Meteorology-based dynamic graph network photovoltaic power station group ultra-short-term prediction method

Abstract

The invention relates to a meteorological-based dynamic graph network photovoltaic power station group ultra-short-term prediction method, which is characterized in that a dynamic space-time graph network is constructed, photovoltaic power stations in a region are regarded as a complex network system connected by meteorological fluctuations, each power station is taken as a node, and the propagation relationship of the meteorological fluctuations among the power stations is represented by edges of a graph. The weight and time delay of the edge are dynamically adjusted according to real-time meteorological data, and dynamic interaction between power stations, the overall trend under the stable meteorological condition and rapid fluctuation caused by sudden weather events are accurately captured in combination with a mixed model of a graph neural network and a recurrent neural network. Compared with a traditional method, the method has the advantages that the prediction precision, the real-time response capability and the physical interpretability are remarkably improved, scattered photovoltaic power stations are integrated into a mutually associated dynamic system, an efficient and reliable solution is provided for power prediction of a power station group, and the operation stability and the power grid dispatching efficiency of a large-scale photovoltaic system are enhanced.

Description

Technical Field

[0001] This invention relates to the fields of new energy power generation technology and power system technology, specifically to a meteorological-based dynamic graph network photovoltaic power station cluster ultra-short-term forecasting method. Background Technology

[0002] Photovoltaic (PV) power generation technology, with its significant advantages of being clean and renewable, plays a crucial role in global energy structure transformation and addressing climate change. With continuous technological advancements and cost reductions, PV power generation has become a core force for global sustainable energy development. Particularly in China, the rapid development of the PV industry has not only promoted the green transformation of the energy structure but also provided strong support for achieving carbon peaking and carbon neutrality goals, setting a benchmark for technological innovation and cost optimization in the global industrial chain. However, with the large-scale deployment of centralized PV power plants, the intermittency, volatility, and uncertainty of their grid-connected power output pose significant challenges to grid operation and dispatch. Therefore, accurate power forecasting for PV power plant clusters, especially ultra-short-term forecasts (e.g., forecasts for the next 15 minutes to 4 hours), has become a key foundation for ensuring the efficient and stable operation of the power system.

[0003] Traditional power forecasting methods primarily focus on statistical and shallow machine learning models, such as autoregressive moving average models and their variants. These models emphasize classical time series analysis. For example, some recent studies have used time series models like ARIMA to predict daily solar power generation, assuming a linear relationship between past and present values. However, complex time series data in the real world are not always linear, leading to suboptimal model prediction accuracy and significant data limitations. Another approach involves extending autoregressive models like SARIMA by incorporating external meteorological data, such as solar irradiance, to achieve short-term photovoltaic power generation forecasts. The performance of this type of method typically depends on the accuracy of the meteorological data; significant errors in the meteorological data directly impact the prediction results.

[0004] While these methods uncover the seasonal and cyclical patterns of photovoltaic (PV) power generation and can be effective in predicting stable power generation patterns when a specific power station is unaffected, in practical applications, PV power stations within a region are typically geographically close, and their power output is also affected by similar weather conditions. Therefore, another research approach attempts to introduce modeling and analysis methods for power station clusters. For example, the Temporal Convolutional Network (TCN) method is used for regional distributed PV power prediction. This type of method can simultaneously process historical data from multiple power stations and use this data to predict the power output of other power stations within the region. However, despite its ability to handle historical data from multiple power stations, in-depth discussions on the selection of representative power stations and their impact are still lacking, and the correlation between power stations has not been fully analyzed.

[0005] From a physics perspective, meteorological events within a given area, such as cloud movement and wind advection, are continuous and interactive. The movement of clouds over a power plant will inevitably affect downstream or adjacent power plants within a certain timeframe. This causal relationship caused by the transmission of meteorological events is physical, not simply a statistical correlation. Traditional forecasting models, due to inherent design flaws, cannot capture this physical causal chain, leading to a significant decrease in prediction accuracy during sudden weather changes or drastic fluctuations—a fundamental defect. To overcome the limitations of traditional methods, some studies have begun to use graph structures to process spatiotemporal data. These methods treat individual units within a defined area as nodes in a graph and utilize the graph structure to capture the correlations between nodes. However, many existing graph neural network models still employ static graph structures, constructing adjacency matrices based on pre-defined topologies, fixed relationships, or correlations at the historical data level. While this static modeling approach has some reference value, it differs significantly from the dynamic meteorological environment encountered by photovoltaic power plants in actual operation. This static modeling approach has its own basis and reference value, but it differs considerably from the rapidly changing meteorological environment experienced by photovoltaic power plants in actual operation. Meteorological conditions evolve over time, and changes in wind direction and speed significantly adjust the interactions between power plants. For example, the relationship between an "upstream" and "downstream" photovoltaic power plant is dynamically influenced by real-time wind direction. Assuming the prevailing wind is northerly, the power plant at the wind's origin can be defined as the upstream plant, and its meteorological fluctuations (such as cloud movement or changes in irradiance) may affect the power output of the downstream plant located downwind. When the wind shifts to southerly, the roles of the "upstream" and "downstream" plants reverse, dynamically adjusting their relationship. Static graphs struggle to capture these dynamic changes caused by meteorological advection and disturbance propagation, making it difficult for models to accurately reflect the true propagation paths of meteorological fluctuations when dealing with sudden weather events. This limits the robustness and interpretability of their predictions. Therefore, effectively characterizing and utilizing the dynamic evolution of correlations between power plants as meteorological conditions change in prediction models remains a critical challenge for current technologies. Summary of the Invention

