernel bandwidth adjustment factor with real-time meteorological transient gradient in the new energy power prediction and uncertainty quantification method provided by this invention.

[0028] Figure 7 This is a bitmap comparing the error probability distribution of static bandwidth and adaptive evolution kernel bandwidth under meteorological transient conditions in the new energy power prediction and uncertainty quantification method provided by this invention.

[0029] Figure 8 This is a schematic diagram illustrating the implementation of the new energy power prediction and uncertainty quantification system provided by this invention;

[0030] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0031] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0032] The following description, with reference to the accompanying drawings, outlines a new energy power prediction and uncertainty quantification system, method, and electronic device according to embodiments of the present invention.

[0033] Example 1:

[0034] like Figure 1 As shown, this embodiment provides a method for predicting new energy power and quantifying uncertainty. This method is implemented based on a three-level distributed architecture consisting of an edge layer, a regional layer, and a cloud layer.

[0035] Specifically, for modern large-scale new energy grid-connected systems, the edge layer is usually deployed at the local control center of a single wind farm or photovoltaic power station, with real-time second-level data acquisition and local computing capabilities; the regional layer is usually deployed at the municipal or provincial power grid dispatch center, responsible for coordinating the operation and dispatch of multiple stations; and the cloud layer is deployed at the national or cross-provincial wide-area power grid supercomputing center, undertaking macro-meteorological analysis and cross-regional resource allocation tasks.

[0036] This embodiment will analyze the operation process of the overall technical solution in conjunction with the three-level physical architecture:

[0037] Specifically, the method in this embodiment first acquires multi-source hierarchical operational data. This multi-source hierarchical operational data includes local unit dynamic characteristic data at the edge layer, regional grid topology data and regional weather forecast data at the regional layer, and cross-regional transmission channel status data at the cloud layer. In actual new energy power plant operation environments, the acquisition of multi-source data forms the data foundation for all prediction and scheduling models. Local unit dynamic characteristic data covers real-time turbine rotational speed, pitch angle, generator stator temperature, and nacelle vibration frequency of wind turbine generators, or DC-side voltage and MPPT (maximum power point tracking) status parameters of photovoltaic inverters. Regional grid topology data reflects the physical connection impedance and switch status from each power plant to the step-up substation and regional hub substation. Regional weather forecast data includes multi-dimensional gridded meteorological data such as temperature, air pressure, wind speed, wind direction, and irradiance issued by the numerical weather prediction (NWP) system. Cross-regional transmission channel status data records the real-time available transmission capacity and cross-sectional power constraints of ultra-high voltage DC or AC tie lines.

[0038] For example, the above-mentioned acquisition of multi-source hierarchical operation data includes: collecting the local unit dynamic characteristic data at the edge layer, performing desensitization processing using local differential privacy technology, removing outliers using a sliding window mid-value filtering algorithm, and generating edge layer preprocessed data after downsampling processing.

[0039] In this step, the acquisition frequency of local unit dynamic characteristic data is typically as high as 1 Hz or even higher. Since these high-frequency signals are easily mixed with transient spike noise caused by sensor electromagnetic interference or extreme weather conditions, the system uses a sliding window mid-range filtering algorithm to sort the sequence and extract the median within a set time window, thereby replacing the abrupt noise points in the original sequence, thus smoothing the high-frequency abrupt data while preserving the basic signal profile.

[0040] Furthermore, considering that the station operation data involves the core business secrets and infrastructure security of power generation companies, local differential privacy technology is used to inject controlled Laplace random perturbations into the original feature gradient before uploading. This prevents attackers from reverse-engineering the absolute operating parameters of a single specific unit, achieving data usability without visibility. After desensitization and filtering, the system performs downsampling processing, compressing and aggregating high-frequency, second-level data into minute-level edge-layer preprocessed data aligned with the resolution of meteorological data, significantly reducing the communication bandwidth pressure from the edge side to the regional side.

[0041] Optionally, this step further includes: aggregating the anonymized edge layer preprocessed data at the regional layer, combining the regional power grid topology data and regional weather forecast data, and using a linear interpolation algorithm to fill in missing values ​​to generate regional layer preprocessed data. When the regional layer receives data streams uploaded concurrently from multiple edge stations, data packet loss often occurs at some timestamps due to fluctuations in the wide area communication network (such as fiber optic link jitter or microwave communication interference from weather). In this case, the linear interpolation algorithm establishes a linear relationship based on the valid observation data points before and after the missing values, estimates and fills in the data gaps, and ensures the continuity of the time series.

[0042] The acquisition process ultimately includes: acquiring the cross-regional power transmission channel status data at the cloud layer, and performing standardization processing on the edge layer preprocessed data, regional layer preprocessed data, and the cross-regional power transmission channel status data. Cross-level and multi-dimensional data exhibit significant differences in physical dimensions (e.g., voltage is in the kilovolt range, while temperature is in the single digit range). Directly mixing these data can lead to subsequent algorithms failing to converge due to uneven gradient distribution. Standardization processing (such as Z-score processing) maps all feature data to a standard space with a mean of zero and a variance of one, eliminating dimensional differences and improving the equality of numerical computation for features of different dimensions.

[0043] Specifically, after acquiring the data, this method constructs a hierarchical power grid business knowledge graph based on the multi-source hierarchical operation data. This hierarchical power grid business knowledge graph consists of a local knowledge graph at the edge layer, a regional knowledge graph at the regional layer, and a global knowledge graph at the cloud layer. A knowledge graph transforms discrete data points into a structured network with semantic associations and physical rule-based logical relationships.

[0044] For example, after constructing the hierarchical power grid business knowledge graph, the method further includes performing a collaborative update mechanism on the hierarchical power grid business knowledge graph. Specifically, the edge layer extracts dynamic attribute features using historical local operating data, updates the local knowledge graph, and generates local knowledge increment packages which are then uploaded to the regional layer. In practical applications, wind turbine blades wear down over time, and photovoltaic modules also experience degradation. These physical changes are reflected in the operating data as drifts in the dynamic characteristics of the equipment. The edge layer analyzes historical operating data to extract dynamic attribute features that reflect the current true efficiency status of the equipment, such as the updated wind power curve fitting coefficients, and updates them to the local knowledge graph that records the equipment inventory. This bottom-up incremental extraction mechanism avoids the waste of computing power and bandwidth caused by uploading the entire graph.

[0045] Optionally, the collaborative update mechanism further includes: the regional layer aggregating local knowledge increment packages within its jurisdiction, performing knowledge consistency verification to update the regional knowledge graph, and generating regional knowledge increment packages which are then uploaded to the cloud layer. After receiving increment packages from dozens or even hundreds of subordinate stations, the regional dispatch center cannot blindly accept them into the database. If a station falsely reports an update to equipment attributes that violates physical principles due to a communication failure, the consistency verification mechanism will cross-verify the data by combining the regional power grid topology with the status of adjacent stations. For example, if a wind farm reports a sudden and abnormal drop in the efficiency of a large area of ​​its units, while neighboring stations less than five kilometers away and regional meteorological data show normal data, the increment package will be marked as failing verification and intercepted, thereby ensuring the security and reliability of the regional knowledge graph as the central decision-making basis.

[0046] It is also important to note that the cloud layer aggregates all regional knowledge increment packages, uses knowledge fusion algorithms to eliminate cross-regional knowledge conflicts, generates a general knowledge benchmark model, and distributes it to each regional layer and edge layer. The cloud layer possesses a global knowledge graph with a macroscopic perspective, including national-level transmission corridor rules and regional climate characteristics. When different regions have differing understandings of the degradation patterns of the same type of equipment (knowledge conflict), the cloud layer uses knowledge fusion algorithms to re-anchor based on massive samples across the entire network from a weighted perspective, refining a more universal and objective general knowledge benchmark model. Subsequently, a top-down synchronous distribution is executed, giving the underlying power stations a global correction capability to resist local data deviations.

[0047] Specifically, this method inputs the multi-source hierarchical operational data and the hierarchical power grid business knowledge graph into a multi-scale causal inference model to generate initial prediction sequences for renewable energy power at each level. Traditional black-box neural networks often only look for co-occurrence correlations in the data, and their prediction accuracy tends to decrease significantly during periods of drastic weather changes. The multi-scale causal inference model, by introducing prior constraints from the physical knowledge graph, forces the model to learn the real physical causal transmission paths (e.g., pressure difference leads to increased wind speed, increased wind speed leads to increased rotor work, and thus increased power), rather than learning superficial data correlations.

[0048] For example, the step of inputting the multi-source hierarchical operation data and the hierarchical power grid business knowledge graph into a multi-scale causal inference model to generate initial prediction sequences for new energy power at each level specifically includes, at the edge layer: performing discrete wavelet transform on the local unit dynamic characteristic data in the multi-source hierarchical operation data to decompose it into trend term feature sequences, fluctuation term feature sequences, and noise term feature sequences. The actual wind power or photovoltaic output power is a non-stationary sequence containing mixed information across multiple frequency bands. Discrete wavelet transform uses scaling and translation wavelet basis functions to decompose the original signal from a single time domain into a multi-scale space combining the frequency and time domains. The trend term feature sequence mainly reflects the low-frequency, long-period physical laws driven by day-night alternation or macroscopic atmospheric circulation; the fluctuation term feature sequence reflects the mid-to-high frequency dynamic fluctuations caused by local topographic eddies or small cloud cover; and the noise term feature sequence removes random and physically meaningless high-frequency sensor noise.

