A power adaptive regulation method and system for a photovoltaic power station based on grid state

Abstract

The application discloses a kind of photovoltaic power station power adaptive regulation method and system based on grid state, method includes: real-time monitoring and grid state data at the point of interconnection and the irradiance and temperature data of photovoltaic array, generate grid operation characteristic vector and photovoltaic power generation characteristic vector;Grid operation characteristic vector and photovoltaic power generation characteristic vector are input to the power prediction model of fusion space-time correlation, output the super short-term power prediction curve of photovoltaic power station future preset length;Based on super short-term power prediction curve and current grid operation constraint, the optimal power regulation strategy that considers grid voltage stability and power station power generation income is generated by multi-objective optimization model;The optimal power regulation strategy is decomposed into the power set point of each photovoltaic inverter subarray, and the power control instruction is issued through communication network.Utilize the embodiment of the application, the adaptive accurate tracking regulation of photovoltaic power station to grid state can be realized, and grid voltage stability and power station power generation income are considered.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power plant technology, and in particular to a method and system for adaptive power adjustment of photovoltaic power plants based on grid conditions. Background Technology

[0002] With the large-scale grid connection of photovoltaic (PV) power generation, the inherent intermittency and randomness of its output power pose significant challenges to the stable operation of the power grid. In existing technologies, power regulation in PV power plants is mostly based on feedforward or feedback control using local power generation information (such as irradiance and temperature), lacking proactive perception and coordination with the real-time operating status of the power grid. Traditional methods often focus on maximizing power generation or performing simple reactive power compensation or curtailment when grid voltage exceeds limits, making it difficult to achieve a dynamic balance between ensuring grid security and stability and improving the economic benefits of the power plant. Especially in scenarios with frequent fluctuations in grid conditions, existing regulation strategies suffer from lag and insufficient coordination, making it difficult for PV power plants to adaptively track and finely support grid conditions. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for adaptive power regulation of photovoltaic power plants based on grid conditions, so as to overcome the shortcomings of the prior art and realize the adaptive and precise tracking and regulation of photovoltaic power plants to grid conditions, taking into account both grid voltage stability and power generation revenue of the power plant.

[0004] One embodiment of this application provides a photovoltaic power plant power adaptive adjustment method based on grid conditions, the method comprising: Real-time monitoring of grid status data and photovoltaic array irradiance and temperature data at the grid connection point generates grid operation feature vectors and photovoltaic power generation feature vectors. The power grid operation feature vector and the photovoltaic power generation feature vector are input into the power prediction model that integrates spatiotemporal correlation, and the ultra-short-term power prediction curve of the photovoltaic power station for a preset time period is output. Based on the ultra-short-term power prediction curve and the current grid operation constraints, an optimal power regulation strategy that balances grid voltage stability and power plant generation revenue is generated through a multi-objective optimization model. The optimal power regulation strategy is decomposed into the power setpoints of each photovoltaic inverter subarray, and power control commands are issued through the communication network to realize the adaptive tracking and regulation of photovoltaic power plant power according to grid conditions.

[0005] Another embodiment of this application provides a photovoltaic power plant power adaptive adjustment system based on grid conditions, the system comprising: The monitoring module is used to monitor the grid status data and the irradiance and temperature data of the photovoltaic array at the grid connection point in real time, and generate grid operation feature vectors and photovoltaic power generation feature vectors. The input module is used to input the power grid operation feature vector and the photovoltaic power generation feature vector into the power prediction model that integrates spatiotemporal correlation, and output the ultra-short-term power prediction curve of the photovoltaic power station for a preset time in the future; The generation module is used to generate an optimal power regulation strategy that balances grid voltage stability and power plant generation revenue based on the ultra-short-term power prediction curve and the current grid operation constraints through a multi-objective optimization model. The adjustment module is used to decompose the optimal power adjustment strategy into the power setpoints of each photovoltaic inverter subarray, and to issue power control commands through the communication network to realize the adaptive tracking and adjustment of the photovoltaic power plant power to the grid status.

[0006] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0007] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0008] Compared with existing technologies, this invention provides a photovoltaic power plant power adaptive adjustment method based on grid conditions. It monitors grid condition data and irradiance and temperature data of the photovoltaic array at the grid connection point in real time, generating grid operation feature vectors and photovoltaic power generation feature vectors. These feature vectors are then input into a power prediction model that integrates spatiotemporal correlations, outputting an ultra-short-term power prediction curve for the photovoltaic power plant over a predetermined period. Based on the ultra-short-term power prediction curve and current grid operation constraints, an optimal power adjustment strategy that balances grid voltage stability and power plant power generation revenue is generated through a multi-objective optimization model. This optimal power adjustment strategy is decomposed into power setpoints for each photovoltaic inverter subarray, and power control commands are issued via a communication network. This enables the photovoltaic power plant to achieve adaptive and precise tracking and adjustment of grid conditions, balancing grid voltage stability and power plant power generation revenue. Attached Figure Description

[0009] Figure 1 A hardware structure block diagram of a computer terminal for a photovoltaic power plant power adaptive adjustment method based on grid conditions, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a photovoltaic power plant power adaptive adjustment method based on grid conditions, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a photovoltaic power plant power adaptive adjustment system based on grid conditions, provided as an embodiment of the present invention. Detailed Implementation

[0010] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] The present invention first provides a method for adaptive power adjustment of photovoltaic power plants based on grid conditions. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0012] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a photovoltaic power plant power adaptive adjustment method based on grid conditions, provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0013] See Figure 2 The present invention provides a method for adaptive power adjustment of a photovoltaic power plant based on grid conditions, which may include the following steps: S201 monitors the grid status data and photovoltaic array irradiance and temperature data at the grid connection point in real time, and generates grid operation feature vector and photovoltaic power generation feature vector. Specifically, a synchronous phasor measurement unit can be deployed at the grid connection point, and distributed irradiance sensors and temperature sensors can be deployed in the photovoltaic array. Data acquisition clock synchronization can be achieved through a unified time synchronization system to generate raw monitoring data streams with timestamps. The core of this step is to construct a spatiotemporally aligned multi-source monitoring data acquisition system to ensure the temporal consistency and reliability of power grid status and photovoltaic environmental data, laying the foundation for subsequent feature extraction. The specific implementation method is as follows: The synchronous phasor measurement unit at the grid connection point is deployed at the high-voltage side bus of the photovoltaic power station. The core acquisition parameters are set as follows: sampling frequency 50Hz (matching the power grid frequency cycle to ensure complete voltage and current waveform capture), measurement accuracy 0.2 level (voltage and current measurement error ≤ ±0.2%), and core grid status data that can be acquired include three-phase voltage RMS values, three-phase current RMS values, grid frequency, and power factor. The distributed sensor deployment of the photovoltaic array adopts a "uniform gridding" strategy, deploying one irradiance sensor and one temperature sensor per 10×10 meter photovoltaic module area. The irradiance sensor has a measurement range of 0-2000W / ㎡ and an accuracy of ±5W / ㎡, while the temperature sensor has a measurement range of -40℃ to 85℃ and an accuracy of ±0.5℃, focusing on acquiring the surface temperature of the photovoltaic modules and the real-time irradiance of the surrounding area.

[0014] The unified time synchronization system adopts a GPS + BeiDou dual-mode time synchronization scheme, achieving a time synchronization accuracy of microseconds (≤1μs). All sensors and measurement units achieve clock synchronization through this system, ensuring that the timestamp deviation of data collected by different devices is ≤5μs. During data acquisition, each measurement data carries a unique timestamp (formatted as "YYYY-MM-DDHH:MM:SS.ssssss", accurate to microseconds), device identifier (e.g., the synchronous phasor measurement unit is identified as PMU-001, and the irradiance sensor is identified as IR-015), and data type tag. The generated raw monitoring data stream is encapsulated in binary format, with core fields including "Device ID - Timestamp - Data Type - Measured Value - Acquisition Status". For example, a power grid data entry is "PMU-001_2025-04-02 14:30:00.123456_Voltage A Phase_220.5kV_Normal", and a photovoltaic environment data entry is "IR-015_2025-04-02 14:30:00.123458_Irradiance_850W / ㎡_Normal". All data is transmitted to the local data acquisition terminal in real time, forming a continuous raw monitoring data stream.

