Optical power prediction method and system based on artificial intelligence

Abstract

The invention provides an artificial intelligence-based optical power prediction method and system, and the method comprises the steps: obtaining multi-data-source information of photovoltaic power generation, and carrying out the preprocessing, and obtaining a standardized multi-mode input tensor; based on the multi-modal input tensor, determining a disturbance criterion at each moment for identifying whether the current weather changes drastically so as to generate a corresponding indication signal; constructing a graph structure by taking the component state data of each photovoltaic component as a graph node according to the component state data constructed by the multi-modal input tensor, screening candidate edges according to a structure / physical / environment coupling relationship between the components, and constructing an edge set and calculating an edge weight according to the candidate edges; performing disturbance propagation based on an improved graph neural network according to the constructed graph structure, and outputting disturbance feature representation of each component node; and carrying out feature integration on the disturbance feature representations of the plurality of assemblies, and outputting a system-level optical power predicted value at the current moment.

Description

Technical Field

[0001] The present invention belongs to the field of optical power prediction, and in particular relates to an optical power prediction method and system based on artificial intelligence. Background Art

[0002] With the rapid development of renewable energy, photovoltaic power generation has been widely used in power systems worldwide. Since photovoltaic power generation is highly dependent on environmental factors such as solar irradiance, cloud changes, wind speed, and temperature, its output power has significant nonlinear fluctuations and uncertainties. To ensure the safe and stable operation of the power grid, high-precision short-term predictions of the output power of photovoltaic systems are required to achieve effective scheduling management and energy optimization. However, existing optical power prediction methods are mostly based on time series modeling, statistical learning, or deep learning technologies. Although they can achieve certain results under ideal or stable climate conditions, the prediction accuracy is still seriously insufficient when faced with non-ideal scenarios such as sudden weather changes, obstruction by complex terrain, and component aging or failure.

[0003] Traditional statistical models (such as ARIMA) have limited ability to cope with non-stationarity and strong volatility. While existing machine learning-based models (such as LSTM, GRU, CNN, Transformer, etc.) have certain fitting capabilities, most methods are static in model structure. That is, regardless of whether the input data undergoes a sudden change, the model prediction path is fixed, and they lack the ability to perceive and dynamically respond to sudden meteorological changes. In addition, most methods still model site-level or system-level power, ignoring the shading propagation effects and electrical coupling relationships between components in the photovoltaic array. In actual applications, when a component is partially blocked or fails, its power drop may affect the output power of the entire component string or even the entire array through the series circuit path, resulting in a nonlinear and sudden drop in system-level power. This phenomenon is often ignored as an outlier in existing models, further reducing the overall prediction accuracy.

[0004] At the same time, there is currently a lack of a method that can systematically perceive both macro-environmental disturbances (such as sudden weather changes) and micro-structural disturbances (such as occlusion propagation), and even less so a prediction system that can make structural modeling adjustments after perception. In complex environments such as sudden weather events, partial occlusion of components, high altitudes, or densely built-up areas, the combined effect of these dual disturbances is particularly pronounced. If this cannot be effectively modeled, the stability, practicality, and credibility of the prediction model will be seriously affected. Therefore, there is an urgent need for a new optical power prediction method that can dynamically model multi-level disturbance characteristics to improve adaptability and accuracy in real complex scenarios. Summary of the Invention

[0005] The purpose of this invention is to propose an artificial intelligence-based optical power prediction method and system, which can simultaneously perceive the mutation risk at the macro-meteorological level and the local disturbance at the photovoltaic array structure level, and make targeted dynamic adjustments to the prediction model structure based on the perception results.

[0006] In a first aspect, an embodiment of the present invention provides an artificial intelligence-based optical power prediction method, the method comprising:

[0007] S1. Acquire multi-data source information of photovoltaic power generation and perform preprocessing to obtain a standardized multi-modal input tensor;

[0008] S2. Based on the multimodal input tensor, determining a disturbance criterion at each moment to identify whether a drastic change has occurred in the current weather, so as to generate a corresponding indication signal; wherein, if the disturbance criterion is less than a preset threshold, using a lightweight model for prediction; if the disturbance criterion is greater than the preset threshold, continuing to execute S3;

[0009] S3. Based on the component status data constructed from the multimodal input tensor, a graph structure is constructed using the component status data of each photovoltaic component as a graph node. Candidate edges are screened based on the structural / physical / environmental coupling relationships between components, and an edge set is constructed based on the candidate edges and edge weights are calculated. The constructed graph structure is initialized to generate a feature vector for each node.

[0010] S4. Perform perturbation propagation based on the improved graph neural network according to the constructed graph structure, and output the perturbation feature representation of each component node;

[0011] S5. Perform feature integration on the disturbance feature representations of several components and output a system-level optical power prediction value at the current moment.

