Spatiotemporal feature missing value detection method and system for photovoltaic power prediction

Abstract

The application provides a spatiotemporal feature missing value detection method and system for photovoltaic power prediction, and relates to the technical field of photovoltaic power prediction.The method comprises the following steps: acquiring an original data sequence of a distributed photovoltaic power station group, performing time alignment and spatial coordinate mapping on the original data sequence to obtain a multi-dimensional feature data cube; based on the multi-dimensional feature data cube, establishing a spatiotemporal correlation constraint, identifying data points that simultaneously violate the time sequence continuity and the spatial consistency, and obtaining a spatiotemporal joint missing candidate set; for the spatiotemporal joint missing candidate set, using a physical coupling rule between meteorological data and power to perform verification, removing data points that meet the physical fluctuation rule, and obtaining a missing positioning result after physical verification correction.The application can realize accurate and reliable missing value detection, and improve the stability and accuracy of photovoltaic power prediction.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and in particular to a method and system for detecting missing spatiotemporal feature values ​​for photovoltaic power prediction. Background Technology

[0002] Distributed photovoltaic (PV) power plant clusters are characterized by dispersed sites, dense layout, and heterogeneous data from multiple sources. Their spatiotemporal characteristic data, such as irradiance, temperature, wind speed, and power, are the core inputs for achieving high-precision power prediction. In practical engineering scenarios, factors such as sensor failures, communication interruptions, data acquisition delays, and sudden changes in regional weather patterns make it difficult to accurately detect and locate missing feature values ​​in spatiotemporal scenarios, becoming a key bottleneck restricting the accuracy of PV power prediction.

[0003] Current methods for detecting missing values ​​mostly rely solely on temporal continuity or spatial consistency, and generally suffer from the following technical shortcomings: They lack joint constraints between the temporal and spatial dimensions, making it difficult to identify missing points that simultaneously violate both temporal patterns and spatial proximity consistency. This leads to misjudging localized meteorological fluctuations as data missing in park-type and contiguous distributed photovoltaic (PV) clusters, or overlooking true missing values ​​in large areas or continuous time periods, resulting in numerous false positives and false negatives. Furthermore, the detection output is merely a simple missing label, without quantifying and classifying the severity based on the physical coupling laws of PV and the degree of missing data. For example, in practical applications such as large-scale contiguous PV cluster scheduling and prediction, and concentrated power prediction in parks, it is difficult to provide the prediction model with adaptively callable missing confidence information. This prevents the model from timely and appropriately reducing the weight of missing data areas during feature extraction and spatial aggregation, ultimately amplifying power prediction errors and affecting the smooth operation of PV power plant scheduling and power trading. Summary of the Invention

[0004] This invention provides a method and system for detecting missing values ​​of spatiotemporal features in photovoltaic power prediction. It can achieve accurate and reliable detection of missing values, improve the stability and accuracy of photovoltaic power prediction, and is suitable for daily monitoring and scheduling prediction scenarios of various large-scale photovoltaic power plants.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for detecting missing spatiotemporal features in photovoltaic power prediction, the method comprising: The original data sequence of the distributed photovoltaic power station group is obtained, and the original data sequence is time-aligned and spatially mapped to obtain a multi-dimensional feature data cube. Based on a multidimensional feature data cube, spatiotemporal correlation constraints are established to identify data points that simultaneously violate temporal continuity and spatial consistency, thus obtaining a spatiotemporal joint missing candidate set. For the spatiotemporal joint missing candidate set, the physical coupling rule between meteorology and power is used for verification. Data points that conform to the physical fluctuation law are removed to obtain the missing location result after physical verification and correction. Based on the missing location results after physical verification and correction, the missing events are classified into levels, and each level is assigned a corresponding confidence weight coefficient to obtain a quantitative map of missing event levels. The missing level quantization map is fused with the spatiotemporal joint missing candidate set and the missing location results after physical verification and correction to construct a dynamic missing mask matrix for use by a preset prediction model. The dynamic missing mask matrix is ​​used as the detection result of spatiotemporal feature missing values.

[0006] Furthermore, the original data sequence of the distributed photovoltaic power station cluster is obtained, and the original data sequence is time-aligned and spatially mapped to obtain a multi-dimensional feature data cube, including: The original data sequence of five types of characteristic data of each power station within a continuous time window with a preset time resolution is obtained, and the original data sequence carries the spatial coordinate information of each power station. Based on the acquired original data sequence, time alignment is performed using a unified standard timestamp as a benchmark. The statistical distribution characteristics of each feature sequence within a continuous time window are analyzed and parameters are fitted to aggregate the five types of feature data from each power station at the same time point under the same time index, resulting in a time-aligned data set. Based on the time-aligned data set, a spatial location mapping relationship is established within a preset spatial range according to the spatial coordinate information of each power station. By fitting the spatial topology and distance measurement parameters between the nodes of each power station, a unique spatial node identifier is assigned to each power station to construct a spatial graph structure. Based on the constructed spatial graph structure, the time-aligned data set is reorganized according to the time dimension, spatial dimension and feature dimension. By parametrically modeling the coupling relationship between the time dimension, spatial dimension and feature dimension, a multidimensional feature data cube is formed with time as the first index, spatial node as the second index and feature type as the third index.

[0007] Furthermore, based on a multidimensional feature data cube, spatiotemporal correlation constraints are established to identify data points that simultaneously violate temporal continuity and spatial consistency, resulting in a spatiotemporal joint missing data candidate set, including: Extract all feature data of each spatial node under each time index from the multidimensional feature data cube to form a spatiotemporal sequence set with spatial nodes as units and time axis as the order; Based on the spatiotemporal sequence set, a spatial topology graph is constructed with each spatial node as a vertex and a preset spatial distance threshold as an edge connection condition. By combining the historical sequence of each node on the time axis, a spatiotemporal association constraint model that simultaneously includes spatial adjacency and temporal dependency is established. Based on the spatiotemporal correlation constraint model, the local outlier factor of each data point in the spatiotemporal neighborhood is calculated. At the same time, combined with the expected value range of each feature dimension within the preset confidence interval, data points that do not meet the historical change pattern in time and are inconsistent with the data of neighboring nodes in space are identified and marked as suspected missing points. Based on the labeled suspected missing points, the missing points are classified and aggregated according to the missing type. Data points with multiple spatial nodes missing the same feature at the same time are grouped into spatial patch missing points, and data points with different features missing at the same spatial node at multiple consecutive time points are grouped into temporal cross-feature missing points. At the same time, the union of spatial patch missing points and temporal cross-feature missing points is used as a candidate set of spatiotemporal joint missing points.

[0008] Furthermore, for the spatiotemporal joint missing candidate set, the physical coupling rules between meteorology and power are used for verification. Data points that conform to physical fluctuation patterns are removed, resulting in the missing location results after physical verification and correction, including: Each suspected missing point in the spatiotemporal joint missing candidate set is obtained, and its spatial node identifier, time index and feature type are extracted. At the same time, the irradiance data, power data and component temperature data of the same spatial node under adjacent time indices are read from the multidimensional feature data cube. Based on the read data, a physical coupling discrimination model is constructed with the ratio of irradiance change rate to power change rate as the core and the correction coefficient of component temperature on power conversion efficiency as the constraint. Substitute each suspected missing point into the physical coupling discrimination model, calculate the actual ratio of its irradiance change rate to power change rate, and combine it with component temperature data to calculate the corrected theoretical ratio range. If the actual ratio falls within the theoretical ratio range, the suspected missing point is determined to be a data point that conforms to the physical fluctuation law; otherwise, it is determined to be a real missing point. Remove all data points that conform to the laws of physical fluctuations from the spatiotemporal joint missing candidate set, retain all real missing points, and reorganize the real missing points according to spatial node identifier, time index and feature type to obtain the missing location result after physical verification and correction.

[0009] Furthermore, based on the missing event localization results corrected by physical verification, the missing events are classified into levels, and each level is assigned a corresponding confidence weight coefficient to obtain a quantitative missing event level map, including: Obtain the missing location results after physical verification and correction, and aggregate the missing points according to the spatial node identifier and time index. This is used to group the missing points that occur in the same spatial node within a continuous time window into a single missing event, thus obtaining a set of missing events. For each missing event in the missing event set, calculate its continuous missing duration in the time dimension, the number of affected power plants that are simultaneously missing in the spatial dimension, and the number of feature types involved in the missing in the feature dimension to obtain the three-dimensional quantitative evaluation index of this missing event. Based on three-dimensional quantitative evaluation indicators, each missing event is divided into three levels: single-point random missing, continuous time period missing, or large area missing. A preset confidence weight coefficient is assigned to each missing event according to its level. The preset confidence weight coefficient is negatively correlated with the severity of the missing event. The spatial node identifier, time index, missing level, and corresponding confidence weight coefficient of each missing event are correlated and mapped to form a missing level quantitative map covering all missing events.

