Artificial intelligence-based photovoltaic power output prediction system and method

Abstract

The present application relates to the technical field of photovoltaic power generation power prediction, specifically to a photovoltaic output prediction system and method based on artificial intelligence, comprising data acquisition and preprocessing, physical driving adaptive time series decomposition, multi-source feature fusion, physical constraint deep learning prediction, uncertainty perception probability prediction and spatial correlation output module, using the solar elevation angle and atmospheric transparency index to construct the complexity index, adaptively adjusting the variational mode decomposition parameters and applying the frequency spectrum physical prior constraint; after fusing the multi-source features, the equivalent circuit, energy conservation and time smoothness constraint loss function are used to train the model to determine the prediction; then the complexity index dynamically adjusts the diffusion step number, generates the prediction residual through the conditional diffusion model to realize the probability prediction, and constructs the dynamic space-time correlation graph for joint prediction. The present application deeply integrates the physical prior and data-driven, significantly improves the prediction accuracy and reliability, and realizes the multi-power station joint probability prediction.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, specifically to a photovoltaic power output prediction system and method based on artificial intelligence. Background Technology

[0002] Photovoltaic power generation output is significantly affected by multiple meteorological factors, including solar irradiance, cloud cover, and temperature variations, exhibiting marked intermittency and volatility. As the installed capacity of photovoltaic power generation in the power system continues to rise, accurate prediction of photovoltaic output has become a crucial aspect of grid dispatching and electricity market transactions. Existing photovoltaic output prediction methods mainly include data-driven methods based on variational mode decomposition, methods that fuse physical models with data-driven approaches, and multi-power station prediction methods based on graph neural networks.

[0003] However, existing technologies have the following shortcomings in practical applications: Chinese patent CN116561567A uses standard variational mode decomposition to perform multi-scale decomposition of photovoltaic power sequences. The number of modes and penalty factors are predetermined through an optimization algorithm and remain fixed throughout the prediction process. Because the frequency components and fluctuation characteristics of photovoltaic power sequences vary significantly with the solar cycle and weather conditions, under clear weather conditions, the power sequence is dominated by smooth diurnal components, requiring fewer decomposition modes and a narrower bandwidth. However, under cloudy to overcast weather conditions, the power sequence is superimposed with a large number of high-frequency, rapid fluctuation components caused by cloud cover, requiring more decomposition modes and a wider frequency coverage. Fixed decomposition parameters cannot adapt to these differences in signal characteristics caused by variations in solar altitude angle and atmospheric transparency, leading to insufficient decomposition or mode aliasing during periods of drastic weather changes. Furthermore, the decomposed mode components lack a correspondence with physical frequencies such as the solar diurnal and semi-diurnal cycles, resulting in insufficient physical interpretability. Chinese patent CN107341569B uses a physical model and a data-driven model as two independent prediction branches, then performs a weighted fusion. The physical model independently calculates the predicted power based on the equivalent circuit equation of the photovoltaic cell, while the data-driven model is independently trained based on historical data and outputs the predicted power. There is no gradient-level information interaction between the two during the model training phase. The data-driven model optimizes parameters solely to minimize prediction error during training, without being constrained by physical laws such as the photovoltaic cell's current-voltage characteristics, the law of conservation of energy, and the rate of power change. This can lead to predictions that violate physical laws, such as negative power values, exceeding the theoretical upper limit of irradiance, or the rate of power change exceeding the physical ramp rate of the photovoltaic array at adjacent times, posing safety hazards in power grid dispatching applications. Chinese patent CN118535901B uses a static influence weight graph as the adjacency matrix of a graph neural network. The connection weights between nodes in the graph structure are not updated with changes in meteorological conditions after construction. In actual operation, when a weather system moves within a region, the meteorological environment of different photovoltaic power stations changes differently, and the correlation of power fluctuations between power stations also changes. Two power stations that were originally highly correlated in power may exhibit significantly different fluctuation patterns due to the passage of the weather system. The fixed graph structure cannot capture this time-varying spatial correlation, resulting in a significant decrease in the accuracy of joint forecasting of multiple power stations during the movement of the weather system.

[0004] Therefore, there is an urgent need in the existing technology for a method that can drive the adaptive adjustment of variational mode decomposition parameters with solar geometric parameters and atmospheric physical state and endow the decomposition results with physical frequency interpretability; a method that can directly embed photovoltaic cell equivalent circuit constraints, energy conservation constraints and time smoothness constraints in the training process of deep learning prediction models in a gradient coupling manner to avoid the prediction results from deviating from physical laws; and a method that can dynamically update the spatial correlation graph structure of multiple power stations and adaptively adjust the probability prediction computing resources according to meteorological conditions. Summary of the Invention

[0005] This invention discloses a photovoltaic power output prediction system and method based on artificial intelligence, which aims to provide high-precision deterministic power prediction and probabilistic prediction results for photovoltaic power plants, and at the same time provide spatial correlation joint prediction capabilities for regions containing multiple photovoltaic power plants.

[0006] To address the above problems, this invention provides a photovoltaic power output prediction system and method based on artificial intelligence: An artificial intelligence-based photovoltaic output prediction system includes a data acquisition and preprocessing module, a physical-driven adaptive temporal decomposition module, a multi-source feature fusion module, a physical-constrained deep learning prediction module, and an uncertainty-aware probability prediction and spatial correlation output module. The data acquisition and preprocessing module transmits the standardized meteorological feature matrix and standardized power sequence obtained after outlier removal and standardization to the physical-driven adaptive temporal decomposition module. The physical-driven adaptive temporal decomposition module calculates the physical-driven meteorological complexity index based on the solar altitude angle, atmospheric transparency index, and irradiance fluctuation rate. This index dynamically determines the number of modes and penalty factor in the variational mode decomposition. After performing physical-driven multi-scale temporal decomposition on the standardized power sequence, the module transmits the multiple intrinsic mode components and one residual component to the multi-source feature fusion module. The multi-source feature fusion module concatenates the multiple intrinsic mode components, residual component, multi-dimensional meteorological feature vectors, and time-period encoding vectors along the feature dimension and maps them to a unified feature dimension space through a linear projection layer. The resulting joint feature matrix is ​​then transmitted to the physically constrained deep learning prediction module and the non-physical-constrained deep learning prediction module. The deterministic perception probabilistic prediction and spatial correlation output module and the physical constraint deep learning prediction module embed the photovoltaic cell single diode equivalent circuit equation, energy conservation constraints, and time smoothness physical regularization term as differentiable penalty terms into the joint loss function for model training. After training the model, the output deterministic power prediction sequence and temporal hidden state vector are transmitted to the uncertainty perception probabilistic prediction and spatial correlation output module. The uncertainty perception probabilistic prediction and spatial correlation output module adaptively determines the diffusion denoising step number based on the physical driving meteorological complexity index. With the deterministic power prediction sequence and temporal hidden state vector as conditional inputs, it generates the probability prediction trajectory and prediction interval based on the conditional denoising diffusion probability model according to the diffusion denoising step number. For areas containing two or more photovoltaic power plants, a dynamic spatiotemporal correlation graph is constructed. After node feature propagation and aggregation through a graph convolutional network, the final power prediction sequence of each photovoltaic power plant is output.

[0007] Furthermore, the physics-driven adaptive temporal decomposition module adds a spectral physics prior constraint term to the objective function of variational mode decomposition. This spectral physics prior constraint term constrains the center angular frequency of each modal component to be near a spectral physics prior center angular frequency that is sequentially assigned from low to high frequency according to the physical meaning of the solar diurnal frequency, semi-diurnal frequency, and seasonal frequency, based on the modal sequence number. In the physics-constrained deep learning prediction module, the physics constraint penalty term of the joint loss function comprises three components: a photovoltaic cell single-diode equivalent circuit constraint term, an energy conservation constraint term, and a time smoothness physics regularization term. The energy conservation constraint term imposes a non-negative constraint on the power prediction value at each prediction time step and is based on irradiance and photovoltaic module theory. The upper bound constraint of irradiance for maximum photoelectric conversion efficiency and the physical regularization term of time smoothness impose an upper bound constraint on the power change rate of the photovoltaic array physical ramp rate on adjacent prediction time steps. In the uncertainty-aware probability prediction and spatial correlation output module, the number of diffusion denoising steps is positively correlated with the physical driving meteorological complexity index. When the physical driving meteorological complexity index is lower than the preset low complexity threshold, the first diffusion step is used. When the physical driving meteorological complexity index is higher than the preset high complexity threshold, the second diffusion step is used, which is greater than the first diffusion step. When the physical driving meteorological complexity index is between the preset low complexity threshold and the preset high complexity threshold, the number of diffusion denoising steps increases linearly with the physical driving meteorological complexity index.

[0008] Furthermore, the data acquisition and preprocessing module obtains multi-source meteorological data and historical active power data from the numerical weather prediction data service system and the photovoltaic power station monitoring data acquisition system. It performs outlier screening and linear interpolation completion based on physical thresholds on the multi-source meteorological data and historical active power data, and outputs a standardized meteorological feature matrix and a standardized power sequence. The multi-source feature fusion module constructs a time-period encoding matrix that includes hourly sine coding, daily periodic coding and seasonal periodic coding. After concatenating the modal component matrix, meteorological feature matrix and time-period encoding matrix along the feature dimension, it maps them to a unified feature dimension space through a linear projection layer. The feature encoder of the physical constraint deep learning prediction module consists of two fully connected network layers. The temporal dependency extraction layer adopts a bidirectional temporal convolutional network structure and performs causal convolution operations on the encoded feature sequence on the forward and reverse paths. Then, it takes the hidden state of the last time step as the temporal hidden state vector. The prediction decoder consists of two fully connected network layers.

