Multi-scale adaptive correction system for wind-solar power

Abstract

The invention relates to the technical field of new energy electric power, and discloses a wind-solar power multi-scale adaptive correction system. A monitoring data fusion module of the system receives wind-light power time sequence data collected by sky-ground network multi-source monitoring equipment, and executes space-time alignment and quality verification of multi-source heterogeneous data; the power characteristic decoupling module is used for separating a periodic fluctuation component and a random fluctuation component of the wind and light power based on the verified time sequence data; the space-time correlation modeling module establishes a regional wind and light equipment topological graph according to the periodic fluctuation component, and calculates a meteorological correlation strength index between the equipment based on the topological graph; and the multi-scale correction module performs cross-device collaborative correction on the random fluctuation component by using the meteorological correlation intensity index to generate a multi-scale wind-solar power prediction correction value. According to the system, through a multi-module synergistic effect, meteorological association between multi-source data values and equipment is fully mined, accurate correction of wind and light power is realized, and fluctuation characteristics of different scales are adapted.

Description

Technical Field

[0001] This invention relates to the field of new energy power technology, specifically to a multi-scale adaptive correction system for wind and solar power. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, wind and solar power, as important renewable energy sources, are seeing their installed capacity in the power system continue to rise. The large-scale grid connection of wind and solar energy has injected new vitality into the energy supply system, but its inherent volatility and randomness also pose numerous challenges to the safe and stable operation of the power system. Accurate wind and solar power forecasting is a crucial prerequisite for achieving efficient renewable energy consumption and optimized dispatch decisions. However, wind and solar power are affected by various factors such as meteorological conditions and geographical environment, exhibiting complex and multi-scale variation patterns. Traditional forecasting methods are insufficient to meet the high accuracy requirements of practical applications. Wind and solar power forecasting primarily relies on single or limited data sources, resulting in insufficient coverage and diversity of monitoring data. Existing data processing methods often fail to effectively align multi-source heterogeneous data in time and space. Differences in sampling frequency and spatiotemporal resolution among different monitoring devices weaken the correlation between data points, making it difficult to guarantee data quality. Furthermore, the absence or inadequacy of data quality verification processes allows noisy data and outliers to directly enter the prediction model, further impacting the reliability of the prediction results. In power characteristic analysis, traditional methods often model wind and solar power sequences as a whole, failing to effectively separate periodic and random fluctuation components. Periodic fluctuation components are often related to regular meteorological factors such as seasonal changes and diurnal cycles, while random fluctuation components are more significantly affected by sudden weather events and local topography. This indiscriminate treatment makes it difficult for models to accurately capture the inherent patterns of different fluctuation components, resulting in insufficient adaptability to both types of components during prediction correction. The wind and solar power equipment within the region does not operate in isolation; their power output is influenced by a shared meteorological system, resulting in significant spatiotemporal correlations between the equipment. However, current technologies for power correction often focus on historical data or local meteorological information for individual equipment, neglecting the meteorological correlations among the wind and solar power equipment clusters within the region. The lack of quantitative analysis of the topological relationships and meteorological correlation strength between equipment makes cross-equipment collaborative correction difficult to achieve, failing to fully utilize the complementarity of meteorological information within the region to improve correction effectiveness. Furthermore, existing correction methods are mostly optimized for a single time scale, making it difficult to adapt to the fluctuation characteristics of wind and solar power at different time scales. This results in significant room for improvement in correction accuracy for both short-term random fluctuations and long-term periodic changes. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-scale adaptive correction system for wind and solar power to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides a multi-scale adaptive correction system for wind and solar power, the system comprising: The monitoring data fusion module, power characteristic decoupling module, spatiotemporal correlation modeling module, and multi-scale correction module are included. The monitoring data fusion module is used to receive wind and solar power time series data collected by multi-source monitoring equipment in the sky and ground network, and to perform spatiotemporal alignment and quality verification of multi-source heterogeneous data; The power characteristic decoupling module separates the periodic fluctuation component and the random fluctuation component of wind and solar power based on the verified time-series data. The spatiotemporal correlation modeling module establishes a regional wind and solar equipment topology map based on the periodic fluctuation components, and calculates the meteorological correlation strength index between the equipment based on the topology map; The multi-scale correction module uses the meteorological correlation strength index to perform cross-device collaborative correction of random fluctuation components, generating multi-scale wind and solar power prediction correction values.

[0005] Preferably, the step of the monitoring data fusion module performing spatiotemporal alignment and quality verification of multi-source heterogeneous data includes: Acquire satellite cloud imagery data, meteorological radar data, ground sensor network data, and historical power data from the power grid dispatching terminal; Time-stamped normalization was performed on the four types of data, and an adaptive sliding window algorithm was used to unify the time resolution. Anomaly data points are identified using a dynamic threshold outlier detection algorithm, and data imputation is performed using spatial interpolation between adjacent sites. Output a fusion monitoring dataset that is aligned in the time dimension and whose integrity meets preset standards.

[0006] Preferably, the step of separating components by the power characteristic decoupling module includes: Extract wind speed, irradiance, and actual power sequences for the target area from the fusion monitoring dataset; The empirical mode decomposition algorithm is used to decompose the actual power sequence into high-frequency components and low-frequency components; Based on the spectral characteristics of wind speed and irradiance sequences, high-frequency components are mapped to a set of random fluctuation components. After phase matching between the low-frequency components and the periodic characteristics of meteorological elements, they are incorporated into the periodic fluctuation component set.