[0010] Furthermore, the missing level quantization map is fused with the spatiotemporal joint missing candidate set and the physically verified and corrected missing location results to construct a dynamic missing mask matrix for use by a preset prediction model. The dynamic missing mask matrix is ​​then used as the detection result of spatiotemporal feature missing values, including: The missing level quantization map, the spatiotemporal joint missing candidate set, and the missing location results after physical verification and correction are aligned to the same time index and spatial node identification system. Using the physical verification and correction results of missing locations as the basic positioning layer, we mark whether there are real missing locations at each spatiotemporal point. At the same time, we use the spatiotemporal joint missing candidate set as the auxiliary verification layer to mark the missing type of each spatiotemporal point in the spatiotemporal joint dimension, thus forming fused positioning information that includes missing location identifiers and missing type identifiers. Based on the fused positioning information, the missing level and confidence weight coefficient in the missing level quantization map are superimposed on the corresponding spatiotemporal points to form a multidimensional missing information matrix containing missing location identifier, missing type identifier, missing level identifier and confidence weight coefficient. The multidimensional missing information matrix is ​​transformed into a dynamic missing mask matrix in a preset format. The dynamic missing mask matrix is ​​used by the preset prediction model to perform adaptive weight reduction processing on the missing regions during the temporal feature extraction and spatial feature aggregation process. At the same time, the dynamic missing mask matrix is ​​used as the detection result of the spatiotemporal feature missing values.

[0011] Secondly, a spatiotemporal feature missing value detection system for photovoltaic power prediction includes: The data preprocessing module is used to obtain the original data sequence of the distributed photovoltaic power station group, perform time alignment and spatial coordinate mapping on the original data sequence, and obtain a multi-dimensional feature data cube. The spatiotemporal missing data filtering module is used to establish spatiotemporal correlation constraints based on multidimensional feature data cubes, identify data points that simultaneously violate temporal continuity and spatial consistency, and obtain a spatiotemporal joint missing data candidate set. The physical verification module is used to verify the spatiotemporal joint missing candidate set using the physical coupling rules between meteorology and power, removing data points that conform to the physical fluctuation law, and obtaining the missing location result after physical verification and correction. The level quantification module is used to classify missing events into levels based on the missing location results after physical verification and correction, and to assign a corresponding confidence weight coefficient to each level to obtain a missing level quantification map. The mask construction module is used to fuse the missing level quantization map with the spatiotemporal joint missing candidate set and the missing location results after physical verification and correction to construct a dynamic missing mask matrix for use by a preset prediction model. The dynamic missing mask matrix is ​​used as the detection result of spatiotemporal feature missing values.

[0012] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0013] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0014] The above-described solution of the present invention has at least the following beneficial effects: By employing techniques such as multi-dimensional feature data cube construction, spatiotemporal correlation constraint modeling and identification, meteorological and power physics coupling rule verification, missing event level quantification assignment, and multi-dimensional missing information fusion dynamic mask matrix generation, this method overcomes the technical problems of traditional missing value detection methods, such as lack of spatiotemporal joint constraints, high false positive and false negative rates, lack of verification in conjunction with the physical laws of photovoltaic power generation, and difficulty in quantifying and classifying missing events. This enables accurate positioning and reliable identification of spatiotemporal feature missing values ​​in distributed photovoltaic power station clusters, reduces the probability of misjudgment caused by local meteorological fluctuations, provides suitable missing confidence weight coefficients for photovoltaic power prediction models, and achieves adaptive weight reduction processing of missing data regions during the prediction process. Ultimately, this improves the accuracy and operational stability of photovoltaic power prediction and is suitable for daily monitoring and scheduling prediction scenarios of various large-scale distributed photovoltaic power station clusters. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a spatiotemporal feature missing value detection method for photovoltaic power prediction provided by an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of a spatiotemporal feature missing value detection system for photovoltaic power prediction provided by an embodiment of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0018] like Figure 1 As shown, embodiments of the present invention propose a method for detecting missing spatiotemporal feature values ​​for photovoltaic power prediction, the method comprising the following steps: Step 1: Obtain the original data sequence of the distributed photovoltaic power station group, perform time alignment and spatial coordinate mapping on the original data sequence to obtain a multi-dimensional feature data cube; Step 2: Based on the multidimensional feature data cube, establish spatiotemporal correlation constraints, identify data points that simultaneously violate temporal continuity and spatial consistency, and obtain a spatiotemporal joint missing candidate set; Step 3: For the spatiotemporal joint missing candidate set, the physical coupling rule between meteorology and power is used for verification. Data points that conform to the physical fluctuation law are removed to obtain the missing location result after physical verification and correction. Step 4: Based on the missing location results after physical verification and correction, the missing events are classified into levels, and each level is assigned a corresponding confidence weight coefficient to obtain a quantitative map of missing event levels. Step 5: The missing level quantization map is fused with the spatiotemporal joint missing candidate set and the missing location results after physical verification and correction to construct a dynamic missing mask matrix for use by a preset prediction model. The dynamic missing mask matrix is ​​used as the detection result of spatiotemporal feature missing values.

[0019] In this embodiment of the invention, a multi-dimensional feature data cube is constructed by performing time alignment and spatial coordinate mapping on the original data of distributed photovoltaic power station groups, laying a data foundation. By combining spatiotemporal correlation constraints to identify suspected missing points, data with both temporal and spatial anomalies can be accurately screened out, effectively reducing false detections and missed detections in missing data. Verification is carried out based on meteorological and power physics coupling rules to eliminate false judgments of normal fluctuations, improving the accuracy of missing data location. Missing events are classified and assigned confidence weight coefficients to form a complete quantitative map, realizing a refined assessment of missing data. A dynamic missing data mask matrix is ​​constructed by integrating missing data from multiple sources, which can provide a reliable basis for calling the prediction model and realize adaptive weight reduction of missing regions. This invention can improve the reliability and accuracy of spatiotemporal feature missing value detection, reduce the interference of data missing data on the prediction process, and effectively improve the accuracy of photovoltaic power prediction.

[0020] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Obtain the original data sequence of five characteristic data types for each power station within a continuous time window, including irradiance, ambient temperature, module temperature, wind speed, and power, at a preset time resolution. The original data sequence carries the spatial coordinate information of each power station. Specifically, for the target distributed photovoltaic power station group, with a preset time resolution of 15 minutes, a continuous 72-hour time window is traced back. The original time-series data of the five characteristics (irradiance, ambient temperature, module temperature, wind speed, and power) for each power station within this window are retrieved in parallel through the data acquisition interface. During the retrieval process, for each characteristic type, the original sequence is arranged in ascending order of time, ensuring that each data point is accompanied by precise... The timestamp is accurate to the second. Simultaneously, based on the unique number of each power station, the corresponding spatial coordinate information is retrieved from the registered power station asset information database. The asset information database is a pre-built and continuously maintained basic data storage system that records the static attribute parameters of each power station, including key spatial coordinate information such as longitude, latitude, and altitude, as well as basic configuration parameters such as the power station's installed capacity, component type, and grid connection method. The retrieved spatial coordinate information is bound to the corresponding power station's original data sequence in the form of key-value pairs. The bound data structure ensures that all subsequent processing steps can trace back to the precise spatial location based on the power station number, laying the foundation for spatial dimension analysis.

[0021] Step 1.2: Based on the acquired original data sequence, time alignment is performed using a unified standard timestamp as a benchmark. The statistical distribution characteristics of each feature sequence within a continuous time window are analyzed and parameters are fitted to aggregate the five types of feature data from each power station at the same time point under the same time index, resulting in a time-aligned dataset. Specifically, this includes: projecting the original data sequence of each power station onto a unified standard time axis according to its timestamp, with the standard time axis scaled at the hour or 15-minute intervals; for data points where the original timestamp deviates from the standard time scale, cubic spline interpolation is used for time alignment. During interpolation, two valid data points before and after the interpolation point are selected as control points to fit a smooth curve, thereby calculating the feature values ​​under the standard time scale and ensuring that the time-aligned sequence maintains its original trend; for each power station... certain features Define its aligned time series as ,in For standard time index, where t =1 , 2 ,…,T ( T (Total number of time indices) correspond to each standard time point under the preset time resolution; calculate the moving mean and moving standard deviation of this series within the time window, specifically by: at time... Take the window length centered at that moment as Time Neighborhood Then the moving average and sliding standard deviation They are represented as follows: ; ; in, Display window The number of time points contained within; based on this, the temporal fluctuation envelope of this feature under normal operating conditions is fitted, that is, its expected value range is defined as follows. ,in The preset confidence interval coefficient is used to control the width of the normal fluctuation range; ultimately, all power plants are indexed at the same standard time. The five types of feature values ​​are grouped into the same data record to form a time-aligned data set. Each record in this set contains a timestamp, a power plant identifier, and the five types of feature values ​​at that time, providing a unified temporal reference for subsequent spatial dimension modeling.

[0022] Step 1.3: Based on the time-aligned dataset, establish a spatial location mapping relationship within a preset spatial range according to the spatial coordinate information of each power station. By fitting the spatial topology and distance metric parameters between the nodes of each power station, assign a unique spatial node identifier to each power station and construct a spatial graph structure. Specifically, this includes: using the spatial coordinate information of all power stations as input, defining a preset spatial range containing all target power stations. Within this range, for any two power stations... a and b Let their spatial coordinates be respectively and The straight-line geographical distance between the two Calculated using the semi-versus formula: ; In the formula, R The average radius of the Earth is taken as 6371 kilometers; It is a power station a latitude, It is a power station b latitude, It is a power station a longitude, It is a power station b Longitude is expressed in radians; For power station a With power station b Latitude difference, For power station a With power station b The longitude difference can be expressed as follows: This distance is used as a measure of spatial proximity to quantify the degree of spatial coupling between power plants.