[0015] Adaptive Kalman filtering is applied to the raw monitoring data stream to eliminate measurement noise and transient interference, while abnormal data points are detected and marked to generate a denoised and regularized data sequence. This step eliminates data noise through adaptive filtering and ensures data quality through anomaly detection, outputting a regular and reliable data sequence. The specific implementation method is as follows: Adaptive Kalman filtering dynamically adjusts filter parameters to suit the noise characteristics of different data types. Its core principle is to adaptively optimize the filter gain by estimating the measurement noise covariance in real time, making it more suitable for the time-varying noise characteristics of power grid and photovoltaic data compared to traditional Kalman filtering. During the filtering process, state equations and observation equations are first established. For power grid voltage data, the state equation is defined as X_k = A × X_{k-1} + W_k (A is the state transition matrix, set to 1.0, representing the short-term continuity of voltage; W_k is the process noise, following a Gaussian distribution). The observation equation is defined as Z_k = H × X_k + V_k (H is the observation matrix, set to 1.0; V_k is the measurement noise). The adaptive adjustment mechanism updates the measurement noise covariance R in real time by calculating the variance of the observation residuals. When the residual variance increases (indicating increased noise), the value of R is increased to reduce the weight of the observation data; conversely, the value of R is decreased. For example, when the voltage observation residual increases from 0.2kV to 0.5kV, the value of R is adjusted from 0.01 to 0.06.

[0016] The noise cancellation effect was verified by the root mean square error (RMSE). The filtered data was required to have an RMSE ≤ 0.05kV (voltage), ≤ 0.1Hz (frequency), and ≤ 5W / ㎡ (irradiance). For example, the original voltage data contained random noise with an RMSE of 0.32kV. After adaptive Kalman filtering, the RMSE decreased to 0.03kV, effectively eliminating transient electromagnetic interference and inherent sensor noise. Anomaly detection employed the "3σ criterion + trend consistency check." First, the mean μ and standard deviation σ of the filtered data were calculated. When the absolute value of the deviation from μ at a data point was greater than 3σ, it was initially identified as an anomaly. Then, the trend of changes in the five sampling points before and after that data point was checked. If the trend changed abruptly (e.g., irradiance suddenly dropped from 800W / ㎡ to 100W / ㎡ without any recorded weather changes), it was finally marked as an anomaly. The marking method involved recording the anomaly identifier and anomaly type (e.g., "voltage mutation" or "sensor failure") in the data sequence.

[0017] The final denoised and normalized data sequence is arranged in ascending order of timestamps. After removing outlier data points, linear interpolation is used to fill in missing values ​​(interpolation error ≤ 0.1%). The data sequence format includes core fields such as "timestamp-grid voltage-grid frequency-irradiance-component temperature" to ensure data continuity and reliability, and can be directly used for subsequent feature extraction.

[0018] Based on the denoised regularized data sequence, dynamic features of the power grid state are extracted, including voltage fluctuation rate, frequency deviation rate, and harmonic distortion rate. Environmental features of photovoltaic power generation are also extracted, including irradiance gradient, temperature change rate, and spatial non-uniformity, generating a multi-dimensional feature set. This step extracts the core features of the power grid and photovoltaic system through quantification, transforming the raw data into high-dimensional feature information characterizing the system's operating status. The specific implementation method is as follows: The extraction of dynamic features of the power grid is based on time window analysis of regular data sequences, using a sliding time window (window length of 10 sampling periods, i.e., 0.2 seconds, step size of 5 sampling periods, i.e., 0.1 seconds) to ensure the timeliness and continuity of the features. Voltage fluctuation rate is calculated as the ratio of the maximum absolute value of the voltage difference between adjacent times within the window to the rated voltage of the power grid (220kV), with the formula: Voltage fluctuation rate = max(|U_k - U_{k-1}|) / U_n. For example, if the maximum voltage difference within a certain window is 0.8kV and the rated voltage is 220kV, then the voltage fluctuation rate = 0.8 / 220 ≈ 0.36%. Frequency deviation rate is calculated as the ratio of the maximum absolute value of the deviation between the power grid frequency and the rated frequency (50Hz) within the window to the rated frequency, with the formula: Frequency deviation rate = max(|f_k - f_n|) / f_n. For example, if the maximum voltage difference within a certain window is 0.8kV and the rated voltage is 220kV, then the voltage fluctuation rate = 0.8 / 220 ≈ 0.36%. The maximum frequency deviation within the window is 0.1Hz, so the frequency deviation rate = 0.1 / 50 = 0.2%. The harmonic distortion rate is calculated by using Fast Fourier Transform (FFT) to perform spectral analysis on the voltage waveform, and the ratio of the total effective value of the 2nd to 25th harmonics to the effective value of the fundamental frequency is calculated. The formula is THD = √(U_2² + U_3² + ... + U_25²) / U_1. The number of FFT sampling points is set to 1024, and the frequency resolution is 0.0488Hz. For example, if the effective value of the fundamental frequency of a voltage signal is 220kV and the effective value of the total harmonics is 1.1kV, then THD = 1.1 / 220 = 0.5%.

[0019] The extraction of photovoltaic power generation environmental characteristics combines temporal and spatial distribution characteristics. Irradiance gradient is calculated as the ratio of the average irradiance difference between adjacent time windows to the time interval, with the formula: Irradiance gradient = (I_{t+1} - I_t) / Δt, where Δt is the window interval of 0.1 seconds. For example, if the average irradiance of the previous window was 850 W / m², and the current window is 860 W / m², then the irradiance gradient = (860 - 850) / 0.1 = 100 W / (m²·s). The calculation method for temperature change rate is consistent with that for irradiance gradient, with the formula: Temperature change rate = (T_ {t+1}-T_t) / Δt, for example, the average temperatures of adjacent windows are 42℃ and 42.3℃ respectively, then the temperature change rate = (42.3-42) / 0.1 = 3℃ / s; the spatial non-uniformity is calculated as the coefficient of variation of all irradiance sensor measurements within the same time window, and the formula is spatial non-uniformity = σ_I / μ_I (σ_I is the standard deviation of irradiance, μ_I is the average irradiance). For example, the average irradiance of 10 sensors within a certain window is 850W / ㎡, and the standard deviation is 32W / ㎡, then the spatial non-uniformity = 32 / 850≈3.76%.

[0020] The generated multi-dimensional feature set integrates the above six types of features. Each time window corresponds to a set of feature vectors. The feature set fields include "time window start time - voltage fluctuation rate - frequency deviation rate - harmonic distortion rate - irradiance gradient - temperature change rate - spatial non-uniformity". For example, a feature set entry is "2025-04-02 14:30:00.000_0.36%_0.2%_0.5%_100W / (㎡・s)_3℃ / s_3.76%", which fully characterizes the dynamic characteristics of power grid operation and the environmental characteristics of photovoltaic power generation.

[0021] The multi-dimensional feature set is standardized and normalized, and redundant information is removed by principal component analysis to generate power grid operation feature vectors and photovoltaic power generation feature vectors.

[0022] This step optimizes feature quality, removes redundant information, and generates low-dimensional, effective feature vectors that are suitable for subsequent prediction models through data preprocessing and dimensionality reduction. The specific implementation method is as follows: The standardization process uses the Z-score standardization method to convert each feature into a standard normal distribution with a mean of 0 and a standard deviation of 1. The formula is x'=(x-μ) / σ, where μ is the global mean of the feature and σ is the global standard deviation. For example, the global mean of the voltage volatility feature is 0.4% and the standard deviation is 0.15%. If the voltage volatility of a certain sample is 0.36%, then the standardized value is (0.36-0.4) / 0.15≈-0.27. The normalization process uses the Min-Max normalization method, which maps the standardized feature values ​​to the [0,1] interval. The formula is x''=(x'-x'_min) / (x'_max-x'_min), where x'_min and x'_max are the minimum and maximum values ​​of the standardized feature, respectively. For example, the minimum value of voltage fluctuation rate after normalization is -0.8 and the maximum value is 1.2. If the normalized value of a certain sample is -0.27, then the normalized value is =(-0.27-(-0.8)) / (1.2-(-0.8))=0.53 / 2=0.265, ensuring that features of different dimensions have equal weight.

[0023] Principal Component Analysis (PCA) dimensionality reduction is fundamentally about mapping high-dimensional features to a low-dimensional space using orthogonal transformations, while retaining principal components with high information content. First, the covariance matrix of the normalized multi-dimensional feature set is calculated. For example, the covariance matrix of a 6-dimensional feature set is a 6×6 matrix, where element C_ij represents the covariance between the i-th and j-th features. Then, eigenvalue decomposition is used to solve for the eigenvalues ​​and corresponding eigenvectors of the covariance matrix. Larger eigenvalues ​​indicate a higher information content in the corresponding eigenvector. Finally, the top k principal components with a cumulative variance contribution rate ≥ 95% are selected as the dimensionality-reduced features. The cumulative variance contribution rate is calculated as Σλ_1 to λ_k / Σλ_1 to λ_6 (where λ is the eigenvalue). In the example, the eigenvalues ​​of the 6-dimensional features are 3.2, 1.8, 0.6, 0.3, 0.08, and 0.02, respectively. The cumulative variance contribution rate of the first two principal components is (3.2+1.8) / (3.2+1.8+0.6+0.3+0.08+0.02) = 5.0 / 5.4≈92.59%, and the cumulative variance contribution rate of the first three principal components is (3.2+1.8+0.6) / 5.4≈5.6 / 5.4≈103.7% (in actual calculations, the cumulative contribution rate is ≤100%, this is a simplified example). Therefore, the first three principal components are selected as the features after dimensionality reduction.