[0006] In view of this, the purpose of this invention is to overcome the shortcomings of the prior art and provide a meteorological-based dynamic graph network method for ultra-short-term forecasting of photovoltaic power plant clusters. This method captures the propagation relationship of meteorological fluctuations among photovoltaic power plants in a region through a dynamic spatiotemporal graph network, and combines a hybrid model of graph neural network and recurrent neural network to achieve ultra-short-term power forecasting of photovoltaic power plant clusters, thereby improving the accuracy, real-time performance and physical interpretability of the forecast.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A meteorological-based dynamic graph network method for ultra-short-term forecasting of photovoltaic power plant clusters includes the following steps:

[0009] S1. Acquire ultra-short-term time-series data of multiple geographically dispersed photovoltaic power stations. The data includes grid-connected power data, irradiance, and regional meteorological data for each power station. The regional meteorological data includes wind speed, wind direction, temperature, humidity, and air pressure. Perform missing value processing, outlier detection, and data preprocessing on the data.

[0010] S2. Define the photovoltaic power station as a node in the graph, and establish a relative static coupling relationship between the nodes based on geographical location similarity, distance relationship between power stations, historical power similarity, or other correlation analysis. At the same time, using the meteorological data, calculate the directed edge weights between the power station nodes based on meteorological physical factors to dynamically construct a dynamic adjacency matrix that reflects the physical propagation effect of meteorological fluctuations. Finally, combine the static coupling relationship and the dynamic adjacency matrix to construct a dynamic graph network G(t)=(V, E(t)), where V represents a set consisting of N nodes and E(t) represents the dynamic edge of the graph structure.

[0011] S3. Construct a hybrid model consisting of a graph neural network and a recurrent neural network. Input the dynamic graph network G(t) and the preprocessed time series data of each power station into the hybrid model for training until the preset training accuracy requirements are met, and complete the model training.

[0012] S4. Determine whether the power prediction result meets the preset prediction accuracy requirements; if not, return to the model building and training stage to retrain or optimize the hybrid model until the prediction accuracy meets the corresponding requirements.

[0013] S5. Once the hybrid model meets the prediction performance accuracy threshold, the latest real-time meteorological data and historical power data are used, combined with the real-time updated dynamic graph network G(t), to predict the power within a specific time period in the future through the hybrid model that meets the accuracy requirements. The power prediction adopts a prediction mechanism, including rolling prediction. The power in the next k steps is calculated by iteratively inputting real-time data, which supports multi-step prediction and can output the corresponding prediction results and confidence intervals, or prediction results of other uncertainty estimation methods.

[0014] As a further improvement of the present invention, the data processing step in step S1 includes:

[0015] Missing value handling:

[0016] Use linear interpolation or forward padding methods to handle missing values ​​in the data, for time series data points x. t Missing value x mIt can be calculated using the following formula (1):

[0017] (1)

[0018] Where, x m x represents the missing value at time m; m-1 x represents the data value at time m-1 before the missing data; m+1 The data value at time m+1 after the missing data;

[0019] Outlier detection:

[0020] Outlier detection is performed on all data, and outliers that clearly exceed the physical range are corrected. The Z-score method is used. If the value is not found, it is considered an anomaly and replaced with the median or other suitable value, where Z is the result of the Z-score; x is the data point being tested; μ is the mean of the dataset; and σ is the standard deviation of the dataset.

[0021] Data preprocessing:

[0022] All feature data are scaled to the required range to eliminate the influence of different units on model training. For irradiance data, the normalization method of formula (2) is used:

[0023] (2)

[0024] Where, x norm The result is after normalization; x is the original data; x min x is the minimum value of this data in the dataset; max This represents the maximum value of the data in the dataset.

[0025] As a further improvement of the present invention, in step S2, the historical power similarity in the static coupling relationship is constructed based on the historical power correlation between power stations according to the correlation coefficient formula, using the following Pearson correlation coefficient formula:

[0026] (3)

[0027] Among them, R ij The power correlation between node i and node j. and , respectively, represent the average power of the power plants represented by nodes i and j; and Let be the power of the power plants represented by nodes i and j at time t, respectively. Based on this, the correlation relationship of the output power between the power plants at each node can be verified. Thus, an adjacency matrix A is constructed. When the correlation exceeds a certain threshold ρ, a connection is established between two power plants at the node. The boundary values ​​of the correlated nodes in the adjacency matrix are set to 1; otherwise, the values ​​are set to 0. The mathematical expression can be simplified as follows:

[0028] (4)

[0029] Where A is the adjacency matrix; a i，j R represents the element values ​​between node i and node j in the adjacency matrix. ij ρ is the historical power correlation coefficient between node i and node j; ρ is the preset correlation threshold.

