A photovoltaic power station ultra-short-term prediction method, system and device based on physical information driven dynamic graph and a storage medium

Abstract

This invention provides a method, system, device, and storage medium for ultra-short-term prediction of photovoltaic power plants based on a physically driven dynamic graph. The method includes the following steps: S1, acquiring irradiance, wind speed, wind direction, geographical location, and time information of multiple photovoltaic power plants to construct a multi-site time-series dataset; S2, calculating the irradiance attenuation ratio based on measured irradiance and corresponding clear-sky irradiance to construct an input feature sequence; S3, combining geographical location meteorological variables, constructing a dynamic adjacency matrix based on a wind field correction model using Ekman layer theory, and incorporating cloud propagation causality constraints; S4, using the dynamic adjacency matrix for spatial correlation modeling to extract spatial correlation features between power plants; S5, fusing spatial features, periodic time features, and time-series observation features, inputting them into a time-series prediction model to complete the modeling; S6, outputting future multi-time-step irradiance prediction results. The method provided by this invention improves prediction accuracy, real-time response capability, and physical interpretability.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation prediction and intelligent information processing technology, specifically to a method, system, device, and storage medium for ultra-short-term prediction of photovoltaic power plants based on physical information-driven dynamic graphs. Background Technology

[0002] With the deepening implementation of the "dual carbon" target, the proportion of photovoltaic power generation in the global energy structure continues to increase, making it one of the fastest-growing clean energy sources in the power system. According to the "2025 China Wind and Solar Energy Resources Annual Bulletin," the national average annual optimal slope total irradiance reaches 1742.2 kWh / m², and the first-year utilization hours of fixed photovoltaic power stations are 1393.8 hours. However, the randomness and intermittency of photovoltaic power generation remain the core bottleneck restricting its large-scale grid connection and consumption. Photovoltaic output is mainly regulated by changes in solar irradiance, which is affected by multiple factors such as dynamic cloud cover, regional meteorological heterogeneity, and spatial location differences. Therefore, achieving high-precision ultra-short-term irradiance forecasting has become a key technical requirement for ensuring stable grid operation and improving photovoltaic absorption capacity.

[0003] Early prediction schemes based on statistical models (such as AR and ARIMA) and traditional machine learning methods (such as SVM and basic ANN) have been gradually replaced by deep learning technology because they are difficult to characterize complex nonlinear spatiotemporal dependencies. In recent years, research in the field of photovoltaic irradiance prediction has focused on the innovative integration of temporal feature mining and spatial correlation modeling. In related research after 2024, hybrid architectures such as TimesNet-GCN and GCN-Informer have become mainstream. By integrating the multi-period extraction capabilities of temporal models (TimesNet and Informer) with the spatial modeling advantages of graph convolutional networks (GCN), they have achieved the collaborative capture of spatiotemporal features, improving the prediction accuracy by 15%-30% compared to traditional LSTM models. End-to-end AI models such as SolarSeer directly map future irradiance through satellite observation data, with a computation speed 1500 times faster than traditional numerical weather prediction (NWP). In the validation at 1800 stations across the United States, the root mean square error was reduced by 15.35% compared to the mainstream NWP model. Another study proposed a physical perception deep learning framework, which separates astronomical periodic components and random meteorological fluctuations through Fourier signal decomposition, achieving an R² value of 0.9281 for 15-minute advance prediction and reducing the maximum absolute error under extreme weather conditions by 15%-20%.

[0004] In the field of multi-power station collaborative prediction, the application of graph neural networks (GNNs) has become a research hotspot. Since 2025, many related studies have adopted dynamic adjacency matrix construction strategies. For example, multi-source spatial adjacency matrices are generated by fusing geographical distance, historical correlation, and graph self-attention networks, or by dynamically encoding the correlation coefficients between nodes based on spatiotemporal attention mechanisms to adapt to the non-stationary characteristics of meteorological conditions. Multi-site collaborative frameworks such as SpatialSolar-Net, through adaptive spatial correlation evaluation mechanisms, have achieved a 64% error reduction under extreme weather conditions such as sandstorms, demonstrating strong robustness. However, existing technologies still have significant shortcomings: on the one hand, most dynamic graph constructions rely on data-driven correlation analysis (such as sliding window Pearson coefficients and KNN algorithms), lacking explicit modeling of the physical mechanisms of cloud propagation, resulting in weak physical interpretability of dynamic correlations; on the other hand, even if some models use meteorological variables such as wind speed and direction as input features, they do not use them to directly construct spatial propagation structures, failing to effectively characterize the transmission effect of cloud movement on irradiance changes, and there is still room for improvement in prediction accuracy under dynamic meteorological scenarios such as rapid cloud passage.