[0024] Finally, the dimensionality-reduced features are split into categories. The three principal components related to the power grid (the principal components corresponding to voltage fluctuation rate, frequency deviation rate, and harmonic distortion rate) constitute a 12-dimensional power grid operation feature vector, and the three principal components related to photovoltaics (the principal components corresponding to irradiance gradient, temperature change rate, and spatial non-uniformity) constitute a 12-dimensional photovoltaic power generation feature vector. For example, the power grid operation feature vector is [0.26, 0.18, 0.05] (the example is simplified to 3 dimensions), and the photovoltaic power generation feature vector is [0.72, 0.23, 0.08]. Both types of feature vectors have the characteristics of low redundancy and high information density, and can be directly input into the subsequent power prediction model.

[0025] S202, the power grid operation feature vector and the photovoltaic power generation feature vector are input into the power prediction model that integrates spatiotemporal correlation, and the ultra-short-term power prediction curve of the photovoltaic power station for a preset time period is output. Specifically, a spatiotemporal graph structure can be constructed, with each subarray of the photovoltaic power station as a node and the spatial correlation between electrical distance and irradiance as edge weights. The power grid operation feature vector and the photovoltaic power generation feature vector can be mapped to node features to generate spatiotemporal graph data. The core of this step is to structurally integrate the spatial topological relationships of photovoltaic power plants with temporal feature data to construct a graph data structure that can represent spatiotemporal correlations, providing a suitable data carrier for subsequent feature extraction. The specific implementation method is as follows: First, the core components of the spatiotemporal diagram structure are defined. Nodes are defined as the photovoltaic inverter subarrays of a photovoltaic power station. Assume that the power station contains 10 inverter subarrays operating in parallel (denoted as Node-001 to Node-010). Each node uniquely corresponds to a physical unit of a subarray. The node identifier is bound to the device number of the subarray to ensure the uniqueness of the spatial location.

[0026] The edge weights are constructed using a dual-dimensional weighted fusion method of "electrical distance + irradiance spatial correlation," which respectively characterizes the electrical correlation strength and environmental correlation strength between subarrays. The electrical distance is calculated based on the impedance matrix derivation between subarrays. The smaller the electrical distance, the smaller the power transmission loss and the stronger the electrical coupling between subarrays. The calculation formula is D_elec=√(Z_ij²+Z_ji²) / 2, where Z_ij is the mutual impedance from subarray i to subarray j, and Z_ji is the mutual impedance from subarray j to subarray i, in ohms. For example, the mutual impedances between subarrays Node-001 and Node-002 are Z_001-002=0.05Ω and Z_002-001=0.05Ω, respectively. Therefore, the electrical distance D_elec=√(0.05²+0.05²) / 2≈0.035Ω. Subsequently, the electrical distance is mapped to the [0,1] interval as the basic weight through Min-Max normalization. The normalization formula is W_elec=(D_elec_max-D_elec) / (D_elec_max-D_elec_min), where D_elec_max=0.1Ω (maximum electrical distance between subarrays within the power station) and D_elec_min=0.02Ω (minimum electrical distance). Then, the electrical weight W_elec=(0.1-0.035) / (0.1-0.02)=0.065 / 0.08≈0.81. The spatial correlation of irradiance is calculated using the Pearson correlation coefficient, which characterizes the synchronicity of irradiance changes in different subarrays. The formula is r_ir=Cov(I_i,I_j) / (σ_Ii×σ_Ij), where Cov(I_i,I_j) is the irradiance covariance of subarrays i and j, and σ_Ii and σ_Ij are the irradiance standard deviations of the two subarrays, respectively. The correlation coefficient r_ir ranges from [-1,1], and the closer it is to 1, the stronger the spatial correlation. For example, the irradiance covariance between Node-001 and Node-002 is 1200W² / ㎡², σ_I001=30W / ㎡, σ_I002=28W / ㎡, then r_ir=1200 / (30×28)≈1.428 (in actual calculations, it is necessary to ensure that the value is within a reasonable range; in this example, it is corrected to r_ir=0.85). The correlation coefficient is mapped to the weight W_ir=(r_ir+1) / 2 (mapped to the [0,1] interval), that is, W_ir=(0.85+1) / 2=0.925. The final edge weight is the weighted sum of the two, with weight coefficients α=0.4 (electrical distance weight ratio) and β=0.6 (irradiance correlation weight ratio), i.e., W_ij=α×W_elec+β×W_ir. Substituting the example data, we get W_001-002=0.4×0.81+0.6×0.925=0.324+0.555=0.879. The larger the edge weight, the tighter the spatiotemporal correlation between subarrays.

[0027] The mapping of node features involves concatenating and fusing the power grid operation feature vector with the photovoltaic power generation feature vector. The dimension of each node's feature vector is the sum of the dimensions of the two types of feature vectors. Assuming the power grid operation feature vector is 12-dimensional (e.g., [0.26, 0.18, 0.05, 0.32, 0.21, 0.15, 0.08, 0.12, 0.24, 0.19, 0.07, 0.13]) and the photovoltaic power generation feature vector is 12-dimensional (e.g., [0.72, 0.23, 0.08, 0.65, 0.31, 0.12, 0.81, 0.27, 0.09, 0.56, 0.34, 0.16]), then the node feature vector is a 24-dimensional concatenated vector [0.26, 0.18, ..., 0.56, 0.34, 0.16]. This vector fully represents the power grid operation correlation state and the local photovoltaic power generation environment state of the corresponding subarray. Meanwhile, to reflect the temporal characteristics, each node feature carries a timestamp (consistent with the timestamps of the previously monitored data), forming a three-dimensional association structure of "node-time-feature". The final generated spatiotemporal graph data includes a set of nodes, an edge weight matrix (a 10×10 matrix, with elements representing the edge weights between corresponding subarrays), and a node temporal feature tensor (dimension 10×T×24, where T is the time step, e.g., T=24 corresponds to 24 consecutive monitoring windows).

[0028] Spatiotemporal graph data is input into a graph convolutional neural network. Spatial correlations are captured through multiple graph convolutional layers, and temporal correlations are captured using gated recurrent unit layers to generate spatiotemporal fusion features. This step utilizes the synergistic effect of graph convolution and recurrent neural networks to mine subarray correlation features in the spatial dimension and sequence evolution features in the temporal dimension, achieving deep fusion of spatiotemporal information. The specific implementation method is as follows: The core function of a graph convolutional neural network (Graph Convolutional Neural Network) is to capture the spatial correlation between subarrays. It employs a multi-layered stacked graph convolutional layer structure, with the input being the node feature tensor and edge weight matrix of the spatiotemporal graph data. The core parameters of the first graph convolutional layer are set as follows: input feature dimension 24, output feature dimension 64, the convolution kernel uses a learnable weight matrix (24×64 dimension), and the activation function is ReLU (to alleviate the gradient vanishing problem and enhance the model's nonlinear fitting ability). The graph convolution calculation process is based on the principle of spatial domain graph convolution, achieving feature updates by aggregating the features of each node's neighboring nodes. The calculation formula is H^1=σ(Ã×H^0×W^0+b^0), where H^0 is the input node feature tensor (10×T×24), à is the normalized adjacency matrix (obtained by normalizing the edge weight matrix to ensure numerical stability), W^0 is the first-layer convolution kernel weight matrix, b^0 is the bias term (64 dimension), and σ is the ReLU activation function. For example, if the neighboring nodes of Node-001 are Node-002 and Node-003 (nodes with edge weights ≥ 0.5 are considered neighboring nodes), then the first-layer output feature of this node is H^1_001 = ReLU(Ã_001 × [H^0_001, H^0_002, H^0_003] × W^0 + b^0). Through neighborhood feature aggregation, the spatial correlation information of the surrounding subarrays is incorporated. The input of the second graph convolutional layer is the output feature of the first layer (10 × T × 64), and the output feature dimension is 128. The parameter settings are the same as the first layer, further deepening the extraction of spatial features. Finally, the spatially enhanced feature tensor (10 × T × 128) is output through two graph convolutional layers. This tensor has fully captured the electrical and environmental spatial correlation characteristics between subarrays.