[0023] Set an adjacency threshold (Typical engineering values: 500m to 3000m, adapted according to the spatial distribution density of the photovoltaic power station group), when the distance between two power stations meets the following conditions... When a spatial adjacency relationship exists between two power stations, a spatial graph structure is constructed by traversing all power station pairs, with power stations as vertices and adjacency relationships as edges. ,in Let be the vertex set consisting of all power station nodes. An edge set that satisfies the adjacency condition; in this graph structure, each power station is assigned a unique spatial node identifier. And record the set of all adjacent nodes of the node. and the corresponding distance metric parameters ,in The diagram fully depicts the topological connections and coupling tightness of distributed photovoltaic power station clusters in physical space, providing a clear adjacency definition for joint constraints in the spatial dimension.

[0024] Step 1.4: Based on the constructed spatial graph structure, the time-aligned data set is reorganized according to the time dimension, spatial dimension, and feature dimension. By parametrically modeling the coupling relationship between the time dimension, spatial dimension, and feature dimension, a multi-dimensional feature data cube is formed, with time as the first index, spatial nodes as the second index, and feature type as the third index. Specifically, this includes: fusing and reorganizing the time-aligned data set with the spatial graph structure, using a standard time index... As the first index dimension, it is identified by all nodes in the spatial graph structure. As the second index dimension, it includes five categories: irradiance, ambient temperature, component temperature, wind speed, and power. ;in Using {irradiance, ambient temperature, component temperature, wind speed, and power} as the third index dimension, a three-dimensional data structure is constructed, denoted as . ,in T The total number of time indices. V Where F is the total number of spatial nodes, and F is the total number of feature types; The set of real numbers indicates that all data values ​​stored in the cube are real numbers; when constructing the cube, the coupling relationship between the three dimensions of time, space and features is parametrically modeled, and the constraint relationship between the dimensions is explicitly defined.

[0025] Temporal dependency modeling in the time dimension includes: for any spatial node and arbitrary features Define its index in standard time. value at The dependency relationship between it and its historical values ​​is in the form of autoregression: ;in, P The autoregressive order is automatically determined during the modeling process using the Akaike Information Criterion; For nodes In features The first The first-order autoregressive coefficients are fitted by minimizing the sum of squared residuals; The residual term is used to characterize random fluctuations that cannot be explained by historical values; the model captures the continuity of each spatiotemporal feature point in previous time steps.

[0026] Spatial adjacency relevance modeling includes: for the same feature f Same standard time index t Define any two adjacent nodesv and u (satisfy The spatial correlation between them is: ;in, These are the spatial weighting coefficients, calculated based on the distance metric parameters in step 1.3. Fitting using a Gaussian kernel function: ;in, This is a spatial bandwidth parameter used to control the rate at which the spatial influence range decays with increasing distance; For nodes v Summing all adjacent nodes and using the sum as the denominator to normalize the weights, ensuring that the nodes... v The sum of the weights of all adjacent nodes is 1; For traversing nodes v A temporary variable used for summation operations when considering all adjacent nodes; It is an exponential function used to convert distance metrics into non-negative weight values; The residual term is used to characterize the nodes. v Random fluctuation components that cannot be explained by the spatial neighborhood at the current point in time and space. Adjacent nodes u In the same standard time index t The same feature f The value of ; this modeling method establishes quantifiable association constraints between each data point in the cube and other data points in its spatial neighborhood.

[0027] Modeling the physical coupling between feature dimensions, including: for the same spatial node v and at the same time t Establish the physical coupling relationship between irradiance, component temperature, and power; let the irradiance value be... The component temperature is The power value is Define the constraint relationship between the three as follows: ;in, For nodes v The photoelectric conversion efficiency benchmark coefficient reflects the basic conversion capability of photovoltaic modules under standard test conditions; The temperature decay coefficient characterizes the negative impact on output power caused by a unit increase in component temperature. For reference temperature, 25 degrees Celsius is usually taken; This is a residual term used to absorb influencing factors other than irradiance and temperature; this modeling method minimizes the error between actual and theoretical power for the parameters. and By performing fitting, the physical coupling law between meteorological characteristics and power characteristics was explicitly characterized within the cube; After the parametric modeling described above, a multidimensional feature data cube X is finally formed, in which each data point is indexed by a standard time. t Spatial nodes v and feature type f It is uniquely determined, and it establishes quantifiable coupling constraints with its temporal neighborhood, spatial neighborhood, and other features of the same node, providing a unified data storage and computational basis for subsequent spatiotemporal joint missing detection.

[0028] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Extract all feature data of each spatial node under each time index from the multidimensional feature data cube to form a spatiotemporal sequence set with spatial nodes as units and time axis as the order. Specifically, this includes: based on the constructed multidimensional feature data cube, the cube organizes data with time as the first index, spatial node as the second index, and feature type as the third index; the extraction operation traverses according to the spatial node identifier, selecting each spatial node as the current processing unit in turn. For the selected spatial node, along the time dimension, from the first time index to the last time index, extract all five types of feature data of the node at that time, specifically including the values ​​of irradiance, ambient temperature, component temperature, wind speed, and power; during the extraction process, strict... The five-dimensional feature vectors at each moment are arranged sequentially in ascending order of time index to form a complete time series. At the same time, this time series is bound to the static attributes of the spatial node, including the node's unique identifier, spatial coordinates, and the list of adjacent nodes determined in step 1.3. After extracting all the time index data of a spatial node, the next spatial node is processed until all spatial nodes have been traversed. Finally, the time series of all spatial nodes are aggregated to form a spatiotemporal sequence set. Each element in this set corresponds to a spatial node, and each element contains all the feature data of that node arranged in chronological order within the entire time window, providing a basic data organization form for establishing spatiotemporal correlation constraints.

[0029] Step 2.2: Based on the spatiotemporal sequence set, construct a spatial topology graph with each spatial node as a vertex and a preset spatial distance threshold as the edge connection condition. By combining the historical sequences of each node on the time axis, establish a spatiotemporal association constraint model that simultaneously includes spatial adjacency relationships and temporal dependencies. Specifically, this includes: constructing a spatial topology graph based on the formed spatiotemporal sequence set; treating each spatial node in the spatiotemporal sequence set as an independent vertex in the graph, and then comparing the calculated geographical distance between each pair of nodes with a preset spatial distance threshold; the spatial distance threshold is set based on the actual spatial distribution characteristics of the distributed photovoltaic power station group, comprehensively considering meteorological spatial correlation. The attenuation scale, geographical differences between power stations, and communication coverage of the data acquisition system in actual engineering are all considered. Typically, this threshold is between several hundred meters and several kilometers, with the specific value determined through statistical analysis of the spatial distribution density of the power station group. This ensures that the threshold can effectively distinguish between spatially significantly correlated neighboring power stations and spatially independent, distant power stations, thus ensuring that each edge in the constructed spatial topology graph represents a pair of nodes with high coupling in meteorological conditions and data acquisition environment. When the geographical distance between two nodes is less than or equal to this threshold, it is determined that these two nodes have a direct spatial adjacency relationship, and an undirected edge is established between the corresponding two vertices. After determining all node pairs, a complete spatial topology graph is formed. This spatial topology graph reflects the physical spatial connection structure of the distributed photovoltaic power station group. Neighboring power station nodes are directly connected by edges in the graph, while distant nodes are not directly connected by edges.

[0030] A spatiotemporal correlation constraint model is established by introducing temporal constraints on the basis of a spatial topology graph. This model is constructed based on the fundamental ideas of graph neural networks, particularly the architecture of spatial graph convolutional networks, extending it to the temporal dimension to form a joint modeling framework that integrates spatial topology and temporal dependencies. Specifically, the model employs an encoder-like architecture, with inputs comprising two core components: first, the adjacency matrix of the spatial topology graph, which records the edge connections between nodes, reflecting the spatial coupling structure of the power plant group; and second, the historical sequence data of each node on the time axis, which constitutes the temporal feature evolution trajectory of each node. By constructing multi-layer spatiotemporal convolutional modules, spatial information aggregation and temporal state transfer are alternately performed in each layer, thereby achieving a deep fusion of spatial adjacency relationships and temporal dependencies.

[0031] The model construction process first defines a spatial dimension aggregation operation. For each spatial node, the information of its neighboring nodes is aggregated through graph convolution. That is, the value of each node at each time step depends not only on its own features at that time step, but also on the features of all its neighboring nodes at the same time step. This aggregation operation controls the degree of influence of different neighboring nodes through learnable weight parameters. Closer nodes are assigned higher weights, and farther nodes are assigned lower weights, so that the model can adaptively learn the law of spatial coupling strength decaying with distance. On this basis, the model introduces a temporal dimension transfer operation. For each spatial node, its state at different time steps is temporally modeled through a gated recurrent unit. This unit can capture the long-term dependencies and changing trends of feature values ​​on the time axis. The gated recurrent unit controls the proportion of historical information transferred to the current time step through the mechanism of update gate and reset gate, so that the model can distinguish between normal smooth changes and sudden abnormal jumps, thereby providing temporal constraints for subsequent missing data detection.