[0049] like Figure 2 The waveform of the multi-scale eigenvalue decomposition signal of new energy power based on discrete wavelet transform is shown. The figure consists of four sub-figures arranged vertically, with sampling time as the horizontal axis (in minutes) and output power as the vertical axis (in megawatts).

[0050] Figure 2 The first subplot from top to bottom first shows the raw power signal from 0 to 1000 minutes with a dark gray curve. This signal exhibits significant mixed fluctuation characteristics, representing the unprocessed local unit dynamic output data collected at the edge layer.

[0051] After discrete wavelet transform decomposition Figure 2 The second subplot uses an orange-red curve to extract the trend term feature sequence, whose value shows a smooth long-period fluctuation between 40 MW and 90 MW. This objectively reflects the low-frequency physical laws driven by macroscopic atmospheric circulation or day-night alternation.

[0052] Figure 2 The third subplot shows the characteristic sequence of the fluctuation term with a blue curve. Its amplitude is concentrated in the mid-frequency oscillation range between -7 MW and +7 MW, revealing the local dynamic power output fluctuation characteristics caused by local topographic eddies or small cloud obstructions.

[0053] Figure 2 The lowest subplot uses a green curve to separate the noise feature sequence, whose high-frequency amplitude is limited to a narrow range of -5 MW to +5 MW, representing the separated random and physically meaningless sensor noise floor.

[0054] By separating waveform features of different colors and comparing the values ​​of horizontal and vertical coordinates, it is demonstrated that the wavelet decomposition step in this embodiment can effectively deconstruct non-stationary complex operating data into independent sequences with clear physical orientations. This provides a clear data foundation for subsequent multi-scale causal inference models to mine the real physical causal transmission path and enhances the anti-interference capability of initial power prediction at the single-site level.

[0055] Optionally, this step further includes: extracting the station equipment ledger and local topology constraint information from the local knowledge graph; using a causal discovery algorithm to combine the station equipment ledger and local topology constraint information to mine the causal path between the trend term feature sequence and the fluctuation term feature sequence and the local power output; and calculating the minute-by-minute power output characteristics at the single station level based on the causal path to generate an initial prediction sequence for edge-layer renewable energy power. At the edge, the causal discovery algorithm (such as the PC algorithm or FCI algorithm) is injected with physical prior information. For example, the local topology constraint information clarifies the maximum number of photovoltaic strings that can be connected to a certain combiner box, and the station equipment ledger clarifies the maximum withstand current of the converter.

[0056] When searching for feature relationships, the algorithm directly prunes and removes spurious correlation paths that violate these hardware physical limits. After filtering out the true physical causal paths, the model follows these directed acyclic graph paths, combining the decomposed trend and fluctuation terms, to positively deduce the ultra-short-term power trend at the single-site level, thereby generating an edge-layer renewable energy power initial prediction sequence with extremely strong anti-interference capabilities.

[0057] It is important to note that the generation of prediction sequences is equally crucial at the macro level. The process of inputting the multi-source hierarchical operational data and the hierarchical power grid business knowledge graph into a multi-scale causal inference model to generate initial prediction sequences for renewable energy power at each level specifically includes the following at the regional and cloud layers: At the regional layer, the initial prediction sequence for renewable energy power output from the edge layer, the regional meteorological forecast data, and the regional knowledge graph are input into the regional causal inference model, outputting a regional-level aggregated power initial prediction sequence. Regional dispatching is not simply an addition of the predicted power of each power station, but rather considers the "spatial-temporal transmission effect" of the regional meteorological system evolution. The regional causal inference model, combining power grid congestion constraints and spatial wind corridor topology in the regional knowledge graph, infers how long it will take for the output attenuation of upwind power stations to be transmitted to downwind power stations, thus outputting a regional-level aggregated power initial prediction sequence with spatial correlation compensation.

[0058] Specifically, at the cloud layer, the regional-level aggregated power initial prediction sequence output from the regional layer, global meteorological forecast data, and the global knowledge graph are input into the global causal inference model at the cloud layer, outputting a network-wide global power initial prediction sequence. This process further introduces a national-level cold wave or monsoon cross-regional migration model and the transmission capacity limitations of inter-provincial corridors, locking in a baseline for the overall new energy output trend from a macro perspective.

[0059] Specifically, the core advantage of this method lies in overcoming the limitations of simple point prediction. It executes initial prediction sequences of new energy power at each level, and then performs three levels of uncertainty quantification collaborative processing sequentially at the edge layer, regional layer, and cloud layer to generate a target power prediction sequence containing uncertainty intervals. Grid dispatch requires not only an absolute prediction value, but also knowledge of the safe upper and lower limits of this value's fluctuations—that is, the uncertainty boundary—in order to rationally allocate thermal power spinning reserve capacity.

[0060] For example, based on the initial prediction sequence of new energy power at each level, a three-level uncertainty quantization collaborative processing is performed sequentially at the edge layer, regional layer, and cloud layer, including: performing causal-statistical dual quantization processing at the edge layer to generate a local initial uncertainty interval. Because micro-meteorological conditions are highly nonlinear, single-dimensional quantization often leads to an excessively narrow uncertainty interval estimate.

[0061] Optionally, the causal-statistical weighting process performed at the edge layer specifically includes: extracting causal path confidence parameters based on the local knowledge graph, and constructing a first-layer confidence interval based on the causal path confidence parameters. In the aforementioned causal inference process, each causal path connection edge has a corresponding confidence score (reflecting the robustness of the physical logic). If the confidence of the causal path on which the prediction mainly relies is low within a certain period (e.g., encountering historically rare extreme weather patterns), the system will expand the first-layer confidence interval to accommodate unknown risks.

[0062] Furthermore, the fuzzy C-means clustering algorithm is used to divide the acquired local historical prediction error data into multiple discrete power error segments. In actual operation, the error distribution of renewable energy power is highly coupled with the current output state. For example, the errors are usually smaller under full-power and extremely low-output conditions, while the error distribution is most severe under intermediate ramp-up conditions. The fuzzy C-means clustering algorithm achieves soft classification and isolation of complex non-Gaussian error features by calculating the membership degree of each error sample to the cluster centers of different output categories.

[0063] Next, within each power error segment, a nonparametric kernel density estimation algorithm is used to fit the error probability distribution map, and the first-level confidence interval is statistically corrected based on the error probability distribution map to generate the local initial uncertainty interval. The nonparametric kernel density estimation algorithm does not rely on any pre-assumed fixed probability distribution shape, but rather uses historical error data points themselves to fit an objective probability density curve. The specific formula of the kernel density estimation algorithm used in this step is as follows:

[0064] ;

[0065] in, The error probability density function evaluation value under the target error range; This represents the total number of historical error samples contained within the currently defined power error segment; This is the basic kernel smoothing bandwidth parameter used to control the smoothness of the probability density distribution curve; This is a pre-defined Gaussian kernel function benchmark calculation unit; The independent variable error scale value is the density to be evaluated; For the first The actual historical prediction error sampling points.

[0066] Using this formula, the system can accurately depict the "long-tail shape" of error divergence under the current physical conditions. Based on this shape, the statistical distribution bandwidth at a specific confidence level (such as 90%) is extracted and corrected by joint intersection or union with the first-level confidence interval. This results in the formation of a local initial uncertainty interval on the edge side that has both physical causal perception and historical statistical regularity.

[0067] It is also important to note the evolution of the three-level collaborative approach: at the regional level, the local initial uncertainty intervals of the subordinate edge layers are integrated, and a causal-statistical-Bayes triple weighting process is performed based on the regional-level prediction data to generate a regional-level initial uncertainty matrix. The regional level faces not only the randomness error of data from a single field station, but also the uncertainty of model understanding brought about by a large number of heterogeneous field stations.

[0068] For example, the causal-statistical-Bayesian weighting process performed at the regional level based on regional prediction data is constructed using a Bayesian deep learning model; the objective function of the Bayesian deep learning model is defined as a loss function including an uncertainty penalty term, and its mathematical expression is:

[0069] ;

[0070] in, The calculated uncertainty loss evaluation index not only measures the magnitude of the absolute error in the prediction, but also rigorously examines the model's confidence in its own output boundaries. The total number of samples in the input region causal inference model; This is the index number of the sample in the current batch processing process; For the first The initial predicted power value for each sample; For the first The actual acquisition power value of each sample; For the first The prediction variance of each sample is the uncertainty quantization width of the model output.

[0071] The mechanism of this formula involves two mutually constraining game logics: the first requires that when the residuals are large, the prediction variance should also be amplified to accommodate the residuals, otherwise a huge penalty loss will occur; the second, the logarithmic function term, acts as a regularization constraint, penalizing the phenomenon of excessive divergence caused by the model setting the variance to infinity. Guided by this rigorous objective function, the initial uncertainty matrix output by the regional model can objectively and accurately reflect the comprehensive risk boundary after the aggregation of multiple sites.