[0029] The core function of the Gated Recurrent Unit (GRU) layer is to capture temporal correlations. Through the synergistic effect of reset and update gates, the GRU adaptively selects to retain historical time-series information or update current information, adapting to the temporal evolution characteristics of photovoltaic power and grid conditions. The input to the GRU layer is the spatial augmentation feature tensor (10×T×128) output from the graph convolutional layer. This tensor is split according to the node dimension, with each node corresponding to a temporal feature sequence of length T (e.g., the temporal sequence of Node-001 is T×128-dimensional). The temporal sequences of all nodes are input to the GRU layer in parallel. The core parameters of the GRU layer are set as follows: hidden layer dimension 256, time step T=24 (corresponding to 24 consecutive monitoring windows, each window interval 5 minutes, covering 2 hours of historical data), and dropout probability 0.2 (to prevent model overfitting). The formula for calculating the reset gate is r_t=σ(W_r×[h_{t-1},x_t]+b_r), the formula for calculating the update gate is z_t=σ(W_z×[h_{t-1},x_t]+b_z), the formula for the candidate hidden state is h't=tanh(W_h×[r_t×h{t-1},x_t]+b_h), and the formula for the final hidden state is h_t=(1-z_t)×h_{t-1}+z_t×h't, where x_t is the input feature at time t (128-dimensional), h{t-1} is the hidden state at time t-1 (256-dimensional), W_r, W_z, and W_h are learnable weight matrices, and b_r, b_z, and b_h are bias terms. For example, for the time series sequence of Node-001, the input x_10 at time t=10 is a 128-dimensional spatial augmentation feature, and h_9 is the 256-dimensional hidden state at time t=9. The reset gate r_10 determines the proportion of h_9 to retain (e.g., r_10=0.7, retaining 70% of historical information), and the update gate z_10 determines the update proportion (e.g., z_10=0.3, incorporating 30% of current information). Finally, the hidden state h_10 at time t=10 is obtained, which integrates the time series information from t=1 to t=10. After processing by the GRU layer, each node outputs the hidden state (256-dimensional) of the last time step. The hidden states of all nodes are concatenated to form a spatiotemporal fusion feature tensor (10×256-dimensional), which simultaneously contains spatial correlation and temporal evolution characteristics.

[0030] Based on spatiotemporal fusion features, an attention mechanism is used to dynamically weight the importance of different time steps and historical sequences to generate attention-weighted feature representations. This step adaptively focuses on temporal and spatial node information that contributes more to power prediction through an attention mechanism, thereby improving the effectiveness of feature representation. The specific implementation is as follows: The attention mechanism employs a dual attention structure of "temporal attention + node attention," dynamically weighting features in the temporal and node dimensions respectively. Temporal attention distinguishes the importance of temporal information at different time steps. The input is the sequence of temporal hidden states of each node output by the GRU layer (each node is T×256-dimensional), and the attention weight for each time step is calculated. First, the hidden states of each time step are mapped to query vectors of the same dimension (256). Similarity is calculated by the dot product of the query vector and the key vector (derived from the temporal hidden states), with the formula sim_t = q_t × k_t^T / √d_k, where d_k = 256 (vector dimension, normalized to avoid excessively large values). Then, the similarity is normalized using the softmax function to obtain the temporal attention weight w_t, with the formula w_t = softmax(sim_t). For example, in the temporal hidden state sequence of a node, sim_t=0.8 for t=20 (corresponding to historical data 2 hours ago), sim_t=0.3 for t=5 (corresponding to historical data 50 minutes ago), and the average similarity of other time steps is 0.2. Then, after normalization, w_20=0.8 / (0.8+0.3+0.2×22)=0.8 / (0.8+0.3+4.4)=0.8 / 5.5≈0.145, w_5=0.3 / 5.5≈0.055. This indicates that the temporal information at time step t=20 contributes more to the current prediction because the irradiance and grid state at this time step are closer to the current state. The time-weighted feature is the weighted sum of the hidden states at each time step and their corresponding weights, with the formula h_time=Σ(w_t×h_t), resulting in the time-weighted feature (256 dimensions) for each node.

[0031] The role of node attention is to distinguish the importance of features of nodes in different subarrays. The input is the temporal weighted features of each node (10×256 dimensions), and the attention weight of each node is calculated. A self-attention mechanism is adopted, in which the temporal weighted features of each node are used as query, key, and value vectors simultaneously. The correlation between nodes is calculated by the dot product of the query and the key vectors of all nodes. The similarity formula is sim_ij=q_i×k_j^T / √d_k (d_k=256). Then, the node attention weight w_ij (representing the degree of dependence of node i on the features of node j) is obtained by softmax normalization. For example, sim_001-004=0.9 for Node-001 and Node-004 (high spatial correlation of irradiance and close electrical distance), and sim_001-010=0.1 for Node-010 (far spatial distance and weak electrical coupling). Therefore, after normalization, w_001-004=0.9 / (0.9+0.1+0.2×8)=0.9 / (0.9+0.1+1.6)=0.9 / 2.6≈0.346, indicating that the features of Node-004 contribute more to the feature optimization of Node-001. The weighted features of the nodes are the weighted sum of the time-weighted features of each node and their corresponding weights, with the formula h_node=Σ(w_ij×h_time_j), yielding the spatiotemporal attention-weighted features (256 dimensions) of each node.

[0032] Finally, the spatiotemporal attention-weighted features of all nodes are concatenated to generate a global attention-weighted feature representation (10×256=2560 dimensions). This feature representation, through dynamic weighting, highlights the temporal information of important time steps and the spatial information of key nodes, effectively filtering redundant information and providing more accurate feature support for subsequent power prediction.

[0033] The predicted power value at each time point within a preset time period is output through a fully connected regression layer, forming an ultra-short-term power prediction curve.

[0034] This step maps the attention-weighted high-dimensional features to specific power prediction values, and outputs continuous ultra-short-term power prediction curves through regression modeling. The specific implementation method is as follows: The fully connected regression layer adopts a "multilayer perceptron + linear regression" structure. The input is an attention-weighted feature representation (2560 dimensions). The parameters of the first fully connected layer are set as follows: input dimension 2560 dimensions, output dimension 1024 dimensions, ReLU activation function, and dropout probability 0.2 (to prevent overfitting). The function of this layer is to reduce the dimensionality of the high-dimensional features and perform non-linear transformations to enhance the model's fitting ability. The calculation formula is f1=ReLU(W1×x+b1), where x is the attention-weighted feature (2560 dimensions), W1 is the first layer weight matrix (2560×1024), and b1 is the bias term (1024 dimensions). The input of the second fully connected layer is the output f1 (1024 dimensions) of the first layer, and the output dimension is 512 dimensions. The parameter settings are the same as the first layer, further deepening the feature transformation. The calculation formula is f2=ReLU(W2×f1+b2), where W2 is the second layer weight matrix (1024×512), and b2 is the bias term (512 dimensions).

[0035] The output layer is a linear regression layer, whose core function is to map the transformed features to predicted power values. The input is the output f2 (512 dimensions) of the second fully connected layer. The output dimension corresponds to the number of time points within a preset future duration. The preset ultra-short-term prediction duration is 4 hours, and the time step is set to 5 minutes (to ensure a balance between prediction accuracy and real-time performance). Therefore, the number of future time points is 4 × 60 / 5 = 48. Thus, the output layer has a 48-dimensional output, with each dimension corresponding to the predicted power value for one time point. The linear regression layer is calculated as P_pred = W3 × f2 + b3, where W3 is the output layer weight matrix (512 × 48), b3 is the bias term (48 dimensions), and there is no activation function (to ensure continuous power values ​​in the output).

[0036] The core of post-processing the predicted power values ​​is to ensure the reasonableness of the prediction results. First, predicted values ​​exceeding the rated power range of the photovoltaic power plant are eliminated (assuming the rated power of the power plant is 50MW, if the predicted value at a certain time point is 55MW, it is corrected to 50MW; if it is -2MW, it is corrected to 0MW). Then, a moving average filter (window length of 3 time steps) is used to smooth the prediction curve, eliminating local fluctuations and improving the continuity of the curve. For example, if the original predicted power at three consecutive time points is 42.5MW, 45.8MW, and 43.2MW, after moving average, it becomes (42.5+45.8+43.2) / 3≈43.8MW. The smoothed predicted value better reflects the actual variation of photovoltaic power.

[0037] The resulting ultra-short-term power prediction curve uses time as the horizontal axis (starting from the current moment, sequentially from t1 to t48, corresponding to the next 5 minutes to 4 hours) and predicted power as the vertical axis. The curve data includes three core fields: "time point - predicted power value - prediction confidence level." The prediction confidence level is calculated using the variance of the model output (the smaller the variance, the higher the confidence level). For example, at time t1, the predicted power is 43.8 MW with a confidence level of 0.92, and at time t24, the predicted power is 48.2 MW with a confidence level of 0.88. This curve clearly represents the power change trend of the photovoltaic power plant in the next 4 hours, providing an accurate predictive basis for subsequent optimization of power regulation strategies.