[0030] As a further improvement of the present invention, in step S2, the geographical location similarity in the static coupling relationship is constructed based on the constant geographical distance between power stations according to the distance calculation formula. The geographical distance between photovoltaic power stations can be calculated using the Haversine formula:

[0031] (5)

[0032] Where φ and λ are longitude and latitude, respectively; subscripts i and j represent nodes i and j, respectively; r is the Earth's radius, if d ij If < δ, where δ is the distance threshold, then an edge from node i to j is established in the static coupling relationship, indicating that there is a significant power correlation between the two power stations. The weight of the edge is calculated using the following formula to reflect the attenuation effect of distance on the coupling strength between the power stations:

[0033] (6)

[0034] Where, ω ij Let β be the edge weight between nodes i and j, and β be the attenuation parameter used to control the influence of distance on the weight. The larger the weight value, the stronger the coupling relationship between the two power stations. The subsequent static adjacency matrix is ​​constructed based on the above edges and weights, providing a foundation for the topology of the dynamic graph network G(t).

[0035] As a further improvement of the present invention, in step S2, the power station node in the constructed graph is defined as follows: N photovoltaic power stations in the region are defined as a set of N nodes V={v1,v2,...,v...} in the graph. N}, each node v i There is a feature vector X at each time step t. i (t), this vector contains the temporal characteristics of the power station itself, including historical power, irradiance, and temperature. The directed edge is defined as any two power station nodes v at any time t.i and v j There exists a directed edge e. ij The directionality of the edge is determined by the direction of meteorological propagation, which includes wind direction, based on the angle θ between the wind direction and the direction of the line connecting power station i to j. ij Calculation of cosθ ij If ρ is a predetermined threshold, then a directed edge is established from i to j, indicating that meteorological fluctuations can propagate from i to j; if cosθ ij If ≤ ρ, then it is determined whether to establish an edge pointing in the opposite direction from j to i, or not to establish an edge, depending on whether the included angle satisfies the threshold condition for back propagation.

[0036] As a further improvement of the present invention, in step S2, a time delay is introduced into the dynamic graph network G(t). To reflect the delayed impact of meteorological fluctuations at upstream power plants on downstream power plants, formulas (7) and (8) are used for calculation:

[0037] (7)

[0038] (8)

[0039] Where, τ ij S is the introduced time delay; t The wind speed at height H0; v t The wind speed is at height H; α is the wind shear index, which is chosen based on local surface conditions; the time delay is based on distance D. ij Equivalent wind speed at high altitude S t Time delay can be used to adjust the edge weights of future time steps, distributing the influence of the current time t to t+k steps, where k is calculated based on the time delay and the time step interval Δt, as shown in the equation:

[0040] (9).

[0041] As a further improvement of the present invention, in step S2, the dynamic adjacency matrix considers the quantification of the impact caused by meteorological fluctuations. Based on the difference between the current grid-connected power and the expected power, the power fluctuation ΔP% when a fluctuation event has occurred is calculated using the following formula (10) to quantify the real-time impact of different meteorological fluctuations on power output:

[0042] (10)

[0043] In the formula, P i (t) represents the output power of power station i at time t; Let be the projected output power of power plant at node i at time t; ΔP% is the quantized power output impact, which can be combined with time delay. The impact on the future is extrapolated to simulate the physical process of wave propagation.

[0044] As a further improvement of the present invention, in step S2, the construction of the dynamic graph network G(t) is achieved by fusing static coupling relationships and dynamic adjacency matrices. First, a static adjacency matrix A is constructed based on geographical location, distance, or historical correlation. static Then, the dynamic weight matrix W is calculated based on real-time weather data and time delay. dynamic (t); the final generated edge matrix E(t) = A static ·W dynamic (t) or other fusion methods, where · represents element-wise multiplication to ensure that the static matrix provides the basic topology, while the dynamic matrix adjusts the real-time weights and directions to form a reasonable spatiotemporal graph network.

[0045] As a further improvement of the present invention, in step S3, the hybrid model is constructed by combining graph neural networks and recurrent neural networks. The GNN variant processes the spatial dependency of the dynamic graph G(t) and aggregates node features; the RNN variant processes the temporal sequence output by the GNN and captures temporal dynamics; the input includes G(t) and preprocessed data, and the training is performed by optimizing the loss function end-to-end.

[0046] As a further improvement of the present invention, in step S4, the prediction accuracy requirement is evaluated by corresponding error indicators, including root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), and the indicators are required to be less than a preset threshold; if the requirements are not met, the model is optimized.