[0005] Existing multi-power station photovoltaic irradiance prediction methods typically describe the spatial relationships between power stations using a fixed adjacency matrix, which fails to reflect the dynamic spatial propagation characteristics caused by changes in meteorological conditions such as wind speed and direction during cloud propagation. Furthermore, while existing methods can utilize historical irradiance data for time-series modeling, they lack sufficient characterization of physical processes such as cloud propagation, thus affecting prediction accuracy. Therefore, this invention addresses the problem of providing an innovative method that integrates meteorological physical mechanisms with dynamic graph modeling. This method directly drives the dynamic evolution of spatially related structures through physical information, achieving accurate characterization of the irradiance propagation process across multiple power stations, further improving the accuracy and robustness of ultra-short-term predictions, and providing technical support for the real-time scheduling of distributed photovoltaic clusters. The goal is to develop a physical information-driven dynamic graph-based ultra-short-term prediction method, system, device, and storage medium for photovoltaic power stations. Summary of the Invention

[0006] The purpose of this invention is to propose a method, system, device, and storage medium for ultra-short-term prediction of photovoltaic power plants based on physical information-driven dynamic graphs. By constructing a dynamic adjacency matrix that changes over time to describe the spatial propagation relationship between power plants, and combining graph convolutional networks with time series prediction models, the invention achieves collaborative prediction of irradiance changes of multiple power plants, thereby improving the prediction model's ability to model cloud propagation processes and the overall prediction accuracy.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] S1. Obtain historical observation data from multiple photovoltaic power stations and construct a multi-site time-series dataset. The observation data includes: irradiance data, wind speed data, wind direction data, time information, and geographical location information for each photovoltaic power station. The data is then processed for missing values, outlier detection, and normalization. Assuming there are N photovoltaic power stations, the measured irradiance sequence data at time t can be represented as:

[0009]

[0010] in: This represents the measured irradiance of the i-th photovoltaic power station at time t;

[0011] Time information uses sine and cosine encoding, with hours encoded as sin(2πt / 24) and cos(2πt / 24), and months encoded as sin(2πt / 12) and cos(2πt / 12), to capture day-night and seasonal periodicity;

[0012] A dynamic adjacency matrix is ​​constructed based on the geographical relationships between photovoltaic power stations and meteorological variables to describe the spatial propagation relationships between photovoltaic power stations over time. Let the geographical vector between power station i and power station j be . Its unit direction vector is:

[0013]

[0014] Among them: when the direction of cloud propagation is... When they are consistent, establish a directed connection from i to j; where, Indicates the geographical coordinates of power station i. Indicates power station By comparing the geographical coordinates of the wind direction with the direction of the line connecting the power stations, it can be determined whether the cloud propagation direction is from power station i to power station j, thereby establishing a directed connection relationship. This represents the Euclidean distance between two points.

[0015] S2. Calculate the irradiance attenuation ratio based on the measured irradiance and the clear-sky irradiance at the corresponding time, and use it as the model input feature. The formula for calculating the irradiance attenuation ratio is:

[0016]

[0017] in: This represents the irradiance attenuation ratio of the i-th photovoltaic power station at time t; This represents the measured irradiance of the i-th photovoltaic power station; Represents the clear sky irradiance at time t. The Ineichen model from the pvlib library is used for calculation, taking into account parameters such as solar altitude angle, atmospheric turbidity, and altitude to ensure the accuracy of irradiance under cloudless conditions.

[0018] To reduce the impact of outliers on model training, the irradiance attenuation ratio is truncated to a preset range:

[0019]

[0020] in: and The upper and lower bounds are preset; clip(·) is a numerical truncation function used to limit the irradiance attenuation ratio within a reasonable range of [0, 1.2]. clip(·) is the clip function followed by the content within the parentheses, such as... ;

[0021] S3. Combining geographical location and meteorological variables, a dynamic adjacency matrix is ​​constructed based on a wind field correction model incorporating Ekman layer theory. A causal constraint on cloud propagation is added to ensure that information propagates unidirectionally from upstream photovoltaic power plants to downstream photovoltaic power plants. The meteorological variables are wind speed and wind direction. To more accurately describe the cloud propagation direction, the meteorological variables including wind speed and wind direction are processed using a wind field correction model incorporating Ekman layer theory to obtain the cloud propagation direction. When performing wind field correction, the corrected cloud propagation vector is set as follows:

[0022]

[0023] in: Wind speed; Wind direction; Here, E represents the cloud height; E(⋅) is the wind field correction function, which is implemented based on the Ekman layer theory. E(⋅) consists of E and the content within the parentheses following it, such as... Cloud height indicator It is obtained by inversion of the irradiance attenuation ratio, and the specific calculation method is as follows:

[0024] Atmospheric transparency parameter is directly taken as the irradiance attenuation ratio: Cloud height estimation formula:

[0025]

[0026] Where: H0 is the local annual average cloud base height, taken as 2000m; k is the atmospheric attenuation coefficient, taken as 1.0; the calculation results are normalized to the [0,1] interval and used as the cloud height input for the wind field correction model, where exp represents the natural exponential function, used to describe the exponential relationship of the cloud attenuation effect, and the estimated cloud height is used as the input for the wind field correction model. After being normalized to the [0,1] interval, it is used as the cloud height input parameter in the wind field correction model and participates in the calculation of the cloud propagation direction in the Ekman wind field correction model;