[0012] Furthermore, the improved graph neural network specifically includes:

[0013] An improved graph neural network with a multi-layer propagation structure is constructed. The goal of each iterative propagation unit is to perceive disturbance information from neighboring nodes and update its own node features. The propagation update method of each layer is as follows:

[0014]

[0015] in, The embedding representation of node i at layer l+1, represents the embedding representation of node i at layer l; is the perturbation propagation weight at the lth layer; is the disturbance difference activation term, reflecting whether the disturbance is pushed from j to i; β (l) Controls the residual strength of the self-maintaining term; σ(·) is a nonlinear activation function; Represents the set of adjacent components of node i.

[0016] Furthermore, the first layer is the disturbance propagation weight The edge weight function is determined based on the joint modeling of historical power collaborative change data and structural edge attributes, which is expressed as:

[0017]

[0018] Among them, ρ ij is the historical disturbance correlation; is the asymmetric response term to disturbance, which measures the following behavior of j when disturbance occurs at i: If the power of component i suddenly drops (i.e. ), the power rise deviation of component j is observed; w ij is the edge weight in the structure graph; γ, η, ζ are the fusion coefficients of the three information sources; softmax j Normalize the neighbor direction of each node to ensure the interpretability and stability of the propagation weight.

[0019] Furthermore, the multi-data source information of photovoltaic power generation includes satellite cloud image sequences, irradiance, wind speed and direction, and working status data of photovoltaic components.

[0020] Furthermore, determining the disturbance criterion at each moment based on the multimodal input tensor specifically includes:

[0021] Obtain the cloud features at the current moment and the previous moment, and generate the corresponding Euclidean distance to represent the magnitude of cloud layer changes;

[0022] Get the wind speed data at the current moment;

[0023] Get the change in irradiance data between the current moment and the previous moment;

[0024] The disturbance criterion at each moment is obtained by weighted summation based on the Euclidean distance, wind speed data and irradiance data changes.

[0025] Furthermore, the screening of candidate edges based on the structural / physical / environmental coupling relationships between components specifically includes:

[0026] If the first component and the second component are in the same series circuit or downstream of the same converter, it is assumed that there is a corresponding graph edge between the first component and the second component, indicating electrical coupling;

[0027] If the first component blocks the second component during certain periods, a graph edge is added to indicate the possible disturbance conduction direction of the blocking path;

[0028] The calculation of edge weights specifically includes:

[0029] Obtain the first-order response correlation of the historical power disturbances of the first component and the second component, and use a sliding window to calculate the Pearson correlation coefficient;

[0030] For the possible disturbance transmission direction of the occlusion path, the occlusion co-occurrence frequency of the first component and the second component is obtained;

[0031] Design a disturbance response directional asymmetry index to reflect whether there is a unidirectional disturbance amplification effect;

[0032] The edge weight is obtained by weighted summing the first-order response correlation, occlusion co-occurrence frequency and disturbance response direction asymmetry index.

[0033] Furthermore, the disturbance response directional asymmetry index represents the average value of the deviation between the actual power of the second component and the expected power when the power of the first component drops sharply; wherein the sharp power drop is when the actual power change rate of the first component is greater than the preset change rate;

[0034] The feature vector of each node is formed by concatenating the multimodal input tensor, the power change rate of the current component at the current moment, and the power standard deviation of the current component in the past T steps.

[0035] Furthermore, the improved graph neural network also includes a perturbation gating regularization term for controlling the propagation of perturbations to propagate only along credible perturbation paths.

[0036] Furthermore, the feature integration of the disturbance feature representations of the plurality of components and output of the system-level optical power prediction value at the current moment specifically includes:

[0037] Design a disturbance-aware structural aggregation network to perform weighted fusion of all components through a set of trainable attention weights to generate the system predicted power value;

[0038] To enhance aggregation stability, a standard mean square error loss is used in the training phase of the structure aggregation network, with the final system predicted power value as input and the current model predicted output as output;

[0039] The structure aggregation network is constructed as follows:

[0040] A mapping function that projects the disturbance feature representations of the plurality of components into a scalar space, and outputs a power estimate of the corresponding component at time t;

[0041] All power estimation values are weighted and aggregated to obtain the system predicted power value.

[0042] In a second aspect, an embodiment of the present invention provides an artificial intelligence-based optical power prediction system, the system comprising:

[0043] The modal tensor acquisition unit is used to obtain multi-data source information of photovoltaic power generation and perform preprocessing to obtain a standardized multi-modal input tensor;

[0044] a disturbance criterion analysis unit, configured to determine a disturbance criterion at each moment based on the multimodal input tensor, for identifying whether a drastic change has occurred in the current weather, and generating a corresponding indication signal; wherein, if the disturbance criterion is less than a preset threshold, a lightweight model is used for prediction; if the disturbance criterion is greater than the preset threshold, step S3 is continued;

[0045] A graph structure construction unit is configured to construct a graph structure based on the component status data constructed from the multimodal input tensor, using the component status data of each photovoltaic component as a graph node, screening candidate edges based on the structural / physical / environmental coupling relationships between components, and constructing an edge set and calculating edge weights based on the candidate edges; initializing the constructed graph structure and generating a feature vector for each node;

[0046] The perturbation propagation unit is used to perform perturbation propagation based on the improved graph neural network according to the constructed graph structure and output the perturbation feature representation of each component node;

[0047] The optical power prediction output unit is used to integrate the disturbance feature representations of several components and output the system-level optical power prediction value at the current moment.