[0007] Preferably, the execution steps of the spatiotemporal correlation modeling module include: A set of equipment nodes is constructed based on the geographical distribution coordinates of regional wind and solar equipment and the power grid topology; The set of periodic fluctuation components is input into a graph neural network to learn the meteorological propagation feature vectors between device nodes; Calculate the device association strength index based on cosine similarity of meteorological propagation feature vectors; Output a topology map of wind and solar power equipment with correlation strength weights.

[0008] Preferably, the correction steps of the multi-scale correction module include: Extract the power residual value of the target device from the set of random fluctuation components; Retrieve adjacent device nodes in the wind and solar equipment topology map that have a strong association with the target device; The random fluctuation components of adjacent devices are weighted and aggregated based on the device correlation strength index; The aggregation result is convolved with the power residual value of the target device to generate a set of correction parameters.

[0009] Preferably, the system further includes a scale recursive feedback module, the execution steps of which include: Obtain minute-level, hourly-level, and day-ahead-level output results of multi-scale wind and solar power prediction correction values; The minute-level correction results are input into the power feature decoupling module to update the set of random fluctuation components; The hourly correction results are input into the spatiotemporal correlation modeling module to iterate the device correlation strength index; The deviation between the day-ahead correction results and the power grid dispatch plan is compared to generate feedback coefficients.

[0010] Preferably, the feedback coefficients of the scale recursive feedback module act on the monitoring data fusion module, specifically including: The step size parameter of the adaptive sliding window is dynamically adjusted based on the feedback coefficient; the sensitivity threshold for outlier detection is optimized based on the deviation comparison results.

[0011] Preferably, the power characteristic decoupling module performs incremental decomposition after receiving updated data, specifically including: The set of periodic fluctuation components in the historical decomposition results is retained, and empirical mode decomposition is performed only on the newly added fused monitoring data; The newly added high-frequency components are merged and reorganized with the set of historical random fluctuation components.

[0012] Preferably, the spatiotemporal correlation modeling module updates the topology graph based on the recombined components, specifically including: Extract the variance characteristics of the random fluctuation components after the merger and reorganization; When the variance characteristic exceeds the preset fluctuation threshold, the device correlation strength index is recalculated. The updated device association strength index will be synchronized to the multi-scale correction module.

[0013] Preferably, the multi-scale correction module further includes: The weight allocation of the weighted aggregation is adjusted according to the updated device association strength index, and a temporal convolutional network is used to extract multi-scale features from the corrected parameter set. The feature extraction results are coupled with real-time weather forecast data to calculate and output a rolling updated multi-scale wind and solar power prediction correction value.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This multi-scale adaptive correction system for wind and solar power provides a comprehensive and refined solution for wind and solar power prediction correction through the synergistic effect of its various modules. The monitoring data fusion module processes data collected by multi-source monitoring devices across the sky and ground networks. Through spatiotemporal alignment, it eliminates differences in sampling time and spatial location between different devices, transforming previously scattered, heterogeneous data into an organic whole and enhancing the correlation between data. Simultaneously, the quality verification process identifies and handles issues such as noise and missing values ​​in the data, reducing the interference of poor-quality data on subsequent analysis and providing a more reliable data foundation for subsequent power characteristic decoupling and modeling. The power characteristic decoupling module separates periodic and random fluctuation components based on verified time-series data, clearly revealing these two types of fluctuations with different patterns of change. Periodic fluctuation components are often related to regular meteorological changes, exhibiting relatively stable trends; random fluctuation components, on the other hand, are more significantly affected by sudden meteorological factors, resulting in more complex changes. This separation process avoids mutual interference between the two types of components, allowing subsequent modeling and correction to be more specifically tailored to the characteristics of different components, thus improving the ability to analyze various types of fluctuations. The spatiotemporal correlation modeling module constructs a regional wind and solar power equipment topology map based on periodic fluctuation components, combining the spatial distribution of equipment with power cycle characteristics to intuitively present the correlation between equipment. Based on this, the calculated meteorological correlation strength index quantifies the degree to which different equipment is affected by common meteorological factors, revealing the intrinsic connections between wind and solar power changes within the region. This process breaks away from the dependence on data from single equipment, fully explores the collaborative information of regional equipment clusters, provides a basis for cross-equipment information sharing and collaborative correction, and enables the correction process to better adapt to the overall trend of regional meteorological changes. The multi-scale correction module utilizes the meteorological correlation strength index to perform cross-device collaborative correction of random fluctuation components. By incorporating the meteorological correlation characteristics between devices within the region, the correction is no longer limited to local information from a single device. Through the coordinated use of monitoring data and fluctuation characteristics from different devices, it can more comprehensively capture the changing patterns of random fluctuations and effectively address the complex manifestations of random fluctuations at different time scales. The generated multi-scale wind and solar power prediction correction values ​​can adapt to the fluctuation characteristics of wind and solar power at short, medium, and long time scales, improving the accuracy of the correction results in reflecting actual power changes and enhancing the applicability of wind and solar power prediction in power system dispatching and operation scenarios. Attached Figure Description