[0072] Finally, uncertainty quantification is achieved through cloud-based aggregation: all regional-level initial uncertainty matrices are integrated at the cloud layer, and global quantification correction is performed by introducing transmission channel uncertainty parameters to generate a target power prediction sequence containing a global confidence index. If a cross-provincial or cross-regional DC transmission corridor experiences extreme thermal voltage drop or bipolar blocking faults, the sudden change in its transmission capacity will directly limit the absorption capacity of new energy sources. Injecting such uncertainties from physical channel constraints into the model ensures that the upper and lower boundaries of the final prediction scheme fully conform to the actual grid's dispatch acceptance limits.

[0073] Specifically, no static model can withstand concept drift over long periods of time. This method acquires actual power data, calculates the real-time prediction error at each level through a hierarchical dynamic ensemble learning mechanism, and performs closed-loop adaptive updates to the multi-scale causal inference model based on these real-time prediction errors. This closed-loop update mechanism is the core link in maintaining the long-term effectiveness and prediction accuracy of the entire system.

[0074] For example, the step of calculating the real-time prediction error at each level through a hierarchical dynamic ensemble learning mechanism and performing closed-loop adaptive updates to the multi-scale causal inference model based on the real-time prediction error includes: constructing a dynamic ensemble model library containing multiple heterogeneous base models at each level, wherein the heterogeneous base models are combined to form the multi-scale causal inference model. The heterogeneous base model library integrates underlying algorithmic foundations such as the LightGBM model, the XGBoost tree model, and long short-term memory artificial neural networks. Different models have their own advantages and disadvantages when dealing with calm weather, severe air pressure fluctuations, or extreme cold conditions. This mechanism of joint decision-making by multiple heterogeneous models itself enhances the robustness of the system.

[0075] Optionally, the update step further includes: independently calculating the real-time root mean square error of each heterogeneous base model within a preset time window at each level; and adjusting the output weight components of each heterogeneous base model in the ensemble decision based on the real-time root mean square error using a preset exponential decay dynamic weight update mechanism to complete the closed-loop adaptive update of the multi-scale causal inference model. With the changing seasons, such as the transition from the calm autumn to the typhoon-prone and highly turbulent summer, some previously high-performing base models may experience severe performance degradation, while the advantages of some deep learning base models that are adept at capturing dramatic fluctuations will become apparent. The real-time root mean square error, as an objective and impartial metric, accurately evaluates the recent prediction errors of each model at the end of each rolling time window.

[0076] It is important to note that the weight adjustment has a clear algorithmic basis: the preset exponential decay dynamic weight update mechanism is adopted, and the output weight components of each heterogeneous basis model in the ensemble decision are adjusted based on the real-time root mean square error. The corresponding weight update formula is as follows:

[0077] ;

[0078] in, Indicates the first After the weight update iteration, the dynamic ensemble model library... The current unnormalized weight assignment values ​​of the heterogeneous base models; Indicates the first After the nth weight update iteration, the th Historical unnormalized weight assignments for heterogeneous base models; This represents the preset constant exponential decay coefficient, which determines the system's sensitivity to new error feedback and is usually set according to the frequency of sudden climate changes at the site. Indicates the first The heterogeneous base model in the first... The real-time root mean square error assessment value is updated within the next time window. Due to the drastic nonlinear decay characteristic of the exponential function, if a base model makes a significantly erroneous prediction that deviates severely from the actual physical laws within the current time window (causing...), the error will be mitigated. (A sudden increase) will immediately cause the formula to significantly reduce the weight it occupies, and the system control will smoothly transfer to the currently better-performing benchmark model. This adaptive update mechanism achieves seamless and smooth parameter reorganization, eliminating the complex and expensive cost of manual offline model retraining.

[0079] Based on existing technology analysis, traditional power prediction schemes typically rely on centralized "large-scale individual models" or decentralized, isolated "small-scale edge algorithms." When local power plants encounter unprecedented extreme micro-meteorological changes, purely data-driven models are often limited by the distribution of historical samples, resulting in severe "extrapolation failures." Furthermore, the single, fixed confidence interval they provide cannot reflect the true physical boundaries of the power grid.

[0080] This embodiment forces the algorithm to learn real physical laws by deeply binding a hierarchical power grid business knowledge graph with physical semantics and a multi-scale causal reasoning model in a three-level cloud-edge collaborative architecture. Then, it establishes a nested three-level uncertainty quantification envelope by combining causal logic confidence, local statistical nonparametric distribution, and Bayesian deep learning. Finally, it eliminates degraded weights in a closed loop with a dynamic integrated learning feedback mechanism.

[0081] This complete set of technical solutions enables the power grid dispatch center to fully protect the privacy and security of massive amounts of downlinked operational data, while also acquiring new energy power prediction curves with rigorous physical logic and high adaptability. It can maintain high-precision safety margin dispatch capability when facing complex and unknown operating conditions, thus achieving a balance between power grid absorption efficiency and operational safety.

[0082] Example 2:

[0083] In practical applications of new energy grid dispatching, edge layer nodes (such as wind farms and photovoltaic power stations) are mostly deployed in remote areas with harsh natural environments. Their local sensors are frequently exposed to severe physical conditions such as extreme high and low temperatures, strong electromagnetic interference, freezing, or sandstorms. This complexity of the physical environment means that the "local knowledge increment packages" uploaded from the edge layer (such as the changing characteristics of equipment dynamic output curves and temporary reconstruction parameters of the local network topology) may contain false and abnormal data caused by sensor short circuits, communication errors, or even malicious network attacks. If the system lacks a rigorous mechanism to verify the authenticity of the knowledge increments, allowing these deceptive or erroneous knowledge increments to directly enter the regional knowledge graph will cause severe distortion of the graph data, leading to severely distorted power prediction results from the multi-scale causal inference model running on that graph.

[0084] Therefore, this embodiment aims to reveal a regional knowledge consistency verification method based on spatiotemporal cross-validation, thereby ensuring the real-time update of the map while building a high-security defense mechanism against underlying data distortion.

[0085] Specifically, the method in this embodiment performs knowledge consistency verification to update the regional knowledge graph, including: constructing a regional node association matrix based on the regional power grid topology data, and extracting a set of neighboring nodes matching the target edge node from the regional node association matrix according to spatial distance and line connection status. In actual physical power grids, there are usually strong meteorological correlations and electrical couplings between new energy power stations within the same geographical area. In order to transform this implicit physical dimension association into an explicit mathematical expression that can be processed by computer algorithms, the regional dispatch center server needs to first read the static geographic information system coordinates and the dynamic energy management system (EMS) wiring diagram. Based on this, the system constructs a high-dimensional regional node association matrix. This matrix not only records the two-dimensional or three-dimensional straight-line physical distance between each edge node in geographic space, but also strictly maps the connection path status between nodes on the electrical network (e.g., whether they are connected to the same busbar, the equivalent AC impedance of the transmission line, the current open / closed state of the tie switch, etc.).

[0086] Building upon this foundation, when the system receives a map update request from a specific field station (i.e., the target edge node), the algorithm uses that target edge node as the origin and performs multi-dimensional distance retrieval optimization within the regional node association matrix. The system sets specific geographical radius thresholds and electrical impedance connectivity thresholds, packaging surrounding field station nodes that simultaneously meet these conditions into a set, formally establishing it as a set of neighboring nodes with strong environmental convergence and electrical synchronization with the target edge node. The significance of this extraction process lies in finding reference nodes with physical reference value for subsequent knowledge verification, ensuring that any isolated anomalies can be revealed through comparison with the surrounding population.

[0087] Optionally, after locating the physical reference group, the system enters the feature extraction stage, specifically: extracting the target knowledge increment features uploaded by the target edge node, and the reference knowledge increment features synchronously uploaded by each neighboring node in the neighboring node set. The target knowledge increment features here are not simply the raw collected data, but highly structured parameters refined through edge-side computation. For example, in a wind power generation scenario, this feature might manifest as the offset of the polynomial fitting coefficients of the power curve within a specific wind speed range; in a photovoltaic power generation scenario, it might manifest as the change in the convergence time constant of the inverter's maximum power point tracking (MPPT) algorithm under a specific irradiance.

[0088] To establish an effective spatiotemporal comparison, the regional layer server simultaneously sends synchronous retrieval commands to the established set of neighboring nodes. These commands require each neighboring node to retrieve incremental feature records whose upload timestamps are within the same data observation window as the target edge node. By parsing these synchronously collected data records, the system constructs multiple sets of incremental features for reference knowledge. Through this preliminary step, the system successfully constructs a structured feature tensor matrix containing the state of the central target and the surrounding environment in the regional layer memory, providing a solid data foundation for deep pattern recognition and bias quantification.

[0089] For example, when faced with highly nonlinear and dimensionally complex feature tensors, the system calculates the feature deviation between the target knowledge increment feature and the reference knowledge increment feature based on a graph attention network, obtaining a dynamic topological deviation index. A graph attention network is an advanced deep learning model capable of processing data with arbitrary graph structures. In the practical application scenario presented in this paper, due to the slight differences between the micro-topography (such as ridges, windward slopes, and leeward slopes) of each neighboring node and the target edge node, the reference value of each neighboring node for verifying the anomaly of the target node is different. The graph attention network, through its built-in self-attention mechanism layer, can autonomously learn and calculate the influence weights of different neighboring node features on the target node features. Subsequently, based on these attention weights, the system performs an aggregation calculation of the weighted distance between the target and reference features, thereby accurately quantifying the degree of dispersion of the target site's knowledge changes from the normal evolutionary patterns of the surrounding population.