[0027] In one implementation, the wind field correction function adjusts the wind direction based on Ekman layer theory, thereby improving the consistency of wind field information between different power plants. The Ekman wind field correction is used to compensate for wind speed attenuation and wind direction deflection caused by surface friction. The specific calculation steps are as follows:

[0028] Coriolis parameter calculation:

[0029]

[0030] Where ω is the angular velocity of Earth's rotation. The latitude of the photovoltaic power station;

[0031] Wind speed amplification correction:

[0032]

[0033] Where, α E This is the Ekman amplification factor, ranging from 0.1 to 0.3, used to compensate for near-surface wind speed attenuation caused by surface friction. The wind speed at time t is amplified and corrected.

[0034] Wind direction rotation correction:

[0035]

[0036] in, Let be the wind direction after rotation correction at time t, and β be the turning coefficient, with a value range of . ~ H(t) is the cloud height at time t of the i-th power station after being inverted from the irradiance attenuation ratio and normalized to the interval [0,1].

[0037] The edge weights pointing from power station i to power station j in the dynamic adjacency matrix are calculated as follows:

[0038]

[0039] in: This is the wind speed modulation coefficient; The direction sensitivity coefficient; This is the corrected wind speed. This is the normalized wind direction vector; Let be the unit vector of the power station connection direction, and exp(·) be the natural exponential function used to calculate the directional similarity weight;

[0040] The initial dynamic adjacency matrix is ​​normalized to obtain the final dynamic graph structure:

[0041]

[0042] in: This is the initial adjacency matrix; It is the identity matrix; For self-loop weights; ReLU is the linear rectified activation function; Softmax is the normalization function;

[0043] To suppress high-frequency fluctuations in the adjacency matrix caused by meteorological noise, the normalized dynamic adjacency matrix is ​​smoothed using an exponential moving average, as shown in the following formula:

[0044]

[0045] in: Let be the final dynamic adjacency matrix after smoothing at time t; This is the adjacency matrix after Softmax and self-loop processing; This is a smoothing coefficient with a value of 0.7, used to preserve historical spatiotemporal correlation features and improve the stability of the graph structure.

[0046] S4. Spatial correlation modeling of multi-power station feature data is performed using graph convolutional networks. The calculation method for the k-th order propagation feature is as follows:

[0047]

[0048] in This represents the propagation characteristics of the k-th order. It is a dynamic adjacency matrix; The weight matrix is ​​trainable. This is the activation function.

[0049] In one implementation, only first-order and second-order propagation are used, and the final spatial representation is obtained through feature concatenation:

[0050]

[0051] S5. Spatial features, periodic temporal features, and time-series observation features are fused and input into a time-series prediction model for time-series modeling. In one embodiment, the time-series prediction model is a time-series hybrid network (TSMixer), which achieves feature interaction through linear transformations of multiple time dimensions and feature dimensions.

[0052] The irradiance prediction results for multiple future time steps are output through a linear mapping layer:

[0053]

[0054] Where: h is the prediction step size. To predict irradiance.

[0055] On the other hand, the present invention provides a photovoltaic power plant ultra-short-term forecasting system based on a physical information-driven dynamic graph, applied to the method described in any one of claims 1-7, comprising:

[0056] The data acquisition and preprocessing module is used to acquire irradiance, wind speed, wind direction, geographical location and time information of multiple photovoltaic power stations to construct a multi-site time series dataset;

[0057] The feature construction module is used to calculate the irradiance attenuation ratio based on the measured irradiance and the corresponding clear sky irradiance, and to construct the input feature sequence.

[0058] The dynamic graph construction module is used to combine geographical location and meteorological variables to construct a dynamic adjacency matrix based on a wind field correction model that includes Ekman layer theory, and to add cloud propagation causality constraints to ensure that information only propagates unidirectionally from upstream photovoltaic power plants to downstream photovoltaic power plants. The meteorological variables are wind speed and wind direction.

[0059] The spatial feature extraction module uses a dynamic adjacency matrix to model spatial associations and extract spatial association features between power stations.

[0060] The time series modeling module is used to integrate spatial features with periodic temporal features and time series observation features, and input them into the time series prediction model to complete the modeling.

[0061] The prediction output module is used to output the irradiance prediction results for future multi-time steps. The irradiance attenuation ratio in step S2 is calculated as follows:

[0062]

[0063] in: This represents the irradiance attenuation ratio of the i-th power station at time t; This represents the measured irradiance of the i-th power station; The clear sky irradiance at time t, the The results were obtained by combining the solar altitude angle, atmospheric turbidity, power station altitude parameters, and the Ineichen model from the pvlib library.

[0064] On the other hand, the present invention also provides a photovoltaic irradiance prediction device, including a processor and a memory, the memory storing a computer program, the computer program including instructions; when the instructions are executed by the processor, the photovoltaic irradiance prediction device performs the method according to any one of claims 1-7.