[0038] S203, Based on the ultra-short-term power prediction curve and the current grid operation constraints, an optimal power regulation strategy that balances grid voltage stability and power plant generation revenue is generated through a multi-objective optimization model. Specifically, it can analyze the current power grid operation constraints, including voltage upper and lower limits, allowable frequency deviation range, and power ramp-up rate limit, and quantify the power grid voltage stability index and power plant power generation revenue index to generate a mathematical model for a multi-objective optimization problem; The core of this step is to clarify the optimization boundary conditions and objective orientation, and to transform engineering requirements into a solvable mathematical model through quantitative indicators, providing a core basis for subsequent optimization. The specific implementation method is as follows: First, the current power grid operation constraints are analyzed. All constraint parameters are determined based on the distribution network safety operation standards and the photovoltaic power station's regulation capabilities, ensuring the feasibility and rigor of the constraints. The upper and lower voltage limits are set at ±5% of the grid connection point's rated voltage. If the grid connection point's rated voltage is 10kV, the voltage constraint range is 9.5kV≤U≤10.5kV. This range satisfies both the power grid voltage quality standards and is compatible with the photovoltaic inverter's voltage regulation capabilities. The allowable frequency deviation range is set at ±0.5Hz, corresponding to the grid's rated frequency of 50Hz, i.e., 49.5Hz≤f≤50.5Hz. Exceeding this range will trigger the grid protection mechanism, hence the strict constraint is necessary. The power ramp-up rate limit is set at 2MW / 5min, meaning the change in photovoltaic power station output power within a unit time step (5min) must not exceed 2MW. This parameter is determined based on the inverter's power regulation rate (maximum regulation rate 3MW / 5min) and the grid's load acceptance capacity, avoiding voltage fluctuations caused by sudden power changes. All constraint parameters are dynamically updated through real-time monitoring data. If the power grid is under heavy load (load rate > 85%), the voltage constraint range will be tightened to ±3% (9.7kV≤U≤10.3kV) to improve power grid stability.

[0039] Subsequently, the grid voltage stability index and the power plant generation revenue index were quantified. The grid voltage stability index was quantified using the sum of squared voltage deviations, with the formula V_stab=Σ(ΔU_t²), where ΔU_t=|U_t-U_n| is the deviation between the grid-connected voltage and the rated voltage at time t, and t is the optimization time step (5 min / step, 48 steps in total, covering a 4-hour prediction duration). The smaller this index, the more stable the voltage. For example, if the voltage deviations from t1 to t3 in a certain time series are 0.1kV, 0.08kV, and 0.12kV respectively, then V_stab=0.1²+0.08²+0.12²=0.01+0.0064+0.0144=0.0308. The power generation revenue indicator for the power plant is quantified using net revenue, with the formula R=Σ(P_t×p_t-C_t), where P_t is the output power of the photovoltaic power plant at time t (taken from the ultra-short-term power forecast curve), p_t is the real-time grid-connected electricity price (e.g., 0.6 yuan / kWh), and C_t is the power regulation cost at time t (positively correlated with the regulation range, formula C_t=k×|P_t-P_{t-1}|, k=0.05 yuan / kW, which is the regulation cost coefficient). The larger this indicator is, the higher the revenue. For example, at time t1, P_1=40MW, p_1=0.6 yuan / kWh, and P_0=38MW, then C_1=0.05×(40-38)×1000=100 yuan, and R_1=40×1000×0.6-100=23900 yuan.

[0040] The multi-objective optimization mathematical model has two objectives: minimizing the grid voltage stability index V_stab and maximizing the power plant's power generation revenue index R. The constraints include the voltage constraint, frequency constraint, and power ramp-up rate constraint analyzed earlier, and supplemented by upper and lower power limits (0≤P_t≤P_max, where P_max is the rated power of the photovoltaic power plant, 50MW). The standard form of the model is: minF(P)=[V_stab(P),-R(P)]; st9.5kV≤U_t(P)≤10.5kV, 49.5Hz≤f_t(P)≤50.5Hz, |P_t-P_{t-1}|≤2MW, 0≤P_t≤50MW, where U_t(P) and f_t(P) are the voltage and frequency values ​​at time t corresponding to the power sequence P, respectively, and are correlated through a power flow calculation model (the power flow calculation uses the Newton-Raphson method with a convergence accuracy of 10^-6).

[0041] Based on the ultra-short-term power prediction curve, a scenario generation method is used to simulate possible future photovoltaic power fluctuation scenarios, and combined with the uncertainty of grid operation, multiple scenario optimization tasks are generated. This step covers the uncertainties of future power and the power grid through scenario simulation, extending single-scenario optimization to multi-scenario robust optimization, and improving the adaptability of the regulation strategy. The specific implementation method is as follows: The scenario generation method employs Monte Carlo simulation, with the core being the construction of a fluctuation model based on the error distribution of the ultra-short-term power prediction curve. First, historical prediction error data is statistically analyzed to determine that the error follows a normal distribution N(μ,σ²), where μ=0 (mean error is 0) and σ=5%×P_pred_t (the standard deviation of the error is positively correlated with the predicted power, taking 5% of the predicted value). This distribution is fitted based on a large amount of historical data, ensuring the realism of the fluctuation scenarios. During the simulation, for each time step t (48 steps in total) of the ultra-short-term power prediction curve, 100 random error samples ε_t,i (i=1 to 100) are generated. The power sequence of the i-th scenario is then P_scene,i,t=P_pred_t×(1+ε_t,i), where P_pred_t is the ultra-short-term predicted power at time t. For example, if the predicted power at time t10 is P_pred_10=45MW and the generated error sample ε_10,23=0.03, then the power at time t10 for this scenario is P_scene,23,10=45×(1+0.03)=46.35MW.

[0042] The scenario selection process employs variance reduction to eliminate extreme scenarios with a probability below 1% (such as scenarios with power fluctuations exceeding ±15%), retaining 50 valid scenarios. The cumulative probability of these valid scenarios is ≥99%, ensuring a balance between representativeness and computational efficiency. After selection, each scenario is assigned a probability weight ω_i, calculated using the formula ω_i = n_i / N, where n_i is the number of occurrences of error samples in that scenario, and N is the total number of valid samples (50 × 48 = 2400). For example, if a scenario has 60 occurrences, then ω_i = 60 / 2400 = 0.025.

[0043] When considering the uncertainties in power grid operation, two main factors are taken into account: load fluctuations and grid impedance changes. Load fluctuations are simulated using a normal distribution similar to that of photovoltaic power, with a fluctuation range of ±8% (based on the statistical characteristics of distribution network load fluctuations), generating a load scenario sequence L_scene,i,t. Grid impedance changes are achieved by correcting the impedance parameters in the power flow calculation model, setting the impedance fluctuation range to ±10%, and obtaining the equivalent grid impedance Z_scene,i under different scenarios. The photovoltaic power scenario, load scenario, and grid impedance scenario are combined to form 50 complete multi-source uncertainty scenarios, each scenario corresponding to a set of "power sequence - load sequence - impedance parameters".

[0044] Finally, multi-scenario optimization tasks are generated, each corresponding to a complete scenario. The optimization objective is to minimize V_stab and maximize R in that scenario, with constraints dynamically adjusted according to the scenario (e.g., voltage constraints are tightened to ±3% in heavy load scenarios). The optimization tasks for all scenarios are aggregated by weights to form a global optimization objective, namely minΣ(ω_i×V_stab,i) and maxΣ(ω_i×R_i), ensuring that the optimization strategy performs well in most scenarios.

[0045] By employing a multi-objective particle swarm optimization algorithm, combined with adaptive inertial weights and Pareto dominance sorting, we solve multi-scenario optimization tasks and generate a Pareto optimal solution set. This step employs an intelligent optimization algorithm to solve the multi-objective problem, improves the solution efficiency through an adaptive mechanism, and uses Pareto sorting to select the optimal solution. The specific implementation method is as follows: The core of the multi-objective particle swarm optimization algorithm is to simulate the cooperation and evolution of a swarm of particles. Each particle corresponds to a power adjustment sequence solution. The particle encoding uses real number encoding with a dimension of 48 (corresponding to the power values ​​at 48 time steps), and the value range of each dimension is [0, 50] MW (matching power constraints). The algorithm initialization parameters are set as follows: particle swarm size of 50 (balancing computational efficiency and solution diversity), maximum number of iterations of 100 (determined based on convergence tests, the change in solution after 100 iterations is < 0.1 MW), and learning factors c1=c2=2.0 (accelerating particles to approach individual and swarm optima).

[0046] The adaptive inertia weight is designed to balance the algorithm's exploration capability (global search) and development capability (local refinement). The weight formula is w = w_max - (w_max - w_min) × iter / max_iter, where w_max = 0.9 (initial weight, enhancing exploration), w_min = 0.4 (weight at the end of the iteration, enhancing development), iter is the current iteration number, and max_iter = 100. For example, at the beginning of the iteration when iter = 10, w = 0.9 - (0.9 - 0.4) × 10 / 100 = 0.85; at the end of the iteration when iter = 90, w = 0.9 - (0.9 - 0.4) × 90 / 100 = 0.45. The search strategy is dynamically adjusted through weight decay.