[0090] Specifically, this calculation process relies on the following feature deviation quantization formula:

[0091] ;

[0092] in: The dynamic topology deviation index, which represents the final calculated output, is larger in scalar value, meaning that the knowledge increment of the target edge node is more inconsistent with the overall performance of the region, and the stronger the spatial heterogeneity. This represents the total number of neighboring nodes contained in the neighboring node set, which determines the breadth of the graph traversed by the algorithm. This is the local index number of the neighboring nodes, used for point-by-point iteration in the summation operation; For the output of the hidden layer of the graph attention network, targeting the first The spatial attention weight coefficient of each neighboring node represents the actual comparative reference value that the network model believes that the neighboring node has for the target node under the current meteorological and topological background. The total dimension of the feature vector representing the incremental features of knowledge determines the richness of the comparison features; This refers to the index number of a specific dimension in the feature vector; The incremental feature of the target knowledge is represented in the first... Quantized extracted values ​​on each feature dimension; Indicates the first The incremental features of the reference knowledge synchronously uploaded by the neighboring nodes in the first... Quantized extracted values ​​on each feature dimension.

[0093] Through the deep calculation of the above formula, the system transforms the complex spatial physical correlation into a single objective index output, effectively eliminating the normal noise caused by terrain differences and accurately capturing the true spatial dimension knowledge mutation.

[0094] like Figure 3 This diagram displays a heatmap showing the regional node association matrix and the weight distribution characteristics of the graph attention network. The horizontal axis represents the reference node sequence (number of nodes), the vertical axis represents the target node sequence (number of nodes), and the color legend on the right represents the graph attention weights (dimensionless).

[0095] exist Figure 3 In the matrix, 10 regional power grid nodes form a 10-row, 10-column association feature matrix. The color intensity in the matrix directly reflects the strength of the spatial topological association between nodes and the magnitude of the feature influence. The closer the color is to bright yellow, the higher the weight value, while the closer the color is to dark blue, the lower the weight value.

[0096] from Figure 3 As can be observed, the diagonal squares are the brightest yellow, with a value of 1, representing that the characteristic state of each node has the highest fundamental influence on its own evaluation. Taking the fifth target node on the vertical axis as an example, the color at its intersection with the fourth reference node on the horizontal axis is bright cyan or light yellow, with a corresponding graph attention weight value of 0.85. This objectively reflects that the two nodes are geographically close or on an electrical network, and have a strong physical environment convergence.

[0097] Therefore, during knowledge consistency verification, the incremental features of the reference knowledge synchronously uploaded by the fourth node play a crucial role in identifying the features of the fifth target node. Conversely, the intersection of the fifth target node and the first reference node on the horizontal axis is dark blue, with a corresponding weight value of only 0.12, indicating that the electrical correlation between the two in spatial topology is extremely low.

[0098] Through this intuitive heatmap data distribution, the graph attention network in this embodiment can autonomously learn and quantify the actual reference value of different neighboring nodes for verifying anomalies at target edge nodes based on objective physical distances and line connection status. This processing method can effectively filter out irrelevant regional noise interference in the spatial dimension, providing reliable data support for calculating accurate dynamic topology deviation indices, thereby improving the accuracy of identifying false anomaly knowledge increment packets at the regional layer.

[0099] Specifically, to avoid misjudgments caused by simultaneous changes in the overall regional climate, such as a large cold wave causing abrupt changes in the characteristics of all stations, this scheme introduces a deeper examination of the temporal dimension in addition to spatial comparison. System execution: Extract the historical knowledge attribute change curves of the target edge node within a preset historical time window, and use a dynamic time warping algorithm to calculate the degree of abrupt change between the target knowledge increment features and the historical knowledge attribute change curves, obtaining the temporal abrupt change difference. The performance degradation of physical equipment or the evolution of the natural environment usually follows a certain historical smoothing inertia. The historical knowledge attribute change curve records the evolution trajectory of the target edge node's operational characteristics over the past few months or even years. Because feature changes caused by real physical laws may exhibit local time distortions and drifts on the timeline due to equipment response delays or communication latency, traditional Euclidean distance calculation methods are prone to generating significant evaluation errors in such time-disaligned scenarios.

[0100] To address this, the system employs a Dynamic Time Warping algorithm. This algorithm finds an optimal warping path that minimizes the cumulative distance within a two-dimensional time-aligned grid, allowing for flexible stretching or compression of data points on the time axis. This enables the algorithm to overcome the errors caused by time drift and rigorously compare the shape profile and evolutionary trend of the target knowledge increment features to ensure there is no significant break from the inherent curve morphology formed over a long historical period. After path cost calculation by this algorithm, the system outputs a temporal abrupt change difference, an indicator used to objectively characterize whether the currently uploaded knowledge increment significantly deviates from the historical physical evolutionary inertia of that node.

[0101] like Figure 4 This graph displays the dynamic time-warped optimal path matching data between the target knowledge increment and the historical attribute evolution curve. The horizontal axis represents the time series of historical attribute changes in days, and the vertical axis represents the time series of target knowledge increment in days. The color legend on the right represents the local feature difference degree in dimensionless units.

[0102] Figure 4The image uses a two-dimensional time-aligned grid to present the feature distance matrix at different time points. A color closer to bright yellow indicates greater local feature difference, while a color closer to dark blue indicates smaller local feature difference. Figure 4 The red solid curve in the diagram represents the optimal regular path found by the algorithm.

[0103] from Figure 4 The distribution of grid data shows that from day 1 to day 30, the red optimal regular path extends close to the grid diagonal and the area it passes through is dark blue. This indicates that the knowledge increment of the target edge node and the historical attribute evolution trajectory maintain a high degree of synchronization and consistency during this stage.

[0104] Between days 30 and 50, bright yellow abrupt color blocks representing high dissimilarity appeared in the diagonal region of the grid. Using absolute time alignment would result in significant evaluation errors. At this point, the optimal red regular path exhibited elastic stretching characteristics, actively deviating from the diagonal and bending to the right and upward, bypassing the high-cost yellow barrier area, and re-establishing the time node mapping along the dark blue low-dissimilarity valley.

[0105] The offset and waveform regularization of this curve intuitively reflect the system's ability to penetrate time drift phenomena such as device response lag or communication delay. The minimum distance obtained by accumulating along this optimal regularization path is the final output's time-series abrupt change difference.

[0106] Through the evolution verification of this two-dimensional data graph, the system can objectively judge whether the currently uploaded knowledge increment seriously deviates from the historical physical evolution inertia of the node itself, thus providing rigorous judgment support for intercepting false and abnormal data.

[0107] It is important to note that after obtaining independent evaluation metrics in both spatial and temporal dimensions, the system enters the final decision fusion stage, which involves: fusing the dynamic topology deviation index with the temporal abrupt change difference, and calculating a comprehensive knowledge credibility score to characterize the security of the target knowledge increment package. The formula for calculating the comprehensive knowledge credibility score is as follows:

[0108] ;

[0109] in: The comprehensive knowledge credibility score, calculated as a whole, serves as the final quantitative basis for the system to make subsequent interception or release decisions. This score is strictly defined mathematically within a numerical range from zero to one. This represents the preset spatial topology weighting coefficient, the value of which is configured offline based on the spatial correlation of the power grid in the region. This represents the preset time evolution weighting coefficient, whose value is related to the expected lifetime and smoothing degradation constant of the physical equipment, and and The scalar sum is always equal to one; This represents the dynamic topology deviation index calculated by the graph attention network in the preceding steps; This represents the degree of difference in temporal abrupt change calculated by the dynamic time warping algorithm.

[0110] The mechanism design of this formula utilizes a negative exponential decay function based on the natural constant. The practical significance of this mathematical choice lies in the fact that when the spatial deviation from the exponential level or the temporal abrupt change in difference is within a very small normal fluctuation range, the output decay of the negative exponential function is very gradual, allowing the system to maintain extremely high tolerance. However, once the deviation from the exponential level or the difference exceeds the inflection point of conventional physical limits, the function's output value will rapidly decline, causing the overall knowledge credibility score to plummet to zero. This sensitive nonlinear mapping mechanism ensures that the system possesses an extremely acute defensive response trigger capability against any potentially high-risk knowledge contamination events.

[0111] like Figure 5 This diagram presents a three-dimensional response surface plot showing how the overall knowledge credibility score varies with the dynamic topological deviation index and the temporal mutation difference. The horizontal axis represents the dynamic topological deviation index (dimensionless), the vertical axis represents the temporal mutation difference (dimensionless), and the color legend on the right side represents the overall knowledge credibility score (dimensionless).

[0112] Figure 5 The system presents a three-dimensional surface that decays nonlinearly from the highest point outwards. The bright yellow area at the top of the surface corresponds to the topological deviation and temporal mutation values ​​that are close to zero. At this point, the comprehensive knowledge credibility score is close to its maximum value of one, indicating that the target knowledge increment package is within the normal fluctuation range in both spatial comparison and temporal evolution, and the system maintains a high degree of tolerance and trust in it.