[0015] Figure 1 This is a timing diagram of the wind and solar power multi-scale adaptive correction system described in this invention; Figure 2 A flowchart illustrating the workflow of the monitoring data fusion module; Figure 3 A flowchart for constructing the topology graph of the spatiotemporal correlation modeling module; Figure 4 This is a flowchart of the multi-scale result processing for the scale recursive feedback module. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides a multi-scale adaptive correction system for wind and solar power, the system comprising: The system comprises a monitoring data fusion module, a power characteristic decoupling module, a spatiotemporal correlation modeling module, and a multi-scale correction module. The monitoring data fusion module receives time-series wind and solar power data collected from multi-source monitoring equipment across the sky and ground network, and performs spatiotemporal alignment and quality verification of the heterogeneous data. The power characteristic decoupling module separates the periodic and random fluctuation components of wind and solar power based on the verified time-series data. The spatiotemporal correlation modeling module establishes a regional wind and solar equipment topology map based on the periodic fluctuation components and calculates the meteorological correlation strength index between equipment based on the topology map. The multi-scale correction module uses the meteorological correlation strength index to perform cross-equipment collaborative correction of the random fluctuation components, generating multi-scale wind and solar power prediction correction values.

[0018] Example 1: See Figure 2The monitoring data fusion module is responsible for processing time-series wind and solar power data from multi-source monitoring equipment in the sky and ground network, including satellite cloud imagery data, meteorological radar data, ground sensor network data, and historical power data from the power grid dispatching terminal. These data have different temporal resolutions, spatial coverages, and acquisition methods, thus requiring spatiotemporal alignment and quality verification. First, time-stamp normalization is performed on the four types of data to ensure that the timestamps of all data are unified to a standard time reference system. An adaptive sliding window algorithm dynamically adjusts the window size according to the sampling frequency of the data source; a smaller window is used for high-frequency data to retain detailed features, while a larger window is used for low-frequency data to reduce noise interference. The step size of the sliding window is adaptively adjusted according to the data update frequency to ensure continuity in the time dimension.

[0019] After time alignment, outlier data points are identified using a dynamic threshold outlier detection algorithm. This algorithm calculates the mean and standard deviation of the data based on a sliding statistical window and sets a dynamic threshold by combining the distribution characteristics of historical data. When a data point deviates from the mean by more than a preset multiple of the standard deviation, it is identified as an outlier. Outlier data point imputation employs a neighboring site spatial interpolation method. This method comprehensively considers geographical proximity and meteorological similarity, using Kriging interpolation to calculate the estimated value of the missing data. During the interpolation process, the spatial weights between sites are dynamically adjusted based on the spatial correlation of historical data to ensure the rationality of the imputed data. The final output fused monitoring dataset meets the preset standards for time dimension alignment and data integrity, providing high-quality input for subsequent modules.

[0020] The power feature decoupling module extracts the wind speed, irradiance, and actual power sequences of the target area from the fused monitoring dataset. The actual power sequence contains both periodic and random fluctuations, which need to be separated using signal decomposition methods. The Empirical Mode Decomposition (EMD) algorithm is suitable for processing non-stationary signals. Through an iterative screening process, it decomposes the actual power sequence into several intrinsic mode function (EMF) components. During the decomposition process, local extrema of the signal are used to construct upper and lower envelopes, and fluctuation components at different time scales are gradually extracted through mean calculation. High-frequency components typically reflect random fluctuations, such as the impact of short-term meteorological changes like turbulence and cloud cover; low-frequency components correspond to periodic fluctuations, such as long-term regularities like day-night cycles and seasonal changes.

[0021] Spectral characteristic analysis of wind speed and irradiance sequences is used to assist in component classification. Fast Fourier Transform (FFT) identifies the dominant frequency components in the signals; high-frequency components matching the spectral characteristics of meteorological elements are grouped into the random fluctuation component set. Low-frequency components require further phase matching with the periodic characteristics of meteorological elements, and time delays are calculated using cross-correlation functions to ensure the synchronization of the decomposed periodic components with meteorological changes. Finally, the power feature decoupling module outputs sets of periodic and random fluctuation components, providing structured feature inputs for spatiotemporal correlation modeling and multi-scale correction.

[0022] In its implementation, the data processing flow of the monitoring data fusion module adopts a parallel computing architecture to improve the efficiency of large-scale data processing. Satellite cloud imagery data and meteorological radar data typically have high spatial resolution but low temporal sampling frequency, thus requiring time interpolation during spatiotemporal alignment. Ground sensor network data has high temporal resolution but limited spatial coverage, necessitating spatial expansion by combining satellite and radar data. Historical power data from the power grid dispatching terminal is primarily used to verify and calibrate the accuracy of other data sources, and its timestamps must be strictly aligned with real-time monitoring data.

[0023] The signal decomposition process of the power characteristic decoupling module adopts an adaptive stopping criterion to ensure that the number of iterations of empirical mode decomposition is sufficient to extract signal features while avoiding over-decomposition that could lead to mode aliasing. High-frequency components are classified based on the short-term fluctuation characteristics of wind speed and irradiance. If the energy concentration area of ​​a high-frequency component coincides with a period of rapid change in meteorological elements, it is classified as a random fluctuation component. Phase matching of low-frequency components employs a sliding time window strategy, dynamically adjusting the size of the matching window to adapt to the cyclical changes of different seasons.

[0024] Example 2: See Figure 3 The spatiotemporal correlation modeling module constructs the correlation relationships between wind and solar power equipment in a region by analyzing periodic fluctuation components. When constructing the set of equipment nodes, spatial location information is determined based on the geographical distribution coordinates of wind turbines and photovoltaic power stations within the target area, while simultaneously integrating electrical connections within the power grid topology. Each equipment node includes attributes such as latitude and longitude coordinates, installed capacity, grid-connected voltage level, and equipment type. The connection relationships between nodes are established based on a dual consideration of electrical distance and geographical distance. Electrical distance is obtained through short-circuit impedance calculation, while geographical distance is converted using spherical trigonometric formulas.