[0047] (11)

[0048] (12)

[0049] (13)

[0050] Where RMSE is the root mean square error; MAE is the mean absolute error; MAPE is the mean absolute percentage error; M is the total number of data samples; y m This represents the actual value of the m-th sample. This is the predicted value for the m-th sample.

[0051] Compared with the prior art, the advantages of the present invention are as follows:

[0052] This invention integrates dispersed photovoltaic power plants into an interconnected dynamic system through the above-mentioned technical solution. Compared with traditional methods, it significantly improves prediction accuracy, real-time response capability, and physical interpretability, providing a reliable solution for power management of large-scale photovoltaic power plant clusters and enhancing the stability and efficiency of actual photovoltaic systems. Attached Figure Description

[0053] Figure 1 : Overall system flowchart of the method of the present invention.

[0054] Figure 2 Heat map of power correlation between power plants.

[0055] Figure 3 Schematic diagram of the correlation of irradiance between power plants.

[0056] Figure 4 : Schematic diagram of the geographical distribution of photovoltaic power stations.

[0057] Figure 5 Figure : Partial experimental results using only recurrent neural networks (RNNs) for prediction.

[0058] Figure 6 Figure 1 shows some experimental results of prediction based on static graph neural networks (GNN).

[0059] Figure 7 The results of the ultra-short-term forecasting method for photovoltaic power plant clusters based on meteorological dynamic graph network proposed in this invention are shown in the figure. Detailed Implementation

[0060] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. It should be understood that the described embodiments are only some examples of the present invention and not all of the present invention. Other embodiments obtained by those skilled in the art without creative effort should also be considered within the scope of protection covered by the present invention.

[0061] The specific implementation of this invention is described in detail using a regional cluster containing three photovoltaic power plants as an example. The core idea of ​​this method is to treat each power plant as a node in a graph and dynamically construct directed weighted edges between nodes using meteorological physics principles. This dynamic graph network is then combined with the actual operating data of the power plants and input into a deep learning model for power prediction. The technical solution will be elaborated below to provide sufficient technical details to ensure operability and meet the corresponding writing requirements.

[0062] Figure 1 This is a flowchart illustrating the overall system process of the method of this invention, showing the complete flow from data preparation, data preprocessing, dynamic graph network construction to the final power prediction output. Figure 1 As shown, this embodiment of the invention provides a meteorological-based dynamic graph network method for ultra-short-term forecasting of photovoltaic power plant clusters, including the following steps:

[0063] Step S1: Obtain ultra-short-term time-series data of multiple geographically dispersed photovoltaic power stations. The data includes grid-connected power data, irradiance, and regional meteorological data (including wind speed, wind direction, temperature, humidity, and air pressure) for each power station. The data is then cleaned, missing values ​​are processed, and normalized.

[0064] Step S2: Define the photovoltaic power station as a node in the graph, and establish a relative static coupling relationship between the nodes based on geographical location, distance, or correlation analysis; at the same time, using the meteorological data, calculate the directed edge weights between the power station nodes based on meteorological physical factors (such as wind direction, wind speed, and cloud layer influence) to dynamically construct a dynamic adjacency matrix that reflects the physical propagation effect of meteorological fluctuations. Finally, combine the static coupling relationship and the dynamic adjacency matrix to construct a dynamic graph network G(t)=(V, E(t)).

[0065] Step S3: Construct a hybrid model consisting of a graph neural network (GNN) and a recurrent neural network (RNN). Input the dynamic graph network G(t) and the preprocessed time series data of each power station into the hybrid model for training until the preset training accuracy requirements are met, and complete the model training.

[0066] Step S4: Determine whether the power prediction result meets the preset prediction accuracy requirements; if not, return to the model building and training stage to retrain or fine-tune the hybrid model until the prediction accuracy meets the specific requirements.

[0067] Step S5: Once the hybrid model meets the prediction performance accuracy threshold, the latest real-time meteorological data and historical power data are used, combined with the real-time updated dynamic graph network G(t), to perform power prediction for a specific time period in the future through the hybrid model that meets the accuracy requirements.

[0068] This invention discloses a meteorological-based dynamic graph network method for ultra-short-term forecasting of photovoltaic (PV) power plant clusters. With the rapid development of distributed PV power generation technology, the scale of PV power plant clusters is continuously expanding. PV power generation is characterized by intermittency and volatility, and is significantly affected by meteorological conditions, leading to increased difficulty in forecasting the power output of PV power plant clusters. To address the problems of low prediction accuracy, poor real-time performance, and weak physical interpretability in existing PV power plant cluster power forecasting methods, this invention constructs a dynamic spatiotemporal graph network, treating PV power plants within a region as a complex system connected by meteorological fluctuations. It dynamically adjusts the edge weights and time delays of the graph using real-time meteorological data, capturing the overall trend under stable conditions and rapid fluctuations caused by sudden events, effectively integrating spatial dependence and temporal dynamics. Compared to existing technologies, this invention is expected to improve prediction accuracy, real-time response capability, and physical interpretability, supporting multi-step rolling forecasting and uncertainty estimation. It can better adapt to the volatile characteristics of large-scale PV systems, improving grid dispatch efficiency and system robustness.