[0065] On the other hand, the present invention also provides a computer-readable storage medium for storing a computer program for performing the method according to any one of claims 1-7.

[0066] Compared with the prior art, the beneficial effects of the present invention are:

[0067] The prediction method provided by this invention integrates dispersed photovoltaic power plants into an interconnected dynamic system, significantly improving prediction accuracy, real-time response capability, and physical interpretability compared to traditional methods. This provides a reliable solution for power management of large-scale photovoltaic power plant clusters, enhancing the stability and efficiency of practical photovoltaic systems. Furthermore, the method provided by this invention exhibits significantly improved prediction accuracy in ultra-short-term forecasting compared to existing methods, with strong physical interpretability, making it suitable for the operation and control of large-scale photovoltaic power plant clusters. Attached Figure Description

[0068] Figure 1 The present invention provides an overall architecture flowchart of a photovoltaic power plant ultra-short-term prediction method based on physical information-driven dynamic graphs.

[0069] Figure 2 This is a cross-correlation analysis diagram of the irradiance attenuation ratio of representative upstream and downstream power plants (such as Ligang and Yangxiang) of this invention;

[0070] Figure 3 This is a schematic diagram illustrating the correlation of irradiance between photovoltaic power plants according to the present invention;

[0071] Figure 4 This is a schematic diagram showing the geographical distribution of the photovoltaic power plants of this invention;

[0072] Figure 5 This is a verification example of cloud propagation based on Himawari-8 satellite cloud images, according to the present invention.

[0073] Figure 6 This is a comparison chart showing the irradiance prediction performance of the method of the present invention (dynamic graph) and the method based on static graph during a typical cloud transit period.

[0074] Figure 7 This is a sensitivity analysis plot showing the impact of the convolution propagation order K on prediction performance. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0076] This invention provides a detailed description of a specific embodiment using a regional cluster of seven photovoltaic power plants as an example, focusing on the Ligang (upstream) and Yangxiang (downstream) power plants as typical representatives. The core idea of ​​this method is to treat each power plant as a node in a dynamic graph. By introducing physical information based on Ekman wind field correction, directed weighted edges between nodes are dynamically constructed. This dynamic graph network is then combined with the actual operating data of the power plants and input into a lightweight time-series hybrid network for irradiance prediction. The technical solution will be elaborated below to provide sufficient technical details to ensure operability and meet the requirements for patent application.

[0077] like Figure 1 As shown, this embodiment of the invention provides a method for ultra-short-term irradiance prediction of photovoltaic power plants based on physically driven dynamic graphs, demonstrating the complete process from data input, physical information-driven dynamic adjacency matrix construction, spatial feature extraction to temporal hybrid network (TSMixer) modeling and output of irradiance prediction results, including the following steps:

[0078] Step S1: Obtain historical observation data from multiple photovoltaic power plants and construct a multi-site time-series dataset:

[0079] In this embodiment, the data comes from seven actually operating photovoltaic power stations in Jiangsu Province, whose geographical distribution is as follows: Figure 4 As shown, the analysis focuses on the Ligang Hydropower Station (upstream) and Yangxiang Hydropower Station (downstream) as typical representatives to demonstrate a significant peak at the positive lag time (e.g., 15 minutes), verifying the directional delay dependency caused by cloud propagation. Figure 2 As shown, the time span is from June 2024 to June 2025, with a data sampling resolution of 15 minutes, and approximately 35,000 synchronous samples were acquired. The observation data for each power station includes: global horizontal irradiance (GHI), DC power output, and wind speed and direction at heights of 10 meters and 100 meters.

[0080] After data collection, preprocessing is performed first, including:

[0081] Missing values ​​are filled using linear interpolation. For time series data points x... t If data at time m is missing, its missing value x m Calculated using the following formula:

[0082]

[0083] 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 This represents the data value at time m+1 after the missing data point. Time information is constructed using sine and cosine encoding. The hour encoding formula is sin(2πt / 24) and cos(2πt / 24), and the month encoding formula is sin(2πt / 12) and cos(2πt / 12). This eliminates the dimensional differences in time features, making it easier for the model to capture periodic patterns.

[0084] Outlier detection can use the Z-score method. If |Z|>3, it is considered an outlier and replaced with the median or other suitable value. Simultaneously, data normalization is performed to scale all feature data to a uniform range, eliminating the influence of different units on model training. For example, for irradiance data, the Min-Max normalization formula can be used.

[0085]

[0086] 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.

[0087] Step S2: Calculate the irradiance attenuation ratio based on the irradiance data and the clear-sky irradiance at the corresponding time, and construct the power station input feature sequence to remove the influence of deterministic diurnal variations and seasonal trends, so that the model focuses on random fluctuations caused by clouds. This method introduces the irradiance attenuation ratio as the core modeling objective. The irradiance attenuation ratio is calculated according to the formula in the invention. The calculation formula is as follows:

[0088]

[0089] in, This represents the irradiance attenuation ratio of the i-th photovoltaic power station at time t; This indicates the measured irradiance of the photovoltaic power station; This indicates the clear-sky irradiance at the corresponding moment.