[0025] The set of periodic fluctuation components is processed by a graph neural network. This network employs a multi-layer message-passing architecture, aggregating feature information from neighboring nodes at each layer. A graph attention mechanism dynamically calculates the interaction weights between nodes, with the attention coefficient determined by the similarity of meteorological features and spatial distance between nodes. The dimension of the meteorological propagation feature vector is consistent with the number of features in the periodic components, with each element representing the intensity of meteorological influence at a specific time scale. After multiple iterations, the device nodes output a fixed-dimensional hidden state vector, which encodes the spatial pattern of meteorological fluctuation propagation between devices.

[0026] The meteorological correlation strength index between devices is calculated based on hidden state vectors. Before calculation, the feature vectors are standardized to eliminate the influence of different physical dimensions on similarity calculation. The cosine similarity algorithm measures the directional consistency of two vectors in space, and the result is mapped to a range of 0 to 1 to represent the correlation strength. The index calculation covers all device node pairs; node pairs with correlation strength exceeding a preset threshold retain their connection in the topology graph. The final output is a weighted adjacency matrix, where matrix elements represent the meteorological coupling strength between device nodes, used to guide subsequent cross-device collaborative correction processes.

[0027] The multi-scale correction module utilizes the aforementioned correlation strength index to collaboratively correct random fluctuation components. The power residual value of the target device is defined as the instantaneous deviation sequence between the actual power measurement value and the baseline prediction value, and is extracted from the random fluctuation component set using a timestamp-aligned index. When searching for adjacent nodes directly connected to the target device in the wind and solar equipment topology map, a minimum threshold for the correlation strength index is set; nodes below this value are not included in the correction process. The correlation strength index is updated in real time to ensure that the topological relationships reflect the current meteorological propagation status.

[0028] Random fluctuation components from adjacent device nodes participate in the weighted aggregation calculation. The aggregation weights are transformed into a probability distribution form by the correlation strength index through a normalization function, ensuring that the sum of all weight coefficients is 1. Devices with high correlation strength have larger weights, and their power fluctuation information contributes more significantly to the correction result. The weighted aggregation result forms a meteorological correlation compensation signal, whose temporal resolution is consistent with the target device residual sequence. This compensation signal is convolved with the original residual value of the target device in one dimension, and the convolution kernel length is dynamically configured according to the correction target. A shorter kernel is used for minute-level correction to capture rapid fluctuations, and a longer kernel is used for hour-level correction to smooth instantaneous noise. A multi-scale correction parameter sequence is generated through the convolution operation, with each parameter corresponding to the power correction amount at a specific time point.

[0029] In the implementation, the device node set is stored in an in-memory database, supporting geospatial indexing and association strength threshold queries. The graph neural network employs an asynchronous training mechanism, pre-training using historical periodic component data and performing online fine-tuning during system runtime. The neighbor node retrieval process implements dual filtering: first, basic topology is filtered based on power grid physical connections, and then a secondary filtering is performed based on real-time association strength indices. The weighted aggregation step adopts a batch computation strategy, implementing grouped parallel processing for large-scale device clusters. The convolutional kernel design considers the continuity of time series, and the kernel weights adopt a causal structure to avoid future information leakage.

[0030] This implementation enhances the dynamic modeling capability of meteorological correlations between devices. The attention mechanism of graph neural networks can accurately capture the path dependence of meteorological fluctuation propagation, especially advantageous in the analysis of propagation shadow effects in complex terrain areas. The combination of weighted aggregation and convolution operations in the multi-scale correction process can introduce spatial smoothing effects while preserving the original fluctuation characteristics, reducing the interference of abnormal fluctuations of a single device on the correction results. The setting of the correlation strength index threshold balances computational accuracy and efficiency, avoiding information dilution effects from low-correlation devices. The multi-scale configuration of convolution kernels adapts to the fluctuation characteristics of different time dimensions; short kernels maintain the sensitivity of minute-level correction, while long kernels enhance the stability of hour-level correction.

[0031] Example 3: See Figure 4 The scale recursive feedback module dynamically optimizes system parameters through a multi-time-dimensional closed-loop adjustment mechanism. This module receives minute-level, hourly-level, and day-ahead-level wind and solar power prediction correction values ​​from the multi-scale correction module, forming a complete output sequence spanning from seconds to days. The minute-level correction results are continuously updated at 5-minute intervals, exhibiting high-frequency fluctuation characteristics, and are directly fed back to the power characteristic decoupling module to trigger incremental updates of the random fluctuation component set. The update process employs a sliding time window mechanism, retaining the latest 30 sampling points as a temporary buffer. When new data arrives, the oldest data points are removed, and the statistical characteristics of the random fluctuation components are recalculated. This real-time rolling update method ensures that the random fluctuation component set always reflects the short-term fluctuation patterns under current meteorological conditions.