[0069] The types of feature variables that can be referenced in this invention are numerous and vary, mainly depending on the specific sensor deployment. In this embodiment, the referenced feature variables are shown in Table 1 below.

[0070] Table 1. Feature variables that can be referenced

[0071]

[0072] The data required for this invention primarily comes from the real-time monitoring systems of three photovoltaic power plants within the region. The data acquisition frequency should be sufficiently high, for example, at intervals of 15 minutes or less, to support ultra-short-term power forecasting.

[0073] Before establishing a static coupling graph using historical power and irradiance data, preprocessing of this historical data is necessary. Preprocessing specifically includes handling missing values, such as using linear interpolation or forward imputation methods to remove null values ​​from the data, for time series data points x. t Missing value x m Through formula For calculation and outlier detection, the Z-score method can be used. If |z| > 3, it is considered an outlier and replaced with the median or other suitable value. Simultaneously, data normalization is performed, scaling all feature data to a uniform range to eliminate the influence of different units on model training. For ease of illustration, Table 2 below shows a partial diagram of the data processing flow for a segment of a day.

[0074] Table 2. Data and Data Processing Diagram

[0075]

[0076] When establishing a static coupling relationship diagram for a distributed photovoltaic power station cluster, a single correlation coefficient index can usually be used as the basis for judgment, such as using the Pearson correlation coefficient formula:

[0077] ;

[0078] Among them, R ij The power correlation between node i and node j. and , respectively, represent the average power of the power plants represented by nodes i and j; and Let be the power outputs of the power plants represented by nodes i and j at time t, respectively. Based on this, we can verify the correlation between the output power of the power plants at each node. Therefore, we construct an adjacency matrix A. If the correlation exceeds a certain threshold ρ, we establish a connection between two power plants, setting the relevant boundary values ​​in the adjacency matrix to 1. Otherwise, we set their values ​​to 0. The mathematical expression can be simplified as follows:

[0079] ;

[0080] Figure 2 A heatmap showing the power correlation between power plants is provided, illustrating the static coupling relationship based on the Pearson correlation coefficient. Figure 2 The heatmap shown can be built based on the correlation coefficient, and an adjacency matrix can be constructed accordingly. If the threshold ρ is 0, then the following adjacency matrix A can be constructed.

[0081] ;

[0082] Not only is there a correlation in power output, but irradiance, as an important indicator of photovoltaic power generation, also often exhibits a certain degree of correlation between regions. Figure 3 This diagram illustrates the correlation of irradiance among power plants, showing the correlation distribution of irradiance characteristics among various photovoltaic power plants. It can serve as another basis for constructing an adjacency matrix. Figure 3 As shown, the same relationship can also be used to construct a more stable adjacency matrix.

[0083] Furthermore, static coupling relationships can also be incorporated into geographic location similarity, using the Haversine formula to calculate geographic distance:

[0084] ;

[0085] Where φ and λ represent longitude and latitude respectively; subscripts i and j represent nodes i and j respectively; r is the Earth's radius, approximately 6371 km. Based on this, if d ijIf the distance threshold is less than δ (δ is the distance threshold, such as 50km), then an edge from node i to j is established in the static coupling relationship, indicating a significant power correlation between the two power stations. The edge weight is calculated using the following formula to reflect the attenuation effect of distance on the coupling strength between the power stations:

[0086] ;

[0087] Where, ω ij Let β be the edge weight between nodes i and j, and β be the attenuation parameter (e.g., 100km) used to control the influence of distance on the weight. The larger the weight value, the stronger the coupling relationship between the two power stations. The subsequent static adjacency matrix is ​​constructed based on the above edges and weights, providing the foundation for the topology of the dynamic graph network G(t).

[0088] Figure 4 This is a schematic diagram showing the geographical distribution of photovoltaic power stations. The specific geographical locations of the three photovoltaic power stations in the example are marked. Using the Haversine formula, distances can be further calculated to form a distance weight matrix, as shown below. Figure 4 Based on the Haversine formula, the geographical locations of the power stations referenced in this embodiment are calculated as follows: the distance between power stations N1 and N2 is 61.51 km, the distance between power stations N2 and N3 is 51.69 km, and the distance between power stations N1 and N3 is 106.73 km. If the attenuation coefficient β is taken as 100, the distance weight matrix can be obtained as follows:

[0089] ;