[0113] As the dynamic topology deviation from the exponent on the horizontal axis or the temporal abrupt change in the vertical axis gradually increases, the surface color transitions from bright yellow to green, and finally drops sharply to a deep blue trough in the region where the deviation from the exponent or the difference exceeds two, at which point the overall score falls below 0.2. This surface morphology intuitively reflects the mapping mechanism of the negative exponential decay function.

[0114] When spatial deviations or temporal abrupt changes exceed the inflection point of normal physical fluctuations, the system's score output exhibits a sharp drop, which can quickly reduce the score below the safety and reliability threshold to trigger the interception action of the anomaly defense mechanism.

[0115] The presentation of this three-dimensional visualized data demonstrates that the comprehensive scoring calculation logic used in this embodiment has a keen ability to defend against and respond to potential high-risk knowledge contamination events. This enables the update decision of the regional knowledge graph to both smoothly accommodate normal local feature shifts and objectively and rigorously isolate false and abnormal data at the underlying level.

[0116] For example, after the scoring is completed, the system executes strict branch control logic. First, when the comprehensive knowledge credibility score is greater than or equal to a preset security and credibility threshold, the target knowledge increment package is determined to have passed the consistency check, allowing it to be integrated into the regional knowledge graph. The security and credibility threshold is a baseline pre-defined by the regional dispatch center based on power grid safety level standards. When the comprehensive score reaches or crosses this threshold, it means that although the target knowledge increment may contain changes in certain characteristic values, these changes have found reasonable spatial verification in the synchronous evolution of surrounding neighboring nodes, and do not significantly deviate from their own historical evolutionary inertia. Therefore, the system has sufficient logical basis to believe that the increment is generated due to real physical environmental changes (such as weather events or normal equipment aging).

[0117] The system then unlocks write access to the knowledge graph database and uses the knowledge graph's native structured update statements to integrate these latest feature nodes and relational edges into the central layer of the regional knowledge graph. This timely fusion action ensures that subsequent multi-scale causal inference models can always perform power prediction deductions based on the freshest and most accurate physical knowledge, maintaining the prediction system's real-time adaptability to the external environment.

[0118] Optionally, the system's anomaly defense mechanism will activate when the score fails to meet the standard: when the comprehensive knowledge credibility score is less than the security and credibility threshold, it is determined that the target knowledge increment package has failed the consistency check, the target knowledge increment package is intercepted and moved into the anomaly knowledge isolation zone, and a sensor device anomaly self-check command is issued to the target edge node. A score below the benchmark threshold indicates that the increment package exhibits significant data deviation characteristics in the spatiotemporal cross-validation network, which is likely due to non-natural physical factors such as obstruction of local anemometer operation, hardware failure of inverter data acquisition module, or illegal intrusion of communication gateway.

[0119] To protect the reliability of the prediction foundation, the interception gateway in the system server will forcibly terminate the process of adding the incremental package to the database. It is worth noting that the system does not simply discard these erroneous data that fail validation; instead, it migrates them entirely to an anomaly knowledge isolation zone located in a non-core storage area. The existence of this anomaly knowledge isolation zone provides power grid maintenance personnel and the backend fault diagnosis expert system with invaluable accident tracing snapshots and offline analysis samples. Simultaneously, the regional layer server will issue high-priority sensor device anomaly self-check commands to the underlying control unit of the target edge node via a secure bidirectional communication protocol.

[0120] Upon receiving the instruction, the edge control unit will trigger the underlying firmware-level diagnostic routines, such as performing overload scans on the input voltages of all peripheral sensors, verifying the integrity of the data communication protocol checksum, and even executing a hardware soft reset. This top-down closed-loop feedback control not only prevents data contamination but also proactively intervenes in the health management process of the underlying hardware.

[0121] In summary, when examined in conjunction with existing technologies, traditional distributed power grid knowledge or model update architectures are often caught in a conflicting set of technical problems:

[0122] On the one hand, the dispatch center urgently needs the underlying stations to frequently upload high-frequency data to maintain the timeliness of the model's perception; on the other hand, the frequent data acquisition failures of the underlying stations are easily transmitted back along the update link and interfere with the upper core prediction network.

[0123] The method disclosed in this embodiment establishes a fully automatic graph anomaly data identification and filtering mechanism without weakening the underlying update frequency by cleverly constructing a spatiotemporal cross-validation model that orthogonally couples the spatial dimension (region node association matrix and graph attention network) and the temporal dimension (historical evolution curve and dynamic time warping) at the region layer.

[0124] This comprehensive technical solution can precisely extract sensor fault codes and distortions hidden within massive incremental data packets. This not only ensures the high objectivity and accuracy of the physical data foundation upon which the knowledge graph and causal reasoning models rely, but also significantly shortens the exposure cycle and investigation costs of underlying hardware vulnerabilities in new energy power plants through its linkage with anomaly isolation and underlying self-checking systems. This has profound practical engineering significance for improving the overall scheduling resilience and predictive robustness of modern high-proportion new energy grid-connected systems.

[0125] Example 3:

[0126] Based on the aforementioned Embodiment 1 and Embodiment 2, this embodiment further elaborates on the micro-adaptive adjustment mechanism of the statistical correction link in the edge layer.

[0127] Specifically, this embodiment focuses on the dynamic optimization process of nonparametric kernel density estimation algorithms when facing extreme or transient weather conditions. In the actual operation of new energy power systems, the output characteristics of wind and solar energy not only depend on large-scale meteorological evolution, but are also affected by the drastic instantaneous fluctuations in local micro-meteorological conditions at the power station. For example, the "sudden wind speed changes" often encountered by wind farms, and the "cloud obstruction transients" experienced by photovoltaic power stations under cloudy weather, can both cause the statistical characteristics of power prediction errors to change drastically in a short period of time. If only a statistical quantification model with fixed parameters is used, it is often impossible to capture the "long-tail effect" of error distribution at transient moments, resulting in insufficient coverage of the uncertainty interval.

[0128] Therefore, this embodiment introduces meteorological transient gradient features to drive the real-time evolution of the bandwidth parameter of the kernel density estimation algorithm, so as to achieve accurate dynamic compensation for the first layer confidence interval.

[0129] Specifically, the process of fitting the error probability distribution map of the nonparametric kernel density estimation algorithm provided in this embodiment, and statistically correcting the first layer confidence interval based on the error probability distribution map to generate the local initial uncertainty interval, firstly executes: acquiring the edge meteorological sensor data at the current observation time, calculating the slope of meteorological element changes within the preset time sliding window, and generating real-time meteorological transient gradient feature values.

[0130] In practical applications, the edge layer control unit collects high-frequency meteorological sensor data in real time through meteorological stations, anemometers, or irradiance meters deployed within the site. For wind farms, the meteorological element mainly refers to the real-time wind speed at the hub height; for photovoltaic power plants, it mainly refers to the total irradiance on a horizontal or inclined plane. It is also important to note that to eliminate random sensing noise and extract effective physical change trends, the system sets a preset time window, such as 5 minutes or 10 minutes. By performing least squares regression or first-order difference calculations on the meteorological element sequence within this window, its slope over time is extracted. The larger the absolute value of this slope, the more drastic the transient change in the current meteorological environment. The resulting real-time meteorological transient gradient feature value becomes a key physical guide for subsequent adjustment of the statistical smoothing scale (i.e., kernel bandwidth).

[0131] For example, after generating the real-time meteorological transient gradient feature values, this method performs the following: based on the real-time meteorological transient gradient feature values, calculates a dynamic kernel bandwidth adjustment factor to adjust the probability distribution width. In statistics, the bandwidth parameter of kernel density estimation directly determines the smoothness of the probability density curve. Too small a bandwidth will cause the distribution curve to be too volatile (overfitting), while too large a bandwidth will mask the detailed features of the distribution (underfitting). During periods of stable weather, prediction errors are usually concentrated around the mean, and a smaller bandwidth should be used to accurately characterize the high-probability interval; however, during periods of meteorological transients, prediction errors often become significantly dispersed, and the variance of the distribution increases dramatically. In this case, the bandwidth must be dynamically increased to accommodate potential "abnormal" error samples.

[0132] Therefore, this embodiment constructs a nonlinear mapping relationship, and the specific formula for calculating the dynamic kernel bandwidth adjustment factor is as follows:

[0133] ;

[0134] in: This represents the calculated dynamic kernel bandwidth adjustment factor, which is a dimensionless multiplier greater than or equal to 1; This represents the preset sensitivity scaling constant, used to adjust the baseline sensitivity of the impact of meteorological fluctuations on bandwidth. This constant is usually set empirically based on the coefficient of variation of historical meteorological data from the station. This refers to the real-time meteorological transient gradient feature value obtained in the aforementioned steps; This represents the preset historical baseline meteorological gradient mean, which serves as a normalized benchmark and reflects the meteorological fluctuation level of the station under normal conditions. This represents the nonlinear amplification exponent constant used to control the intensity of the mutation response, when When the value is greater than 1, the system will respond exponentially to drastic weather changes, forcibly expanding the breadth of the probability distribution to ensure safety.

[0135] This formula allows the system to convert real-time sensed transient meteorological pressure into statistically significant regulatory kinetic energy. It is also important to note that in practical engineering applications, to prevent issues caused by sensor pulse malfunctions... Abnormal maxima can cause excessive bandwidth expansion, leading to the loss of scheduling constraints within intervals. The system will typically... Set an upper limit threshold for engineering security, such as no more than 3 times the baseline bandwidth, and implement mandatory truncation protection.