[0032] The hourly correction results are summarized and analyzed at 15-minute intervals, exhibiting a mid-frequency fluctuation pattern. Upon receiving this level of data, the spatiotemporal correlation modeling module initiates an iterative optimization process for the device correlation strength index. The iterative algorithm employs a gradient descent-based parameter adjustment strategy, and the weight matrix W of the graph neural network is updated according to the following rules:

[0033] in: Let represent the weight matrix for the t-th iteration. The learning rate is dynamically adjusted. This is the gradient of the loss function with respect to the weights. The loss function calculates the mean squared error between the actual and predicted association strengths. The initial learning rate is set to 0.01 and decreases exponentially with each iteration. After each iteration, the meteorological propagation feature vectors between device nodes are recalculated to generate an updated exponential distribution of association strength. This process runs continuously, but the computation time for each iteration is limited to no more than 2 minutes to avoid affecting the real-time response performance of the system.

[0034] The deviation between the day-ahead correction results and the power grid dispatch plan, after time-series alignment, is compared to generate system-level feedback adjustment coefficients. The deviation comparison employs a sliding window correlation coefficient analysis method, with a window width set to 24 hours to cover the complete daily cycle. The correlation coefficient calculation considers the comprehensive matching degree between predicted and planned values ​​in three dimensions: waveform similarity, phase deviation, and amplitude difference. The feedback coefficients are dynamically generated by a proportional-integral-derivative controller. The controller's input variable is the 24-hour cumulative deviation, and the output variable is the adjustment coefficient within the range of 0.8 to 1.2. These coefficients are updated hourly and broadcast to the monitoring data fusion module via a message queue.

[0035] The feedback coefficient directly affects the core parameter configuration of the monitoring data fusion module. This includes the step size parameter of the adaptive sliding window. Dynamic scaling is applied based on a feedback coefficient. When the coefficient is greater than 1, the step size increases linearly to enhance data smoothing; when the coefficient is less than 1, the step size decreases to improve temporal resolution. The specific adjustment relationship is as follows: ,in It is the reference step size. For feedback coefficients, To adjust the sensitivity parameters, the sensitivity threshold of the dynamic threshold outlier detection algorithm is adjusted synchronously, using the moving percentile method to recalculate the upper and lower boundaries of the normal data range. When the feedback coefficient indicates a persistent negative deviation in the system, the sensitivity threshold is automatically widened to reduce false positives; when a positive deviation occurs, the threshold is tightened to enhance anomaly detection capabilities.

[0036] When receiving updated fused data, the power characteristic decoupling module employs an incremental signal decomposition strategy to optimize computational efficiency. The historical periodic fluctuation component set is stored in a circular buffer, with a buffer capacity set to 288 data points (corresponding to a 24-hour, 5-minute interval) based on a typical meteorological cycle. The empirical mode decomposition process for newly added data points inherits the residual components of the signal at the end of the buffer as initial conditions, avoiding the computational burden of recalculating the complete signal. The merging and recombining of high-frequency components uses a time-varying weighting method; the weight of new data points decays exponentially over time, while the weight of historical data increases accordingly, ensuring smooth component transitions without abrupt changes.

[0037] The spatiotemporal correlation modeling module monitors changes in the statistical characteristics of the recombined components. When the sliding variance of the random fluctuation component exceeds a preset threshold, it triggers a recalculation of the correlation strength. The variance threshold is set as a baseline based on historical meteorological data for the region and dynamically adjusted according to the feedback coefficient. The recalculation process prioritizes device nodes with sudden increases in variance, employing a local subgraph update strategy rather than global reconstruction, and only re-runs graph neural network inference on the affected nodes and their direct neighbors. The updated correlation strength index is synchronized to all relevant modules through a distributed caching system, with synchronization latency controlled within milliseconds to meet real-time requirements.

[0038] The parameter generation process of the multi-scale correction module incorporates the deep feature extraction capabilities of a temporal convolutional network. The network structure consists of three dilated convolutional layers, with each layer's output processed by a gated linear unit activation function. The dilation coefficients of the dilated convolutions are configured in a geometric progression of 1, 2, and 4, progressively expanding the receptive field to capture fluctuation features at different time scales. The network input is a tensor concatenation of the correction parameter set and real-time weather forecast data, with the output dimension matching the correction time scale. The minute-level output channel focuses on fluctuation details within a 5-minute time window, the hour-level output channel extracts trend features within a 60-minute range, and the day-ahead output channel focuses on the overall pattern of the 24-hour cycle.

[0039] The system's runtime data flow adopts a layered processing architecture. The bottom real-time data processing layer processes raw sensor data with a second-level latency; the middle feature fusion layer performs signal decomposition and correlation calculations at minute-level intervals; and the high-level decision feedback layer optimizes system parameters with hourly granularity. Data exchange between layers is achieved through message middleware, and the message protocol includes metadata such as timestamps, data dimensions, and quality flags. The resource scheduler monitors the computational load of each module and automatically reduces the priority of non-critical tasks during peak periods to ensure the real-time performance of the core correction process.

[0040] The feedback adjustment mechanism is designed with the inertial characteristics of the meteorological system in mind. Minute-level feedback employs a rapid response strategy, focusing on suppressing sudden fluctuations; hourly-level feedback emphasizes trend correction, gradually approaching optimal parameters through iterative optimization; and day-ahead-level feedback implements global adjustment to prevent systemic deviations caused by error accumulation. Priority marking is applied to the feedback signals at all three time scales during transmission to ensure that critical updates reach the target modules in a timely manner. The system's status monitoring interface displays the real-time operational status of each feedback loop, including key indicators such as data freshness, computational latency, and resource utilization.