[0009] Furthermore, an artificial intelligence-based photovoltaic power output prediction method includes the following steps: Step S1: Collect numerical weather forecast data and historical power data of the photovoltaic power station in the area where the photovoltaic power station is located. Perform data cleaning and standardization preprocessing on the numerical weather forecast data and historical power data of the photovoltaic power station to obtain a standardized meteorological feature matrix and a standardized power sequence. Step S2: Based on the meteorological parameters in the standardized meteorological feature matrix, calculate the physical driving meteorological complexity index that integrates the solar elevation angle and atmospheric transparency index. Use the physical driving meteorological complexity index to determine the number of modes and penalty factor of variational mode decomposition. Add a spectral physical prior constraint term to the objective function of the constrained variational problem to decompose the standardized power sequence into multiple intrinsic mode components and a residual component. Step S3: Perform multi-source feature fusion on multiple intrinsic mode components, residual components, multi-dimensional meteorological feature vectors in the standardized meteorological feature matrix, and time period coding vectors to obtain a joint feature matrix; Step S4: Construct a deep learning prediction network that integrates photovoltaic physical conversion constraints. The deep learning prediction network includes a feature encoder, a temporal dependency extraction layer, and a prediction decoder. Input the joint feature matrix into the deep learning prediction network for model training. During the training process, the photovoltaic cell equivalent circuit equation, energy conservation constraints, and time smoothness physical regularization term are embedded as differentiable hard constraints into the joint loss function to obtain a trained deterministic prediction model. At the same time, the temporal dependency extraction layer outputs the hidden state vector. Step S5: After processing the meteorological data of the period to be predicted through the same process as in steps S1 to S4, input it into the trained deterministic prediction model to obtain the deterministic power prediction value. Adaptively determine the diffusion denoising step number based on the current physical-driven meteorological complexity index. Using the deterministic power prediction value and the hidden state vector output by the temporal dependency extraction layer as conditional information, perform reverse iterative denoising sampling through the conditional denoising diffusion probability model according to the diffusion denoising step number to generate the probability prediction trajectory and prediction interval. Step S6: For areas containing two or more photovoltaic power plants, construct a dynamic spatiotemporal correlation graph based on the geographical location information, meteorological conditions information and historical power information of each power plant. Use a graph convolutional network to propagate and aggregate node features of the dynamic spatiotemporal correlation graph and output the joint power prediction results of multiple power plants.

[0010] Further, in step S1, data cleaning and standardization preprocessing are performed on the numerical weather prediction data and historical power data of photovoltaic power plants to obtain a standardized meteorological feature matrix and a standardized power sequence, including the following steps: Step S11: Obtain multi-source meteorological data of the area where the photovoltaic power station is located from the numerical weather prediction data service system. The multi-source meteorological data includes total irradiance, diffuse irradiance, ambient temperature, relative humidity, wind speed, wind direction, cloud cover and air pressure. Obtain historical active power data of the photovoltaic power station from the photovoltaic power station monitoring data acquisition system. Arrange the multi-source meteorological data according to time steps to construct the original meteorological feature matrix. Arrange the historical active power data of the photovoltaic power station according to time steps to construct the original power sequence. Step S12: Perform a physical validity check on each data point in the original power sequence. Data points exceeding the preset upper and lower power thresholds are marked as outliers and removed. For data gaps caused by outlier removal, linear interpolation is performed on the arithmetic mean of the two adjacent valid data points at that position to complete the data gaps, resulting in a cleaned power sequence. The same outlier removal and linear interpolation operation is performed on the original meteorological feature matrix to obtain a cleaned meteorological feature matrix. Z-score standardization is then applied to both the cleaned meteorological feature matrix and the cleaned power sequence to obtain a standardized meteorological feature matrix. and normalized power sequence ,in The total number of time steps. This represents the total number of meteorological dimensions.

[0011] Further, in step S2, based on the meteorological parameters in the standardized meteorological feature matrix, a physical driving meteorological complexity index that integrates the solar altitude angle and atmospheric transparency index is calculated. This index is used to determine the number of modes and penalty factor in the variational mode decomposition. A spectral physical prior constraint term is added to the objective function of the constrained variational problem, decomposing the standardized power sequence into multiple intrinsic mode components and a residual component. This includes the following sub-steps: Step S21: From the standardized meteorological feature matrix Extract the irradiance, temperature, and cloud cover sequences for the current processing period, and calculate the solar altitude angle and atmospheric transparency index sequences. The formula for calculating the solar altitude angle is: ; in, For the first The solar altitude angle corresponding to each time step. This refers to the geographical latitude of the location of the photovoltaic power station. For the first The solar declination angle corresponding to each time step For the first The solar hour angle corresponding to each time step; The formula for calculating the atmospheric transparency index is: ; in, For the first Atmospheric transparency index corresponding to each time step. For the first Measured values ​​of total horizontal irradiance at each time step The solar constant is the constant outside the atmosphere. To prevent regularization constants with denominators of zero; Step S22: Calculate the physical driving meteorological complexity index based on the solar altitude angle sequence, atmospheric transparency index sequence, and irradiance sequence. The calculation formula is: ; in, As a physical-driven meteorological complexity index, This represents the standard deviation of the solar altitude angle sequence during the current processing period. This is the arithmetic mean of the solar altitude angle sequence within the current processing period. This represents the standard deviation of the atmospheric transparency index series during the current processing period. This is the arithmetic mean of the atmospheric transparency index series within the current processing period. The standard deviation of the irradiance sequence during the current processing period. This is the arithmetic mean of the irradiance sequence during the current processing period. To prevent regularization constants with denominators of zero; Step S23: Based on the physical-driven meteorological complexity index Calculate the number of modes in variational mode decomposition. and penalty factor The calculation formula is: ; ; in, The number of modes in the variational mode decomposition. Number of basic modes, This is the modality number adjustment coefficient. For floor operations, As a penalty factor, Based on the basic penalty factor value, The penalty factor decay coefficient, As a physical-driven meteorological complexity index, It is a natural exponential function; Step S24, with modal number and penalty factor As input parameters to the variational mode decomposition model, a spectral physics prior constraint term is added to the objective function of the constrained variational problem to construct an enhanced constrained variational problem. The objective function of the enhanced constrained variational problem is: ; The constraints are: ; in, For the first Each modal component in time The amplitude at that point, For the first The center angular frequency of each modal component For the number of modes, For the Dirac function, The imaginary unit, This represents the convolution operation. This represents taking the partial derivative with respect to the time variable. for Norm, For standardized power sequences In time The power value at that location, These are the weighting coefficients for the spectral physics prior constraint term. For the first The spectral physical prior center angular frequency of each modal component is assigned sequentially from low frequency to high frequency according to the physical meaning of the solar diurnal frequency, semi-diurnal frequency and seasonal frequency. Step S25: Iteratively solve the enhanced constraint variational problem using the alternating direction multiplier method to obtain... Each intrinsic mode component and a residual component The residual component is equal to the normalized power sequence minus the sum of all intrinsic mode components.

[0012] Further, in step S3, multiple intrinsic mode components, residual components, and multi-dimensional meteorological feature vectors and time period coding vectors in the standardized meteorological feature matrix are fused to obtain a joint feature matrix, including the following steps: Step S31, will The intrinsic modal components and residual components are vertically concatenated along the feature dimension to form the modal component matrix. From the standardized meteorological feature matrix Extract the multidimensional meteorological feature vector corresponding to each time step to construct a meteorological feature matrix. ; Step S32: Construct the time period encoding matrix The time cycle encoding matrix contains three dimensions: hourly sine encoding, daily cycle encoding, and seasonal cycle encoding. The hourly sine encoding value is obtained by mapping the number of hours corresponding to the time step onto the 24-hour cycle using a sine function. The daily cycle encoding value is obtained by mapping the day number of the time step in the current month onto the total number of days in the current month using a sine function. The seasonal cycle encoding value is obtained by mapping the month number corresponding to the time step onto the 12-month cycle using a cosine function. Step S33: Convert the modal component matrix Meteorological feature matrix and time period encoding matrix Horizontal concatenation is performed along the feature dimension to obtain a concatenated feature matrix. This concatenated feature matrix is ​​then input into a linear projection layer for dimension mapping to obtain a joint feature matrix. ,in is the feature dimension after projection.