[0032] Spatial aggregation operations and temporal propagation operations are stacked alternately to form spatiotemporal convolutional layers. Within each spatiotemporal convolutional layer, a spatial dimension graph convolution is first performed to incorporate the information of neighboring nodes into the current node. Then, a temporal dimension gating loop is performed to propagate the information from the previous time step to the current time step. By stacking multiple such spatiotemporal convolutional layers, the model can gradually expand its receptive field, enabling each node to indirectly perceive the influence of multi-hop neighborhoods in space and capture longer-span temporal dependencies in time. The model training process adopts a self-supervised learning approach, using historical data in a multi-dimensional feature data cube as training samples. The training objective is to minimize the model's reconstruction error of normal data points. That is, based on the input historical sequence and spatial neighborhood information, the model outputs the predicted values ​​of each feature value of each node at the current moment. During the training process, the model parameters are optimized through the backpropagation algorithm so that the predicted values ​​are as close as possible to the actual observed values. In this way, the model gradually learns the normal variation patterns of distributed photovoltaic power station clusters in the spatiotemporal dimensions during the training process, including temporal continuity constraints and spatial consistency constraints.

[0033] Compared to traditional standalone temporal or spatial detection methods, the proposed model architecture achieves learnable modeling of spatial adjacency relationships through graph convolution, avoiding the insufficient adaptability of fixed thresholds. It captures dynamic temporal changes through gated recurrent units, enhancing the model's robustness to normal weather fluctuations. The alternating stacking of spatiotemporal convolutions achieves joint spatiotemporal constraints, simultaneously handling temporal dependencies and spatial interactions. After training and convergence, the spatiotemporal correlation constraint model can provide an expected value range for each data point in the multidimensional feature data cube, integrating temporal and spatial patterns. This range considers both the point's historical trajectory and the current values ​​of its neighboring nodes in its spatial neighborhood, providing a joint constraint basis for identifying suspected missing points.

[0034] Step 2.3: Based on the spatiotemporal correlation constraint model, calculate the local outlier factor of each data point in its spatiotemporal neighborhood. Simultaneously, combine the expected value range of each feature dimension within a pre-set confidence interval to identify data points that do not conform to historical change patterns in time and are inconsistent with neighboring node data in space. Mark these as suspected missing points. Specifically, this includes: using the established spatiotemporal correlation constraint model as the detection basis, evaluating each data point in the multidimensional feature data cube one by one; during the evaluation process, first define the spatiotemporal neighborhood of the current data point to be evaluated. In the time dimension, take the time index where the data point is located as the center, and take several consecutive time indices forward and backward to form a time neighborhood window; in the spatial dimension, take the spatial node where the data point is located as the benchmark, and obtain all directly adjacent nodes of the node through the constructed spatial topology graph. These adjacent nodes constitute the spatial neighborhood; the spatiotemporal neighborhood is the combination of the above-mentioned time neighborhood and spatial neighborhood, containing the historical data before and after this data point in time and the contemporaneous data of neighboring nodes in space.

[0035] Based on the defined spatiotemporal neighborhood, the local outlier factor of the data point is calculated: The calculation of the local outlier factor first measures the deviation of the data point from its historical sequence within its temporal neighborhood. Specifically, this is done by comparing the value of the data point with the mean and fluctuation range of its historical data within its temporal neighborhood to determine whether the value exceeds the normal range of variation allowed by historical patterns. Next, the difference between the data point and the values ​​of the same feature from its neighboring nodes within its spatial neighborhood is measured. Specifically, this is done by calculating the deviation between the value of the data point and the mean of the same feature values ​​of all spatial neighborhood nodes at the same time, and determining whether this deviation exceeds the tolerance range allowed by spatial consistency. The deviations of the above two dimensions are combined to obtain a comprehensive outlier factor. Simultaneously, combined with the expected value range of each feature dimension calculated in step 1.2 within a preset confidence interval, this preset confidence interval... The expected value range is a pre-calculated normal fluctuation envelope for each spatial node, each type of feature, and each time index, reflecting the reasonable range of values ​​for that feature under normal meteorological and operational conditions. A data point is identified as a suspected missing point when it simultaneously meets the following two conditions: the first condition is that the value of the data point exceeds the expected value range of the corresponding feature at the corresponding time determined in step 1.2, indicating that the data point has an abnormal deviation in time series and does not meet the historical change pattern; the second condition is that the comprehensive outlier factor of the data point exceeds the preset threshold, indicating that the data point has a significant inconsistency with the values ​​of the same feature of multiple neighboring nodes in its spatial neighborhood. By jointly determining these two conditions, it is ensured that the marked suspected missing points exhibit abnormal characteristics in both the time and spatial dimensions, thereby effectively reducing the misjudgment that may be caused by a single-dimensional judgment.

[0036] Step 2.4: Based on the marked suspected missing points, classify and aggregate them according to the missing type. Data points with multiple spatial nodes missing the same feature at the same time are grouped into spatially fragmented missing data points, while data points with different features missing at consecutive time points at the same spatial node are grouped into temporally cross-feature missing data points. The union of spatially fragmented and temporally cross-feature missing data points is then used as a spatiotemporal joint missing data candidate set. Specifically, this includes: systematically classifying and aggregating all marked suspected missing points. The classification and aggregation are divided into two independent dimensions based on the spatiotemporal distribution of the missing data points; the first dimension is the aggregation of spatially fragmented missing data points. During the processing, all... Suspected missing points are identified by filtering out data points with the same time index and feature type. These data points are then grouped according to their spatial nodes. For each group, it is determined whether the number of spatial nodes contained in the group reaches a preset spatial patching threshold. If it does, all related suspected missing points under this time index and feature type are grouped into one category and marked as spatial patching missing. This type of missing reflects the situation where multiple power stations in a specific area simultaneously exhibit data anomalies on the same feature type at a specific time. This missing pattern is often related to regional communication interruptions or local extreme weather events.

[0037] The second dimension is the aggregation of temporal cross-feature missing data. During processing, all suspected missing points are traversed, and data points belonging to the same spatial node are selected and sorted according to their time indices. Simultaneously, time periods with consecutive missing data within this spatial node are identified, i.e., multiple adjacent time indices are marked as missing. For each consecutive missing time period, the types of features involved in the missing data within that time period are counted. If the number of missing feature types within that time period reaches a preset cross-feature threshold (i.e., the missing data involves two or more different types of features), then all relevant suspected missing points of that spatial node within this consecutive time period are grouped into one category and marked as temporal cross-feature missing data. This type of missing data reflects a certain... When a power plant experiences simultaneous or alternating anomalies in multiple types of feature data over a continuous period, this missing data pattern is often related to sensor malfunctions or data acquisition equipment abnormalities within the power plant itself. After completing the classification and aggregation of the above two dimensions, the union of spatially fragmented missing data and temporally cross-feature missing data is taken: that is, all suspected missing points marked as spatially fragmented missing data and all suspected missing points marked as temporally cross-feature missing data are merged together, and duplicates are removed to form the final spatiotemporal joint missing data candidate set. The candidate set not only includes all suspected missing points identified under spatiotemporal joint constraints, but also performs preliminary classification according to the morphological characteristics of the missing data, providing a precise input data range for physical verification and missing data level classification in subsequent steps.

[0038] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Obtain each suspected missing point in the spatiotemporal joint missing candidate set, extract its spatial node identifier, time index, and feature type. Simultaneously, read the irradiance data, power data, and component temperature data of the same spatial node under adjacent time indices from the multidimensional feature data cube. Specifically, this includes: for the spatiotemporal joint missing candidate set, sequentially traversing each suspected missing point in the set; for the currently processed suspected missing point, extracting three key elements from the point's record information: the unique identifier of the spatial node to which the point belongs, the specific time index within the time window where the point is located, and the specific feature type corresponding to the point. These three elements together determine the precise location of the suspected missing point in the multidimensional feature data cube; then, based on the extracted spatial node identifier, time index, and feature type... The system uses the node identifier and time index to retrieve relevant data from the multidimensional feature data cube that belongs to the same spatial node as the suspected missing point. Specifically, the retrieved data includes three parts: first, irradiance, power, and component temperature data at the time preceding the current time index; second, irradiance, power, and component temperature data at the time following the current time index; and third, if the feature type of the suspected missing point is precisely one of irradiance, power, or component temperature, then the actual observed value of that point at the current time is also retrieved. Through these operations, a complete set of spatiotemporal neighborhood data is obtained for each suspected missing point, including the trajectory of changes in irradiance, power, and component temperature at adjacent times. This data will serve as the basic input for physical verification.