[0090] In this embodiment, the clear-sky irradiance is calculated using the Ineichen model in the pvlib library. This model takes into account factors such as solar altitude angle, atmospheric turbidity, and power station altitude, and can accurately describe the theoretical irradiance under cloudless conditions.

[0091] To prevent outliers caused by sensor noise or extreme weather from affecting model training, the calculated attenuation ratio ai(t) is truncated and limited to a physically reasonable range [0, 1.2]. That is: .

[0092] Finally, the normalized irradiance attenuation ratio, wind speed, wind direction, time coding and other features are concatenated to form the input feature vector for each time step, thus constructing a multivariate time series dataset.

[0093] Step S3: As Figure 5 As shown, the cloud imagery from the Himawari-8 satellite provides a clear visual verification of the cloud propagation process from the upstream Ligang Hydropower Station to the downstream Yangxiang Hydropower Station. This provides a physical basis for constructing the dynamic adjacency matrix of this invention, demonstrating the cloud's movement from the upstream power station (Ligang) to the downstream power station (Yangxiang), and offering intuitive evidence for physical information-driven graph structures. Furthermore, Figure 3 The correlation distribution of irradiance among photovoltaic power plants is shown, further confirming the existence of upstream-downstream dependence, which can serve as a reference for constructing an adjacency matrix.

[0094] This step is one of the core innovations of the method of this invention, which aims to construct a dynamic graph structure that can truly reflect the physical process of cloud propagation.

[0095] Determine the fixed geographical relationships between power stations. Let the geographical vector between power station i and power station j be dij. Its unit direction vector is: When the direction of cloud propagation is... When they are consistent, establish a directed connection from i to j; where, Indicates the geographical coordinates of power station i. Indicates power station Geographic coordinates It represents the Euclidean distance between two points.

[0096] Physical corrections are applied to the measured meteorological variables. To more accurately estimate the propagation direction of cloud height, a corrected model based on Ekman layer theory is introduced:

[0097]

[0098] in: Wind speed; Wind direction; E(t) is a cloud height indicator inferred based on the irradiance attenuation ratio ai(t), used to distinguish different cloud advection states in the Ekman correction; The wind field correction function estimates the true advection direction of the cloud cluster by appropriately amplifying the near-surface wind speed and applying a certain wind direction rotation (Ekman turn) based on the cloud height. The coefficients of this modified model are shared across different power plants and at different times, ensuring spatial consistency and preventing overfitting.

[0099] In this embodiment, the cloud height The irradiance attenuation ratio is estimated using an exponential mapping method, and the specific formula is as follows:

[0100]

[0101] in: The average cloud base height is 2000m in this embodiment; k is the atmospheric attenuation coefficient, which is 1.0 in this embodiment. The calculated irradiance attenuation ratio; This represents the natural exponential function. This formula dynamically reflects the strength of cloud cover based on the degree of irradiance attenuation: the lower the irradiance attenuation ratio, the stronger the cloud cover, and the lower the estimated cloud height; the higher the irradiance attenuation ratio, the higher the atmospheric transparency, and the higher the corresponding cloud height indication value.

[0102] Then, the estimated Normalized to the [0,1] interval and used as input parameters for the Ekman wind field correction model, it is used to adjust the degree of wind direction deflection, thereby improving the ability of the dynamic graph structure to describe the physical consistency of cloud propagation.

[0103] Directed edge weights are constructed based on the corrected wind field. When the cloud propagation direction is consistent with the power station connection direction, a directed connection is established from power station i to power station j. Its original edge weights... The calculation is as follows:

[0104]

[0105] in: This is the wind speed modulation coefficient; The direction sensitivity coefficient; This is the corrected wind speed. This is the normalized wind direction vector; This is the unit vector representing the direction of the power station connection.

[0106] By incorporating causality constraints and constructing an initial adjacency matrix, an indicator function is introduced to ensure the causality of propagation.

[0107] 1[·] is an indicator function. When the expression is true, it takes the value of 1, establishing a directed connection from photovoltaic power station i to j; when the expression is false, it takes the value of 0, suppressing the reverse edge to ensure the causality of cloud propagation and ensuring that information flows only from upstream to downstream.

[0108] Finally, construct the initial adjacency matrix. Then, normalization and self-loop processing are performed:

[0109]

[0110]

[0111] Where, ω b =exp(wb) represents the baseline connectivity strength, ω s Let I be the self-loop weight, I be the identity matrix, and ReLU be the linear rectified activation function. Softmax is the normalization function; the Softmax operation is performed along the row dimension to ensure that the sum of the outgoing edge weights of each node is 1, forming a probabilistic routing matrix. Finally, the adjacency matrix is ​​smoothed using an exponential moving average to suppress high-frequency jitter.

[0112] Step S4: Use the dynamic adjacency matrix to perform spatial association modeling on the feature data of multiple power stations.