[0047] The particle velocity and position update formulas are: v_iter+1=w×v_iter+c1×r1×(p_best-x_iter)+c2×r2×(g_best-x_iter), x_iter+1=x_iter+v_iter+1, where v is the particle velocity (value range [-2,2]MW / iteration, matching power ramp-up rate constraint), r1 and r2 are [0,1] random numbers, p_best is the individual particle optimal solution, and g_best is the swarm optimal solution (initially a randomly generated feasible solution). After the update, the feasibility of the particle position needs to be verified. If the power value at a certain time step exceeds [0,50]MW or the ramp rate exceeds 2MW / 5min, it should be corrected to the constraint boundary value. For example, the power of a particle x_iter+1 at time t15 is 52MW, which is corrected to 50MW; the power at time t16 is 47MW, which is 3MW different from the power at time t15, and is corrected to 48MW (ensuring that the difference is ≤2MW).

[0048] Pareto dominance sorting is used to filter non-dominated solutions. The dominance relation is defined as follows: if the V_stab of solution A ≤ the V_stab of solution B and the R of A ≥ the R of B, and at least one of these conditions is strictly true, then A dominates B. After each iteration, all particle solutions are sorted by dominance, eliminating dominated solutions and retaining non-dominated solutions to form a temporary Pareto solution set. During iteration, an external archive is used to store the best non-dominated solutions from each iteration. The archive capacity is limited to 100. If the capacity is exceeded, solutions in crowded regions are deleted using crowding calculations to maintain the diversity of the solution set. Crowding is calculated as the sum of distances between a solution and its neighboring solutions in the target space; a larger distance indicates better solution diversity. For example, after 100 iterations, the external archive ultimately retains 32 non-dominated solutions, forming the Pareto optimal solution set. Each solution corresponds to a power regulation sequence that balances voltage stability and power generation benefits.

[0049] The most balanced solution is selected from the Pareto optimal solution set using fuzzy decision theory, and then converted into a time series command for power regulation to generate the optimal power regulation strategy.

[0050] This step achieves a balance and trade-off among multiple objectives through fuzzy decision-making, transforming the discrete Pareto solution set into executable adjustment instructions. The specific implementation method is as follows: The application process of fuzzy decision theory includes four steps: establishing an evaluation index system, constructing a fuzzy evaluation matrix, determining the weight vector, and calculating the comprehensive evaluation score. The evaluation index system consists of the grid voltage stability index V_stab and the power plant power generation revenue index R mentioned earlier. First, the two indices are normalized (mapped to the [0,1] interval). V_stab is normalized in reverse (x'=(x_max-x) / (x_max-x_min), the larger the value, the more stable the voltage), and R is normalized in forward (x'=(x-x_min) / (x_max-x_min), the larger the value, the higher the revenue). Here, x_max and x_min are the maximum and minimum values ​​of the corresponding indices in the Pareto solution set, respectively. For example, if the maximum value of V_stab in the solution set is 0.05 and the minimum value is 0.01, and the V_stab value of a certain solution is 0.02, then after normalization, it is (0.05-0.02) / (0.05-0.01)=0.75; if the maximum value of R is 300,000 yuan and the minimum value is 220,000 yuan, and the R value of a certain solution is 260,000 yuan, then after normalization, it is (260,000-220,000) / (300,000-220,000)=0.5.

[0051] When constructing the fuzzy evaluation matrix, the normalized index values ​​are used as membership degrees to establish a fuzzy relation matrix R=[r_ij], where r_ij is the membership degree of the i-th solution on the j-th index (j=1 is V_stab, j=2 is R). When determining the weight vector, the analytic hierarchy process (AHP) is used to invite three power grid operation experts and two photovoltaic power plant operation and maintenance experts to score the importance of the two indicators. After passing the consistency test (consistency ratio CR=0.06<0.1), the weight vector W=[0.6,0.4] is determined, meaning that the weight of grid voltage stability is higher than that of power generation revenue, which conforms to the power grid operation principle of "safety first, with consideration for revenue."

[0052] The comprehensive evaluation score is calculated using fuzzy matrix synthesis, with the formula S_i = W × R_i, where R_i is the fuzzy evaluation vector of the i-th solution, and S_i is the comprehensive score. The solution with the highest score is the most balanced compromise solution. For example, if the fuzzy evaluation vector of one solution is R_i = [0.75, 0.5], then S_i = 0.6 × 0.75 + 0.4 × 0.5 = 0.45 + 0.2 = 0.65; and if another solution has R_i = [0.6, 0.7], then S_i = 0.6 × 0.6 + 0.4 × 0.7 = 0.36 + 0.28 = 0.64. The former has a higher score and is selected as the optimal solution.

[0053] The conversion of the optimal solution requires transforming its corresponding 48-dimensional power regulation sequence into time-series instructions. Each instruction contains three core fields: "timestamp - power setpoint - constraint conditions." The timestamp corresponds one-to-one with the time step of the ultra-short-term prediction curve (e.g., t1 corresponds to the current time + 5 minutes, and the timestamp format is "YYYY-MM-DDHH:MM:SS"). The power setpoint is retained to two decimal places, and the constraint conditions indicate the voltage, frequency, and ramp rate limits for that time step. For example, a certain entry in the time-series instruction might be "2025-04-02 15:05:00_45.23MW_U≥9.7kV、f∈[49.5,50.5]Hz、ΔP≤2MW". The final generated optimal power regulation strategy integrates all time-series instructions and adds strategy execution instructions (such as the issuance cycle of regulation instructions and emergency adjustment thresholds under abnormal conditions) to ensure that the strategy can be directly used for subsequent power allocation and execution.

[0054] S204, the optimal power regulation strategy is decomposed into the power setpoint of each photovoltaic inverter subarray, and power control commands are issued through the communication network to realize the adaptive tracking and regulation of the photovoltaic power plant power to the grid status.

[0055] Specifically, based on the real-time available power and health status of each photovoltaic inverter subarray, the total power command of the optimal power regulation strategy can be proportionally allocated to generate the initial power setting value for each subarray. The core of this step is to dynamically allocate total power commands based on the actual operating capabilities of the subarrays, ensuring the feasibility and balance of power allocation, and laying the foundation for subsequent coordination and optimization. The specific implementation method is as follows: First, obtain the real-time available power and health status data of each photovoltaic inverter subarray. The calculation of real-time available power combines the irradiance and temperature data of the corresponding photovoltaic modules of the subarray with the inverter efficiency. The formula is P_available,i=P_max,i×(G_i / G_std)×(1-k×(T_i-T_std))×η_inv,i, where P_max,i is the rated maximum output power of the i-th subarray (e.g., 5MW / subarray, 10 subarrays in total), G_i is the real-time irradiance of the subarray (unit: W / m²), G_std is the standard test irradiance (1000W / m²), k is the temperature decay coefficient (value: 0.005 / ℃, based on the photovoltaic module characteristics fitting), T_i is the real-time temperature of the subarray modules (unit: ℃), T_std is the standard test temperature (25℃), and η_inv,i is the real-time inverter efficiency (based on load rate fitting, η_inv,i≥98% when the load rate is 50%-100%). For example, if the subarray Node-001 has P_max,001=5MW, real-time G_001=850W / ㎡, T_001=32℃, and η_inv,001=98.2%, then P_available,001=5×(850 / 1000)×(1-0.005×(32-25))×0.982=5×0.85×0.965×0.982≈4.08MW.

[0056] The health status is quantified using a multi-indicator weighted score, which includes subarray failure rate, cumulative runtime, and most recent maintenance interval, with a weight vector W_health=[0.5,0.3,0.2]. The failure rate f_i = number of failures in the past 3 months / total runtime (in times / hour), standardized as f'_i=1-min(f_i / f_max,1) (f_max=0.001 times / hour, the failure threshold); the cumulative runtime score t'_i=1-min(t_i / t_design,1) (t_i is the cumulative runtime, t_design=25000 hours, the design life); and the maintenance interval score m'_i=1-min(m_i / m_max,1) (m_i is the most recent maintenance interval, m_max=30 days, the maximum allowable maintenance interval). The health status score S_i = W_health・[f'_i,t'_i,m'_i]^T, with a value range of [0,1]. The higher the score, the better the health status. For example, for Node-001, f_i = 0.0003 times / hour, f'_i = 1 - 0.0003 / 0.001 = 0.7; t_i = 12000 hours, t'_i = 1 - 12000 / 25000 = 0.52; m_i = 15 days, m'_i = 1 - 15 / 30 = 0.5; then S_001 = 0.5 × 0.7 + 0.3 × 0.52 + 0.2 × 0.5 = 0.35 + 0.156 + 0.1 = 0.606.