[0039] Step 3.2: Based on the read data, construct a physical coupling discrimination model with the ratio of irradiance change rate to power change rate as the core and the correction coefficient of module temperature on power conversion efficiency as the constraint. Specifically, this includes: based on the adjacent time data corresponding to each suspected missing point, construct a physical coupling discrimination model to determine whether the data point conforms to the physical laws of photovoltaic power generation. The design of this model is rooted in the basic physical principles of photovoltaic power generation, namely, the output power of photovoltaic modules is jointly determined by incident irradiance and module temperature, and the three follow a definite physical relationship: power increases with increasing irradiance and decreases with increasing temperature, and its rate of change is constrained by the photoelectric conversion efficiency of the photovoltaic system. The core idea of ​​this model is to transform this physical relationship into a quantifiable discrimination criterion, and to determine whether there is an anomaly in the data point by comparing whether the proportional relationship between the observed power change and the irradiance change is within the theoretically reasonable range. This physical coupling discrimination model adopts a physical-guided discrimination architecture, which is based on the parameterized expression of photovoltaic physical mechanisms and combined with a discrimination learning mechanism based on physical constraints.

[0040] The model's architecture comprises three core layers. The first layer is the physical parameter baseline layer. This layer pre-fits the photoelectric conversion efficiency baseline coefficient and temperature decay coefficient for each spatial node in the distributed photovoltaic power station cluster. The fitting process utilizes historical normal data stored in the multi-dimensional feature data cube from step 1.4. By minimizing the deviation between theoretical and actual power, the conversion efficiency of each node under standard conditions and the decay law of efficiency with temperature variation are obtained. This baseline layer ensures that the model can be customized for the photovoltaic module characteristics of different power stations, rather than using uniform fixed parameters. The second layer is the physical relationship derivation layer. This layer, based on the physical mechanism of photovoltaic power generation, constructs the coupling relationship between irradiance variation and power variation. Specifically, this layer expresses photovoltaic power as the product of irradiance and conversion efficiency, where the conversion efficiency is linearly corrected by module temperature. By applying this physical relationship in the time dimension... The first layer performs dynamic decomposition to deduce the ratio between the power change rate and the irradiance change rate, which includes a real-time correction term for the conversion efficiency due to module temperature. This process extends the static physical parameter relationship to a dynamic rate-of-change constraint, enabling the model to capture the transient response characteristics of the photovoltaic system during changes in operating conditions. The third layer is the discrimination boundary generation layer: based on the physical parameter benchmark of the spatial node where each suspected missing point is located, and the module temperature observation value of that node at adjacent times, the upper and lower boundaries of the theoretical ratio range are dynamically calculated. This layer introduces a physical confidence interval parameter to extend a reasonable tolerance range based on the theoretical ratio, in order to absorb small deviations caused by measurement errors, local cloud shading, and other factors. In this way, the discrimination boundary generation layer generates a reasonable range of variation for each suspected missing point that conforms to the physical laws of photovoltaics and takes into account actual engineering errors.

[0041] This physical coupling discrimination model employs a semi-supervised training strategy. Its training process fully utilizes the normal historical data within the multi-dimensional feature data cube constructed in step 1.4, eliminating the need for manual annotation of missing points. The training objective is to calibrate the physical parameters of each spatial node, ensuring that the model's discrimination results for normal data points stably fall within the theoretical ratio range. The specific training process is divided into two stages: the first stage is the initial fitting of physical parameters. Using historical data within a long time window, an optimization algorithm is used to solve for the photoelectric conversion efficiency baseline coefficient and temperature decay coefficient of each node, minimizing the overall error between the theoretical power curve and the measured power curve. The second stage is the fine-tuning of the discrimination boundary parameters. By statistically analyzing the distribution of the actual ratio of power change rate to irradiance change rate in historical normal data, the physical confidence interval parameters for each node are determined, ensuring that under normal operating conditions, the vast majority of actual ratios fall within the boundary range jointly determined by the physical parameters and the confidence interval.

[0042] Compared to traditional data-driven discrimination methods, this physically coupled discrimination model has the following advantages: It uses the physical equations of photovoltaic power generation as its core, ensuring precise physical interpretability of its discrimination criteria and avoiding misjudgments caused by training data distribution shifts in purely data-driven models; by independently fitting physical parameters to each spatial node, it can adaptively adapt to different types and aging levels of photovoltaic modules, solving the problem of a unified physical model being difficult to adapt to diverse power plants; the introduction of physical confidence interval parameters can absorb measurement noise and local environmental fluctuations in actual engineering while maintaining strict physical constraints, reducing the risk of misjudgments due to minor fluctuations; it complements the established spatiotemporal correlation constraint model, with the former providing discrimination criteria from the physical law dimension and the latter from the spatiotemporal statistical dimension, and their combined use effectively improving the accuracy and robustness of missing data detection; thus, the physically coupled discrimination model can provide a discrimination result that integrates photovoltaic physical laws for each suspected missing point, providing a reliable physical basis for distinguishing between true missing data and normal fluctuations.

[0043] Step 3.3: Substitute each suspected missing point into the physical coupling discrimination model to calculate the actual ratio of its irradiance change rate to power change rate, and combine this with component temperature data to calculate the corrected theoretical ratio range. If the actual ratio falls within the theoretical ratio range, the suspected missing point is determined to be a data point that conforms to the physical fluctuation law; otherwise, it is determined to be a real missing point. Specifically, this includes: substituting the adjacent time data extracted for each suspected missing point in Step 3.1 into the constructed physical coupling discrimination model for calculation and judgment; for the currently processed suspected missing point, let its spatial node be... v The current time index is t The feature type corresponding to this point is denoted as f From the read data, obtain the spatial node at the previous moment. Irradiance observation values Power observations and component temperature observations and at the next moment Irradiance observation values Power observations and component temperature observations ; Calculate the rate of change of irradiance It is defined as the difference in irradiance between the previous and subsequent moments divided by the time interval Δ. t ,Right now Similarly, calculate the rate of change of power. Defined as ; where the time interval Δ t The time difference between adjacent time indices is determined according to a preset time resolution, for example, Δ corresponds to 15 minutes. tIt takes 0.25 hours.

[0044] Start calculating the actual ratio That is, the ratio of the power change rate to the irradiance change rate, calculated using the following formula: This ratio reflects the degree of power response to changes in irradiance as actually observed under the current operating conditions. Simultaneously, based on the observed module temperature values ​​of this node at adjacent time points, the theoretical correction coefficient for the power conversion efficiency under the corresponding temperature conditions is calculated. Let the average module temperature of this node at adjacent time points be... Based on the temperature decay coefficient fitted for this node Calculate the temperature correction factor ,in Use the reference temperature as a basis; based on this, determine the lower bound of the theoretical ratio range. and the Upper Realm The formula for calculating the theoretical ratio range is: and ,in θ is the reference coefficient for the photoelectric conversion efficiency of this node, and θ is a determined physical confidence interval parameter (typical engineering value: 0.05 to 0.1), used to absorb the effects of measurement errors and local environmental fluctuations.

[0045] Completed actual ratio The range of calculated and theoretical ratios Once determined, the actual ratio is compared with the theoretical ratio range; if If the actual ratio falls within the theoretical ratio range, the suspected missing point is determined to conform to the physical fluctuation law of photovoltaic power generation, indicating that the change in the time series at this point is a reasonable physical response, rather than an anomaly caused by missing data. Therefore, it is marked as a data point conforming to the physical fluctuation law. Conversely, if the actual ratio exceeds this range, i.e. or If the suspected missing point does not conform to the laws of physical fluctuations, it means that the anomaly at this point cannot be explained by normal photovoltaic physical processes, and therefore it is marked as a true missing point.

[0046] Step 3.4 removes all data points conforming to physical fluctuation laws from the spatiotemporal joint missing candidate set, retaining all true missing points. Simultaneously, the true missing points are reorganized according to spatial node identifiers, time indices, and feature types to obtain the missing point location results after physical verification and correction. Specifically, this includes: classifying all suspected missing points in the spatiotemporal joint missing candidate set; based on the judgment results of each point in Step 3.3, removing all data points marked as conforming to physical fluctuation laws from the candidate set, because although these points exhibit anomalies under spatiotemporal joint constraints, physical verification confirms that they conform to the normal physical laws of photovoltaic power generation, and they are normal data points that were misjudged as missing and should not be included in the final missing point location results; at the same time, retaining all data points marked as true missing, these points exhibit anomalies under spatiotemporal joint constraints but do not meet the physical fluctuation laws. The physical laws governing photovoltaic power generation represent the true data gaps. After screening, the retained true gaps are reorganized according to three dimensions: spatial node identifier, time index, and feature type. During reorganization, multiple feature gaps under the same spatial node and time index are merged together to form a record of the missing feature combination at that node at that moment. Gaps of the same spatial node and the same feature type on consecutive time indices are linked together to form a record of consecutive missing time periods for that feature. Through this multi-dimensional reorganization operation, the originally scattered individual gaps are integrated into a structured gap location result. This result clearly identifies the spatial node involved in each gap event, the time of occurrence, the affected feature type, and the spatiotemporal distribution of the gaps, forming a gap location result that has been physically verified and corrected, providing accurate and reliable input for gap level classification.