[0041] Example 4: After receiving updated data from the monitoring data fusion module, the power characteristic decoupling module uses an incremental processing method to optimize computational efficiency. Taking a wind and solar power cluster in a certain region as an example, this cluster includes 12 wind turbines and 8 photovoltaic power stations, with a data sampling interval of 5 minutes. The historical periodic fluctuation component set maintained by the module is stored in a circular buffer with a capacity of 288 data points, corresponding to 24 hours of continuous monitoring data. When a new batch of fused monitoring data arrives, the system first verifies the continuity of the timestamps. After confirming that no data is lost, the earliest historical data is removed from the buffer, and the new data is appended to the end of the queue in chronological order.

[0042] The random fluctuation components are processed using a dynamic recombination strategy. High-frequency components generated from newly added data after empirical mode decomposition are merged with the historical random fluctuation component set. During the merging process, the system assigns a time-varying weight to each data point. The initial weight of a new data point is set to 0.7, decreasing by 0.02 every 5 minutes over time; the weights of historical data are increased accordingly, maintaining a total weight of 1. This weighting method ensures that recent fluctuation characteristics dominate while retaining the reference value of historical data. The recombination process of random fluctuation components in a photovoltaic power station during typhoon weather is illustrated in Table 1. The time column represents the relative temporal position of the data point, the new data marker column indicates whether it is a newly added data point, and the weight column reflects the influence of the point in the recombination calculation.

[0043] Table 1: Data on the recombination process of random fluctuation components of a photovoltaic power station under typhoon weather are as follows:

[0044] The spatiotemporal correlation modeling module continuously monitors the statistical characteristics of the recombined random fluctuation components. For the photovoltaic power station cluster in the example above, the system calculates the sliding window variance every 15 minutes. When the variance value of a device exceeds a preset threshold, a local topology graph update process is initiated. This threshold is dynamically set based on the device type: the baseline threshold for wind turbines is 25kW², and for photovoltaic power stations it is 16kW², fluctuating within ±30% in actual operation based on the feedback coefficient. The update process prioritizes selecting the three devices with the largest sudden increase in variance as central nodes, extracts their first-degree neighbor nodes in the topology graph to form a subgraph, and recalculates the correlation strength index only for nodes within this subgraph.

[0045] During subgraph updates, the graph neural network loads pre-trained weight parameters, with input data limited to the periodic fluctuation components of these nodes over the past 4 hours. The meteorological propagation feature vectors of neighboring nodes are generated through two rounds of message passing: the first round aggregates features from directly connected nodes, and the second round extends to the indirect influence of second-degree neighbors. After the new association strength index is calculated, it is distributed to the multi-scale correction module via incremental update packages marked with version numbers. The version control system records the device scope and timestamp of each update, supporting rollback to historical stable versions when necessary.

[0046] The multi-scale correction module receives the updated correlation strength index and immediately adjusts the weighted aggregation strategy. For device nodes with a sudden increase in variance, the system temporarily raises the participation threshold of their adjacent nodes, increasing the required correlation strength index from the usual 0.5 to 0.6. Simultaneously, an enhanced correction mode is activated for these devices, expanding the convolution kernel length from the standard 5 time points to 7 to improve fluctuation smoothing. After the correction parameters are generated, the module compares the output differences between the old and new versions. When the average change in the minute-level correction value exceeds 10%, an early warning signal is triggered to notify maintenance personnel to review the device status.

[0047] At the implementation level, the circular buffer employs memory-mapped file technology to improve access efficiency, supporting over 100,000 data point accesses per second. Weight coefficient calculation uses fixed-point arithmetic instead of floating-point arithmetic, saving 30% of computing resources on embedded devices. The subgraph update process is parallelized, with the update time for a single subgraph controlled within 200 milliseconds, ensuring that the system can complete all correlation strength index updates within 2 seconds even when 20 wind turbines experience simultaneous fluctuations. Version control uses distributed key-value storage; each update package is compressed to less than 500 bytes, enabling index synchronization of thousands of devices within 1 second on a 100Mbps network bandwidth.

[0048] The anomaly handling mechanism is designed to account for edge cases in actual operation. When the power deviation between new data and historical data exceeds three times the standard deviation for six consecutive sampling points, the system suspends the automatic reassembly process and switches to manual confirmation mode. If a change in the power grid topology is detected during the topology map update process (such as offline equipment maintenance), the correlation strength index of the relevant nodes is automatically frozen until the topology returns to stability. Version conflict resolution adopts a last-write-first strategy, while retaining conflict logs for post-event analysis.

[0049] The data visualization interface dynamically displays the evolution of the recombined components. Operators can observe the weight distribution heatmap of the random fluctuation components of specific equipment, as well as the spatiotemporal propagation animation of the correlation strength index. The interface provides a historical backtracking function, supporting the comparison of differences in component recombination under different meteorological conditions, such as comparing the distribution changes of photovoltaic module fluctuation characteristics in sunny and rainy weather. The system records a complete processing log, including the timestamp of each data update, the number of affected devices, calculation time, and other metadata, for subsequent operation optimization and fault diagnosis.

[0050] This implementation significantly reduces the computational burden on large-scale wind and solar clusters through incremental updates and local optimization strategies. Actual operational data from a certain region shows that, compared to the global reconstruction method, this scheme reduces the computational resources required for topology updates by 72%, while maintaining the accuracy of the correlation strength index within an acceptable range. Dynamic adjustment of weight coefficients enables the system to adapt to rapid changes in meteorological conditions, maintaining stable correction performance even during extreme weather events such as typhoons. The versioned update mechanism provides maintenance personnel with a clear operational traceability path, facilitating the analysis of the system's response characteristics to various meteorological scenarios.