[0013] Further, in step S4, a deep learning prediction network integrating photovoltaic physical conversion constraints is constructed. The joint feature matrix is ​​input into the deep learning prediction network for model training. During the training process, the photovoltaic cell equivalent circuit equation, energy conservation constraints, and time smoothness physical regularization term are embedded as differentiable hard constraint terms into the joint loss function to obtain the trained deterministic prediction model, including the following steps: Step S41: Construct a deep learning prediction network. The deep learning prediction network sequentially includes a feature encoder, a temporal dependency extraction layer, and a prediction decoder. The feature encoder consists of two fully connected layers, which will jointly generate the feature matrix. Mapped to encoded feature sequences The mapping calculation formula is: ; in, To encode feature sequences, For the joint characteristic matrix, This is the weight matrix of the first layer of the feature encoder. This is the weight matrix of the second layer of the feature encoder. This is the bias vector of the first layer of the feature encoder. This is the bias vector of the second layer of the feature encoder. To modify the activation function of the linear unit, The dimension of the encoded feature; Step S42: The temporal dependency extraction layer adopts a bidirectional temporal convolutional network structure to encode feature sequences on the forward path. Perform causal convolution operations to extract positive temporal dependency information from past to present time, and encode feature sequences on the reverse path. Causal convolution operations are performed to extract the inverse temporal dependency information from future time to the current time. The outputs of the forward path and the inverse path are concatenated along the feature dimension to obtain a bidirectional temporal feature sequence. The hidden state at the last time step is taken as the hidden state vector output by the temporal dependency extraction layer. ; Step S43: The prediction decoder consists of two fully connected layers, which convert the bidirectional temporal feature sequence... Mapped to deterministic power prediction sequence The mapping calculation formula is: ; in, For deterministic power prediction sequences, To predict the number of time steps, It is a bidirectional time-series characteristic sequence. To predict the weight matrix of the first layer of the decoder, To predict the weight matrix of the second layer of the decoder, To predict the bias vector of the first layer of the decoder, To predict the bias vector of the second layer of the decoder; Step S44: Embed the equivalent circuit equation of a single diode in the photovoltaic cell as a differentiable physical constraint term into the joint loss function. The equivalent circuit equation of a single diode in the photovoltaic cell is: ; in, To output current for photovoltaic cells, The output voltage of the photovoltaic cell. For photocurrent, This is the reverse saturation current of the diode. The electron charge constant, For series equivalent resistance, For diode ideality factor, Boltzmann's constant, The absolute temperature of the photovoltaic cell. For the area of ​​photovoltaic cells, For parallel equivalent resistance, It is a natural exponential function; Step S45: Construct a joint loss function that includes a data-driven loss term and a physical constraint penalty term. The calculation formula is: ; in, For the joint loss function value, For data-driven loss terms, For the equivalent circuit constraints of a single diode in a photovoltaic cell, For energy conservation constraints, For time smoothness physical regularization, For the constraint terms of the single-diode equivalent circuit of a photovoltaic cell, the weighting coefficients are... These are the weighting coefficients for the energy conservation constraint term. These are the weighting coefficients for the physical regularization term of time smoothness; Data-driven loss term The Huber loss function is used, and the calculation formula is as follows: ; ; in, For deterministic power prediction sequences The Middle Power prediction values ​​for each prediction time step For the first The actual power observations corresponding to each predicted time step To predict the number of time steps, To predict residuals, These are the inflection point parameters of the Huber loss function; Photovoltaic cell single diode equivalent circuit constraint terms The calculation formula is: ; in, For the first The maximum power point voltage of the photovoltaic array corresponding to each predicted time step. To be The calculated value of the photovoltaic cell output current is obtained by substituting the meteorological conditions corresponding to the current time step into the equivalent circuit equation of the single diode of the photovoltaic cell and solving it. Energy conservation constraint The calculation formula is: ; in, This represents the theoretical maximum photoelectric conversion efficiency of a photovoltaic module. For the first Total irradiance prediction for each prediction time step The effective light-receiving area of ​​the photovoltaic array. This is an operation that takes the larger of two parameters; Time smoothness physical regularization term The calculation formula is: ; in, For the first Power prediction values ​​for each prediction time step For the first Power prediction values ​​for each prediction time step This represents the physical maximum ramp rate of the photovoltaic array. This represents the time interval between adjacent prediction time steps.

[0014] Further, in step S5, the meteorological data for the period to be predicted is processed using the same procedures as in steps S1 to S4 and then input into the trained deterministic prediction model to obtain deterministic power prediction values. The number of diffusion denoising steps is adaptively determined based on the current physical-driven meteorological complexity index. Using the deterministic power prediction values ​​and the hidden state vector output by the temporal dependency extraction layer as conditional information, reverse iterative denoising sampling is performed according to the number of diffusion denoising steps using the conditional denoising diffusion probability model to generate a probability prediction trajectory and prediction interval. This includes the following steps: Step S51: Deterministic power prediction sequence The hidden state vector output at the last time step of the temporal dependency extraction layer Concatenation forms a condition vector The calculation formula is: ; in, For conditional vectors, For deterministic power prediction sequences, This is the hidden state vector output at the last time step of the temporal dependency extraction layer. This represents a vertical concatenation operation of vectors; Step S52: Based on the current physical-driven meteorological complexity index Adaptive determination of diffusion denoising steps The calculation formula is: ; in, To the number of diffusion denoising steps, To minimize the number of diffusion denoising steps, To maximize the number of diffusion denoising steps, As a physical-driven meteorological complexity index, To preset a low complexity threshold, To preset a high complexity threshold, when Less than Time to take ,when Greater than Time to take ; Step S53: Construct a conditional denoising network Conditional denoising network It consists of a three-layer residual convolutional network, with a diffusion time step Current noisy prediction residual sequence and condition vector As input, the output is the predicted noise. ,in This is the diffusion time step index, with a value range of 1- , To the number of diffusion denoising steps, For the set of trainable parameters of a conditional denoising network; Step S54: In the reverse denoising sampling stage, sample the initial noise sequence from the standard Gaussian distribution. According to The reverse denoising iterations are performed sequentially, decreasing to one. The calculation formula for each iteration is as follows: ; in, For the first The denoising results of each diffusion step For the first Noisy prediction residual sequence of diffusion steps, For the first Each diffusion step adds a coefficient to the forward noise. Add coefficients to the forward noise from step one to step two. The cumulative product of steps, The predicted noise output by the conditional denoising network. For the first Random noise scaling factor for each diffusion step. For a random noise sequence sampled from a standard Gaussian distribution, when hour, ,when hour, ; Step S55: The final output after the reverse denoising iteration is completed. As the prediction residual sequence Predicting residual sequences With deterministic power prediction sequence By adding each element, a probability prediction trajectory is obtained. Repeat steps S54 and this step. This time Probability prediction trajectory, for The predicted trajectory is calculated by taking the arithmetic mean at each prediction time step to obtain the predicted correction point value. The predicted value at the 6th time step is then taken. Quantiles are used to obtain the lower bound of the prediction interval, and the quantile is taken as the lower bound. Quantiles are used to determine the upper bound of the prediction interval, where The preset significance level parameter, This represents the total number of samples for probabilistic trajectory prediction.

[0015] Further, in step S6, for areas containing two or more photovoltaic power plants, a dynamic spatiotemporal correlation graph is constructed based on the geographical location information, meteorological condition information, and historical power information of each power plant. A graph convolutional network is then used to propagate and aggregate node features from the dynamic spatiotemporal correlation graph, outputting the joint power prediction result for multiple power plants. This includes the following steps: Step S61: Treat each photovoltaic power station within the region as a node in a dynamic spatiotemporal correlation graph, with the total number of nodes denoted as . Each node Node feature vectors The joint characteristic matrix of the power station The vector after global average pooling and the deterministic power prediction sequence of the power plant. It is pieced together; Step S62: Calculate every two nodes in the dynamic spatiotemporal correlation graph. With nodes Dynamic edge weights between The calculation formula is: ; in, For nodes With nodes Dynamic edge weights between them For the first The photovoltaic power station and the first Geographic Euclidean distance between photovoltaic power stations For distance attenuation parameters, For the first The mean vector of meteorological characteristics of each photovoltaic power station during the current processing period. For the first The mean vector of meteorological characteristics of each photovoltaic power station during the current processing period. for The transpose of , For vectors Norm, For the first The first photovoltaic power station and the first Covariance between the historical standardized power series of individual photovoltaic power plants For the first Standard deviation of the historical standardized power series of a photovoltaic power plant For the first Standard deviation of the historical standardized power series of a photovoltaic power plant This is a weighting coefficient for geographical distance similarity. The weighting coefficients for the similarity of meteorological conditions. Let be the weighting coefficients for power correlation, and satisfy . , It is a natural exponential function; Step S63: Using dynamic edge weights Construct an adjacency matrix , using node feature vectors Constructing the node feature matrix ,in The dimension of the node feature vector is used to process the node feature matrix through a graph convolutional network. Perform two-layer graph convolution propagation operations, the first... The formula for updating node features in layered graph convolutional propagation is: ; in, For the first The node feature matrix of the layer, For the first The node feature matrix of the layer, when hour , The adjacency matrix after adding self-loops, for An identity matrix of order 1. for The corresponding degree matrix and The diagonal elements of a diagonal matrix , for A diagonal matrix with a power of -1 / 2. For the first The trainable weight matrix of layer graph convolution, To modify the activation function of the linear unit; Step S64: The node feature matrix after two layers of graph convolution propagation The input is a fully connected output layer, which generates the final power prediction sequence for each photovoltaic power plant. The calculation formula is as follows: ; in, For the first The final power prediction sequence for each photovoltaic power plant. Node feature matrix The Middle The node feature vector corresponding to the row, This is the weight matrix of the fully connected output layer. This is the bias vector for the fully connected output layer.

[0016] The advantages of this invention compared to the prior art are: This invention constructs a physical-driven meteorological complexity index by introducing solar elevation angle and atmospheric transparency index, which dynamically adjusts the number of modes and penalty factor of variational mode decomposition with solar geometric parameters and atmospheric physical state. Furthermore, it adds spectral physical prior constraints corresponding to the solar diurnal and semi-diurnal cycles to the objective function, aligning the decomposed modal components with the physical frequencies of solar motion, thereby improving the decomposition quality fluctuation problem caused by fixed parameters in traditional variational mode decomposition.

[0017] This invention incorporates the equivalent circuit equation of a single diode in a photovoltaic cell, the energy conservation constraint, and the physical regularization term for time smoothness into a joint loss function of a deep learning prediction network as differentiable hard constraints. During model training, this ensures that the prediction results are continuously constrained by the photovoltaic cell's current-voltage characteristics, power non-negativity, irradiance upper bound, and physical ramp rate, thus preventing the prediction results from deviating from physical laws.

[0018] This invention uses a conditional denoising diffusion probability model by concatenating deterministic power prediction values ​​with temporal hidden state vectors as conditional vectors, and adaptively determines the number of diffusion denoising steps based on the physical-driven meteorological complexity index. This reduces the computational resource consumption under simple weather conditions while providing probability prediction results and prediction intervals.

[0019] This invention dynamically calculates the edge weights between multiple power stations based on three factors: geographical distance, similarity of meteorological conditions, and correlation of historical power, and constructs a time-varying adjacency matrix. This enables the graph structure to be automatically updated as spatial relationships change due to the movement of the meteorological system, thus improving the problem that static graph structures cannot reflect time-varying spatial relationships. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0021] In the attached diagram: Figure 1 This is a system framework diagram of an artificial intelligence-based photovoltaic power output prediction system in Example 1.