[0113] This step uses multi-order directed graph convolution to extract spatial correlation features between power stations. For the input feature tensor... (Where B is the batch size, L is the backtracking window length, N is the number of power stations, and F is the feature dimension), in this embodiment, the batch size B is set to 32, the backtracking window length L is set to 12 (corresponding to 3 hours of time-series data), the optimizer is the Adam optimizer, and the learning rate is set to 0.001. First, it is mapped to the hidden layer dimension d through a linear compression layer:

[0114]

[0115] Subsequently, multi-order propagation is performed using the dynamic adjacency matrix At constructed in step S3. In this embodiment, considering the locality and short-term nature of cloud propagation, only first-order and second-order propagation are used to avoid the oversmoothing problem caused by higher-order propagation.

[0116]

[0117] in, Let be the trainable weight matrix, and σ be the ReLU activation function.

[0118] Figure 7 The results of sensitivity analysis on the impact of different propagation orders K on prediction performance (MAE) are presented. The performance is optimal when K=2. When K>2, the error increases due to excessive smoothing. Therefore, K=2 is selected in this embodiment.

[0119] To prevent numerical amplification due to repeated multiplication by At, the adjacency matrix is ​​truncated after each propagation to ensure its value is no greater than 1. Finally, the first-order and second-order features are concatenated to obtain the final spatial feature representation:

[0120]

[0121] Step S5: The spatial features are fused with the temporal features and time series observation features, and then input into the time series prediction model for time series modeling.

[0122] The spatial features Zspatial extracted in step S4 are concatenated with the temporal information constructed in step S1 to construct the enhanced input sequence:

[0123]

[0124] The enhanced sequence is fed into a lightweight time-series mixing network (TSMixer) for temporal modeling. TSMixer efficiently captures long-term dependencies in the sequence by alternately performing MLP mixing of the time and feature dimensions through stacked Mixer layers. Its structure includes layer normalization, residual connections, and channel mixing, enabling effective extraction of temporal dynamic features while maintaining low computational complexity.

[0125] Step S6: As Figure 6 As shown, during typical cloud transit periods, the prediction results of the dynamic graph method of this invention are closer to the measured values ​​than those of the static graph method, and can effectively capture the rapid dynamic changes in GHI. It outputs the photovoltaic irradiance prediction results for multiple future time steps. The output of TSMixer is mapped to the target prediction length H through a linear mapping head:

[0126]

[0127] Here This is the predicted irradiance attenuation ratio sequence for the next H time steps. After obtaining the attenuation ratio prediction results, multiplying them by the clear-sky irradiance at the corresponding time step yields the absolute irradiance prediction value: ;

[0128] In this embodiment, the prediction step size H is set to 4, which represents the irradiance for the next hour (4 × 15 minutes). The model is trained end-to-end, and the loss function is the mean absolute error (MAE) between the predicted attenuation ratio and the actual attenuation ratio.

[0129]

[0130] Experimental results show that, under complex meteorological conditions such as the passage of dynamic clouds, this method has higher prediction accuracy and physical interpretability compared to models using only recurrent neural networks or static graph neural networks. Validation using Himawari-8 satellite cloud images demonstrates that the dynamic graph structure inferred by the model is highly consistent with the actual cloud propagation direction, proving the effectiveness and reliability of this method.

[0131] 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.

[0132] Table 1 shows the feature variables that can be referenced.

[0133]

[0134] In constructing the multi-site time-series dataset, this embodiment uses six actually operating photovoltaic power stations in Jiangsu Province as examples, spanning from June 2024 to June 2025, with a data sampling resolution of 15 minutes, acquiring approximately 35,000 sets of synchronous samples. To more intuitively demonstrate the data format, Table 2 below shows a portion of the preprocessed data samples from the Ligang power station on a typical day.

[0135] Table 2 Data Illustration Table

[0136]

[0137] To verify the effectiveness of each key module in this invention, an ablation experiment was designed in this embodiment. The dynamic adjacency matrix was removed (replaced with a static geographic map) and physical information guidance (i.e., pure data-driven graph learning without physical constraints) respectively. The results are shown in Table 3 below. The experimental results show that both the dynamic graph structure and physical information guidance significantly contribute to the model accuracy.

[0138] Table 3 Ablation Experiment Results

[0139]

[0140] Furthermore, to comprehensively evaluate the predictive performance of the proposed method, it was compared with several representative baseline models, including recurrent neural networks (GRU), convolutional long short-term memory networks (CNN-LSTM), linear models (DLinear), and Transformer-based models (Informer). All models were trained and tested under the same dataset and experimental settings, and the comparison results are shown in Table 4 below. The results show that the proposed PIDG-TSMixer model achieves the best results on all evaluation metrics, verifying its superiority in the ultra-short-term photovoltaic irradiance prediction task.