[0090] To construct a dynamic spatiotemporal graph network, the three photovoltaic power stations in the region are first defined as a node set V = {v1, v2, v3} in the graph, with each time step being 15 minutes. Each of the three nodes has a feature vector composed of the processed data shown in Table 2, containing the power station's own temporal characteristics (such as historical power, irradiance, temperature, etc.). Directed edges are defined at any time t, where any two power station nodes v1, v2, v3, v3, v4, v5, v6, v7, v8, v9, v1, v2, v3 ...9, v1, v2, v3, v3, v6, v9, v1, v2, v3, v3, v6, v1, v2, v3, v3, v6, v1, v2, v3, v3, v6, v i and v j There exists a directed edge e. ij Its directionality is determined by the direction of meteorological propagation (especially wind direction), for example, based on the angle θ between the wind direction and the direction of the line connecting power station i to j. ij Calculation of cosθ ij If ρ (ρ is a set threshold), then a directed edge is established from i to j, indicating that meteorological fluctuations can propagate from i to j; if cosθ ij If ≤ ρ, then it is determined whether to establish an edge pointing in the opposite direction from j to i, or not to establish an edge, based on the wind direction. The specific decision depends on whether the included angle satisfies the threshold condition for backpropagation. For example... Figure 4The diagram shows the specific geographical locations of the power stations. Based on the wind direction data, the edge direction matrix can be completed at any given time.

[0091] Table 3 Wind direction data and side direction diagram

[0092]

[0093] ;

[0094] The construction of dynamic graph networks further introduces time delays to reflect the delayed impact of meteorological fluctuations at upstream power plants on downstream power plants. For example, time delays can be calculated using formulas such as:

[0095] ;

[0096] ;

[0097] Where, τ ij S is the introduced time delay; t The wind speed at height H0; v t The wind speed is at height H; α is the wind shear index, which is chosen based on local surface conditions; the time delay is based on distance D. ij Equivalent wind speed at high altitude S t The time delay can be used to adjust the edge weights of future time steps, for example, to distribute the influence of the current time t to t+k steps, where k is based on the time delay and the time step interval Δt.

[0098] Meanwhile, the dynamic adjacency matrix considers the inversion of meteorological fluctuations by calculating the percentage difference ΔP% between the grid-connected power of upstream power plants currently affected by meteorological fluctuations and the expected power.

[0099] ;

[0100] Then, by combining the impact of time delay extrapolation on the future, the physical process of wave propagation is simulated. In the example, power plant N1 at noon is taken as an example, with a power fluctuation ΔP% of 2.4%, and the time delay result τ is combined with the actual situation. 12 With a time step interval Δt of 15 minutes, it will have a similar impact on power station N2 after 2 hours, or 8 time steps. Similarly, the fluctuation situation at this moment is also affected by the other two power stations k time steps ago, which can be constructed as a matrix:

[0101] ;

[0102] The final construction of the dynamic graph network G(t) is achieved by fusing static coupling relationships and a dynamic adjacency matrix. For example, a static adjacency matrix A is first constructed based on geographical location, distance, or historical relevance. static Then calculate the dynamic weight matrix W.dynamic (t) (based on real-time weather and time lag):

[0103] ;

[0104] The final generated edge matrix is ​​E(t) = A static ·W dynamic (t), thus ensuring that the static matrix provides the basic topology, while the dynamic matrix adjusts the real-time weights and directions to form a reasonable spatiotemporal graph network.

[0105] Furthermore, in the hybrid model construction and training phase, the model can be implemented through a combination of graph neural networks and recurrent neural networks. For example, GNN variants (such as GCN) handle the spatial dependencies of the dynamic graph G(t) and aggregate node features; RNN variants (such as LSTM) process the temporal sequence output of GNN and capture temporal dynamics. The input includes G(t) and preprocessed data, and training is performed by optimizing the loss function end-to-end.

[0106] Model validation and tuning evaluate prediction accuracy using multiple error metrics, such as root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).

[0107] ;

[0108] ;

[0109] ;

[0110] If the evaluation metrics are better than the preset threshold, training is complete; otherwise, it returns to the training phase for retraining or optimization. These metrics can also be used to demonstrate the network's performance. In this example, predictions were performed using only a recurrent network, and some results are shown below. Figure 5 As shown, the prediction accuracy of the model without introducing spatial dependence is demonstrated; predictions using a static graph neural network are presented in the following figures. Figure 6 As shown, the prediction performance of the model under a fixed adjacency matrix condition is illustrated; and the prediction is performed using the network constructed by the meteorological-based dynamic graph network photovoltaic power plant cluster ultra-short-term power prediction method proposed in this invention, with the results as follows. Figure 7 As shown, verification was performed, and the experimental results are listed in Table 4 below for comparison. Figure 5 and Figure 6 This demonstrates a significant improvement in prediction accuracy and robustness.

[0111] As can be seen, the present invention integrates dispersed photovoltaic power plants into an interconnected dynamic system through the above technical solution. Compared with traditional methods, it significantly improves prediction accuracy, real-time response capability and physical interpretability, provides a reliable solution for power management of large-scale photovoltaic power plant groups, and enhances the stability and efficiency of actual photovoltaic systems.