[0022] Figure 2 This is a flowchart of an artificial intelligence-based photovoltaic power output prediction method in Example 1. Detailed Implementation

[0023] The following detailed description of the embodiments is used to exemplify the principles of this application, but should not be used to limit the scope of this application. That is, the photovoltaic output prediction system and method based on artificial intelligence in this application are not limited to the described embodiments.

[0024] The present invention will be further described below with reference to embodiments.

[0025] Example 1 Step S1: Collect numerical weather forecast data and historical power data of the photovoltaic power station in the area where the photovoltaic power station is located, and perform data cleaning and standardization preprocessing. In a specific embodiment, step S1 obtains multi-source meteorological data for the area where the three photovoltaic power stations are located from the numerical weather prediction data service system provided by the provincial meteorological department. The meteorological data dimensions include eight dimensions: total irradiance, diffuse irradiance, ambient temperature, relative humidity, wind speed, wind direction, cloud cover, and air pressure. Simultaneously, historical active power data is obtained from the data acquisition and monitoring system of each photovoltaic power station. The meteorological data is arranged in 15-minute time steps to construct an original meteorological feature matrix, and the power data is arranged in the same time steps to construct an original power sequence. Taking 96 time steps per day as an example, the original meteorological feature matrix has a dimension of 96×8, and the original power sequence has a dimension of 96×1.

[0026] Physical validity checks were performed on each data point in the original power sequence. Data points with negative power values ​​or exceeding the rated installed capacity of the power station by 50 MW were marked as outliers and removed. For data gaps resulting from outlier removal, linear interpolation was performed using the arithmetic mean of the two adjacent valid data points at that position. The same outlier removal and linear interpolation were performed on the original meteorological feature matrix. Z-score standardization was then applied to the cleaned meteorological feature matrix and the power sequence. This involved calculating the mean and standard deviation for each dimension, subtracting the mean of the corresponding dimension from each data point, and then dividing by the standard deviation to obtain the standardized meteorological feature matrix. , dimension ,in The total number of time steps. The total number of meteorological dimensions is obtained, along with a standardized power sequence. , dimension Physical plausibility testing can eliminate erroneous data caused by sensor failure or communication interruption. Z-score standardization eliminates the differences in numerical ranges between different physical dimensions, bringing all dimensional features to the same numerical scale.

[0027] Step S2: Calculate the physical-driven meteorological complexity index, dynamically determine the variational mode decomposition parameters, and add spectral physical prior constraints. In a specific embodiment, step S2 extracts the irradiance sequence, temperature sequence, and cloud cover sequence for the current processing period from the standardized meteorological feature matrix, and calculates the solar altitude angle sequence and atmospheric transparency index sequence. Taking the first photovoltaic power station as an example, the geographical latitude of the power station location is given. The formula for calculating the solar altitude angle is: where is the solar altitude angle at the i-th time step, is the solar declination angle, and is the solar hour angle.

[0028] Substituting specific values, the solar declination angle around the vernal equinox on a certain observation day The solar hour angle corresponding to 10:00 AM Calculated Corresponding solar altitude angle .

[0029] The formula for calculating the atmospheric transparency index is: ,in For the first Atmospheric transparency index at each time step This represents the measured value of total irradiance on the horizontal plane. The solar constant is the constant outside the atmosphere. To prevent the regularization constant from being zero in the denominator, substitute the current time-to-time data. , Calculated This indicates that the current atmospheric conditions are relatively transparent.

[0030] Physics-driven meteorological complexity index The calculation formula is ,in The standard deviation of the solar altitude angle sequence. The arithmetic mean of the solar altitude angle sequence. The standard deviation of the atmospheric transparency index series. This is the arithmetic mean of the atmospheric transparency index series. The standard deviation of the irradiance sequence. The arithmetic mean of the irradiance sequence. To prevent the regularization constant from having a denominator of zero, the standard deviation of the irradiance series here corresponds to the irradiance volatility. Taking the data of the current processing period as an example, the mean of the solar altitude angle series... Standard deviation The mean of the atmospheric transparency index series Standard deviation Mean of irradiance sequence Standard deviation Substituting into the calculation, we get This value corresponds to a medium level of weather complexity.

[0031] Number of modes The calculation formula is ,in Number of basic modes, This is the modality number adjustment coefficient. This is for floor function. Penalty factor. The calculation formula is ,in Based on the basic penalty factor value, This is the penalty factor decay coefficient. Substitute it into... Calculated , Under sunny conditions hour, , Under cloudy to overcast weather conditions hour, , Traditional variational mode decomposition methods, which use fixed parameters, cannot adapt to the differences in signal characteristics under different weather conditions. This invention uses a physical-driven meteorological complexity index to drive the adaptive adjustment of decomposition parameters, so that the number of decomposed modes matches the complexity of the actual signal.

[0032] Add a spectral physics prior constraint to the objective function of variational mode decomposition, and the objective function is: The constraints are ,in For the first Each modal component in time The amplitude at that point, For the first The center angular frequency of each modal component For the number of modes, These are the weighting coefficients for the spectral physics prior constraint term. For the first The physical prior center angular frequency of the spectrum of each modal component To standardize the power sequence in time The power value at that location.

[0033] The a priori central angular frequencies of the spectrum are assigned sequentially from low to high frequency according to the physical meaning of the solar diurnal, semi-diurnal, and seasonal frequencies, based on their mode numbers. The a priori central angular frequency of the lowest frequency mode corresponds to the seasonal variation cycle. The intermediate frequency mode corresponds to the daily cycle. Higher frequency modes correspond to half-day cycles The prior center angular frequencies of the remaining high-frequency modes are set to a uniform distribution to allow for free learning of the rapid fluctuation components caused by cloud shading. The reason for choosing the daily and semi-daily cycles as prior frequencies is that photovoltaic power generation is driven by solar motion, and its power sequence must contain oscillating components with 24-hour and 12-hour cycles.

[0034] The augmented constraint variational problem is solved iteratively using the alternating direction multiplier method, with each modal waveform updated sequentially in each iteration. Center frequency The Lagrange multipliers converged after 200 iterations, yielding the following result. Each intrinsic mode component and a residual component Residual components It equals the normalized power sequence minus the sum of all intrinsic mode components, i.e. .by For example, the decomposition yields four intrinsic mode components and one residual component. The spectral physical prior constraint term constrains the center frequency of each mode to be near the physical prior frequency, ensuring that the center frequency of the diurnal mode does not deviate from the true solar diurnal frequency. The decomposed mode components have clear physical meanings, corresponding to components such as seasonal trends, diurnal variations, semi-diurnal fluctuations, and rapid changes in cloud cover, thus avoiding the problems of mode aliasing and unclear physical meaning that are prone to occur in traditional variational mode decomposition.

[0035] Step S3: Perform multi-source feature fusion on modal components, residual components, meteorological features, and time period codes. In a specific embodiment, step S3 involves vertically concatenating the intrinsic mode components and residual components along the feature dimension to form a mode component matrix with dimension 5. For example, the mode component matrix has a dimension of 96×5. Then, an 8-dimensional meteorological feature vector corresponding to each time step is extracted from the standardized meteorological feature matrix to form a meteorological feature matrix with a dimension of 96×8.

[0036] Constructing a time-period encoding matrix , dimension It includes three dimensions: hourly sine code, daily cycle code, and seasonal cycle code. The hourly sine code is calculated as follows: ,in For the number of hours corresponding to this time step, when Time-coded value The calculation method for daily cycle encoding is as follows: ,in This is the day number of the current month in which the time step is located. The total number of days in the month. and Time-coded value The calculation method for seasonal cycle codes is as follows: ,in The month number corresponding to this time step, when Time-coded value The time-period encoding matrix explicitly encodes the periodic information of solar motion into numerical features, enabling the prediction network to directly perceive the periodic modulation effect of day-night alternation and seasonal changes on photovoltaic power generation.

[0037] modal component matrix Meteorological feature matrix and time period encoding matrix Perform horizontal concatenation along the feature dimension to obtain a concatenated feature matrix with dimension . The concatenated feature matrix is ​​input into a linear projection layer for dimension mapping to obtain the joint feature matrix. , dimension ,in This represents the unified feature dimension after projection. The linear projection layer uses a trainable weight matrix. and bias vector The mapping calculation method is as follows The linear projection layer aligns feature information from different sources into a unified semantic space, enabling features of different dimensions to interact on the same numerical scale. The modal component matrix contains the decomposition components of the power sequence at different time scales, the meteorological feature matrix provides external driving factors affecting photovoltaic power generation, and the time period encoding matrix explicitly expresses the periodicity information of solar motion. After the fusion of these three types of information, the prediction network can simultaneously perceive the power evolution law, meteorological influence, and astronomical periodic characteristics.

[0038] Step S4: Construct a deep learning prediction network that integrates photovoltaic physical conversion constraints and embeds a joint loss function of triple physical constraints for training. like Figure 1 As shown in the physical constraint deep learning prediction module, in a specific embodiment, step S4 constructs a deep learning prediction network that integrates photovoltaic physical conversion constraints. This network sequentially includes a feature encoder, a temporal dependency extraction layer, and a prediction decoder. The feature encoder consists of two fully connected layers. The first layer processes the joint feature matrix... A nonlinear transformation is performed, and the second layer further maps the result to an encoded feature sequence. , dimension ,in This represents the dimension of the encoded features. The mapping calculation formula is as follows: ,in This is the weight matrix of the first layer of the feature encoder, with dimensions of 128×64. This is the weight matrix of the second layer of the feature encoder, with dimensions of 64×64. and This represents the corresponding bias vector. Taking the first photovoltaic power station as an example, the joint feature matrix... The matrix has a dimension of 96×128. After transformation by the first fully connected network layer, an intermediate feature matrix with a dimension of 96×64 is obtained. After introducing nonlinearity through the ReLU activation function, it undergoes a linear transformation by the second fully connected network layer, outputting the encoded feature sequence. The dimension is 96×64. The role of the feature encoder is to compress the high-dimensional joint features into a lower-dimensional representation space, thereby reducing computational complexity while retaining key information that is discriminative for power prediction.