[0051] Example 5: The multi-scale correction module couples deep feature extraction with real-time meteorological data to dynamically optimize wind and solar power predictions. This module receives updated equipment correlation strength indices from the spatiotemporal correlation modeling module and first adjusts the weighting strategy. After normalization, the correlation strength indices are combined with a time decay factor to generate the final weighting coefficients. The time decay factor is calculated based on data freshness, with recently updated indices receiving higher weights and historical data experiencing exponentially decreasing weights. This dynamic weighting mechanism enables the system to adapt to changes in meteorological conditions and respond quickly to sudden weather events.

[0052] The processing of the modified parameter set employs a temporal convolutional network for multi-scale feature extraction. The network structure comprises three layers, each using convolutional kernels of different sizes to capture fluctuation patterns within a corresponding time range. The first layer focuses on short-term fluctuations, with kernels covering a 5-minute time window to extract rapid changes from the second to the minute scale. The second layer focuses on medium-term trends, with kernels extending to a 30-minute range to identify power fluctuation patterns from the ten-minute to the hourly scale. The third layer analyzes long-term patterns, with kernels spanning up to 2 hours to capture the overall trend from the hourly to the day-ahead scale. The output of each convolutional layer is filtered through a gating mechanism to retain features relevant to the current meteorological conditions and suppress noise interference.

[0053] Real-time weather forecast data and feature extraction results are spatiotemporally aligned before being input into the coupled computation unit. The weather data includes 15-minute forecasts of wind speed, irradiance, and cloud cover for the next 6 hours, as well as auxiliary parameters such as temperature and humidity. The data alignment process considers the spatial mapping relationship between equipment locations and the meteorological grid, using bilinear interpolation to match the grid forecast data to the coordinates of each wind and solar equipment. The coupled computation employs a feature-level fusion strategy, concatenating meteorological parameters as additional channels with convolutional features and inputting them together into the fully connected layer to generate correction values. During computation, the importance weights of meteorological parameters are dynamically allocated through an attention mechanism, automatically increasing the decision weights of key meteorological factors under complex conditions such as severe convective weather.

[0054] The correction values ​​are output using a rolling update mechanism. Minute-level corrections are released every 5 minutes, predicting power changes for the next 5-15 minutes based on the latest measured data and short-term characteristics. Hourly-level corrections are updated every 15 minutes, integrating medium-term characteristics and weather forecasts to provide power trends for the next 1-3 hours. Day-ahead corrections are adjusted hourly, combining long-term characteristics and weather pattern analysis to output the power profile for the next 24 hours. Correction results at all time scales are published through a unified interface, maintaining strict timestamp synchronization to avoid forecast conflicts between different scales.

[0055] The system implements multi-level quality control during operation. Integrity checks are performed during the raw data input phase; a data source switching mechanism is triggered if more than 10% of the data is missing. During feature extraction, the signal-to-noise ratio of the convolution output is monitored; when feature quality falls below a threshold, the system automatically downgrades to a simplified calculation mode. Before the corrected results are released, a rationality check is performed, comparing them with physical constraints to ensure that power values ​​are within the equipment's rated capacity and that ramp-up rates meet unit performance limits. Abnormal situations trigger automatic alarms, and detailed diagnostic information is recorded for subsequent analysis.

[0056] The computing resource management employs a dynamic allocation strategy. During periods of stable weather, the computation frequency of convolutional networks is reduced to conserve processing power. When sudden changes in wind speed or rapid cloud movement are detected, computation priority is automatically increased, shortening the feature extraction cycle. The resource scheduler balances the load of each module in real time, prioritizing resource allocation for minute-level correction channels when computationally intensive tasks are queued. The distributed computing framework supports horizontal scaling, dynamically adding container instances during peak periods to handle traffic surges.

[0057] The human-computer interface provides multi-dimensional operation monitoring functions. Operators can view the real-time correction values ​​and their composition analysis for each device, and understand the impact of different meteorological factors. Historical correction records support categorized retrieval by time scale, facilitating comparative analysis of the performance differences of different algorithms. The system health status panel displays key indicators such as feature extraction time, data freshness, and resource utilization, assisting in operation and maintenance decisions. The configuration interface allows expert users to adjust core parameters such as convolution kernel parameters and weight allocation rules to adapt to the specific operating characteristics of a particular site.

[0058] This implementation significantly improves the spatiotemporal consistency of correction values ​​by deeply integrating temporal features and weather forecasts. The multi-level structure of the temporal convolutional network effectively captures the scale characteristics of wind and solar power fluctuations, avoiding the limitations of single-time-scale analysis. The dynamic weighting mechanism enables the system to balance historical patterns with real-time changes, maintaining stable output during periods of meteorological stability and promptly tracking actual fluctuations during periods of rapid change. The meteorological attention mechanism of the coupled computational units enhances the characterization capability of key weather processes, particularly improving correction accuracy under complex meteorological conditions such as frontal passage and thunderstorm development.