[0047] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Obtain the missing data location results after physical verification and correction. Aggregate the missing points according to the spatial node identifier and time index. This is used to group missing points occurring within a continuous time window of the same spatial node into a single missing event, resulting in a missing event set. Specifically, this includes: obtaining the missing data location results after physical verification and correction. These results record all data points determined to be truly missing in a structured form. Each missing point contains three core attributes: the unique identifier of the spatial node to which it belongs, the specific time index of the missing event, and the affected feature type. Before starting the aggregation operation, all missing points in the missing data location results are grouped according to the spatial node identifier. All missing points belonging to the same spatial node are grouped into the same group, forming a missing point set based on spatial nodes. For each spatial node group, the missing points within the group are further sorted in ascending order of the time index to obtain the missing point sequence of that node on the time axis.

[0048] The missing point sequence is traversed and scanned to identify missing point sequences that are consecutive or spaced within a preset tolerance range in terms of time index. The preset tolerance range is set to handle the problem of missing segments being broken due to data sampling intervals or minor discontinuities. For example, when the difference between the time indices of two missing points is less than or equal to the preset tolerance threshold, they are still considered to belong to the same consecutive missing event, even if they are not strictly adjacent in time. Each set of missing point sequences that meets the conditions of being consecutive or adjacent is identified and grouped into a missing event, and a unique identifier is generated for the event. During the grouping process, the start time index, end time index, and all feature types involved in the event are recorded simultaneously. After scanning and grouping all spatial nodes, the missing events obtained from each grouping are summarized as independent units to form a missing event set. Each event in this set fully records the spatial location, time range, and affected feature range of the missing event.

[0049] Step 4.2: For each missing event in the missing event set, calculate its continuous missing duration in the time dimension, the number of affected power plants simultaneously missing in the spatial dimension, and the number of feature types involved in the missing event in the feature dimension to obtain a three-dimensional quantitative evaluation index for this missing event. Specifically, for the formed missing event set, iterate through each missing event in turn and calculate the three-dimensional quantitative evaluation index for each event. These three indicators together constitute a comprehensive evaluation vector of the severity of the event. The first dimension is the continuous missing duration in the time dimension. This index is calculated by counting the number of time indices covered by the missing event on the time axis. Specifically, the time span between the start time index and the end time index of the event is calculated. This span directly reflects the degree of persistence of the missing event in the time direction. For non-strictly continuous events formed by merging within a preset tolerance range, the continuous missing duration is still calculated according to the total time span from start to end, because even if there is a brief gap in available data in the middle, the overall reliability of the data within this time period is still substantially affected by the missing event.

[0050] The second dimension is the number of power plants affected by the simultaneous occurrence of missing events in the spatial dimension. This indicator is calculated by counting the number of other spatial nodes that are simultaneously missing within the same time window as the current missing event. In the specific calculation process, the time window range of the current missing event is first determined, that is, the time period covered from the start time index to the end time index. Then, all other events in the missing event set are traversed to filter out those events whose time windows overlap with the current event's time window, and the spatial node identifiers involved in these events are collected. Finally, the collected spatial node identifiers are deduplicated and counted. The resulting number is the number of power plants affected. This indicator reflects the spatial diffusion range and synchronicity of the missing events. The third dimension is the number of feature types involved in the missing data. This indicator is calculated by counting the number of different feature types involved in the missing data points of the current missing event. Specifically, feature type information of all missing data points of the event is extracted from the event records, the feature types are deduplicated, and the number of deduplicated feature types is counted. This indicator reflects the breadth of the impact of the missing data on the feature dimension, that is, whether the missing event is limited to the loss of data of a certain type of feature or affects multiple types of features at the same time. Through the above calculation, each missing event obtains a three-dimensional quantitative evaluation indicator vector consisting of continuous missing duration, number of affected power stations, and number of feature types, providing a quantitative basis for subsequent classification.

[0051] Step 4.3: Based on the three-dimensional quantitative evaluation index, each missing event is divided into three levels: single-point random missing, continuous time-period missing, or large-area missing. A preset confidence weight coefficient is assigned to each missing event according to its level. The preset confidence weight coefficient is negatively correlated with the severity of the missing event. Specifically, this includes: classifying each event in the missing event set according to the calculated three-dimensional quantitative evaluation index; the level classification uses a tiered threshold determination method, comprehensively considering the specific values ​​of three indicators: time dimension, spatial dimension, and feature dimension. First, threshold boundaries are set for each of the three dimensions, including a long-term missing threshold in the time dimension, a large-area impact threshold in the spatial dimension, and a multi-feature threshold in the feature dimension. The specific logic of the level classification is as follows: when the continuous missing duration of a missing event is short and the number of affected power plants is less than or equal to one preset threshold... When the minimum value of the missing data and the number of feature types is one, the event is classified as a single-point random missing data level. This type of missing data usually manifests as the loss of a single feature data at an isolated moment by an individual power station, exhibiting randomness and locality. When the continuous missing duration of the missing event exceeds the long-term missing data threshold, but the number of affected power stations is still less than the area-wide affected data threshold, the event is classified as a continuous-time missing data level. This type of missing data usually manifests as data loss at a power station within a continuous period of time, exhibiting temporal persistence but spatial limitation. When the number of affected power stations of the missing event reaches or exceeds the area-wide affected data threshold, or the number of feature types reaches or exceeds the multi-feature threshold, and the continuous missing duration simultaneously reaches a certain level, the event is classified as an area-wide missing data level. This type of missing data usually manifests as data loss of multiple features at multiple power stations simultaneously within the same time period, exhibiting spatial diffusion and feature breadth.

[0052] After classifying the missing events, a preset confidence weight coefficient is assigned to each missing event based on its level. The preset value of the confidence weight coefficient follows the principle of negative correlation with the severity of the missing event; that is, the more severe the missing event, the lower its corresponding confidence weight coefficient. Specifically, single-point random missing events are assigned a higher confidence weight coefficient, such as a value close to full weight, indicating that this type of missing event has a small impact on the overall data confidence, and other data in this area can be trusted in subsequent processing. Continuous time-period missing events are assigned a medium confidence weight coefficient, such as a value within a medium range, indicating that this type of missing event has a moderate impact on the data confidence, and the dependence on the data in this area should be appropriately reduced during feature extraction and spatial aggregation. Large-area missing events are assigned a lower confidence weight coefficient, such as a value within a lower range, indicating that this type of missing event has a large impact on the data confidence, and the trust in the data in this area should be significantly reduced. In extreme cases, the data in this area can even be temporarily shielded from the model input. Through this mapping relationship between level and weight coefficient, the severity of missing events is quantified into numerical weights that can be directly called by the preset prediction model.

[0053] Step 4.4 involves associating and mapping the spatial node identifier, time index, missing level, and corresponding confidence weight coefficient for each missing event to form a quantitative missing level map covering all missing events. Specifically, this includes: extracting the core attribute information of each missing event after level classification and weight allocation for association and mapping; extracting four key pieces of information for each missing event, including the spatial node identifier involved in the event, the start and end time indices of the event's time window, the missing level of the event, and the confidence weight coefficient corresponding to that level; and organizing and mapping this information according to spatial and temporal dimensions. The specific mapping method is as follows: using spatial nodes as the spatial node identifier, time index, missing level, and corresponding confidence weight coefficient. Using the node identifier as the first-level index and the time index as the second-level index, a two-dimensional graph structure is constructed. The horizontal axis is the time axis, covering all time indices within the entire time window, and the vertical axis is the spatial axis, covering all spatial nodes. For each spatiotemporal point in the graph, it is filled according to the records in the missing event set: if the spatiotemporal point belongs to the time window range of a missing event and belongs to the spatial node affected by the event, then the missing level and confidence weight coefficient of the event are filled into the corresponding graph position; for normal spatiotemporal points that have not experienced missing events, they are marked as normal in the graph and assigned a full confidence weight coefficient, that is, the weight coefficient is the maximum value.

[0054] During the filling process, if a spatiotemporal point falls within the influence range of multiple missing events, the most severe missing level and its corresponding lowest confidence weight coefficient are used for filling, following the principle of strict handling. Through this correlation mapping operation, the information of all missing events is integrated into a unified structured map. This map fully covers the missing status of all spatial nodes within the entire time window, clearly showing the missing level label and corresponding confidence weight coefficient of each spatiotemporal point. The final missing level quantification map contains not only the spatial distribution information of missing events, but also the temporal evolution trajectory of missing events, as well as the quantitative assessment results of missing severity, providing a structured data input foundation for constructing a dynamic missing mask matrix.