[0039] The temporal dependency extraction layer employs a bidirectional temporal convolutional network structure. Along the forward path, causal convolutions are used to extract temporal dependency information layer by layer from the first time step to the last. The kernel size of the causal convolution is 3, and zero-padding is used to maintain the output sequence length. Each time step is a separate step. On the reverse path, the same causal convolutional structure is used to perform causal convolution operations from the last time step to the first time step, extracting reverse temporal dependency information. The outputs of the forward and reverse paths are concatenated along the feature dimension to obtain a bidirectional temporal feature sequence. , dimension The hidden state at the last time step is taken as the hidden state vector output by the temporal dependency extraction layer. , dimension This vector is the temporal hidden state vector. The reason for choosing a bidirectional temporal convolutional network is that photovoltaic power generation is not only affected by past meteorological conditions, but also constrained by future meteorological trends. The forward path captures the evolution of historical meteorological conditions, while the reverse path provides reference information on future trends. The two are combined to form a more complete temporal representation.

[0040] The predictive decoder consists of two fully connected layers that process bidirectional temporal feature sequences. Mapped to deterministic power prediction sequence , dimension ,in This represents the number of time steps to predict, corresponding to the prediction range for the next hour. The mapping calculation formula is as follows: ,in The weight matrix for the first layer of the predictor decoder has dimensions of 128×64. The weight matrix for the second layer of the predictor decoder has a dimension of 64×1. and This is the corresponding bias vector.

[0041] During training, the equivalent circuit equation of a single diode in a photovoltaic cell is embedded as a differentiable physical constraint term into the joint loss function. The equivalent circuit equation of a single diode is: ,in To output current for photovoltaic cells, The output voltage of the photovoltaic cell. For photocurrent, This is the reverse saturation current of the diode. The electron charge constant, For series equivalent resistance, For diode ideality factor, Boltzmann's constant, The absolute temperature of the photovoltaic cell. For the area of ​​photovoltaic cells, This is the parallel equivalent resistance. Taking the monocrystalline silicon photovoltaic modules used in the first photovoltaic power station as an example, under standard test conditions, i.e., an irradiance of 1000 W / m... 2 At a battery temperature of 25°C, the equivalent circuit parameters obtained through parameter identification are the photocurrent. Diode reverse saturation current Series equivalent resistance Parallel equivalent resistance Diode ideality factor Under non-standard conditions, the photocurrent changes linearly with irradiance, while the diode reverse saturation current changes exponentially with temperature. Taking the current moment as an example, the maximum power point voltage... Substituting the meteorological conditions such as irradiance and temperature at the time into the equivalent circuit equation of a single diode, the calculated value of the photovoltaic cell output current is obtained. This calculated value is used to calculate the physical constraint loss term.

[0042] Joint loss function The calculation formula is ,in For data-driven loss terms, For the equivalent circuit constraints of a single diode in a photovoltaic cell, For energy conservation constraints, For time smoothness physical regularization, These are the weighting coefficients for the equivalent circuit constraint terms. These are the weighting coefficients for the energy conservation constraint term. The weighting coefficients for the physical regularization term of time smoothness.

[0043] Data-driven loss term The Huber loss function is used, and the calculation formula is as follows: ,in exist Time equals ,exist Time equals , For the first Power prediction values ​​for each prediction time step For the corresponding actual power observation value, This is the inflection point parameter for the Huber loss function. Squared loss is used when the absolute value of the prediction residual is less than or equal to 1.0, and linear loss is used when the absolute value of the prediction residual is greater than 1.0. Compared to mean squared error loss, the Huber loss function is more robust to outliers. The contribution of occasional outliers in the power data to the loss function is limited to linear growth rather than squared growth, reducing the excessive influence of outliers on the gradient direction.

[0044] Photovoltaic cell single diode equivalent circuit constraint terms The calculation formula is ,in For the first The maximum power point voltage of the photovoltaic array corresponding to each predicted time step. To be The calculated output current is obtained by substituting the meteorological conditions corresponding to the current time step into the single-diode equivalent circuit equation. This term measures the deviation between the predicted power and the physical power determined by the equivalent circuit equation.

[0045] Energy conservation constraint The calculation formula is ,in This represents the theoretical maximum photoelectric conversion efficiency of a photovoltaic module. For the first Total irradiance prediction for each prediction time step This represents the effective light-receiving area of ​​the photovoltaic array. The first term imposes a non-negative constraint on the predicted power value, penalizing negative predicted power. The second term imposes an upper bound constraint on irradiance; the theoretical upper limit power equals irradiance multiplied by photoelectric conversion efficiency multiplied by the effective light-receiving area. Substituting specific values, when the irradiance... At that time, the theoretical upper limit of power is A penalty term is introduced if the model predicts power exceeding 80 kW. When the irradiance... At that time, the theoretical upper limit of power is This constraint ensures that the predicted power remains within a physically reasonable range.

[0046] Time smoothness physical regularization term The calculation formula is ,in This represents the physical maximum ramp rate of the photovoltaic array. This is the time interval between adjacent prediction time steps. When the absolute value of the power change between adjacent time steps exceeds the physically permissible maximum change... A penalty is incurred at certain times. Substituting specific values, if the predicted power values ​​for two adjacent prediction time steps are 35.2 MW and 39.0 MW, the power change is... If the allowable limit of 2.5 MW is exceeded, a penalty value will be applied to this item. This constraint reflects the physical fact that photovoltaic arrays are limited by the rate of change of solar radiation and the response speed of the inverter.

[0047] Triple physical constraints impose constraints on the prediction results from different dimensions. The equivalent circuit constraint ensures that the predicted power conforms to the volt-ampere characteristic curve of the photovoltaic cell; the energy conservation constraint ensures that the predicted power is within a reasonable range from the perspective of macroscopic energy balance; and the time smoothness constraint constrains the temporal continuity of the prediction sequence from the perspective of the power change rate. These triple physical constraints directly affect the model parameter updates through gradient backpropagation, allowing physical laws to continuously guide model learning during training. This method of directly embedding physical constraints into the loss function and coupling them through gradients, compared to the scheme of fusing the physical model and the data-driven model after they run independently, ensures that the model is continuously guided by physical laws during training.

[0048] Step S5: Adaptively determine the number of diffusion denoising steps based on the physical-driven meteorological complexity index, and generate probability prediction results through the conditional denoising diffusion probability model. like Figure 2 As shown in step S5, in a specific embodiment, step S5 first processes the meteorological data for the period to be predicted through the same process as steps S1 to S4, and then inputs it into the trained deterministic prediction model to obtain the deterministic power prediction sequence. Then the deterministic power prediction sequence The hidden state vector output at the last time step of the temporal dependency extraction layer Concatenation forms a condition vector The calculation formula is: ,in This represents the vertical concatenation operation of vectors. Taking a prediction period comprising 4 time steps totaling 1 hour as an example, the deterministic power prediction sequence... The dimension is 4×1, and the hidden state vector is... The dimension is 128×1, and the concatenated condition vector The dimension is 132×1. The conditional vector contains both deterministic power prediction information for future periods and a compressed representation of historical time series features, providing sufficient prediction context for the conditional denoising network.

[0049] Based on the current physical-driven meteorological complexity index Adaptive determination of diffusion denoising steps The calculation formula is: ,in To minimize the number of diffusion denoising steps, To maximize the number of diffusion denoising steps, To preset a low complexity threshold, A high complexity threshold is preset. When Less than Time to take ,when Greater than Time to take Substitute Calculated Step. Under sunny conditions. Below the low complexity threshold, the diffusion denoising step count is set to 50 steps. Under complex weather conditions... If the complexity exceeds the high complexity threshold, the diffusion denoising step count is set to 500 steps.

[0050] Constructing a conditional denoising network The network consists of three layers of residual convolutional networks, with a diffusion time step. Current noisy prediction residual sequence and condition vector As input, the output is the predicted noise. Residual convolutional networks employ skip connections to mitigate the vanishing gradient problem in deep networks. Each residual block contains two convolutional operations and a ReLU activation function. (Conditional vector) After being mapped to the same dimension as the input sequence through a fully connected layer, it is concatenated with the noisy sequence along the channel dimension. The reason for choosing a residual convolutional network is that convolution operations can effectively capture the local temporal dependencies in the predicted residual sequence, while residual connections enable the network to learn deeper feature representations.

[0051] In the reverse denoising sampling stage, the initial noise sequence is sampled from a standard Gaussian distribution. According to The reverse denoising iterations are performed sequentially, decreasing to 1. The calculation formula for each iteration is as follows: ,in For the first The denoising results of each diffusion step For the first Noisy prediction residual sequence of diffusion steps, For the first Each diffusion step adds a coefficient to the forward noise. Add coefficients to the forward noise from step one to step two. The cumulative product of steps, The predicted noise output by the conditional denoising network. For the first Random noise scaling factor for each diffusion step. This is a random noise sequence sampled from a standard Gaussian distribution. In the first few iterations, the noisy sequence... It is mainly dominated by large-amplitude random noise, and the conditional denoising network is based on the conditional vector. The deterministic power prediction information is used to identify and remove major noise components. As the number of iterations decreases, structural features related to the prediction residuals gradually emerge in the noisy sequence. In the final step... At that time, without adding additional random noise, It outputs a deterministic prediction residual sequence.