[0059] The system architecture design takes into account practical engineering constraints, achieving a balance between computational accuracy and real-time performance. The modular processing flow allows for individual optimization of each functional unit, facilitating subsequent algorithm upgrades. Layered quality control mechanisms reduce the propagation of anomalies, enhancing overall robustness. The flexible resource scheduling strategy adapts to the deployment needs of sites of varying sizes, effectively supporting everything from distributed small-scale photovoltaic power plants to centralized large-scale wind farms. The layered information presentation of the visualization interface satisfies the simplicity requirements of daily monitoring while preserving complete data traceability capabilities for in-depth analysis.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-scale adaptive correction system for wind and solar power, characterized in that, include: The monitoring data fusion module, power characteristic decoupling module, spatiotemporal correlation modeling module, and multi-scale correction module are included. The monitoring data fusion module is used to receive wind and solar power time series data collected by multi-source monitoring equipment in the sky and ground network, and to perform spatiotemporal alignment and quality verification of multi-source heterogeneous data; The power characteristic decoupling module separates the periodic fluctuation component and the random fluctuation component of wind and solar power based on the verified time-series data. The spatiotemporal correlation modeling module establishes a regional wind and solar equipment topology map based on the periodic fluctuation components, and calculates the meteorological correlation strength index between the equipment based on the topology map; The multi-scale correction module uses the meteorological correlation strength index to perform cross-device collaborative correction of random fluctuation components, generating multi-scale wind and solar power prediction correction values.

2. The wind and solar power multi-scale adaptive correction system according to claim 1, characterized in that, The steps of the monitoring data fusion module in performing spatiotemporal alignment and quality verification of multi-source heterogeneous data include: Acquire satellite cloud imagery data, meteorological radar data, ground sensor network data, and historical power data from the power grid dispatching terminal; Time-stamped normalization was performed on the four types of data, and an adaptive sliding window algorithm was used to unify the time resolution. Anomaly data points are identified using a dynamic threshold outlier detection algorithm, and data imputation is performed using spatial interpolation between adjacent sites. Output a fusion monitoring dataset that is aligned in the time dimension and whose integrity meets preset standards.

3. The wind and solar power multi-scale adaptive correction system according to claim 2, characterized in that, The power characteristic decoupling module separates components by the following steps: Extract wind speed, irradiance, and actual power sequences for the target area from the fusion monitoring dataset; The empirical mode decomposition algorithm is used to decompose the actual power sequence into high-frequency components and low-frequency components; Based on the spectral characteristics of wind speed and irradiance sequences, high-frequency components are mapped to a set of random fluctuation components. After phase matching between the low-frequency components and the periodic characteristics of meteorological elements, they are incorporated into the periodic fluctuation component set.

4. The wind and solar power multi-scale adaptive correction system according to claim 3, characterized in that, The execution steps of the spatiotemporal correlation modeling module include: A set of equipment nodes is constructed based on the geographical distribution coordinates of regional wind and solar equipment and the power grid topology; The set of periodic fluctuation components is input into a graph neural network to learn the meteorological propagation feature vectors between device nodes; Calculate the device association strength index based on cosine similarity of meteorological propagation feature vectors; Output a topology map of wind and solar power equipment with correlation strength weights.

5. The wind and solar power multi-scale adaptive correction system according to claim 4, characterized in that, The correction steps of the multi-scale correction module include: Extract the power residual value of the target device from the set of random fluctuation components; Retrieve adjacent device nodes in the wind and solar equipment topology map that have a strong association with the target device; The random fluctuation components of adjacent devices are weighted and aggregated based on the device correlation strength index; The aggregation result is convolved with the power residual value of the target device to generate a set of correction parameters.

6. The wind and solar power multi-scale adaptive correction system according to claim 5, characterized in that, The system also includes a scale recursive feedback module, whose execution steps include: Obtain minute-level, hourly-level, and day-ahead-level output results of multi-scale wind and solar power prediction correction values; The minute-level correction results are input into the power feature decoupling module to update the set of random fluctuation components; The hourly correction results are input into the spatiotemporal correlation modeling module to iterate the device correlation strength index; The deviation between the day-ahead correction results and the power grid dispatch plan is compared to generate feedback coefficients.

7. The wind and solar power multi-scale adaptive correction system according to claim 6, characterized in that, The feedback coefficients of the scale recursive feedback module act on the monitoring data fusion module, specifically including: The step size parameter of the adaptive sliding window is dynamically adjusted based on the feedback coefficient; the sensitivity threshold for outlier detection is optimized based on the deviation comparison results.

8. The wind and solar power multi-scale adaptive correction system according to claim 7, characterized in that, The power characteristic decoupling module performs incremental decomposition after receiving updated data, specifically including: The set of periodic fluctuation components in the historical decomposition results is retained, and empirical mode decomposition is performed only on the newly added fused monitoring data; The newly added high-frequency components are merged and reorganized with the set of historical random fluctuation components.

9. The wind and solar power multi-scale adaptive correction system according to claim 8, characterized in that, The spatiotemporal correlation modeling module updates the topology graph based on the recombined components, specifically including: Extract the variance characteristics of the random fluctuation components after the merger and reorganization; When the variance characteristic exceeds the preset fluctuation threshold, the device correlation strength index is recalculated. The updated device association strength index will be synchronized to the multi-scale correction module.

10. The wind and solar power multi-scale adaptive correction system according to claim 9, characterized in that, The multi-scale correction module also includes: The weight allocation of the weighted aggregation is adjusted according to the updated device association strength index, and a temporal convolutional network is used to extract multi-scale features from the corrected parameter set. The feature extraction results are coupled with real-time weather forecast data to calculate and output a rolling updated multi-scale wind and solar power prediction correction value.

Images