[0057] The power allocation ratio is calculated using "available power × health status" as the weight to ensure that the allocation priority is tilted towards subarrays with high available power and good health status. The allocation weight w_i=(P_available,i×S_i) / Σ(P_available,j×S_j) (j=1 to N, N is the total number of subarrays). The initial power setting value P_init,i=w_i×P_total, where P_total is the total power command of the optimal power adjustment strategy. Example: N=10, P_total=45MW, Σ(P_available,j×S_j)=38.5, P_available,i×S_i for Node-001=4.08×0.606≈2.47, then w_001=2.47 / 38.5≈0.064, P_init,001=0.064×45≈2.88MW; P_available,i=4.2MW, S_i=0.65, P_available,i×S_i=2.73, w_002=2.73 / 38.5≈0.071, P_init,002=0.071×45≈3.195MW, and so on, to obtain the initial power settings for all subarrays.

[0058] Considering the power balance and communication delay between subarrays, a distributed consensus algorithm is used to coordinate and optimize the initial power setting value to generate a coordinated power setting value. This step eliminates the initial allocation deviation through a distributed consensus algorithm, ensuring power balance between subarrays and compensating for the impact of communication delays. The specific implementation method is as follows: The criterion for determining power balance between subarrays is that the sum of the absolute values ​​of the deviations between the actual output power and the allocated power of each subarray is ≤0.5MW, and the power difference between any two adjacent subarrays is ≤0.1MW, to avoid electrical imbalances (such as circulating current and voltage fluctuations) caused by local power concentration. The impact of communication delay is handled by real-time monitoring of the transmission delay τ_i,j (transmission time between subarrays i and j). A delay threshold τ_max = 50ms is set (based on the transmission capacity of industrial Ethernet). If τ_i,j > τ_max, historical communication data (the average delay of the three most recent effective communications) is used for estimation, and the communication link is marked as a "weak connection," reducing its weight in the algorithm.

[0059] The distributed consensus algorithm employs a first-order consensus protocol based on neighbor communication. The core idea is that each subarray iteratively updates its own power setting value by exchanging power setting value information with neighboring subarrays until the setting values ​​of all subarrays converge to a consistent value. The algorithm's update formula is P_opt,i(k+1)=P_opt,i(k)+α×Σ(w_i,j×(P_opt,j(k)-P_opt,i(k)))-β×τ_i,j×(P_opt,i(k)-P_opt,i(k-1)), where k is the iteration number, α is the consensus gain (valued at 0.15, controlling the convergence speed), w_i,j is the communication weight between subarrays i and j (adjacent subarrays w_i,j=0.2, non-adjacent subarrays w_i,j=0, ensuring distributed characteristics), β is the delay compensation coefficient (valued at 0.08, compensating for update lag caused by communication delay), and P_opt,i(k) is the power setting value of subarray i in the k-th iteration.

[0060] The communication delay compensation mechanism is implemented by introducing a delay term β×τ_i,j×(P_opt,i(k)-P_opt,i(k-1)). As the delay τ_i,j increases, the delay term increases, accelerating the update speed of the current set value and offsetting the information lag caused by the delay. The convergence condition of the algorithm is |Σ(P_opt,i(k)-P_opt,i(k-1))|≤0.05MW (the iterative update amount is sufficiently small), and the power difference between any adjacent subarrays is ≤0.1MW. The maximum number of iterations is set to 20 (balancing convergence accuracy and computational efficiency).

[0061] Example: After initial allocation, Node-001's P_init,001 = 2.88MW, the adjacent subarray Node-002's P_init,002 = 3.195MW, Node-003's P_init,003 = 2.95MW, and the communication delay τ_001-002 = 35ms, τ_001-003 = 28ms. First iteration: P_opt,001(1)=2.88+0.15×(0.2×(3.195-2.88)+0.2×(2.95-2.88))-0.08×(35×10^-3×(2.88-2.88)+28×10^-3×(2.88-2.88))=2.88+0.15×(0.063+0.014)-0=2.88+0.01155≈2. 8916MW; After Node-002 iteration, P_opt,002(1)=3.195+0.15×(0.2×(2.88-3.195)+0.2×(3.02-3.195))-0.08×35×10^-3×(3.195-3.195)≈3.195+0.15×(-0.063-0.035)≈3.195-0.0147≈3.1803MW. After 12 iterations, the sum of the updated power settings of all subarrays is 0.04MW≤0.05MW, and the maximum power difference between adjacent subarrays is 0.08MW≤0.1MW. Convergence is completed, and the coordinated power settings are obtained, such as Node-001=2.92MW, Node-002=2.95MW, Node-003=2.93MW, etc., thus achieving power balance between subarrays.

[0062] The coordinated power setpoints are encapsulated into control commands via low-latency industrial Ethernet, and timestamps and check codes are added to generate standardized power control commands. These commands are then executed by each photovoltaic inverter subarray to form adaptive tracking regulation.

[0063] This step enables standardized transmission and precise execution of power setpoints, and ensures adaptive tracking of power adjustment through closed-loop feedback. The specific implementation method is as follows: The control commands are encapsulated in a standardized binary format. The frame structure is defined as "Frame Header (4 bytes) - Subarray ID (2 bytes) - Timestamp (8 bytes) - Power Setting Value (4 bytes) - Checksum (4 bytes) - Frame Trailer (2 bytes)," for a total length of 24 bytes. The frame header is fixed at 0x5A5A5A5A (used for frame synchronization to ensure the receiver correctly identifies the start of the command); the subarray ID is a unique identifier (e.g., 0x0001 for Node-001, 0x0002 for Node-002); the timestamp uses UTC time, accurate to microseconds (in hexadecimal encoding of "YYYYMMDDHHMMSSssssss", e.g., 20250402163000123456 corresponds to hexadecimal 0x07E9). 0222101E0001E240) ensures the timing consistency of the instructions; the power setting value is in kW and uses 32-bit floating-point encoding (e.g., 2920kW corresponds to hexadecimal 0x44788000); the check code is calculated using the CRC32 algorithm, and the calculation range is the "subarray ID-timestamp-power setting value" field, which is used to check for bit errors during the instruction transmission process (bit error rate requirement ≤10^-6); the frame tail is fixed at 0xA5A5 (used for frame integrity verification).

[0064] The low-latency industrial Ethernet is configured with the following parameters: a transmission rate of 1Gbps, a TCP / IP communication protocol, and the TCP_NODELAY option to disable the Nagle algorithm (to reduce transmission latency). A dedicated communication link (isolated from other data transmission links) is used to ensure end-to-end transmission latency ≤10ms (meeting the real-time requirements of power regulation). Command transmission uses a "broadcast + unicast confirmation" mode. The control center first broadcasts the command to all subarrays. Each subarray receives the command, calculates the checksum, and if the checksum passes, returns a unicast confirmation frame (containing its own ID and command timestamp). If the checksum fails, a retransmission request is returned. If the control center does not receive a confirmation frame (timeout period 20ms), it automatically retransmits the command, with a maximum of 3 retransmissions. If the retransmission still fails, the subarray is marked as having a "communication failure" and an alarm is triggered.

[0065] The execution of power control commands is completed by the inverter's digital control module. After receiving the command, the module parses the power setpoint and adjusts the switching frequency of the IGBT using pulse width modulation (PWM) technology (the switching frequency is set to 10kHz to ensure smooth adjustment of the output power). This controls the output current and voltage of the photovoltaic modules, ensuring that the actual output power tracks the setpoint. During execution, the inverter collects the output power in real time (sampling frequency 100Hz) and feeds it back to the control center via industrial Ethernet, forming a closed-loop control.

[0066] The verification standard for adaptive tracking adjustment is: the absolute value of the deviation between the actual output power and the set value ≤ 0.5% × P_opt,i (i.e., relative deviation ≤ 0.5%). If the deviation exceeds the threshold, the control center triggers secondary optimization (re-executes the allocation and coordination process of steps one and two). For example, the coordinated set value of Node-001 is 2.92MW (2920kW), and the maximum allowable deviation is 2920 × 0.5% = 14.6kW. If the actual output power is 2908kW, the deviation is 12kW ≤ 14.6kW, and the tracking is qualified; if the actual output power is 2890kW, the deviation is 30kW > 14.6kW, and secondary optimization is triggered, the power set value of the subarray is recalculated, and instructions are issued. At the same time, the control center monitors the grid status data (voltage, frequency) in real time. If the grid status changes abruptly (such as a voltage deviation exceeding ±3%), the current adjustment strategy is immediately interrupted, and the entire power adaptive adjustment process is re-executed to ensure that the photovoltaic power station power always adapts to the grid status.

[0067] As can be seen, real-time monitoring of grid status data and irradiance and temperature data of the photovoltaic array at the grid connection point generates grid operation feature vectors and photovoltaic power generation feature vectors. These feature vectors are then input into a power prediction model that integrates spatiotemporal correlations, outputting an ultra-short-term power prediction curve for the photovoltaic power station for a predetermined duration. Based on the ultra-short-term power prediction curve and current grid operation constraints, an optimal power regulation strategy that balances grid voltage stability and power station power generation revenue is generated through a multi-objective optimization model. This optimal power regulation strategy is decomposed into power setpoints for each photovoltaic inverter subarray, and power control commands are issued through a communication network. This enables the photovoltaic power station to adaptively and accurately track and regulate the grid status, balancing grid voltage stability and power station power generation revenue.