[0052] The output after the reverse denoising iteration As the prediction residual sequence ,Will With deterministic power prediction sequence Element-by-element summation yields a probability prediction trajectory. This sampling process is repeated for a total of... This process yields 50 probability prediction trajectories. The arithmetic mean of these 50 trajectories is then taken at each prediction time step to obtain the predicted correction point value. The corrected point prediction value is then calculated by taking the 1st... Quantiles are used to obtain the lower bound of the prediction interval, and the quantile is taken as the lower bound. Quantiles are used to determine the upper bound of the prediction interval, where The preset significance level parameter corresponds to a 95% confidence level for the prediction interval. The corrected point prediction value eliminates the randomness of a single sampling by averaging multiple probability prediction trajectories, exhibiting higher robustness compared to a single deterministic prediction. The prediction interval provides quantitative information on the uncertainty of power prediction, enabling grid dispatching departments to rationally allocate spinning reserve capacity. The diffusion denoising steps are adaptively adjusted based on the physics-driven meteorological complexity index. In simple scenarios such as clear weather, only 50 steps are needed to obtain an effective probability distribution estimate, reducing computational time compared to a fixed-step scheme. Under complex weather conditions, the number of steps is increased to 500 to fully characterize the uncertainty features of power prediction.

[0053] Step S6: Construct a dynamic spatiotemporal correlation graph and output the joint power prediction results of multiple power plants through a graph convolutional network. like Figure 2 As shown in step S6, in a specific embodiment, step S6 constructs a dynamic spatiotemporal relationship graph for the three photovoltaic power plants. Each power plant is treated as a node in the graph, with a total number of nodes... Each node Node feature vectors The joint characteristic matrix of the power station Vector and deterministic power prediction sequence after global average pooling Concatenated, global average pooling will divide the dimension into... The joint feature matrix is ​​averaged along the time dimension to obtain a matrix of dimension 1. The feature vectors of dimension are then combined with the feature vectors of dimension . The deterministic power prediction sequences are concatenated to obtain a sequence with dimension . The node feature vectors.

[0054] Calculate every two nodes With nodes Dynamic edge weights between The calculation formula is: ,in For the first The photovoltaic power station and the first Geographic Euclidean distance between photovoltaic power stations For distance attenuation parameters, For the first The mean vector of meteorological characteristics of each photovoltaic power station during the current processing period. For the first The mean vector of meteorological characteristics of a photovoltaic power station For the first The and the first Covariance between the historical standardized power series of individual power plants For the first Standard deviation of the historical standardized power series of a power plant For the first Standard deviation of the historical standardized power series of a power plant This is a weighting coefficient for geographical distance similarity. The weighting coefficients for the similarity of meteorological conditions. These are the weighting coefficients for power correlation, and the sum of the three is 1.

[0055] The dynamic edge weights are composed of three weighted factors. The first factor is geographical distance similarity, calculated using a Gaussian kernel function; the closer the power stations are, the higher the similarity. The second factor is meteorological condition similarity, calculated using the cosine similarity of the mean vectors of the meteorological characteristics of the two power stations; the closer the meteorological conditions, the higher the similarity between the power stations. The third factor is power correlation, calculated using the Pearson correlation coefficient between the historical standardized power series of the two power stations; the more similar the power fluctuation patterns, the stronger the correlation between the power stations.

[0056] Taking the first and second power stations as examples, the geographical distance between the two power stations Geographical distance similarity is The mean vector of meteorological characteristics of the first power station The mean vector of meteorological characteristics of the second power station The cosine similarity, calculated as the dot product of two vectors divided by the product of their respective L2 norms, is 0.998. The Pearson correlation coefficient between the historical power series of the two power stations is 0.75. Substituting the weighting coefficients, the result is... .

[0057] Taking the first and third power stations as examples, the geographical distance between the two power stations Geographical distance similarity is The Pearson correlation coefficient for the historical power series is 0.55, and the cosine similarity for meteorological conditions is 0.92. Substituting these values ​​into the weighting coefficients, the calculated... The comparison shows that the edge weight of 0.814 between the first and second power stations, which are closer and have similar meteorological conditions, is significantly greater than the edge weight of 0.495 between the first and third power stations, which are farther apart. The calculation results of the dynamic edge weight are consistent with the actual impact of geographical distance and meteorological conditions on the relationship between power stations.

[0058] With dynamic edge weights Construct a 3×3 adjacency matrix , using node feature vectors Constructing the node feature matrix The dimension is 3×132. Node features are updated through two layers of graph convolution propagation operations. The formula for updating node features in layer graph convolution is as follows: ,in For the first The node feature matrix of the layer, when hour , The adjacency matrix after adding self-loops, It is a 3-order identity matrix. for The corresponding degree matrix, For the first The trainable weight matrix for layered graph convolution. The significance of adding self-loops is that each node retains its own feature information while aggregating neighbor information. Degree matrix. It is a diagonal matrix whose diagonal elements are equal to its adjacency matrix. The sum of elements in the corresponding row. Symmetric normalized form. This ensures that the aggregation of node features is unaffected by node degree, preventing nodes with high height from having excessive weight during feature propagation. Taking the first layer of graph convolution as an example... The graph has a dimension of 132×64. After graph convolution propagation, the features of each node are fused with its own features and the features of its neighboring nodes. The node features of the first power station incorporate more feature information from the second power station due to the larger edge weight between it and the second power station, while the node features of the third power station incorporate relatively less feature information from the third power station due to the smaller edge weight between them.

[0059] The node feature matrix after two layers of graph convolution propagation The fully connected input layer generates a power prediction sequence for each photovoltaic power plant. The calculation formula is as follows: ,in For the first Power prediction sequence for a photovoltaic power plant Node feature matrix The Middle The node feature vector corresponding to the row, This is the weight matrix of the fully connected output layer, with dimensions 64×4. This is the bias vector of the fully connected output layer, with a dimension of 4×1.

[0060] Compared to static graph structures, dynamic spatiotemporal correlation graphs constructed using three-factor dynamic edge weights more accurately reflect the actual relationships between power plants. Geographical distance, meteorological condition similarity, and power correlation characterize the spatial correlation between power plants from different dimensions. As weather systems move within a region, meteorological condition similarity and power correlation change accordingly, and the dynamic edge weights automatically track this time-varying characteristic. Graph convolutional networks achieve information exchange among multiple power plants through feature propagation on the adjacency matrix, enabling power prediction for each power plant to not only utilize local information but also integrate correlation information from other power plants within the region. This has a significant advantage in capturing large-scale power fluctuations caused by weather system movement.

[0061] The power prediction sequences and probability prediction ranges of the three photovoltaic power plants are transmitted to the energy management system of the provincial power grid dispatch center for dispatchers to arrange day-ahead power generation plans and spinning reserve capacity.

[0062] It should be noted that the combination of the technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0063] It should also be noted that the embodiments listed above are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and similar changes or modifications made thereto are those that can be directly derived or easily conceived by those skilled in the art from the content disclosed in the present invention, and should all fall within the protection scope of the present invention.

Claims

1. An artificial intelligence based photovoltaic power output prediction system, characterized in that, It includes a data acquisition and preprocessing module, a physics-driven adaptive temporal decomposition module, a multi-source feature fusion module, a physics-constrained deep learning prediction module, and an uncertainty-aware probability prediction and spatial correlation output module. The data acquisition and preprocessing module transmits the standardized meteorological feature matrix and standardized power sequence obtained after outlier removal and standardization to the physical-driven adaptive temporal decomposition module. The physical-driven adaptive temporal decomposition module calculates the physical-driven meteorological complexity index based on the solar altitude angle, atmospheric transparency index, and irradiance fluctuation rate. This index dynamically determines the number of modes and penalty factor in the variational mode decomposition. After performing physical-driven multi-scale temporal decomposition on the standardized power sequence, the module transmits the multiple intrinsic mode components and one residual component obtained to the multi-source feature fusion module. The multi-source feature fusion module concatenates the multiple intrinsic mode components, residual component, multi-dimensional meteorological feature vector, and time period encoding vector along the feature dimension and maps them to a unified feature dimension space through a linear projection layer. The resulting joint feature matrix is ​​then transmitted to the physically constrained deep learning prediction module. The uncertainty-aware probability prediction and spatial correlation output module, in which the physical constraint deep learning prediction module embeds the photovoltaic cell single diode equivalent circuit equation, energy conservation constraint, and time smoothness physical regularization term as differentiable penalty terms into the joint loss function for model training, transmits the output deterministic power prediction sequence and temporal hidden state vector to the uncertainty-aware probability prediction and spatial correlation output module. The uncertainty-aware probability prediction and spatial correlation output module adaptively determines the diffusion denoising step number based on the physical driving meteorological complexity index. With the deterministic power prediction sequence and temporal hidden state vector as conditional inputs, it generates probability prediction trajectories and prediction intervals based on the conditional denoising diffusion probability model according to the diffusion denoising step number. For regions containing two or more photovoltaic power stations, it constructs a dynamic spatiotemporal correlation graph, and outputs the final power prediction sequence for each photovoltaic power station after node feature propagation and aggregation through a graph convolutional network.

2. The photovoltaic power output prediction system based on artificial intelligence according to claim 1, characterized in that, The physical-driven adaptive temporal decomposition module adds a spectral physical prior constraint term to the objective function of variational mode decomposition. This spectral physical prior constraint term constrains the center angular frequency of each modal component to a range from low to high frequency, allocated according to the physical meaning of the solar diurnal, semi-diurnal, and seasonal cycles, based on the modal sequence number. In the physical-constrained deep learning prediction module, the physical constraint penalty term of the joint loss function comprises three components: a photovoltaic cell single-diode equivalent circuit constraint term, an energy conservation constraint term, and a time smoothness physical regularization term. The energy conservation constraint term applies a non-negative constraint to the power prediction value at each prediction time step and is based on the theoretical maximum of irradiance and photovoltaic modules. The upper bound constraint of irradiance for photoelectric conversion efficiency, and the physical regularization term of time smoothness apply an upper bound constraint of the photovoltaic array physical ramp rate to the power change rate between adjacent prediction time steps; in the uncertainty perception probability prediction and spatial correlation output module, the number of diffusion denoising steps is positively correlated with the physical driving meteorological complexity index. When the physical driving meteorological complexity index is lower than the preset low complexity threshold, the first diffusion step is used; when the physical driving meteorological complexity index is higher than the preset high complexity threshold, the second diffusion step is used, which is greater than the first diffusion step; when the physical driving meteorological complexity index is between the preset low complexity threshold and the preset high complexity threshold, the number of diffusion denoising steps increases linearly with the physical driving meteorological complexity index.