[0136] like Figure 6 A two-dimensional curve of the nonlinear evolution of the dynamic kernel bandwidth adjustment factor with real-time meteorological transient gradient is shown. The horizontal axis of the graph represents the characteristic value of the real-time meteorological transient gradient, in dimensionless units, and the vertical axis represents the dynamic kernel bandwidth adjustment factor, in dimensionless units.

[0137] Figure 6 The image shows three upward-extending nonlinear curves in different colors, along with a red horizontal dashed line. The blue, green, and orange curves represent the bandwidth-adjustable dynamic response paths at amplification exponents of 1.5, 2.0, and 2.5, respectively.

[0138] As can be seen from the evolution trend of the curves, when the real-time meteorological transient gradient characteristic value of the horizontal axis is in a small range of 0 to 2, the vertical axis values ​​of the three curves are all maintained smoothly around 1.0. This indicates that during periods of relatively stable meteorological conditions, the system maintains a basic smoothing bandwidth that remains essentially unchanged to ensure the precision of the probability distribution.

[0139] As the horizontal axis values ​​gradually increase, representing a rise in the intensity of meteorological transients, the curves begin to exhibit a non-linear, accelerating upward trend. The larger the amplification index of the orange curve, the faster its upward climb, reflecting the system's high sensitivity to drastic meteorological changes.

[0140] also, Figure 6 The red horizontal dashed line at scale 3 on the vertical axis represents the upper limit threshold line for engineering safety. As the value on the horizontal axis continues to increase, causing the adjustment factor to rise along the curve and touch the red dashed line, the curve then turns into a horizontally extending straight state, and the value is strictly limited to 3 and no longer increases.

[0141] For example, the orange curve reaches its truncation limit when the horizontal axis value is approximately 6.96, and the green curve reaches its truncation limit when the horizontal axis value is 8.0. This graphical feature intuitively illustrates the forced truncation protection mechanism executed by the system when facing extreme weather anomalies or abnormal sensor data.

[0142] The data distribution and trends of this set of two-dimensional curves objectively demonstrate that the system can transform real-time perceived meteorological changes into statistical adjustment quantities, and achieve a reasonable constraint relationship between nonlinear adaptive expansion and engineering physical constraints, providing stable and reliable control logic support for dynamic correction of uncertainty intervals.

[0143] Specifically, this method further performs the following step: multiplying the dynamic kernel bandwidth adjustment factor into the basic smoothing bandwidth parameter of the nonparametric kernel density estimation algorithm to obtain the adaptive kernel bandwidth. The basic smoothing bandwidth parameter is a static parameter pre-calculated based on historical full error data using classical bandwidth selection rules (such as the Silverman reference criterion), and it represents the statistical resolution capability of the system under average operating conditions.

[0144] Optionally, by Multiplying by the base bandwidth, the system achieves "adaptive evolution" of the bandwidth parameter. During periods of extremely stable wind speed or irradiance, Approaching zero Maintaining a value near 1, the adaptive kernel bandwidth remains at the baseline level to ensure finer prediction intervals; once a surge in transient gradients caused by severe convective weather or passing air currents is detected, As the number of cells increases rapidly, the adaptive kernel bandwidth expands accordingly. This process enables the uncertainty quantification model to adaptively and dynamically adjust to changes in weather conditions.

[0145] It is also important to note that after obtaining the adaptive kernel bandwidth, the system performs the following steps: It uses a non-parametric kernel density estimation algorithm incorporating the adaptive kernel bandwidth to fit the data corresponding to the power error segment, generating a dynamic error probability distribution map. In this step, the system retrieves locally stored historical power error samples that match the current power output level and performs density function fitting using the improved kernel density estimation formula. Because of the introduction of bandwidth that varies with the meteorological gradient, the generated dynamic error probability distribution map exhibits a wider support set and a thicker tail during transient events. This means that far-end error points, which were originally considered to have extremely low probability under the static model, will be assigned more reasonable probability weights under transient conditions. Specifically, the fitting formula for the dynamic error probability distribution map is as follows:

[0146] ;

[0147] in: This represents the dynamic error probability distribution density value generated by fitting under the current adaptive bandwidth; This refers to the aforementioned basic kernel smoothing bandwidth parameter; This represents the dynamic kernel bandwidth adjustment factor calculated in the aforementioned steps. The product of these two factors constitutes the adaptive kernel bandwidth.

[0148] Through real-time calculation of this formula, the system can extract dynamic error probability features that closely follow the physical transient logic from the dimension of pure data.

[0149] For example, after obtaining the dynamic distribution map, the system enters the final correction stage and performs the following: extracting the upper and lower dynamic confidence boundaries of the dynamic error probability distribution map under the preset probability quantile, using the upper and lower dynamic confidence boundaries as compensation thresholds, and performing correction operations on the upper and lower limits of the first layer confidence interval to obtain the local initial uncertainty interval.

[0150] In this embodiment, preset probability quantiles, such as the 5% and 95% quantiles, define the reliability level desired by the user. Since the dynamic error probability distribution map scales with the adaptive bandwidth, its corresponding quantile positions also shift towards the sides or center of the error axis. The extracted upper and lower dynamic confidence boundaries are essentially risk redundancy suggestions given from a statistical perspective for current meteorological conditions. The first-level confidence interval (the interval constructed based on the confidence level of the causal inference path) reflects the uncertainty in the physical logic dimension, while the dynamic confidence boundaries reflect the real-time correction amount in the historical statistical dimension.

[0151] Specifically, the correction operation involves offsetting the boundary of the first-level interval towards the direction of the dynamic confidence boundary to ensure that the local initial uncertainty interval of the final output remains converged during the stable weather period to improve scheduling economy, and expands in a timely manner during the meteorological transient period to avoid scheduling risks.

[0152] like Figure 7 This image presents a bitmap comparing the error probability distributions of static bandwidth and adaptive evolution kernel bandwidth under transient weather conditions. The horizontal axis represents power prediction error in megawatts, and the vertical axis represents error probability density in dimensionless units. Figure 7 The image shows two probability density curves of different colors and their corresponding semi-transparent bottom fill areas.

[0153] The blue curve represents the static bandwidth error probability distribution map fitted using the traditional static smoothing bandwidth parameter. Its shape is tall and narrow, with error data highly concentrated near the zero point of the horizontal axis. This static quantization model struggles to cover the divergent actual error samples when encountering meteorological conditions such as sudden changes in wind speed or cloud cover.

[0154] In contrast, the orange-red curve represents the adaptive kernel bandwidth error probability distribution generated by adaptive kernel bandwidth fitting after sensing a surge in meteorological transient gradients in this embodiment. It can be clearly observed that the central peak of the orange-red curve is significantly reduced, and the probability distribution expands dramatically to both sides of the horizontal axis, exhibiting a distinct and gentle long-tail characteristic.

[0155] Figure 7 The vertical dashed lines indicate the upper and lower dynamic confidence boundaries extracted based on the 95% probability quantile. Under the action of the adaptive evolution mechanism, the confidence boundary corresponding to the blue curve is located in the range of -4.9 MW to +4.9 MW, while the confidence boundary corresponding to the orange-red curve is automatically compensated and expands outward to the range of -9.8 MW to +9.8 MW.

[0156] By dynamically extending the interval width, the system successfully encompasses the far-end error points that diverge drastically under meteorological transient conditions within the new safety boundary. This data comparison objectively reflects the necessity and effectiveness of introducing meteorological gradient features to adjust the statistical smoothing scale in real time. It demonstrates that the generated local initial uncertainty interval can spontaneously expand risk redundancy during micro-meteorological changes, thereby effectively reducing the risk of missed reports in actual power grid dispatching operations and ensuring the overall operational safety of the power grid.

[0157] It should also be noted that this correction operation is not a simple replacement, but a process of integrating physical causal laws with statistical dynamic characteristics.

[0158] Optionally, if the error divergence range indicated by the dynamic confidence boundary is much larger than the range of the physical causal derivation, the system will adopt a conservative principle and use the dynamic confidence boundary as the dominant compensation amount to prevent the risk of missed reports due to the simplification of the physical model.

[0159] Based on existing technological analysis, traditional statistical quantification methods, such as the Gaussian distribution assumption or fixed-bandwidth kernel density estimation, typically assume that the error process is time-homogeneous, failing to recognize the drastic distortion of the statistical distribution of errors caused by meteorological transients. When faced with the "long-tail" error characteristics of renewable energy power, traditional methods are prone to failure at critical turning points, leading to insufficient reserve capacity in grid dispatching.

[0160] This embodiment constructs a nonlinear linkage mechanism between the real-time meteorological transient gradient and the kernel bandwidth adjustment factor at the edge layer, enabling the kernel density estimation algorithm to spontaneously adjust the smoothing scale according to real-time changes in the physical environment. In practical applications, this technical solution can accurately capture the error divergence characteristics caused by meteorological transients, and the generated dynamic error probability distribution map has strong environmental adaptability.