[0068] Another embodiment of the present invention provides a photovoltaic power plant power adaptive regulation system based on grid conditions, see [link to relevant documentation]. Figure 3 The system may include: The monitoring module 301 is used to monitor the grid status data and the irradiance and temperature data of the photovoltaic array at the grid connection point in real time, and generate grid operation feature vectors and photovoltaic power generation feature vectors. Input module 302 is used to input the power grid operation feature vector and the photovoltaic power generation feature vector into the power prediction model that integrates spatiotemporal correlation, and output the ultra-short-term power prediction curve of the photovoltaic power station for a preset time in the future; The generation module 303 is used to generate an optimal power regulation strategy that balances grid voltage stability and power plant generation revenue based on the ultra-short-term power prediction curve and the current grid operation constraints through a multi-objective optimization model. The adjustment module 304 is used to decompose the optimal power adjustment strategy into the power setpoint of each photovoltaic inverter subarray, and send power control commands through the communication network to realize the adaptive tracking and adjustment of the photovoltaic power plant power to the grid status.

[0069] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0070] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0071] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0072] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for adaptive power adjustment of a photovoltaic power plant based on grid conditions, characterized in that, The method includes: Real-time monitoring of grid status data and photovoltaic array irradiance and temperature data at the grid connection point generates grid operation feature vectors and photovoltaic power generation feature vectors. The power grid operation feature vector and the photovoltaic power generation feature vector are input into the power prediction model that integrates spatiotemporal correlation, and the ultra-short-term power prediction curve of the photovoltaic power station for a preset time period is output. Based on the ultra-short-term power prediction curve and the current grid operation constraints, an optimal power regulation strategy that balances grid voltage stability and power plant generation revenue is generated through a multi-objective optimization model. The optimal power regulation strategy is decomposed into the power setpoints of each photovoltaic inverter subarray, and power control commands are issued through the communication network to realize the adaptive tracking and regulation of photovoltaic power plant power according to grid conditions.

2. The method according to claim 1, characterized in that, The real-time monitoring of grid status data at the grid connection point and irradiance and temperature data of the photovoltaic array generates grid operation feature vectors and photovoltaic power generation feature vectors, including: A synchronous phasor measurement unit is deployed at the grid connection point, and distributed irradiance sensors and temperature sensors are deployed on the photovoltaic array. Data acquisition clock synchronization is achieved through a unified time synchronization system to generate a raw monitoring data stream with timestamps. Adaptive Kalman filtering is applied to the raw monitoring data stream to eliminate measurement noise and transient interference, while abnormal data points are detected and marked to generate a denoised and regularized data sequence. Based on the denoised regularized data sequence, dynamic features of the power grid state are extracted, including voltage fluctuation rate, frequency deviation rate, and harmonic distortion rate. Environmental features of photovoltaic power generation are also extracted, including irradiance gradient, temperature change rate, and spatial non-uniformity, generating a multi-dimensional feature set. The multi-dimensional feature set is standardized and normalized, and redundant information is removed by principal component analysis to generate power grid operation feature vectors and photovoltaic power generation feature vectors.

3. The method according to claim 2, characterized in that, The step of inputting the power grid operation feature vector and the photovoltaic power generation feature vector into a power prediction model that integrates spatiotemporal correlation, and outputting an ultra-short-term power prediction curve for a preset duration for the future of the photovoltaic power plant, includes: A spatiotemporal graph structure is constructed, with each subarray of the photovoltaic power station as a node and the spatial correlation between electrical distance and irradiance as edge weights. The power grid operation feature vector and the photovoltaic power generation feature vector are mapped to node features to generate spatiotemporal graph data. Spatiotemporal graph data is input into a graph convolutional neural network. Spatial correlations are captured through multiple graph convolutional layers, and temporal correlations are captured using gated recurrent unit layers to generate spatiotemporal fusion features. Based on spatiotemporal fusion features, an attention mechanism is used to dynamically weight the importance of different time steps and historical sequences to generate attention-weighted feature representations. The predicted power value at each time point within a preset time period is output through a fully connected regression layer, forming an ultra-short-term power prediction curve.

4. The method according to claim 3, characterized in that, The process of generating an optimal power regulation strategy that balances grid voltage stability and power plant generation revenue, based on the ultra-short-term power prediction curve and current grid operation constraints, through a multi-objective optimization model, includes: The current power grid operation constraints are analyzed, including voltage upper and lower limits, allowable frequency deviation range, and power ramp-up rate limit. The power grid voltage stability index and power plant power generation revenue index are quantified, and a mathematical model for a multi-objective optimization problem is generated. Based on the ultra-short-term power prediction curve, a scenario generation method is used to simulate possible future photovoltaic power fluctuation scenarios, and combined with the uncertainty of grid operation, multiple scenario optimization tasks are generated. By employing a multi-objective particle swarm optimization algorithm, combined with adaptive inertial weights and Pareto dominance sorting, we solve multi-scenario optimization tasks and generate a Pareto optimal solution set. The most balanced solution is selected from the Pareto optimal solution set using fuzzy decision theory, and then converted into a time series command for power regulation to generate the optimal power regulation strategy.

5. The method according to claim 4, characterized in that, The step of decomposing the optimal power regulation strategy into power setpoints for each photovoltaic inverter subarray and issuing power control commands via a communication network to achieve adaptive tracking and regulation of the photovoltaic power plant's power to the grid status includes: Based on the real-time available power and health status of each photovoltaic inverter subarray, the total power command of the optimal power regulation strategy is proportionally allocated to generate the initial power setting value for each subarray. Considering the power balance and communication delay between subarrays, a distributed consensus algorithm is used to coordinate and optimize the initial power setting value to generate a coordinated power setting value. The coordinated power setpoints are encapsulated into control commands via low-latency industrial Ethernet, and timestamps and check codes are added to generate standardized power control commands. These commands are then executed by each photovoltaic inverter subarray to form adaptive tracking regulation.

6. A photovoltaic power plant power adaptive adjustment system based on grid conditions, characterized in that, The system includes: The monitoring module is used to monitor the grid status data and the irradiance and temperature data of the photovoltaic array at the grid connection point in real time, and generate grid operation feature vectors and photovoltaic power generation feature vectors. The input module is used to input the power grid operation feature vector and the photovoltaic power generation feature vector into the power prediction model that integrates spatiotemporal correlation, and output the ultra-short-term power prediction curve of the photovoltaic power station for a preset time in the future; The generation module is used to generate an optimal power regulation strategy that balances grid voltage stability and power plant generation revenue based on the ultra-short-term power prediction curve and the current grid operation constraints through a multi-objective optimization model. The adjustment module is used to decompose the optimal power adjustment strategy into the power setpoints of each photovoltaic inverter subarray, and to issue power control commands through the communication network to realize the adaptive tracking and adjustment of the photovoltaic power plant power to the grid status.

7. The system according to claim 6, characterized in that, The monitoring module is specifically used for: A synchronous phasor measurement unit is deployed at the grid connection point, and distributed irradiance sensors and temperature sensors are deployed on the photovoltaic array. Data acquisition clock synchronization is achieved through a unified time synchronization system to generate a raw monitoring data stream with timestamps. Adaptive Kalman filtering is applied to the raw monitoring data stream to eliminate measurement noise and transient interference, while abnormal data points are detected and marked to generate a denoised and regularized data sequence. Based on the denoised regularized data sequence, dynamic features of the power grid state are extracted, including voltage fluctuation rate, frequency deviation rate, and harmonic distortion rate. Environmental features of photovoltaic power generation are also extracted, including irradiance gradient, temperature change rate, and spatial non-uniformity, generating a multi-dimensional feature set. The multi-dimensional feature set is standardized and normalized, and redundant information is removed by principal component analysis to generate power grid operation feature vectors and photovoltaic power generation feature vectors.

8. The system according to claim 7, characterized in that, The input module is specifically used for: A spatiotemporal graph structure is constructed, with each subarray of the photovoltaic power station as a node and the spatial correlation between electrical distance and irradiance as edge weights. The power grid operation feature vector and the photovoltaic power generation feature vector are mapped to node features to generate spatiotemporal graph data. Spatiotemporal graph data is input into a graph convolutional neural network. Spatial correlations are captured through multiple graph convolutional layers, and temporal correlations are captured using gated recurrent unit layers to generate spatiotemporal fusion features. Based on spatiotemporal fusion features, an attention mechanism is used to dynamically weight the importance of different time steps and historical sequences to generate attention-weighted feature representations. The predicted power value at each time point within a preset time period is output through a fully connected regression layer, forming an ultra-short-term power prediction curve.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images