[0048] The beneficial technical effects of the present invention are at least as follows:

[0049] First, by constructing a multimodal mutation recognition mechanism, based on satellite cloud images, meteorological sequences and spatial information, it is possible to detect in real time whether there is a trend of drastic changes in lighting conditions, thereby identifying in advance system-level disturbances that may cause model prediction failures. Secondly, when a mutation is identified or when operating in a highly sensitive scenario, the present invention further introduces a structural perception mechanism, which models the propagation path of the disturbance in the array based on the shading relationship and electrical coupling between photovoltaic components, thereby more fine-grainedly characterizing the nonlinear impact of local anomalies on system power. Through this overall architecture of perception-modeling-response, the present invention can quickly respond, accurately model and dynamically adjust model prediction strategies in emergency environments, significantly improving the robustness and credibility of optical power predictions, and is particularly suitable for deployment in power station scenarios under complex environments, with broad engineering value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The present invention is further described with reference to the accompanying drawings. However, the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative effort.

[0051] Figure 1 This is a flow chart of an optical power prediction method based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0052] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0053] In one embodiment, Figure 1 As shown, an artificial intelligence-based optical power prediction method is provided, comprising the following steps:

[0054] S1. Obtain multi-data source information of photovoltaic power generation and perform preprocessing to obtain a standardized multi-modal input tensor.

[0055] Specifically, in this step, the goal of the present invention is to integrate and standardize the raw data from different sources and finally construct a unified input tensor x t . Since the power output of photovoltaic power generation is affected by multiple factors, such as meteorological changes, component status, and changes in light intensity, the data source is relatively complex, so it is necessary to effectively fuse the multimodal data. First, the present invention collects information from multiple data sources. Typical input data include satellite cloud image sequences, irradiance, wind speed and direction, and working status data of photovoltaic components. These data reflect the weather conditions, light intensity, and the operating status of each component respectively. Specifically, the satellite cloud image sequence records the movement of clouds and is an important basis for predicting weather changes; irradiance data is the main driving factor of photovoltaic power generation and directly affects the power output of photovoltaic components; wind speed and direction information helps to further infer weather trends; the working status of photovoltaic components, such as temperature and current, provides operating efficiency information for each component.

[0056] Among them, satellite cloud image sequence preprocessing:

[0057] Data Source: Satellite cloud images are typically acquired by remote sensing satellites. Each image represents the distribution of clouds over a specific time period. The dimensions of each image are typically (H, W, C), where H and W represent the image's height and width, respectively, and C represents the number of channels. For example, for an RGB image, C = 3, representing the red, green, and blue channels.

[0058] Normalization: To ensure that each image has the same scale and range, the image pixel values are first normalized in the following way: each pixel value is mapped from the original range (usually [0,255]) to the range [0,1]. The specific processing method is:

[0059]

[0060] Among them I original is the pixel value in the original image, min(I original ) and max(I original ) are the minimum and maximum values of the image, I normalized is the normalized image.

[0061] Time series alignment: Due to the different temporal resolutions of satellite cloud images and other data, this method requires aligning the image data to the same time step as the other time series data. For example, if the temporal resolution of the cloud image data is 10 minutes, and the temporal resolution of the irradiance data is 1 minute, the time step of the cloud image data needs to be adjusted to 1 minute through interpolation (such as linear interpolation or cubic interpolation) to achieve time series alignment.

[0062] Furthermore, the irradiance and wind speed data are normalized:

[0063] Data Source: Irradiance data is acquired in real time from photovoltaic power plants or weather stations. It is typically a one-dimensional time series, representing the solar irradiance at each moment. Wind speed data is also typically a time series, representing the current wind speed.

[0064] Minimum-maximum normalization: To ensure that different types of data have the same scale, irradiance and wind speed data need to be normalized. Here, the minimum-maximum normalization method is used to make the value range of all data between [0,1]. The formula is as follows:

[0065]

[0066] where r original Represents the original irradiance or wind speed data, min(r original ) and max(r original ) are the minimum and maximum values of the data, r normalized is the normalized data.

[0067] Furthermore, the working status data of the photovoltaic modules is processed:

[0068] Data Source: The operating status of each PV module includes information such as temperature, current, and voltage. This data typically comes from a real-time monitoring system. Each module's status data can be represented as a multidimensional vector with a shape of (M, 1), where M is the number of features per module. For example, if temperature, voltage, and current are three features, then M = 3.