[0161] By applying this statistical correction quantity with dynamic physical awareness to the first-level confidence interval, the resulting local initial uncertainty interval effectively suppresses the false width of the interval during the stationary period while ensuring high confidence coverage. This achieves a high degree of unity between safety and economy in the quantification of uncertainty in renewable energy power prediction. This has significant engineering practical value for improving the power grid's resilience to sudden weather risks.

[0162] Example 4:

[0163] like Figure 8 As shown, this embodiment provides a new energy power prediction and uncertainty quantification system.

[0164] In the current context of renewable energy power prediction technologies, prediction systems typically employ either a centralized cloud processing architecture or an isolated local prediction mode at power plants. In actual power grid dispatching scenarios, this presents a conflict between the real-time response requirements of the underlying systems and the coordination of global physical constraints. A centralized cloud architecture requires the underlying power plants to upload massive amounts of high-frequency operational data, which not only introduces communication delays and data privacy and security risks but also fails to meet the rapid control needs of the underlying units in response to emergencies. On the other hand, isolated local prediction modes, lacking macro-level operational knowledge such as regional power grid topology, large-scale meteorological trends, and the status of inter-regional transmission channels, often produce quantitative results that deviate from the operational constraints of the actual physical power grid, easily leading to severe prediction biases during meteorological transients or power grid topology switching.

[0165] The system provided in this embodiment aims to resolve the aforementioned conflicts through distributed collaborative computing at the physical layer. The system enables multi-scale causal inference models to mine causal prediction sequences based on physical topology and operational characteristics, and corrects error boundaries layer by layer through three-level uncertainty quantification collaborative processing. Ultimately, while ensuring the privacy of underlying data and the timeliness of business response, it improves the safety margin and refined scheduling level of the entire network's renewable energy power prediction.

[0166] Specifically, this embodiment discloses a new energy power prediction and uncertainty quantification system, which is implemented based on a three-level distributed architecture consisting of an edge layer, a regional layer, and a cloud layer. In the actual power grid physical topology, this three-level distributed architecture corresponds to different hardware deployment entities: the "edge layer" is usually composed of industrial control computers (IPCs) or edge computing boxes (such as industrial control equipment equipped with embedded GPUs or NPU modules) deployed in the local computer room of wind farms or photovoltaic power plants; the "regional layer" is usually composed of high-performance server clusters deployed in the computer room of municipal or provincial power grid dispatch centers; and the "cloud layer" is deployed on national or inter-provincial power grid cloud platforms, relying on computing node clusters of supercomputing centers or large distributed data centers.

[0167] Specifically, the system includes the following core functional modules:

[0168] The multi-source data acquisition module is used to acquire multi-source hierarchical operational data. This multi-source hierarchical operational data includes local unit dynamic characteristic data at the edge layer, regional power grid topology data and regional weather forecast data at the regional layer, and cross-regional transmission channel status data at the cloud layer. In practical applications, at the edge layer, this module functions as a communication gateway for the Supervisory Control and Data Acquisition (SCADA) system. It reads dynamic parameters from the programmable logic controller (PLC) registers of wind turbine generators or photovoltaic inverters in real time via industrial Ethernet, RS-485 bus, or fiber optic ring network, using industrial standard protocols such as IEC61850 and Modbus. At the regional layer, this module acts as a data interface server for the Energy Management System (EMS), responsible for receiving numerical weather prediction (NWP) data streams from the meteorological bureau and topology messages uploaded by the substation automation system. At the cloud layer, this module, relying on a data exchange front-end, interfaces with the dispatch management system of the cross-regional DC transmission channel to obtain the channel's available transmission capacity (ATC) and real-time congestion signals. The module also includes a built-in hardware encryption machine and desensitization algorithm unit to ensure privacy and communication security when data is transmitted across layers.

[0169] The knowledge graph management module is used to construct a hierarchical power grid business knowledge graph based on the multi-source hierarchical operational data. This hierarchical power grid business knowledge graph consists of a local knowledge graph at the edge layer, a regional knowledge graph at the regional layer, and a global knowledge graph at the cloud layer. This module primarily relies on a graph database engine (such as Neo4j or JanusGraph) distributed across the three layers and corresponding storage arrays. At the edge layer, the solid-state drives (SSDs) within the industrial control computer are allocated dedicated lightweight graph database space for storing the single-line graph structure and equipment ledger parameters of wind turbines or photovoltaic arrays. At the regional layer, the distributed file system in the server cluster is responsible for storing and maintaining the regional-level graph structure containing substations, transmission lines, and their impedance parameters. At the cloud layer, the large-scale distributed graph database not only stores the connection relationships of key nodes across the entire network but also has a dedicated graph computing engine server for performing consistency checks and knowledge fusion algorithm calculations on the knowledge graph.

[0170] This module is responsible for transforming scattered device parameters and operating status into graph-structured data with topological relationships, providing structured physical rule inputs for subsequent reasoning.

[0171] The multi-scale causal prediction module is used to input the multi-source hierarchical operational data and the hierarchical power grid business knowledge graph into the multi-scale causal inference model to generate initial prediction sequences for new energy power at each level. In terms of hardware and computing power allocation, this module is supported by AI inference acceleration cards (such as GPU computing cards and tensor processing units) at each level. The edge layer inference module utilizes the artificial intelligence chip within the edge computing box to retrieve wind speed or irradiance time series and local knowledge graphs from local memory, execute wavelet transform and causal discovery algorithms, and quickly generate ultra-short-term prediction sequences for a single power station. The regional and cloud layer inference modules rely on the powerful parallel computing capabilities of the server cluster to process the multi-scale causal inference main model, which includes large-scale graph neural networks (GNNs) or temporal convolutional networks (TCNs).

[0172] This module interacts with the high-speed bus (such as PCIe) of the central processing unit through dedicated video memory, loading physical topology rules into the loss function or hidden layer structure of the neural network, ensuring that the output initial prediction sequence strictly follows the causal transmission logic of the real power grid.

[0173] The uncertainty quantification and collaboration module is used to perform three levels of uncertainty quantification and collaboration processing sequentially at the edge layer, regional layer, and cloud layer based on the initial prediction sequences of new energy power at each level, generating a target power prediction sequence containing uncertainty intervals. In the actual physical device, this module is manifested as a statistical analysis engine running within each computing node and a cross-level message queue middleware (such as Kafka or RabbitMQ). The statistical processing unit at the edge layer performs matrix operations of the kernel density estimation algorithm based on historical error data in local memory to generate local confidence interval parameters. The regional layer and cloud layer asynchronously receive the local confidence interval parameters reported from the lower layer through the message queue middleware and input them into a Bayesian deep learning network deployed on a server with a large amount of running memory. The servers at the regional layer and cloud layer perform Markov chain Monte Carlo (MCMC) sampling or variational inference calculations, integrating the safety margin of the transmission channel with the cognitive uncertainty of the model itself. Finally, the target power prediction sequence curve with clearly defined upper and lower bound envelopes is rendered on the display terminal screen of the dispatch center, providing dispatchers with an intuitive decision reference range.

[0174] The closed-loop adaptive update module acquires actual power data, calculates the real-time prediction error at each level through a hierarchical dynamic ensemble learning mechanism, and performs closed-loop adaptive updates to the multi-scale causal inference model based on the real-time prediction error. This module relies on a model parameter server and a log storage and analysis system. Actual power data is periodically transmitted back to the time-series databases at each level via the power data acquisition terminal. The ensemble learning scheduler within the module triggers timed tasks within a preset time window, compares the predicted sequences in the database with the actual sequences, and calculates evaluation metrics such as the root mean square error (RMSE).

[0175] Based on the error results, the weight update script inside the parameter server will adjust the combined weight parameters of each heterogeneous base model. For model weights that need to be updated, the system uses the Internet of Things (IoT) device management platform and the Remote Procedure Call (RPC) protocol under the microservice architecture to send the updated weight file or network parameters to the memory of the edge layer computing box or regional layer inference server, thereby completing the closed-loop adaptive evolution of the model without manual intervention.

[0176] Through the coordinated operation of the aforementioned physical modules and logical units, this system constructs a highly automated prediction and quantization pipeline in actual power grid operation. Edge layer devices are responsible for local processing / computation of high-frequency data and physical causal filtering; regional layer servers are responsible for the quantitative fusion of topological correlations and spatial uncertainties; and cloud layer devices coordinate global constraints and long-term macroscopic evolution. This system effectively overcomes the response lag caused by concentrated computing power and the lack of physical boundary constraints in individual models in existing technologies. It possesses the ability to continuously output objective, stable prediction results with interpretable safety boundaries under complex and variable extreme weather and power grid operating conditions.