[0055] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1 aligns the missing level quantization map, the spatiotemporal joint missing candidate set, and the physically verified and corrected missing location results to the same time index and spatial node identifier system. Specifically, this includes: obtaining the missing level quantization map, the spatiotemporal joint missing candidate set, and the physically verified and corrected missing location results. These three data sources come from different processing stages, and their data structures and index systems may differ. To enable subsequent fusion operations, they need to be unified under the same time index and spatial node identifier system. The alignment operation is based on the original index system of the multidimensional feature data cube, which has defined a standard time axis sequence and spatial node identifier set. For the missing level quantization map, its time... The index range is checked to ensure it covers the entire standard timeline. If any time indices are missing, they are filled in, and uncovered spatiotemporal points are marked as normal. Simultaneously, the spatial node identifiers are checked to ensure they are completely consistent with the baseline system, guaranteeing that each spatial node has a corresponding position in the map. For the spatiotemporal joint missing candidate set, each suspected missing point is mapped to its corresponding position in the baseline index system according to its carried time index and spatial node identifier, forming a candidate set representation aligned with the baseline system. For the missing location results corrected through physical verification, each actual missing point is also mapped to its corresponding position in the baseline index system. After completing the above alignment operations, all three data sources use the same set of time indices and spatial node identifiers for organization and indexing.

[0056] Step 5.2: Using the physically verified and corrected missing location results as the basic positioning layer, mark whether each spatiotemporal point has a true missing location. Simultaneously, use the spatiotemporal joint missing candidate set as the auxiliary verification layer to mark the missing type of each spatiotemporal point under the spatiotemporal joint dimension, forming fused positioning information containing missing location identifiers and missing type identifiers. Specifically, this includes: constructing fused positioning information under the aligned unified index system; first, using the physically verified and corrected missing location results as the basic positioning layer, which is a set of true missing points confirmed after physical law verification and has the highest credibility; for each spatiotemporal point, mark whether the point has a true missing location based on whether there is a corresponding true missing record in the basic positioning layer, with the mark value represented by a binary state, i.e., true missing or not true missing; then, use the spatiotemporal joint missing candidate set as the auxiliary verification layer, which records all suspected missing points identified as abnormal under spatiotemporal joint constraints, including points subsequently judged as true missing by physical verification and points judged as normal fluctuations.

[0057] For each spatiotemporal point, based on whether there is a corresponding suspected missing record in the auxiliary verification layer, the missing type of the point in the spatiotemporal joint dimension is marked. Specifically, if a spatiotemporal point has a record in the spatiotemporal joint missing candidate set, it is further classified according to the missing form to which the record belongs in the candidate set, and marked as spatial patch missing type or temporal cross-feature missing type, or both types. If a spatiotemporal point does not have a record in the spatiotemporal joint missing candidate set, it is marked as no spatiotemporal anomaly. The marking information of the basic positioning layer and the auxiliary verification layer are associated and integrated to form fused positioning information containing missing location identifier and missing type identifier. This information provides both the existence identifier of the real missing point and the anomaly type identifier in the spatiotemporal joint dimension for each spatiotemporal point, providing a complete positioning basis for the weight superposition in subsequent steps.

[0058] Step 5.3: Based on the fused positioning information, the missing level and confidence weight coefficient in the missing level quantization map are superimposed onto the corresponding spatiotemporal points to form a multidimensional missing information matrix containing missing location identifiers, missing type identifiers, missing level identifiers, and confidence weight coefficients. Specifically, this includes: superimposing the quantization information in the missing level quantization map onto the corresponding spatiotemporal points based on the fused positioning information; for each spatiotemporal point, obtaining the actual missing existence identifier and spatiotemporal joint missing type identifier of that point from the fused positioning information; and then searching for the missing level and confidence weight coefficient corresponding to that spatiotemporal point from the missing level quantization map. The search rules for the missing level quantization map are as follows: Below: If the spatiotemporal point belongs to the time window of a missing event and is a spatial node affected by the event, the map records the missing level and confidence weight coefficient of the point; if the spatiotemporal point does not belong to any missing event, the map marks it as normal and assigns it a full confidence weight coefficient; integrate the information from the above four dimensions to construct a complete information record for the spatiotemporal point, including the missing location identifier (whether there is a true missing point), the missing type identifier (spatial block missing point, temporal cross-feature missing point, or no anomaly), the missing level identifier (single point random missing point, continuous time period missing point, block area missing point, or normal missing point), and the confidence weight coefficient (specific value).

[0059] Traverse all time indices and all spatial nodes to generate corresponding information records for each spatiotemporal point. Organize these records into a multidimensional information matrix in the order of time and space dimensions. The dimensions of the multidimensional information matrix are consistent with those of the multidimensional feature data cube in both time and space. Each element in the matrix contains a complete description of the above four types of information, forming a structured multidimensional missing information matrix.

[0060] Step 5.4: Convert the multidimensional missing information matrix into a dynamic missing mask matrix in a preset format. The dynamic missing mask matrix is ​​used by the preset prediction model to adaptively reduce the weight of missing regions during temporal feature extraction and spatial feature aggregation. Simultaneously, the dynamic missing mask matrix serves as the detection result for spatiotemporal feature missing values. Specifically, this includes: converting the multidimensional missing information matrix into a dynamic missing mask matrix in a preset format; the conversion process adapts the format according to the interface requirements of the prediction model, encoding the four types of information at each spatiotemporal point in the multidimensional missing information matrix into a data structure that the prediction model can directly read and parse; the preset format dynamic missing mask matrix adopts the same spatiotemporal index structure as the multidimensional feature data cube, i.e., the time dimension is used as the first index, the spatial dimension as the second index, and mask information is stored in the third dimension; the specific details of the mask information... The encoding method is as follows: For each spatiotemporal point, its missing level identifier and confidence weight coefficient are combined and encoded to form the mask value for that point. The missing level identifier determines the basic type of the mask, and the confidence weight coefficient determines the numerical weight of the mask. For normal points, the mask value is set to full weight and the type identifier is normal. For missing points, the mask value is set with corresponding weight coefficients according to the severity of the missing level, and a corresponding type identifier is attached. During the encoding process, the missing location identifier and missing type identifier are encapsulated as metadata information of the mask, so that the prediction model can not only obtain the weight coefficients when reading the mask, but also understand the specific type and nature of the missing data. After the encoding is completed, the resulting dynamic missing mask matrix is ​​fully aligned with the multidimensional feature data cube in the spatiotemporal dimension, and can be used as an independent input channel to be input into the preset prediction model along with the feature data.

[0061] This pre-defined prediction model adopts an architecture based on spatiotemporal graph convolutional networks. Its design originates from the fusion and improvement of classic graph convolutional networks and temporal convolutional networks. The core architecture of the model consists of multiple stacked spatiotemporal convolutional modules. Each module contains two core components: one is a spatial graph convolutional layer, which is used to aggregate the feature information of adjacent nodes in the spatial dimension. This layer uses the spatial topology graph constructed in step 2.2 as the graph structure prior and realizes the adaptive transmission of spatial information through a learnable adjacency weight matrix; the other is a temporal convolutional layer, which is used to capture the dependency relationship of feature sequence in the temporal dimension. This layer adopts the design of dilated convolution to expand the temporal receptive field, so that the model can learn the pattern of power change over a longer time span. The two components are arranged alternately to form an end-to-end trainable structure, which can simultaneously extract the spatial correlation and temporal dependency in spatiotemporal features.

[0062] The training process of this prediction model employs a multi-task learning strategy, using historical data from a multi-dimensional feature data cube as training samples. The training objectives include two parts: first, a power prediction task, which predicts the power value at future moments based on feature data within a historical time window, with the loss function using mean squared error to measure the deviation between the predicted and true values; second, a mask reconstruction task, which uses a dynamic missing mask matrix to weight the input features, enabling the model to learn during training how to adaptively adjust the focus on missing regions based on mask information. Both tasks share the underlying spatiotemporal graph convolutional network parameters, and through joint optimization, the model improves prediction accuracy while enhancing its ability to handle missing data. The model exhibits robustness; during training, the dynamic missing mask matrix is ​​input as a fixed input along with the feature data. In the spatiotemporal graph convolutional layer, the model weights neighborhood aggregation based on the mask weights, and in the temporal convolutional layer, it attenuates the historical sequence based on the mask weights, thereby achieving adaptive weight reduction processing for missing regions. After training and convergence, the prediction model can automatically adjust the weight allocation of feature extraction and spatial aggregation when faced with missing data of varying degrees, effectively reducing the negative impact of missing data on the prediction results. Finally, the dynamic missing mask matrix is ​​output as the detection result of spatiotemporal feature missing values, providing a complete description of the missing state and an adaptive weight reduction basis for subsequent photovoltaic power prediction tasks.

[0063] like Figure 2 As shown, embodiments of the present invention also provide a spatiotemporal feature missing value detection system for photovoltaic power prediction, including: The data preprocessing module is used to obtain the original data sequence of the distributed photovoltaic power station group, perform time alignment and spatial coordinate mapping on the original data sequence, and obtain a multi-dimensional feature data cube. The spatiotemporal missing data filtering module is used to establish spatiotemporal correlation constraints based on multidimensional feature data cubes, identify data points that simultaneously violate temporal continuity and spatial consistency, and obtain a spatiotemporal joint missing data candidate set. The physical verification module is used to verify the spatiotemporal joint missing candidate set using the physical coupling rules between meteorology and power, removing data points that conform to the physical fluctuation law, and obtaining the missing location result after physical verification and correction. The level quantification module is used to classify missing events into levels based on the missing location results after physical verification and correction, and to assign a corresponding confidence weight coefficient to each level to obtain a missing level quantification map. The mask construction module is used to fuse the missing level quantization map with the spatiotemporal joint missing candidate set and the missing location results after physical verification and correction to construct a dynamic missing mask matrix for use by a preset prediction model. The dynamic missing mask matrix is ​​used as the detection result of spatiotemporal feature missing values.