3. The photovoltaic output prediction system based on artificial intelligence according to claim 1, characterized in that, The data acquisition and preprocessing module obtains multi-source meteorological data and historical active power data from the numerical weather prediction data service system and the photovoltaic power station monitoring data acquisition system. It performs outlier screening and linear interpolation completion based on physical thresholds on the multi-source meteorological data and historical active power data, and outputs a standardized meteorological feature matrix and a standardized power sequence. The multi-source feature fusion module constructs a time-period encoding matrix including hourly sine coding, daily periodic coding and seasonal periodic coding. The modal component matrix, meteorological feature matrix and time-period encoding matrix are concatenated along the feature dimension and mapped to a unified feature dimension space through a linear projection layer. The feature encoder of the physical constraint deep learning prediction module consists of two fully connected network layers. The temporal dependency extraction layer adopts a bidirectional temporal convolutional network structure and performs causal convolution operations on the encoded feature sequence on the forward and reverse paths. The hidden state of the last time step is taken as the temporal hidden state vector. The prediction decoder consists of two fully connected network layers.

4. A photovoltaic power output prediction method based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Collect numerical weather forecast data and historical power data of the photovoltaic power station in the area where the photovoltaic power station is located. Perform data cleaning and standardization preprocessing on the numerical weather forecast data and historical power data of the photovoltaic power station to obtain a standardized meteorological feature matrix and a standardized power sequence. Step S2: Based on the meteorological parameters in the standardized meteorological feature matrix, calculate the physical driving meteorological complexity index that integrates the solar elevation angle and atmospheric transparency index. Use the physical driving meteorological complexity index to determine the number of modes and penalty factor of variational mode decomposition. Add a spectral physical prior constraint term to the objective function of the constrained variational problem to decompose the standardized power sequence into multiple intrinsic mode components and a residual component. Step S3: Perform multi-source feature fusion on multiple intrinsic mode components, residual components, multi-dimensional meteorological feature vectors in the standardized meteorological feature matrix, and time period coding vectors to obtain a joint feature matrix; Step S4: Construct a deep learning prediction network that integrates photovoltaic physical conversion constraints. The deep learning prediction network includes a feature encoder, a temporal dependency extraction layer, and a prediction decoder. Input the joint feature matrix into the deep learning prediction network for model training. During the training process, the photovoltaic cell equivalent circuit equation, energy conservation constraints, and time smoothness physical regularization term are embedded as differentiable hard constraints into the joint loss function to obtain a trained deterministic prediction model. At the same time, the temporal dependency extraction layer outputs the hidden state vector. Step S5: After processing the meteorological data of the period to be predicted through the same process as in steps S1 to S4, input it into the trained deterministic prediction model to obtain the deterministic power prediction value. Adaptively determine the diffusion denoising step number according to the current physical-driven meteorological complexity index. Using the deterministic power prediction value and the hidden state vector output by the temporal dependency extraction layer as conditional information, perform reverse iterative denoising sampling through the conditional denoising diffusion probability model according to the diffusion denoising step number to generate the probability prediction trajectory and prediction interval. Step S6: For areas containing two or more photovoltaic power plants, construct a dynamic spatiotemporal correlation graph based on the geographical location information, meteorological condition information and historical power information of each power plant. Perform node feature propagation and aggregation on the dynamic spatiotemporal correlation graph through a graph convolutional network, and output the joint power prediction result of multiple power plants.

5. The photovoltaic power output prediction method based on artificial intelligence according to claim 4, characterized in that, In step S1, data cleaning and standardization preprocessing are performed on numerical weather forecast data and historical power data of photovoltaic power plants to obtain a standardized meteorological feature matrix and a standardized power sequence, including the following steps: Step S11: Obtain multi-source meteorological data of the area where the photovoltaic power station is located from the numerical weather prediction data service system. The multi-source meteorological data includes total irradiance, diffuse irradiance, ambient temperature, relative humidity, wind speed, wind direction, cloud cover and air pressure. Obtain historical active power data of the photovoltaic power station from the photovoltaic power station monitoring data acquisition system. Arrange the multi-source meteorological data according to time steps to construct the original meteorological feature matrix. Arrange the historical active power data of the photovoltaic power station according to time steps to construct the original power sequence. Step S12: Perform a physical validity check on each data point in the original power sequence. Data points exceeding preset upper and lower power thresholds are marked as outliers and removed. For data gaps caused by outlier removal, linear interpolation is performed on the arithmetic mean of the two adjacent valid data points at that position to complete the data gaps, resulting in a cleaned power sequence. The same outlier removal and linear interpolation operation is performed on the original meteorological feature matrix to obtain a cleaned meteorological feature matrix. Z-score standardization is then performed on the cleaned meteorological feature matrix and the cleaned power sequence to obtain a standardized meteorological feature matrix. and normalized power sequence ,in The total number of time steps. This represents the total number of meteorological dimensions.

6. The photovoltaic power output prediction method based on artificial intelligence according to claim 4, characterized in that, In step S2, based on the meteorological parameters in the standardized meteorological feature matrix, a physical driving meteorological complexity index that integrates the solar altitude angle and atmospheric transparency index is calculated. This index is used to determine the number of modes and penalty factor in the variational mode decomposition. A spectral physical prior constraint term is added to the objective function of the constrained variational problem, decomposing the standardized power sequence into multiple intrinsic mode components and a residual component. This includes the following sub-steps: Step S21: From the standardized meteorological feature matrix Extract the irradiance, temperature, and cloud cover sequences for the current processing period, and calculate the solar altitude angle and atmospheric transparency index sequences. The formula for calculating the solar altitude angle is: ； in, For the first The solar altitude angle corresponding to each time step. This refers to the geographical latitude of the location of the photovoltaic power station. For the first The solar declination angle corresponding to each time step For the first The solar hour angle corresponding to each time step; The formula for calculating the atmospheric transparency index is as follows: ； in, For the first Atmospheric transparency index corresponding to each time step. For the first Measured values ​​of total horizontal irradiance at each time step The solar constant is the constant outside the atmosphere. To prevent regularization constants with denominators of zero; Step S22: Calculate the physical driving meteorological complexity index based on the solar altitude angle sequence, atmospheric transparency index sequence, and irradiance sequence. The calculation formula is: ； in, As a physical-driven meteorological complexity index, This represents the standard deviation of the solar altitude angle sequence during the current processing period. This is the arithmetic mean of the solar altitude angle sequence within the current processing period. This represents the standard deviation of the atmospheric transparency index series during the current processing period. This is the arithmetic mean of the atmospheric transparency index series within the current processing period. The standard deviation of the irradiance sequence during the current processing period. This is the arithmetic mean of the irradiance sequence during the current processing period. To prevent regularization constants with denominators of zero; Step S23: Based on the physical-driven meteorological complexity index Calculate the number of modes in variational mode decomposition. and penalty factor The calculation formula is: ； ； in, The number of modes in the variational mode decomposition. Number of basic modes, This is the modality number adjustment coefficient. For floor operations, As a penalty factor, Based on the basic penalty factor value, The penalty factor decay coefficient, As a physical-driven meteorological complexity index, It is a natural exponential function; Step S24, with modal number and penalty factor As input parameters to the variational mode decomposition model, a spectral physics prior constraint term is added to the objective function of the constrained variational problem to construct an enhanced constrained variational problem. The objective function of the enhanced constrained variational problem is: ； The constraints are: ； in, For the first Each modal component in time The amplitude at that point, For the first The center angular frequency of each modal component For the number of modes, For the Dirac function, The imaginary unit, This represents the convolution operation. This represents taking the partial derivative with respect to the time variable. for Norm, For standardized power sequences In time The power value at that location, These are the weighting coefficients for the spectral physics prior constraint term. For the first The spectral physical prior center angular frequency of each modal component is allocated sequentially from low frequency to high frequency according to the physical meaning of the solar diurnal frequency, semi-diurnal frequency and seasonal frequency. Step S25: Iteratively solve the enhanced constraint variational problem using the alternating direction multiplier method to obtain... Each intrinsic mode component and a residual component The residual component is equal to the normalized power sequence minus the sum of all intrinsic mode components.

7. The photovoltaic power output prediction method based on artificial intelligence according to claim 4, characterized in that, In step S3, multiple intrinsic mode components, residual components, multi-dimensional meteorological feature vectors in the standardized meteorological feature matrix, and time period coding vectors are fused to obtain a joint feature matrix, including the following steps: Step S31, will The intrinsic modal components and residual components are vertically concatenated along the feature dimension to form the modal component matrix. From the standardized meteorological feature matrix Extract the multidimensional meteorological feature vector corresponding to each time step to construct a meteorological feature matrix. ; Step S32: Construct the time period encoding matrix The time cycle encoding matrix includes three dimensions: hourly sine encoding, daily cycle encoding, and seasonal cycle encoding. The hourly sine encoding value is obtained by mapping the number of hours corresponding to the time step onto a 24-hour cycle using a sine function. The daily cycle encoding value is obtained by mapping the day number of the time step in the current month onto a sine function onto a cycle of the total number of days in the current month using a sine function. The seasonal cycle encoding value is obtained by mapping the month number corresponding to the time step onto a 12-month cycle using a cosine function. Step S33: Convert the modal component matrix Meteorological feature matrix and time period encoding matrix Horizontal concatenation is performed along the feature dimension to obtain a concatenated feature matrix. This concatenated feature matrix is ​​then input into a linear projection layer for dimension mapping to obtain a joint feature matrix. ,in The feature dimension is the projected dimension.