[0177] Example 5:

[0178] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0179] like Figure 9 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0180] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0181] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0182] The memory 103 stores a computer program corresponding to the new energy power prediction and uncertainty quantification method of the above embodiments of the present invention. The computer program is controlled and executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0183] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 9 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0184] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting new energy power and quantifying uncertainty, characterized in that, The implementation is based on a three-tier distributed architecture consisting of an edge layer, a regional layer, and a cloud layer, including: Acquire multi-source hierarchical operation data; the multi-source hierarchical operation data includes local unit dynamic characteristic data at the edge layer, regional power grid topology data and regional weather forecast data at the regional layer, and cross-regional power transmission channel status data at the cloud layer; A hierarchical power grid business knowledge graph is constructed based on the multi-source hierarchical operation data; the hierarchical power grid business knowledge graph consists of a local knowledge graph at the edge layer, a regional knowledge graph at the regional layer, and a global knowledge graph at the cloud layer. The multi-source hierarchical operation data and the hierarchical power grid business knowledge graph are input into the multi-scale causal reasoning model to generate initial prediction sequences of new energy power at each level. Based on the initial prediction sequences of new energy power at each level, three-level uncertainty quantification collaborative processing is performed sequentially at the edge layer, regional layer, and cloud layer to generate a target power prediction sequence containing uncertainty intervals. The actual power data is acquired, and the real-time prediction error of each level is calculated through a hierarchical dynamic ensemble learning mechanism. Based on the real-time prediction error, the multi-scale causal inference model is updated in a closed loop. The initial power prediction sequence of new energy sources based on each level undergoes a three-level uncertainty quantification collaborative processing at the edge layer, regional layer, and cloud layer. This includes: performing causal-statistical binary quantization processing at the edge layer to generate a local initial uncertainty interval; fusing the local initial uncertainty intervals of the edge layers under the jurisdiction of the regional layer and performing causal-statistical-Bayes triple quantization processing based on the regional-level prediction data to generate a regional-level initial uncertainty matrix; and fusing all regional-level initial uncertainty matrices at the cloud layer and performing global quantization correction by introducing transmission channel uncertainty parameters to generate a target power prediction sequence containing a global confidence index. The causal-statistical weighting process performed at the edge layer specifically includes: extracting causal path confidence parameters based on the local knowledge graph, and constructing a first-layer confidence interval based on the causal path confidence parameters; using a fuzzy C-means clustering algorithm to divide the acquired local historical prediction error data into multiple discrete power error segments; within each power error segment, using a non-parametric kernel density estimation algorithm to fit the error probability distribution map, and statistically correcting the first-layer confidence interval based on the error probability distribution map to generate the local initial uncertainty interval; When fusing the local initial uncertainty interval of the edge layers under the jurisdiction of the regional layer, and performing causal-statistical-Bayesian triple weighting processing based on the regional-level prediction data, a Bayesian deep learning model is used for construction; the objective function of the Bayesian deep learning model is defined as a loss function including an uncertainty penalty term, and its mathematical expression is: in, The calculated uncertainty loss evaluation index; The total number of input data samples; k is the index number of the current sample; This represents the initial predicted power value for the k-th sample; This represents the actual power value collected for the k-th sample. Let be the prediction variance of the k-th sample.

2. The method according to claim 1, characterized in that, The acquisition of multi-source hierarchical operation data includes: The local unit dynamic characteristic data is collected at the edge layer, and local differential privacy technology is used for desensitization processing. Outliers are removed by a sliding window mid-value filtering algorithm, and the edge layer preprocessed data is generated after downsampling. The desensitized edge layer preprocessed data is aggregated at the regional layer, and combined with the regional power grid topology data and regional weather forecast data, the missing values ​​are filled in by a linear interpolation algorithm to generate regional layer preprocessed data; The status data of the cross-regional power transmission channel is obtained at the cloud layer, and the preprocessed data at the edge layer, the preprocessed data at the regional layer, and the status data of the cross-regional power transmission channel are subjected to standardization processing.

3. The method according to claim 1, characterized in that, After constructing the hierarchical power grid business knowledge graph, the method further includes performing a collaborative update mechanism on the hierarchical power grid business knowledge graph, specifically including: The edge layer uses historical local operating data to extract dynamic attribute features, updates the local knowledge graph, and generates a local knowledge increment package which is then uploaded to the region layer. The regional layer aggregates local knowledge increment packages within its jurisdiction, performs knowledge consistency verification to update the regional knowledge graph, and generates regional knowledge increment packages to be uploaded to the cloud layer. The cloud layer aggregates knowledge increment packages from all regions, uses knowledge fusion algorithms to eliminate cross-regional knowledge conflicts, generates a general knowledge benchmark model, and distributes it to each regional layer and edge layer.

4. The method according to claim 1, characterized in that, The step of inputting the multi-source hierarchical operation data and the hierarchical power grid business knowledge graph into the multi-scale causal inference model to generate initial prediction sequences for new energy power at each level, specifically includes the following at the edge layer: The local unit dynamic characteristic data in the multi-source hierarchical operation data are subjected to discrete wavelet transform to decompose the trend term feature sequence, fluctuation term feature sequence and noise term feature sequence. Extract the site equipment ledger and local topology constraint information from the local knowledge graph; A causal discovery algorithm is used to combine the station equipment ledger with local topology constraint information to mine the causal path of the trend term feature sequence and fluctuation term feature sequence to local power output. Based on the causal path, the minute-by-minute power output characteristics at the single power station level are calculated to generate the initial prediction sequence of new energy power in the edge layer.

5. The method according to claim 1, characterized in that, The step of inputting the multi-source hierarchical operation data and the hierarchical power grid business knowledge graph into the multi-scale causal reasoning model to generate initial prediction sequences for new energy power at each level, specifically includes the following at the regional and cloud layers: At the regional layer, the edge layer new energy power initial prediction sequence output from the edge layer, the regional meteorological forecast data, and the regional knowledge graph are input into the regional causal reasoning model of the regional layer, and the regional aggregated power initial prediction sequence is output. At the cloud layer, the regional aggregated power initial prediction sequence output from the regional layer, the global weather forecast data, and the global knowledge graph are input into the global causal reasoning model at the cloud layer, and the network-wide global power initial prediction sequence is output.

6. The method according to claim 1, characterized in that, The step of calculating the real-time prediction error of each level through a hierarchical dynamic ensemble learning mechanism, and performing closed-loop adaptive updates to the multi-scale causal inference model based on the real-time prediction error, includes: A dynamic integrated model library containing multiple heterogeneous base models is constructed at each level, and the combination of the heterogeneous base models constitutes the multi-scale causal inference model; Within each preset time window, the real-time root mean square error of each heterogeneous basis model is calculated independently. A preset exponential decay dynamic weight update mechanism is adopted to adjust the output weight components of each heterogeneous base model in the integrated decision based on the real-time root mean square error, so as to complete the closed-loop adaptive update of the multi-scale causal inference model.

7. A new energy power prediction and uncertainty quantification system, characterized in that, The implementation is based on a three-tier distributed architecture consisting of an edge layer, a regional layer, and a cloud layer, including: The multi-source data acquisition module is used to acquire multi-source hierarchical operation data; the multi-source hierarchical operation data includes local unit dynamic characteristic data at the edge layer, regional power grid topology data and regional weather forecast data at the regional layer, and cross-regional power transmission channel status data at the cloud layer. The knowledge graph management module is used to construct a hierarchical power grid business knowledge graph based on the multi-source hierarchical operation data; the hierarchical power grid business knowledge graph consists of a local knowledge graph at the edge layer, a regional knowledge graph at the regional layer, and a global knowledge graph at the cloud layer. The multi-scale causal prediction module is used to input the multi-source hierarchical operation data and the hierarchical power grid business knowledge graph into the multi-scale causal inference model to generate the initial prediction sequence of new energy power at each level. The uncertainty quantification and coordination module is used to perform three-level uncertainty quantification and coordination processing sequentially in the edge layer, regional layer and cloud layer based on the initial prediction sequence of new energy power at each level, to generate a target power prediction sequence containing uncertainty intervals; The closed-loop adaptive update module is used to acquire actual power data, calculate the real-time prediction error of each level through a hierarchical dynamic ensemble learning mechanism, and perform closed-loop adaptive update of the multi-scale causal inference model based on the real-time prediction error. The initial power prediction sequence of new energy sources based on each level undergoes a three-level uncertainty quantification collaborative processing at the edge layer, regional layer, and cloud layer. This includes: performing causal-statistical binary quantization processing at the edge layer to generate a local initial uncertainty interval; fusing the local initial uncertainty intervals of the edge layers under the jurisdiction of the regional layer and performing causal-statistical-Bayes triple quantization processing based on the regional-level prediction data to generate a regional-level initial uncertainty matrix; and fusing all regional-level initial uncertainty matrices at the cloud layer and performing global quantization correction by introducing transmission channel uncertainty parameters to generate a target power prediction sequence containing a global confidence index. The causal-statistical weighting process performed at the edge layer specifically includes: extracting causal path confidence parameters based on the local knowledge graph, and constructing a first-layer confidence interval based on the causal path confidence parameters; using a fuzzy C-means clustering algorithm to divide the acquired local historical prediction error data into multiple discrete power error segments; within each power error segment, using a non-parametric kernel density estimation algorithm to fit the error probability distribution map, and statistically correcting the first-layer confidence interval based on the error probability distribution map to generate the local initial uncertainty interval; When fusing the local initial uncertainty interval of the edge layers under the jurisdiction of the regional layer, and performing causal-statistical-Bayesian triple weighting processing based on the regional-level prediction data, a Bayesian deep learning model is used for construction; the objective function of the Bayesian deep learning model is defined as a loss function including an uncertainty penalty term, and its mathematical expression is: in, The calculated uncertainty loss evaluation index; The total number of input data samples; k is the index number of the current sample; This represents the initial predicted power value for the k-th sample; This represents the actual power value collected for the k-th sample. Let be the prediction variance of the k-th sample.

Images