[0141] Table 4. Model Comparison Results

[0142]

[0143] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

[0144] 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 method for ultra-short-term forecasting of photovoltaic power plants based on physical information-driven dynamic graphs, characterized in that, Includes the following steps: S1. Obtain irradiance, wind speed, wind direction, geographical location and time information of multiple photovoltaic power stations, and construct a multi-site time series dataset; S2. Calculate the irradiance attenuation ratio based on the measured irradiance and the corresponding clear sky irradiance, and construct the input feature sequence; S3. Combining geographical location and meteorological variables, a dynamic adjacency matrix is ​​constructed based on a wind field correction model incorporating Ekman layer theory, and cloud propagation causality constraints are added to ensure that information propagates unidirectionally from upstream photovoltaic power station to downstream photovoltaic power station. The meteorological variables are wind speed and wind direction. S4. Use dynamic adjacency matrix to model spatial associations and extract spatial association features between power stations; S5. Integrate spatial features, periodic temporal features, and time series observation features, and input them into the time series prediction model to complete the modeling; S6. Output the future multi-timestep irradiance prediction results; The irradiance attenuation ratio in step S2 is calculated as follows: ； in: This represents the irradiance attenuation ratio of the i-th photovoltaic power station at time t; This represents the measured irradiance of the i-th photovoltaic power station; The clear sky irradiance at time t, the The results were obtained by combining the solar altitude angle, atmospheric turbidity, power station altitude parameters, and the Ineichen model from the pvlib library.

2. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 1, characterized in that, The dynamic adjacency matrix is ​​constructed based on the geographic vector, wind speed, and wind direction of the photovoltaic power station's geographical location, and the directed connection relationship between photovoltaic power stations is determined by the cloud propagation direction. Let the geographical vector between power station i and power station j be . Its unit direction vector is When the direction of cloud propagation is... When they are consistent, establish a directed connection from i to j; where, Indicates the geographical coordinates of power station i. Indicates power station Geographic coordinates It represents the Euclidean distance between two points.

3. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 2, characterized in that, The meteorological variables in step S3 are processed by a wind field correction model incorporating Ekman layer theory to obtain the cloud propagation direction: ； in: Let be the measured wind speed of the i-th power station at time t; Let be the measured wind direction of the i-th power station at time t; Where is the cloud height; E(⋅) is the wind field correction function; the wind field correction function E(⋅) is implemented based on the Ekman layer theory; Wind field correction includes wind speed amplification correction and wind direction rotation correction. The formula for wind speed amplification correction is: ； in, Let α be the wind speed after amplification and correction at time t. E The value is the Ekman amplification factor, ranging from 0.1 to 0.3, used to compensate for the near-surface wind speed attenuation caused by surface friction; The formula for wind direction rotation correction is: ； in, Let t be the wind direction after rotation correction, β be the turning coefficient with a value range of 10°-30°, and H(t) be the cloud height at time t of the i-th power station after inversion of the irradiance attenuation ratio and normalization to the interval [0,1].

4. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 3, characterized in that, In the dynamic adjacency matrix, the edge weights pointing from power station i to power station j are calculated as follows: ； in: This is the wind speed modulation coefficient; The direction sensitivity coefficient; This is the corrected wind speed. is the normalized wind direction vector, and exp(·) is the natural exponential function used to calculate the direction similarity weight.

5. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 3, characterized in that, The initial adjacency matrix is ​​normalized as follows to obtain the final dynamic graph structure: ； in, This is the initial adjacency matrix; It is the identity matrix; For self-loop weights; ReLU is the linear rectified activation function; Softmax is the normalization function; The normalized dynamic adjacency matrix is ​​smoothed by an exponential moving average, as shown in the following formula: ； in, Let be the final dynamic adjacency matrix after smoothing at time t; This is the adjacency matrix after Softmax and self-loop processing; This is the smoothing coefficient.

6. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 1, characterized in that, In step S1, the time information is processed by sine and cosine encoding: the hour is encoded as sin(2πt / 24) and cos(2πt / 24); the month is encoded as sin(2πt / 12) and cos(2πt / 12). The encoded data is then concatenated with the time series observation features as the model input.

7. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 3, characterized in that, cloud height The following is derived from the inversion of the irradiance attenuation ratio: The irradiance attenuation ratio characterizes atmospheric transparency, and is converted into a cloud height indicator value through exponential mapping. The irradiance attenuation ratio is then used as a parameter of atmospheric transparency. Through the formula: ； Where H0 is the local average annual cloud base height, taken as 2000m; k is the atmospheric attenuation coefficient, taken as 1.0; exp represents the natural exponential function, used to describe the exponential relationship of cloud attenuation effect, and is used to estimate the cloud height. After being normalized to the [0,1] interval, it is used as the cloud height input parameter in the wind field correction model and participates in the calculation of the cloud propagation direction in the Ekman wind field correction model.

8. The method for ultra-short-term forecasting of photovoltaic power plants based on physical information-driven dynamic graphs according to claim 1, characterized in that, ... There are N photovoltaic power stations. The measured irradiance sequence data at time t can be expressed as: ； in: This represents the measured irradiance of the i-th photovoltaic power station at time t. Let N represent an N-dimensional real vector space, where the vector dimension is the same as the total number of photovoltaic power plants, N.

9. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 1, characterized in that, The irradiance attenuation ratio is truncated to a preset range: ； in: and The upper and lower bounds are preset; clip(·) is a numerical truncation function used to limit the irradiance attenuation ratio within a reasonable range of [0, 1.2].

10. The method for ultra-short-term forecasting of photovoltaic power plants based on physically driven dynamic graphs according to claim 2, characterized in that, Ekman wind field correction is used to compensate for wind speed attenuation and wind direction deflection caused by surface friction. The specific calculation steps are as follows: Coriolis parameter calculation: ； Where ω is the angular velocity of Earth's rotation. The latitude of the photovoltaic power station.

11. The method for ultra-short-term forecasting of photovoltaic power plants based on physically driven dynamic graphs according to claim 5, characterized in that, In step S4, the spatial correlation modeling of multi-power station feature data is performed using a graph convolutional network. The calculation method for the k-th order propagation feature is as follows: ； in This represents the propagation characteristics of the k-th order. It is a dynamic adjacency matrix; The weight matrix is ​​trainable. This is the activation function.

12. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 11, characterized in that, First-order and second-order propagation are used, and the final spatial representation is obtained through feature concatenation: ； in, and These are the first-order and second-order graph convolution propagation features, respectively; Concat is the feature concatenation operation.

13. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 12, characterized in that, The irradiance prediction results for multiple future time steps are output through a linear mapping layer: ； Where: h is the prediction step size. To predict irradiance.

14. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 2, characterized in that, After obtaining the time information in step S1, for the time series data point x t If data at time m is missing, its missing value x m Calculated using the following formula: ； 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 value at time m+1 after the missing data.

15. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 5, characterized in that, For irradiance data, the Min-Max normalization formula is used: ； 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.

16. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 5, characterized in that, To ensure the causality of propagation, an initial adjacency matrix is ​​constructed by incorporating causal constraints. Introducing indicator functions: ; Where 1[·] is an indicator function, which takes the value of 1 when the expression is true, establishing a directed connection from photovoltaic power station i to j; when the expression is false, it takes the value of 0, suppressing the reverse edge to ensure the causality of cloud propagation.

17. The method for ultra-short-term forecasting of photovoltaic power plants based on physically driven dynamic graphs according to claim 2, characterized in that, The spatial features extracted in step S4 are concatenated with the temporal encoded features constructed in step S1 to construct the enhanced input sequence: ； in, for The dimensional tensor space represents the dimensionality type of the fused and enhanced input sequence; B is the batch size; L is the temporal backtracking window length, the number of steps of historical time series data input to the model. To enhance the overall dimensionality of features, spatial features With time characteristics The total feature dimension after splicing.

18. The ultra-short-term forecasting method for photovoltaic power plants based on physically driven dynamic graphs according to claim 1, characterized in that, The time series prediction model in step S5 uses a lightweight temporal hybrid network, which mixes the time dimension and feature dimension alternately, and outputs the prediction results by combining layer normalization and residual connections.

19. A photovoltaic power plant ultra-short-term forecasting system based on a physical information-driven dynamic graph, characterized in that, Applied to the method of any one of claims 1-18, comprising: The data acquisition and preprocessing module is used to acquire irradiance, wind speed, wind direction, geographical location and time information of multiple photovoltaic power stations to construct a multi-site time series dataset; The feature construction module is used to calculate the irradiance attenuation ratio based on the measured irradiance and the corresponding clear sky irradiance, and to construct the input feature sequence. The dynamic graph construction module is used to combine geographical location and meteorological variables to construct a dynamic adjacency matrix based on a wind field correction model that includes Ekman layer theory, and to add cloud propagation causality constraints to ensure that information only propagates unidirectionally from upstream photovoltaic power plants to downstream photovoltaic power plants. The meteorological variables are wind speed and wind direction. The spatial feature extraction module uses a dynamic adjacency matrix to model spatial associations and extract spatial association features between power stations. The time series modeling module is used to integrate spatial features with periodic temporal features and time series observation features, and input them into the time series prediction model to complete the modeling. The prediction output module is used to output the irradiance prediction results for future multi-time steps. The irradiance attenuation ratio in step S2 is calculated as follows: ； in: This represents the irradiance attenuation ratio of the i-th power station at time t; This represents the measured irradiance of the i-th power station; The clear sky irradiance at time t, the The results were obtained by combining the solar altitude angle, atmospheric turbidity, power station altitude parameters, and the Ineichen model from the pvlib library.

20. A photovoltaic irradiance prediction device, characterized in that, The device includes a processor and a memory, the memory storing a computer program, the computer program including instructions; when the instructions are executed by the processor, the photovoltaic irradiance prediction device causes the photovoltaic irradiance prediction device to perform the method according to any one of claims 1-18.

21. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method according to any one of claims 1-18.

Images