8. The photovoltaic power output prediction method based on artificial intelligence according to claim 4, characterized in that, In step S4, a deep learning prediction network integrating photovoltaic physical conversion constraints is constructed. The joint feature matrix is ​​input into the deep learning prediction network for model training. During the training process, the photovoltaic cell equivalent circuit equation, energy conservation constraints, and time smoothness physical regularization term are embedded as differentiable hard constraints into the joint loss function to obtain a trained deterministic prediction model. This includes the following steps: Step S41: Construct a deep learning prediction network, which sequentially includes a feature encoder, a temporal dependency extraction layer, and a prediction decoder. The feature encoder consists of two fully connected layers and combines the joint feature matrix. Mapped to encoded feature sequences The mapping calculation formula is: ； in, To encode feature sequences, For the joint characteristic matrix, This is the weight matrix of the first layer of the feature encoder. This is the weight matrix of the second layer of the feature encoder. This is the bias vector of the first layer of the feature encoder. This is the bias vector of the second layer of the feature encoder. To modify the activation function of the linear unit, The dimension of the encoded feature; Step S42: The temporal dependency extraction layer adopts a bidirectional temporal convolutional network structure, and processes the encoded feature sequence on the forward path. Perform causal convolution operations to extract positive temporal dependency information from past to present time, and process the encoded feature sequence on the reverse path. Causal convolution operations are performed to extract the inverse temporal dependency information from future time to the current time. The outputs of the forward path and the inverse path are concatenated along the feature dimension to obtain a bidirectional temporal feature sequence. The hidden state at the last time step is taken as the hidden state vector output by the temporal dependency extraction layer. ; Step S43: The prediction decoder consists of two fully connected layers, which convert the bidirectional temporal feature sequence... Mapped to a deterministic power prediction sequence The mapping calculation formula is: ； in, For deterministic power prediction sequences, To predict the number of time steps, The bidirectional temporal feature sequence, To predict the weight matrix of the first layer of the decoder, To predict the weight matrix of the second layer of the decoder, To predict the bias vector of the first layer of the decoder, To predict the bias vector of the second layer of the decoder; Step S44: Embed the photovoltaic cell single-diode equivalent circuit equation as a differentiable physical constraint term into the joint loss function. The photovoltaic cell single-diode equivalent circuit equation is as follows: ； in, To output current for photovoltaic cells, The output voltage of the photovoltaic cell. For photocurrent, This is the reverse saturation current of the diode. The electron charge constant, For series equivalent resistance, For diode ideality factor, Boltzmann's constant, The absolute temperature of the photovoltaic cell. For the area of ​​photovoltaic cells, For parallel equivalent resistance, It is a natural exponential function; Step S45: Construct a joint loss function that includes a data-driven loss term and a physical constraint penalty term. The calculation formula is: ； in, For the joint loss function value, For data-driven loss terms, For the equivalent circuit constraints of a single diode in a photovoltaic cell, For energy conservation constraints, For time smoothness physical regularization, For the constraint terms of the single-diode equivalent circuit of a photovoltaic cell, the weighting coefficients are... These are the weighting coefficients for the energy conservation constraint term. These are the weighting coefficients for the physical regularization term of time smoothness; The data-driven loss term The Huber loss function is used, and the calculation formula is as follows: ； ； in, For deterministic power prediction sequences The Middle Power prediction values ​​for each prediction time step. For the first The actual power observations corresponding to each predicted time step To predict the number of time steps, To predict residuals, These are the inflection point parameters of the Huber loss function; The constraints of the photovoltaic cell single diode equivalent circuit The calculation formula is: ； in, For the first The maximum power point voltage of the photovoltaic array corresponding to each predicted time step. To be The calculated value of the photovoltaic cell output current is obtained by substituting the meteorological conditions corresponding to the current time step into the equivalent circuit equation of the single diode of the photovoltaic cell and solving it. The energy conservation constraint term The calculation formula is: ； in, This represents the theoretical maximum photoelectric conversion efficiency of a photovoltaic module. For the first Total irradiance prediction for each prediction time step The effective light-receiving area of ​​the photovoltaic array. This is an operation that takes the larger of two parameters; The time smoothness physical regularization term The calculation formula is: ； in, For the first Power prediction values ​​for each prediction time step. For the first Power prediction values ​​for each prediction time step. This represents the physical maximum ramp rate of the photovoltaic array. This represents the time interval between adjacent prediction time steps.

9. The photovoltaic power output prediction method based on artificial intelligence according to claim 4, characterized in that, In step S5, the meteorological data for the period to be predicted is processed using the same procedures as in steps S1 to S4 and then input into the trained deterministic prediction model to obtain deterministic power prediction values. The number of diffusion denoising steps is adaptively determined based on the current physical-driven meteorological complexity index. Using the deterministic power prediction values ​​and the hidden state vector output by the temporal dependency extraction layer as conditional information, the model performs reverse iterative denoising sampling according to the number of diffusion denoising steps through the conditional denoising diffusion probability model to generate a probability prediction trajectory and prediction interval. This includes the following steps: Step S51: Deterministic power prediction sequence The hidden state vector output at the last time step of the temporal dependency extraction layer Concatenation forms a condition vector The calculation formula is: ； in, For conditional vectors, For the deterministic power prediction sequence, This refers to the hidden state vector output at the last time step of the temporal dependency extraction layer. This represents a vertical concatenation operation of vectors; Step S52: Based on the current physical-driven meteorological complexity index Adaptive determination of diffusion denoising steps The calculation formula is: ； in, To the number of diffusion denoising steps, To minimize the number of diffusion denoising steps, To maximize the number of diffusion denoising steps, As a physical-driven meteorological complexity index, To preset a low complexity threshold, To preset a high complexity threshold, when Less than Time to take ,when Greater than Time to take ; Step S53: Construct a conditional denoising network The conditional denoising network It consists of a three-layer residual convolutional network, with a diffusion time step Current noisy prediction residual sequence and condition vector As input, the output is the predicted noise. ,in This is the diffusion time step index, with a value range of 1- , To the number of diffusion denoising steps, For the set of trainable parameters of a conditional denoising network; Step S54: In the reverse denoising sampling stage, sample the initial noise sequence from the standard Gaussian distribution. According to The reverse denoising iterations are performed sequentially, decreasing to one. The calculation formula for each iteration is as follows: ； in, For the first The denoising results of each diffusion step For the first Noisy prediction residual sequence of diffusion steps, For the first Each diffusion step adds a coefficient to the forward noise. Add coefficients to the forward noise from step one to step two. The cumulative product of steps, The predicted noise output by the conditional denoising network. For the first Random noise scaling factor for each diffusion step. For a random noise sequence sampled from a standard Gaussian distribution, when hour, ,when hour, ; Step S55: The final output after the reverse denoising iteration is completed. As the prediction residual sequence Predicting residual sequences With deterministic power prediction sequence By adding each element, a probability prediction trajectory is obtained. Repeat steps S54 and this step. This time Probability prediction trajectory, for The predicted trajectory is calculated by taking the arithmetic mean at each prediction time step to obtain the predicted correction point value. The predicted value at the 6th time step is then taken. Quantiles are used to obtain the lower bound of the prediction interval, and the quantile is taken as the lower bound. Quantiles are used to determine the upper bound of the prediction interval, where The preset significance level parameter, This represents the total number of samples for probabilistic trajectory prediction.

10. The photovoltaic power output prediction method based on artificial intelligence according to claim 4, characterized in that, In step S6, for areas containing two or more photovoltaic power plants, a dynamic spatiotemporal correlation graph is constructed based on the geographical location information, meteorological condition information, and historical power information of each power plant. A graph convolutional network is then used to propagate and aggregate node features from the dynamic spatiotemporal correlation graph, outputting the joint power prediction result for multiple power plants. This includes the following steps: Step S61: Treat each photovoltaic power station within the region as a node in a dynamic spatiotemporal correlation graph, with the total number of nodes denoted as . Each node Node feature vectors The joint characteristic matrix of the power station The vector after global average pooling and the deterministic power prediction sequence of the power plant. It is pieced together; Step S62: Calculate every two nodes in the dynamic spatiotemporal correlation graph. With nodes Dynamic edge weights between The calculation formula is: ； in, For nodes With nodes Dynamic edge weights between them For the first The photovoltaic power station and the first Geographic Euclidean distance between photovoltaic power stations For distance attenuation parameters, For the first The mean vector of meteorological characteristics of each photovoltaic power station during the current processing period. For the first The mean vector of meteorological characteristics of each photovoltaic power station during the current processing period. for The transpose of , For vectors Norm, For the first The first photovoltaic power station and the first Covariance between the historical standardized power series of individual photovoltaic power plants For the first Standard deviation of the historical standardized power series of a photovoltaic power plant For the first Standard deviation of the historical standardized power series of a photovoltaic power plant This is a weighting coefficient for geographical distance similarity. The weighting coefficients for the similarity of meteorological conditions. Let be the weighting coefficients for power correlation, and satisfy . , It is a natural exponential function; Step S63: Using dynamic edge weights Construct an adjacency matrix , using node feature vectors Constructing the node feature matrix ,in The dimension of the node feature vector is used to process the node feature matrix through a graph convolutional network. Perform two-layer graph convolution propagation operations, the first... The formula for updating node features in layered graph convolutional propagation is: ； in, For the first The node feature matrix of the layer, For the first The node feature matrix of the layer, when hour , The adjacency matrix after adding self-loops, for An identity matrix of order 1. for The corresponding degree matrix and The diagonal elements of a diagonal matrix , for A diagonal matrix with a power of -1 / 2. For the first The trainable weight matrix of layer graph convolution, To modify the activation function of the linear unit; Step S64: The node feature matrix after two layers of graph convolution propagation The input is a fully connected output layer, which generates the final power prediction sequence for each photovoltaic power plant. The calculation formula is as follows: ； in, For the first The final power prediction sequence for each photovoltaic power plant. Node feature matrix The Middle The node feature vector corresponding to the row, This is the weight matrix of the fully connected output layer. This is the bias vector for the fully connected output layer.

Images