[0069] Standardization: For component status data, the present invention performs standardization on each feature (such as temperature, current, and voltage) individually. The present invention uses a zero mean unit variance method, and the formula is:

[0070]

[0071] where x original is the original data, μ x and σ x are the mean and standard deviation of the feature, x normalized is the standardized data.

[0072] Furthermore, the construction of the unified input tensor:

[0073] After all data are preprocessed and normalized, the present invention combines them into an input tensor x t . This tensor contains all the processed data:

[0074]

[0075] Where: I t-k:t represents the satellite cloud image sequence of k time steps from time point tk to t; r t-k:t is the irradiance time series of the past k steps; v t is the wind speed data at the current moment; is the working status data of each PV module, and N is the number of PV modules.

[0076] Furthermore, data format and alignment:

[0077] The present invention ensures that the time steps of all data sources are aligned. If some data have different time resolutions, the present invention will align them through interpolation methods (such as linear interpolation or cubic interpolation) to ensure that each data source has a valid value at the same time point, avoiding errors caused by time deviation.

[0078] Through these processes, the present invention generates a standardized multimodal input tensor x t , this tensor can be used as input for subsequent steps such as perturbation perception and component modeling, ensuring that data from different sources are effectively fused at a unified time step and the same scale.

[0079] S2. Based on the multimodal input tensor, determine the disturbance criterion at each moment to identify whether the current weather has undergone drastic changes and generate a corresponding indicator signal; if the disturbance criterion is less than a preset threshold, use a lightweight model for prediction; if the disturbance criterion is greater than the preset threshold, continue to execute S3. The lightweight model is a set of four-layer fully connected neural networks, whose input is the multimodal input tensor constructed in step 1. The model first extracts a low-dimensional representation through the first-layer fully connected network and improves the expression ability through a nonlinear activation function. The second-layer fully connected network then outputs a predicted value to estimate the current power generation of the entire station.

[0080] The model has a simple structure and low computational cost. The training process is based on operating data under historical stable weather conditions. It can achieve efficient, fast, and near real-time power prediction in the absence of significant disturbances.

[0081] Specifically, in this step, the present invention will be based on the input tensor χ obtained from step 1 t , identifies sudden meteorological changes and, based on the intensity of the disturbance, decides whether to switch to the enhanced path and execute S3. The design of this module not only ensures a rapid response to sudden meteorological changes, but also effectively reduces the computational complexity under normal circumstances.

[0082] Furthermore, the present invention calculates the disturbance criterion Δ at each moment t t , which is used to identify whether the current weather has undergone drastic changes. The design of the criterion needs to take into account the changes in cloud sequence and the influence of wind speed, because rapid changes in clouds and strong winds often mean sudden changes in weather. Specifically, the calculation formula of the disturbance criterion is:

[0083] Δ t =||z t -z t-1 ||+α·||v t ||+λ·||r t -r t-1 || (5)

[0084] Among them, z t and z t-1 are the cloud features of the current moment and the previous moment, respectively, ||z t -z t-1 || is the Euclidean distance between them, indicating the magnitude of cloud layer changes; v t is the wind speed data at the current moment, ||v t || is the magnitude of the wind speed, α is the weight of the wind speed on the criterion; r t and r t-1 are the changes in irradiance data at the current moment and the previous moment, respectively, ||r t -rt-1 || reflects the severity of the change in solar irradiance, and λ is the weight of the impact of irradiance change on the disturbance criterion.

[0085] It should be noted that, in the case of disturbance, the present invention needs to be based on the criterion Δ t The size of determines whether to switch model paths. The purpose of path scheduling is to dynamically adjust the complexity of the model. If the meteorological disturbance is large (such as fast-moving clouds), a more refined model path is switched for fine-grained prediction; if the disturbance is small, a simpler and faster model is used for prediction.

[0086] Among them, the decision-making mechanism:

[0087] If Δ t >δ indicates that the system is facing strong meteorological disturbances and needs to switch to a more complex model for refined prediction. In this case, the model needs to consider the occlusion propagation effect between components and make predictions through component-level propagation modeling.

[0088] If Δ t ≤δ, indicating that the weather changes are relatively stable, and the system uses a lightweight model for rapid prediction.

[0089] Furthermore, the core of this decision-making mechanism lies in dynamically adjusting the forecast path and deciding whether to use a complex modeling path based on the complexity of the current meteorological conditions, thereby improving computational efficiency and forecast accuracy.

[0090] The output of this module is an indication signal s t , whose values are:

[0091] s t =1, indicating that the enhanced path (i.e., component-level disturbance propagation modeling) needs to be enabled;

[0092] s t =0, indicating the use of the conventional lightweight prediction model.