[0112] Table 4 Comparison of prediction performance of different methods

[0113]

[0114] Furthermore, in the final real-time power prediction stage, the trained network can be saved, and rolling prediction can be used to perform multi-step prediction of the power in the next k steps by iteratively inputting real-time data. The corresponding prediction results and confidence intervals can be output to complete the complete workflow of the system.

[0115] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A meteorology-based dynamic graph network photovoltaic power station group ultra-short-term prediction method, characterized in that, The method comprises the following steps: S1, obtaining ultra-short time series data of a plurality of photovoltaic power stations dispersed in geographical positions, the data comprising grid-connected power data, irradiance and regional meteorological data of each power station, the regional meteorological data comprising wind speed, wind direction, temperature, humidity and air pressure, and performing missing value processing, abnormal value detection and data preprocessing on the data; S2, defining the photovoltaic power stations as nodes in a graph, and establishing a relatively static coupling relationship between the nodes based on geographical position similarity, power station distance relationship, historical power similarity or other correlation analysis; at the same time, using the meteorological data, calculating the directed edge weight between the power station nodes based on meteorological physical factors to dynamically construct a dynamic adjacency matrix reflecting the physical propagation effect of meteorological fluctuations, and finally combining the static coupling relationship and the dynamic adjacency matrix to construct a dynamic graph network G(t) = (V, E(t)), wherein V represents a set composed of N nodes, and E(t) represents a dynamic edge of the graph structure; S3, constructing a hybrid model composed of a graph neural network and a recurrent neural network, inputting the dynamic graph network G(t) and the preprocessed time series data of each power station into the hybrid model for training until a preset training accuracy requirement is met, and completing model training; S4, judging whether the power prediction result meets a preset prediction accuracy requirement; if not, returning to the model construction and training stage to retrain or optimize the hybrid model until the prediction accuracy meets the corresponding situation requirement; S5, when the hybrid model meets the prediction performance accuracy threshold, using the latest real-time meteorological data and historical power data, combining the real-time updated dynamic graph network G(t), and performing power prediction in a specific time period in the future through the hybrid model meeting the accuracy requirement, the power prediction adopting a prediction mechanism, including rolling prediction, future k-step power calculated by iterative input of real-time data, supporting multi-step prediction, and being capable of outputting corresponding prediction results and confidence intervals or prediction results of other uncertainty estimation methods. 2.The weather-based dynamic graph network photovoltaic power station group ultra-short-term prediction method according to claim 1, characterized in that: In the step S1, the data processing steps comprise: missing value processing: Using linear interpolation or forward fill method to handle missing values in data, for time series data point x t , missing value x m may be calculated by the following equation (1): (1) Wherein, x m is the missing value at time m; x m-1 is the data value at the previous time m-1 of the missing data; x m+1 is the data value at the next time m+1 of the missing data; abnormal value detection: Data outliers detection is performed on all data and abnormal data that obviously exceed the physical range are corrected. Z-score method is adopted. If then it is considered as abnormal and replaced by median or other appropriate value, wherein Z is the result of Z-score value; x is the data of the detected data point; μ is the mean of the data set; σ is the standard deviation of the data set. data preprocessing: scaling all feature data to a required range to eliminate the influence of different dimensions on model training, and using the following formula (2) for normalization for irradiance data: (2) where x norm is the normalized result; x is the original data; x min is the minimum value of the data in the data set; x max is the maximum value of the data in the data set. 3.The weather-based dynamic graph network photovoltaic power station group ultra-short-term prediction method according to claim 2, characterized in that: In the step S2, the historical power similarity in the static coupling relationship is constructed based on the historical power correlation between power stations according to a correlation coefficient formula, and the Pearson correlation coefficient formula is used as follows: (3) wherein R ij is the power correlation of node i and node j, is the average power of the power station represented by node i and node j, respectively; is the power of the power station represented by node i and node j at time t, respectively; is the power of the power station represented by node i and node j at time t, respectively; Based on this, the correlation relationship between the output powers of the nodes and power stations can be verified, and an adjacency matrix A is constructed. In the case where the correlation exceeds a certain threshold value p, a connection between the two nodes and power stations is established, and the relevant edge value in the adjacency matrix is set to 1, otherwise, the value is set to 0. The mathematical expression relationship can be simplified as follows: (4) wherein A is an adjacency matrix; a i，j is an element value between node i and node j in the adjacency matrix; R ij is a historical power correlation coefficient between node i and node j; and p is a preset correlation threshold. 4.The weather-based dynamic graph network photovoltaic power station group ultra-short-term prediction method according to claim 3, characterized in that: In the step S2, the geographical position similarity in the static coupling relationship is constructed based on the invariable geographical distance between power stations according to a distance calculation formula, and the Haversine formula can be used to calculate the geographical distance between photovoltaic power stations: (5) where φ and λ are the longitude and latitude, respectively; the subscripts i and j represent nodes i and j, respectively; r is the earth radius, and d is the distance between nodes i and j. ij where δ is the distance threshold, then an edge between nodes i and j is established in the static coupling relationship, indicating that there is a significant power correlation between the two power plants, and the weight of the edge is calculated by the following formula to reflect the attenuation effect of distance on the coupling strength between power plants: (6) where ω ij is the edge weight between nodes i and j, and β is the decay parameter used to control the strength of the distance effect on the weight. The larger the weight value, the stronger the coupling relationship between the two power plants. The subsequent static adjacency matrix is based on the above edge and weight, providing a basis for the topology of the dynamic graph network G(t).