[0064] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0065] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0066] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0067] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting missing spatiotemporal features in photovoltaic power prediction, characterized in that, The method includes: Step 1: Obtain the original data sequence of the distributed photovoltaic power station group, perform time alignment and spatial coordinate mapping on the original data sequence to obtain a multi-dimensional feature data cube; Step 2: Based on the multidimensional feature data cube, establish spatiotemporal correlation constraints, identify data points that simultaneously violate temporal continuity and spatial consistency, and obtain a spatiotemporal joint missing candidate set; Step 3: For the spatiotemporal joint missing candidate set, a physical coupling rule between meteorology and power is used for verification. Data points that conform to physical fluctuation laws are removed, resulting in a physically verified and corrected missing location result. This includes: acquiring each suspected missing point in the spatiotemporal joint missing candidate set, extracting its spatial node identifier, time index, and feature type; simultaneously reading irradiance data, power data, and component temperature data of the same spatial node under adjacent time indices from the multidimensional feature data cube; based on the read data, constructing a system with the ratio of irradiance change rate to power change rate as the core, and the effect of component temperature on power conversion efficiency. A physical coupling discrimination model with correction coefficients as constraints is used. Each suspected missing point is substituted into the physical coupling discrimination model to calculate the actual ratio of its irradiance change rate to power change rate. Combined with component temperature data, the corrected theoretical ratio range is calculated. If the actual ratio falls within the theoretical ratio range, the suspected missing point is determined to be a data point that conforms to the physical fluctuation law; otherwise, it is determined to be a real missing point. All data points that conform to the physical fluctuation law in the spatiotemporal joint missing candidate set are removed, and all real missing points are retained. At the same time, the real missing points are reorganized according to spatial node identifier, time index, and feature type to obtain the missing location result after physical verification and correction. Step 4: Based on the missing location results after physical verification and correction, the missing events are classified into levels, and each level is assigned a corresponding confidence weight coefficient to obtain a quantitative map of missing event levels. Step 5: The missing level quantization map is fused with the spatiotemporal joint missing candidate set and the missing location results after physical verification and correction to construct a dynamic missing mask matrix for use by a preset prediction model. The dynamic missing mask matrix is ​​used as the detection result of spatiotemporal feature missing values.

2. The method for detecting missing spatiotemporal features in photovoltaic power prediction according to claim 1, characterized in that, The original data sequence of the distributed photovoltaic power station group is obtained, and the original data sequence is time-aligned and spatially mapped to obtain a multi-dimensional feature data cube, including: The original data sequence of five types of characteristic data of each power station within a continuous time window with a preset time resolution is obtained, and the original data sequence carries the spatial coordinate information of each power station. Based on the acquired original data sequence, time alignment is performed using a unified standard timestamp as a benchmark. The statistical distribution characteristics of each feature sequence within a continuous time window are analyzed and parameters are fitted to aggregate the five types of feature data from each power station at the same time point under the same time index, resulting in a time-aligned data set. Based on the time-aligned data set, a spatial location mapping relationship is established within a preset spatial range according to the spatial coordinate information of each power station. By fitting the spatial topology and distance measurement parameters between the nodes of each power station, a unique spatial node identifier is assigned to each power station to construct a spatial graph structure. Based on the constructed spatial graph structure, the time-aligned data set is reorganized according to the time dimension, spatial dimension and feature dimension. By parametrically modeling the coupling relationship between the time dimension, spatial dimension and feature dimension, a multidimensional feature data cube is formed with time as the first index, spatial node as the second index and feature type as the third index.

3. The method for detecting missing spatiotemporal features in photovoltaic power prediction according to claim 2, characterized in that, Based on a multidimensional feature data cube, spatiotemporal correlation constraints are established to identify data points that simultaneously violate temporal continuity and spatial consistency, resulting in a spatiotemporal joint missing data candidate set, including: Extract all feature data of each spatial node under each time index from the multidimensional feature data cube to form a spatiotemporal sequence set with spatial nodes as units and time axis as the order; Based on the spatiotemporal sequence set, a spatial topology graph is constructed with each spatial node as a vertex and a preset spatial distance threshold as an edge connection condition. By combining the historical sequence of each node on the time axis, a spatiotemporal association constraint model that simultaneously includes spatial adjacency and temporal dependency is established. Based on the spatiotemporal correlation constraint model, the local outlier factor of each data point in the spatiotemporal neighborhood is calculated. At the same time, combined with the expected value range of each feature dimension within the preset confidence interval, data points that do not meet the historical change pattern in time and are inconsistent with the data of neighboring nodes in space are identified and marked as suspected missing points. Based on the labeled suspected missing points, the missing points are classified and aggregated according to the missing type. Data points with multiple spatial nodes missing the same feature at the same time are grouped into spatial patch missing points, and data points with different features missing at the same spatial node at multiple consecutive time points are grouped into temporal cross-feature missing points. At the same time, the union of spatial patch missing points and temporal cross-feature missing points is used as a candidate set of spatiotemporal joint missing points.

4. The method for detecting missing spatiotemporal features in photovoltaic power prediction according to claim 3, characterized in that, Based on the missing event localization results after physical verification and correction, the missing events are classified into levels, and each level is assigned a corresponding confidence weight coefficient to obtain a quantitative missing event level map, including: Obtain the missing location results after physical verification and correction, and aggregate the missing points according to the spatial node identifier and time index. This is used to group the missing points that occur in the same spatial node within a continuous time window into a single missing event, thus obtaining a set of missing events. For each missing event in the missing event set, calculate its continuous missing duration in the time dimension, the number of affected power plants that are simultaneously missing in the spatial dimension, and the number of feature types involved in the missing in the feature dimension to obtain the three-dimensional quantitative evaluation index of this missing event. Based on three-dimensional quantitative evaluation indicators, each missing event is divided into three levels: single-point random missing, continuous time period missing, or large area missing. A preset confidence weight coefficient is assigned to each missing event according to its level. The preset confidence weight coefficient is negatively correlated with the severity of the missing event. The spatial node identifier, time index, missing level, and corresponding confidence weight coefficient of each missing event are correlated and mapped to form a missing level quantitative map covering all missing events.

5. The method for detecting missing spatiotemporal features in photovoltaic power prediction according to claim 4, characterized in that, The missing level quantization map is fused with the spatiotemporal joint missing candidate set and the physically verified and corrected missing location results to construct a dynamic missing mask matrix for use by a preset prediction model. This dynamic missing mask matrix is ​​used as the detection result of spatiotemporal feature missing values, including: The missing level quantization map, the spatiotemporal joint missing candidate set, and the missing location results after physical verification and correction are aligned to the same time index and spatial node identification system. Using the physical verification and correction results of missing locations as the basic positioning layer, we mark whether there are real missing locations at each spatiotemporal point. At the same time, we use the spatiotemporal joint missing candidate set as the auxiliary verification layer to mark the missing type of each spatiotemporal point in the spatiotemporal joint dimension, thus forming fused positioning information that includes missing location identifiers and missing type identifiers. Based on the fused positioning information, the missing level and confidence weight coefficient in the missing level quantization map are superimposed on the corresponding spatiotemporal points to form a multidimensional missing information matrix containing missing location identifier, missing type identifier, missing level identifier and confidence weight coefficient. The multidimensional missing information matrix is ​​transformed into a dynamic missing mask matrix in a preset format. The dynamic missing mask matrix is ​​used by the preset prediction model to perform adaptive weight reduction processing on the missing regions during the temporal feature extraction and spatial feature aggregation process. At the same time, the dynamic missing mask matrix is ​​used as the detection result of the spatiotemporal feature missing values.

6. A spatiotemporal feature missing value detection system for photovoltaic power prediction, wherein the system implements the method as described in any one of claims 1 to 5, characterized in that, include: The data preprocessing module is used to obtain the original data sequence of the distributed photovoltaic power station group, perform time alignment and spatial coordinate mapping on the original data sequence, and obtain a multi-dimensional feature data cube. The spatiotemporal missing data filtering module is used to establish spatiotemporal correlation constraints based on multidimensional feature data cubes, identify data points that simultaneously violate temporal continuity and spatial consistency, and obtain a spatiotemporal joint missing data candidate set. The physical verification module is used to verify the spatiotemporal joint missing candidate set using the physical coupling rules between meteorology and power, removing data points that conform to the physical fluctuation law, and obtaining the missing location result after physical verification and correction. The level quantification module is used to classify missing events into levels based on the missing location results after physical verification and correction, and to assign a corresponding confidence weight coefficient to each level to obtain a missing level quantification map. The mask construction module is used to fuse the missing level quantization map with the spatiotemporal joint missing candidate set and the missing location results after physical verification and correction to construct a dynamic missing mask matrix for use by a preset prediction model. The dynamic missing mask matrix is ​​used as the detection result of spatiotemporal feature missing values.

7. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images