[0093] In actual deployment, based on this indication signal, the system will flexibly switch the model path according to different meteorological disturbance conditions, thereby achieving more accurate and efficient optical power prediction.

[0094] S3. Based on the component status data constructed from the multimodal input tensor, a graph structure is constructed with the component status data of each photovoltaic component as a graph node, candidate edges are screened based on the structural / physical / environmental coupling relationship between components, and an edge set is constructed and edge weights are calculated based on the candidate edges; the constructed graph structure is initialized to generate a feature vector for each node.

[0095] Specifically, the goal of this step is to construct a graph structure G = (V, E) that reflects the coupling relationship between components in the photovoltaic array and the disturbance propagation mechanism, and assign an initialization feature vector with physical meaning to each component node.

[0096] Further, construct the node set V of the graph. According to the unified input tensor χ generated in step 1 t Component status data in The present invention takes each photovoltaic module as a graph node v i The number, location and identity of the nodes can be directly imported through the PV array structure file (such as the device topology configuration table of the power station SCADA system) to ensure physical consistency.

[0097] Furthermore, an edge set E is constructed to reflect the structural / physical / environmental coupling relationship between components. The present invention introduces a multi-factor coupling mechanism to define edge e ij Whether it exists, and the weight w ij The value of is derived from:

[0098] If components i and j are in the same series circuit or downstream of the same converter, then an edge e exists by default. ij , indicating electrical coupling;

[0099] If component i blocks component j during certain periods (e.g., modeled by shading analysis), an edge is added to indicate the possible disturbance conduction direction of the shading path;

[0100] For all candidate edges, perturbation propagation correlation is introduced to calculate edge weights.

[0101] Among them, the core edge weight design is as follows:

[0102]

[0103] ρ ij is the first-order response correlation of the historical power disturbances of components i and j, and the Pearson correlation coefficient is calculated using a sliding window;

[0104] κ ij represents the co-occurrence frequency of occlusions between components i and j, which can be derived from historical occlusion labels or estimated through spatial modeling (e.g., shadow casting);

[0105] It is an indicator of the directional asymmetry of the disturbance response, which is used to reflect whether there is a unidirectional disturbance amplification effect and is defined as:

[0106]

[0107] Indicates that a sharp drop in power occurs at component i When the actual power of component j is With expected power The average deviation of the three weights is denoted by the average deviation of the perturbation from i to j. This metric reflects the physical feedback phenomenon of whether the perturbation diffuses from i to j. The contribution strength of the three weights is controlled by the hyperparameters γ, η, and μ, respectively, and can be optimized through cross-validation or end-to-end training.

[0108] Furthermore, initialize the node features The present invention uses the current status of each component obtained in step 1 The state includes a vector of observed values such as temperature, voltage, and current. To enhance the disturbance propagation modeling capability, the present invention adds two disturbance-related indicators to the feature vector:

[0109] Indicates the power change rate of component i at the current moment, used to indicate the source of the disturbance;

[0110] It represents the standard deviation of the power of component i in the past T steps, which measures its stability.

[0111] Among them, the power change rate This refers to the change in a component's output power between the current sampling moment and the previous sampling moment, measured in watts per minute. The system first records the current power value from the previous value, subtracts the two, and then divides them by a fixed sampling interval (e.g., one minute) to determine the component's power change per unit time. Subsequently, the mean and standard deviation of the power change rate samples for all components in the station during the historical stable period are calculated, and the indicator is standardized to maintain a consistent numerical scale in the model input.

[0112] Power standard deviation It refers to the fluctuation amplitude of a component's power value within a fixed time window, measured in watts. During the calculation, the system selects the power values of the component at the most recent consecutive sampling moments (e.g., the past ten minutes) from the historical power sequence, calculates the average of these power values, and then calculates the average of the squared deviations between each sampled value and the average. Finally, the square root of this average is taken to obtain the power standard deviation of the component during that time period. This standard deviation is then normalized using the Z-score to give it a uniform numerical scale, serving as one of the inputs for node features in disturbance propagation modeling.

[0113] Therefore, the node features can be constructed as concatenated vector forms:

[0114]

[0115] This vector will serve as the input of the perturbation propagation network in subsequent steps, determining how each component responds to neighbor perturbations in the initial state.

[0116] Furthermore, to avoid overfitting historical perturbation patterns during graph structure modeling, this paper introduces a "perturbation attention regularization term" during graph construction to constrain graph edge weights, guiding the model to focus on high-confidence perturbation paths and weakening low-correlation but structurally existing edge connections. This regularization term is used in subsequent training stages and can be expressed as:

[0117]

[0118] where conf ij is a confidence score based on edge weight source data (such as occlusion frequency, disturbance similarity, etc.), w ij It is edge e ij This term encourages the model to focus on credible edges and weaken spurious correlation paths.