5. The dynamic graph network meteorology-based ultra-short-term prediction method for a photovoltaic power plant group according to claim 4, characterized in that: The definition of power station nodes in the graph constructed in step S2 is: N photovoltaic power stations in the region are defined as N node set V = {v1, v2,..., vN} in the graph. N} each node v i In each time step t, there is a feature vector X i (t) containing the time series characteristics of the power station itself including historical power, irradiance and temperature. The directed edge is defined as: at any time t, there is a directed edge e i between any two power station nodes v j and v ij The directionality of the edge is determined by the meteorological propagation direction, which includes the wind direction. The cosine value based on the angle θ ij between the wind direction and the connecting direction of the power station i to j is calculated. If cosθ ij > ρ, ρ is a certain threshold value, then a directed edge from i to j is established, indicating that meteorological fluctuations can propagate from i to j. If cosθ ij ≤ ρ, then according to the wind direction, it is judged whether to establish an opposite direction edge from j to i or not to establish an edge, depending on whether the angle meets the threshold condition of reverse propagation. 6.The weather-based dynamic graph network photovoltaic power station group ultra-short-term prediction method according to claim 5, characterized in that: In the step S2, the dynamic graph network G(t) introduces time delay To reflect the delayed influence of the weather fluctuation of the upstream power station on the downstream power station, the formula (7) and the formula (8) are calculated: (7) (8) where τ ij is the introduced time lag; S t is the wind speed at height H0; v t is the wind speed at height H; a is the wind shear exponent, which is taken as an appropriate value according to local ground conditions; the time lag is based on the distance D ij and the equivalent wind speed at high altitude S t ; the time lag can be used to adjust the edge weight of the future time step, distributing the influence of the current time t to the t+k step, where k is calculated based on the time lag and the time step interval At, as follows: (9)。 7.The weather-based dynamic graph network photovoltaic power station group ultra-short-term prediction method according to claim 6, characterized in that: In the step S2, the dynamic adjacency matrix considers the quantification of the influence caused by meteorological fluctuations, and calculates the power fluctuation ΔP% when a fluctuation event has occurred based on the difference between the current grid-connected power and the expected power as follows (10) to quantify the real-time influence of different meteorological fluctuations on power output: (10) where P i (t) is the output power of the power plant at node i at time t; is the forecasted output power of the power plant at node i at time t; ΔP% is the quantized power output impact, based on which the future impacts can be derived to simulate the physical process of the wave propagation. to derive the future impacts to simulate the physical process of the wave propagation. 8.The weather-based dynamic graph network photovoltaic power station group ultra-short-term prediction method according to claim 7, characterized in that: In the step S2, the construction of the dynamic graph network G(t) is achieved by fusing the static coupling relationship and the dynamic adjacency matrix. First, a static adjacency matrix A is constructed based on geographical location, distance or historical correlation static ; then a dynamic weight matrix W dynamic (t) is calculated based on real-time weather and time lag; finally, an edge matrix E(t) = A static ·W dynamic (t) or other fusion methods are generated, wherein · represents element-level multiplication, to ensure that the static matrix provides the basic topology, while the dynamic matrix adjusts the real-time weight and direction to form a reasonable space-time graph network. 9.The weather-based dynamic graph network photovoltaic power station group ultra-short-term prediction method of claim 8, wherein: In the step S3, the hybrid model is constructed by combining a graph neural network and a recurrent neural network. The GNN variant processes the spatial dependence of the dynamic graph G(t) and aggregates node features. The RNN variant processes the time sequence output by the GNN and captures time dynamics. The input includes G(t) and preprocessed data, and the training is performed by end-to-end optimization of a loss function. 10.The weather-based dynamic graph network photovoltaic power station group ultra-short-term prediction method of claim 9, wherein: In the step S4, the prediction accuracy requirement is evaluated by a corresponding error indicator, including root mean square error RMSE, mean absolute error MAE, and mean absolute percentage error MAPE. The requirement indicator is required to be less than a preset threshold. If not satisfied, the model is tuned. (11) (12) (13) Wherein, RMSE is the root mean square error; MAE is the mean absolute error; MAPE is the mean absolute percentage error; M is the total number of data samples; y m is the actual value of the mth sample; is the predicted value of the mth sample.

Images