[0119] Finally, the output is a graph structure G = (V, E), which contains the feature vector of each node and the perturbation propagation weight w of each edge ij This graph structure will be passed as input to the perturbation propagation modeling network in the subsequent steps.

[0120] S4. Perform perturbation propagation based on the improved graph neural network according to the constructed graph structure, and output the perturbation feature representation of each component node.

[0121] Specifically, this step simulates the dynamic diffusion process of disturbances between photovoltaic modules through structure perception, and further obtains deep disturbance perception representation from the evolution of module-level features. Its input is the module graph structure G output in step 3, and the initial node features of each module are and the disturbance propagation edge weight w ij , whose goal is to generate the final perturbation representation For use by the next step system-level power aggregation module.

[0122] This step plays a core role in connecting the preceding and following stages of the patented system: it follows the abstract representation of the inter-component structure and disturbance transmission mechanism from the previous step, and outputs a feature embedding capable of understanding dynamic disturbances, which is the key to the robustness and causal explanatory power of system-level power forecasting. The Disturbance Propagation Graph Network (DPGN) proposed in this invention is not only a variant of a graph neural network, but also an innovative model designed by combining the physical constraints of photovoltaic scenarios, heterogeneous disturbance propagation mechanisms, and power response structures.

[0123] The network adopts a multi-layer propagation structure (the number of layers is denoted as L). The goal of each layer of iterative propagation unit is to perceive disturbance information from neighboring nodes and update its own node features. Unlike the mean or weighted aggregation method of traditional GCN and GAT, this invention specifically introduces a disturbance difference-driven propagation mechanism to capture the "unbalanced diffusion pattern" of disturbances. This invention defines the propagation update method of each layer as follows:

[0124]

[0125] represents the embedding representation of node i at layer l;

[0126] is the perturbation propagation weight at layer l (see the definition below);

[0127] is the disturbance difference activation term, reflecting whether the disturbance is “pushed” from j to i;

[0128] β (l) Control the residual strength of the self-maintaining term;

[0129] σ(·) is a nonlinear activation function (such as ReLU);

[0130] Represents the set of adjacent components of node i.

[0131] Furthermore, this propagation mechanism embodies a core concept: the propagation of disturbances is not determined by the neighboring states themselves, but rather by the difference between the neighboring states and the system's own. For example, a sudden drop in power in a neighboring component will cause disturbances to connected components, a behavior that cannot be modeled in traditional graph models.

[0132] The present invention proposes a perturbation causal weighted function Used to determine whether the disturbance is propagating from a neighbor. It is an edge weight function based on the joint modeling of historical power collaborative change data and structural edge attributes. The present invention defines:

[0133]

[0134] Among them, ρ ij is the historical disturbance correlation (such as the Pearson correlation coefficient under the sliding window);

[0135] is the asymmetric response term to disturbance, which measures the following behavior of j when disturbance occurs at i:

[0136]

[0137] If the power of component i suddenly drops (i.e. ), then the power rise deviation of component j is observed;

[0138] w ij is the basic edge weight calculated in step 3 in the structure graph;

[0139] γ, η, ζ are the fusion coefficients of the three information sources;

[0140] softmax j Normalize the neighbor direction of each node to ensure the interpretability and stability of the propagation weight.

[0141] Furthermore, the disturbance propagation weight is one of the most innovative designs of this step, allowing the present invention to learn the "causal driving force" of the disturbance based on physical structure + data response, capturing the directional diffusion behavior, and is a representative modeling method that integrates GNN theory with power plant fault / occlusion mechanism.

[0142] Furthermore, in the above propagation update process, to prevent information from "infinitely diffusing" in the graph and causing oversmoothing, the present invention introduces a perturbation gating regularization term as part of the loss function to control the model to propagate only along the credible perturbation path. It is defined as follows:

[0143]

[0144] where conf ij Score the perturbation confidence defined in step 3, representing the edge e ij As the credibility of the perturbation path; is the perturbation propagation weight of the current layer. This term acts as an attention gating for the perturbation path, further improving the stability and interpretability of perturbation modeling by encouraging high-confidence paths to have high weights and low-confidence paths to have low weights.

[0145] It can be understood that the final output of the model is the perturbation feature representation of each component node It integrates the historical state evolution of the component itself, the neighbor disturbance response structure, the propagation path weight and the disturbance asymmetry, forming a deep understanding of the future state of the component. These outputs will be weighted and integrated in the next power aggregation module to generate system-level power prediction. Key input.

[0146] S5. Perform feature integration on the disturbance feature representations of several components and output a system-level optical power prediction value at the current moment.

[0147] Specifically, this step is to propagate the disturbance characteristics of all component nodes obtained in the previous step. Through structured and weight-adaptive aggregation, the system-level optical power prediction value at the current moment is output.

[0148] Furthermore, the input is the disturbance propagation characteristics of the N components output in step 4 Each This is a high-dimensional vector that incorporates the component's physical state changes, information about the disturbance response chain, and the impact of neighboring disturbances. These features have been deeply abstracted through disturbance propagation modeling and can be considered an encoding of the component's impact on overall power in the current environment.

[0149] In order to reasonably map these component-level representations to the overall station power output, this paper designs a disturbance-aware structure aggregation network, which uses a set of trainable attention weights α i Perform weighted fusion of all components:

[0150]

[0151] It will A mapping function (such as a perceptron or linear layer) that projects data into a scalar space, whose output represents the power estimate of the component at time t.

[0152] α i is the aggregate weight of each component, generated by an attention function (e.g. followed by a softmax layer);

[0153] is the system predicted power value.

[0154] This aggregation process is structurally adaptive: that is, under different disturbance scenarios, the model can automatically learn the components that have the greatest impact on the system, thereby dynamically adjusting the aggregation path and improving the overall system's robustness to disturbances.

[0155] Furthermore, to enhance aggregation stability, the present invention adopts a standard mean square error loss during the training phase:

[0156]

[0157] in is the actual collected system output power;

[0158] is the current model prediction output;

[0159] This loss is back-propagated during the training phase to optimize α i Parameters related to f(·).

[0160] final, It can be used as input for terminal modules such as power station dispatching systems, upper-level grid controllers, and energy management platforms, completing the data closed loop from disturbance data perception to actual control decision-making.

[0161] In one embodiment, an artificial intelligence-based optical power prediction system is provided, the system comprising:

[0162] The modal tensor acquisition unit is used to obtain multi-data source information of photovoltaic power generation and perform preprocessing to obtain a standardized multi-modal input tensor;

[0163] a disturbance criterion analysis unit, configured to determine a disturbance criterion at each moment based on the multimodal input tensor, for identifying whether a drastic change has occurred in the current weather, and generating a corresponding indication signal; wherein, if the disturbance criterion is less than a preset threshold, a lightweight model is used for prediction; if the disturbance criterion is greater than the preset threshold, step S3 is continued;

[0164] A graph structure construction unit is configured to construct a graph structure based on the component status data constructed from the multimodal input tensor, using the component status data of each photovoltaic component as a graph node, screening candidate edges based on the structural / physical / environmental coupling relationships between components, and constructing an edge set and calculating edge weights based on the candidate edges; initializing the constructed graph structure and generating a feature vector for each node;

[0165] The perturbation propagation unit is used to perform perturbation propagation based on the improved graph neural network according to the constructed graph structure and output the perturbation feature representation of each component node;

[0166] The optical power prediction output unit is used to integrate the disturbance feature representations of several components and output the system-level optical power prediction value at the current moment.

[0167] Unless otherwise specifically stated, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0168] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the system described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0169] In the description of this application, it should be noted that the terms "upper" and "lower" etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or the orientations or positional relationships in which the invented product is usually placed when in use. These are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they should not be understood as limitations on this application.

[0170] It should also be noted that, in the description of this application, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on the specific circumstances.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An optical power prediction method based on artificial intelligence, characterized in that: The method comprises: S1. Obtain multi-data source information of photovoltaic power generation and perform preprocessing to obtain a standardized multi-modal input tensor; S2. Based on the multimodal input tensor, determining a disturbance criterion at each moment to identify whether a drastic change has occurred in the current weather, so as to generate a corresponding indication signal; wherein, if the disturbance criterion is less than a preset threshold, using a lightweight model for prediction; if the disturbance criterion is greater than the preset threshold, executing S3; S3. Based on the component status data constructed from the multimodal input tensor, a graph structure is constructed using the component status data of each photovoltaic component as a graph node. Candidate edges are screened based on the structural / physical / environmental coupling relationships between components, and an edge set is constructed based on the candidate edges and edge weights are calculated. The constructed graph structure is initialized to generate a feature vector for each node. S4. Perform perturbation propagation based on the improved graph neural network according to the constructed graph structure, and output the perturbation feature representation of each component node; S5. Perform feature integration on the disturbance feature representations of several components and output a system-level optical power prediction value at the current moment.

2. The optical power prediction method based on artificial intelligence according to claim 1, characterized in that: The improved graph neural network specifically includes: An improved graph neural network with a multi-layer propagation structure is constructed. The goal of each iterative propagation unit is to perceive disturbance information from neighboring nodes and update its own node features. The propagation update method of each layer is as follows: in, The embedding representation of node i at layer l+1, represents the embedding representation of node i at layer l; is the perturbation propagation weight at layer l; is the disturbance difference activation term, reflecting whether the disturbance is pushed from j to i; β (l) Controls the residual strength of the self-maintaining term; σ(·) is a nonlinear activation function; Represents the set of adjacent components of node i.

3. The optical power prediction method based on artificial intelligence according to claim 2, characterized in that: The perturbation propagation weight at the lth layer The edge weight function is determined based on the joint modeling of historical power collaborative change data and structural edge attributes, which is expressed as: Among them, ρ ij is the historical disturbance correlation; is the asymmetric response term to disturbance, which measures the following behavior of j when disturbance occurs at i: If the power of component i suddenly drops, Then the power rise deviation of component j is observed; w ij is the edge weight in the structure graph; γ, η, ζ are the fusion coefficients of the three information sources; softmax j Normalize the neighbor direction of each node to ensure the interpretability and stability of the propagation weight.

4. The optical power prediction method based on artificial intelligence according to claim 1, characterized in that: The multi-data source information of photovoltaic power generation includes satellite cloud image sequences, irradiance, wind speed and direction, and working status data of photovoltaic components.

5. The optical power prediction method based on artificial intelligence according to claim 4, characterized in that: Determining a disturbance criterion at each moment based on the multimodal input tensor specifically includes: Obtain the cloud features at the current moment and the previous moment, and generate the corresponding Euclidean distance to represent the magnitude of cloud layer changes; Get the wind speed data at the current moment; Get the change in irradiance data between the current moment and the previous moment; The disturbance criterion at each moment is obtained by weighted summation based on the Euclidean distance, wind speed data and irradiance data changes.

6. The optical power prediction method based on artificial intelligence according to claim 1, characterized in that: The screening of candidate edges based on the structural / physical / environmental coupling relationships between components specifically includes: If the first component and the second component are in the same series circuit or downstream of the same converter, it is assumed that there is a corresponding graph edge between the first component and the second component, indicating electrical coupling; If the first component blocks the second component during certain periods, a graph edge is added to indicate the possible disturbance conduction direction of the blocking path; The calculation of edge weights specifically includes: Obtain the first-order response correlation of the historical power disturbances of the first component and the second component, and use a sliding window to calculate the Pearson correlation coefficient; For the possible disturbance transmission direction of the occlusion path, the occlusion co-occurrence frequency of the first component and the second component is obtained; Design a disturbance response directional asymmetry index to reflect whether there is a unidirectional disturbance amplification effect; The edge weight is obtained by weighted summing the first-order response correlation, occlusion co-occurrence frequency and disturbance response direction asymmetry index.

7. The optical power prediction method based on artificial intelligence according to claim 6, characterized in that: The disturbance response directional asymmetry index represents the average value of the deviation between the actual power of the second component and the expected power when the power of the first component drops sharply; wherein the sharp power drop is when the actual power change rate of the first component is greater than the preset change rate; The feature vector of each node is formed by concatenating the multimodal input tensor, the power change rate of the current component at the current moment, and the power standard deviation of the current component in the past T steps.

8. The optical power prediction method based on artificial intelligence according to claim 1, characterized in that: The improved graph neural network also includes a perturbation gating regularization term for controlling the propagation of perturbations to propagate only along credible perturbation paths.

9. The optical power prediction method based on artificial intelligence according to claim 1, characterized in that: The feature integration of the disturbance feature representations of the plurality of components and output of the system-level optical power prediction value at the current moment specifically includes: Design a disturbance-aware structural aggregation network to perform weighted fusion of all components through a set of trainable attention weights to generate the system predicted power value; To enhance aggregation stability, a standard mean square error loss is used in the training phase of the structure aggregation network, which takes the system predicted power value as input and outputs the current model predicted output; The structure aggregation network is constructed as follows: A mapping function that projects the disturbance feature representations of the plurality of components into a scalar space, and outputs a power estimate of the corresponding component at time t; All power estimation values are weighted and aggregated to obtain the system predicted power value.

10. An artificial intelligence-based optical power prediction system, characterized in that: The system comprises: The modal tensor acquisition unit is used to obtain multi-data source information of photovoltaic power generation and perform preprocessing to obtain a standardized multi-modal input tensor; a disturbance criterion analysis unit, configured to determine a disturbance criterion at each moment based on the multimodal input tensor, for identifying whether a drastic change has occurred in the current weather, and generating a corresponding indication signal; wherein, if the disturbance criterion is less than a preset threshold, a lightweight model is used for prediction; and if the disturbance criterion is greater than the preset threshold, S3 is executed; A graph structure construction unit is configured to construct a graph structure based on the component status data constructed from the multimodal input tensor, using the component status data of each photovoltaic component as a graph node, screening candidate edges based on the structural / physical / environmental coupling relationships between components, and constructing an edge set and calculating edge weights based on the candidate edges; initializing the constructed graph structure and generating a feature vector for each node; The perturbation propagation unit is used to perform perturbation propagation based on the improved graph neural network according to the constructed graph structure and output the perturbation feature representation of each component node; The optical power prediction output unit is used to integrate the disturbance feature representations of several components and output the system-level optical power prediction value at the current moment.

Images