Photovoltaic cluster day-ahead climbing prediction method based on multi-dimensional information in snowy weather

Abstract

The invention belongs to the technical field of power systems, and particularly relates to a photovoltaic cluster day-ahead climbing prediction method based on multi-dimensional information in snowy weather. The method comprises the following steps: acquiring related information of each station in a photovoltaic cluster in a preset area; dividing snowy weather labels; respectively constructing an irradiance satellite remote sensing data inversion model of each power station; utilizing an irradiance satellite remote sensing data inversion model to invert historical actually-measured surface irradiance of areas near each power station in the cluster; according to the inversion earth surface irradiance data, aiming at each station in the cluster, respectively constructing a snowstorm weather prediction model of an area where each power station is located; constructing a photovoltaic cluster generation power day-ahead climbing prediction model under the snowstorm weather condition based on the multi-dimensional observation information; and carrying out comprehensive judgment, and carrying out day-ahead climbing early warning on the photovoltaic cluster generation power. According to the method, a more accurate power prediction result is obtained in the time range of the snowy weather, and the reliability of power climbing prediction and early warning in the snowy weather is improved.

Description

Technical Field

[0001] This invention belongs to the field of power system technology, and particularly relates to a method for predicting the day-ahead ramp-up of photovoltaic clusters under blizzard conditions based on multi-dimensional information. More specifically, it relates to a method for predicting and warning the day-ahead ramp-up of photovoltaic cluster power generation under blizzard conditions based on multi-dimensional observation information. Background Technology

[0002] Currently, the photovoltaic power generation industry is developing rapidly. By the end of 2024, the national installed photovoltaic capacity had reached approximately 890 million kilowatts, accounting for 26.6% of the total installed power generation capacity, an increase of 45.2% year-on-year. However, due to the influence of meteorological conditions, the output of photovoltaic power is highly random and volatile. The large-scale integration of new energy sources into the power system will greatly increase the uncertainty on the power system's source side, posing a significant challenge to the dispatch and operation of the power system.

[0003] Existing research primarily focuses on photovoltaic (PV) power generation under normal weather conditions. However, extreme weather significantly impacts PV output, causing power ramp-up events. During blizzards, accurate measurement of solar irradiance is difficult; during snowfall, snow cover on PV array surfaces reduces the effective irradiance received by the panels, resulting in power loss. Existing PV power prediction methods lack specific research for blizzard scenarios. Regarding data utilization, some methods consider using spatiotemporal data such as satellite cloud images and numerical weather prediction to predict PV power, but most only study one or two types of observation data, failing to effectively utilize multi-dimensional observation data from PV power plants.

[0004] Effective mining of multidimensional data can provide more comprehensive information support for rapid ramp-up prediction and early warning modeling. Therefore, it is urgent to propose a method for predicting and warning of large-scale photovoltaic power generation ramp-up under blizzard weather based on multidimensional observation information, focusing on rapid ramp-up prediction and early warning of photovoltaic power output under blizzard weather. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention provides a method for predicting the day-ahead ramp-up of photovoltaic (PV) clusters under blizzard conditions based on multi-dimensional information. The purpose of this invention is to achieve accurate prediction of PV power generation and precise early warning of ramp-up under blizzard conditions, thereby providing effective support for grid operation and dispatch.

[0006] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0007] A method for predicting the day-ahead ramp-up of photovoltaic clusters under blizzard conditions based on multi-dimensional information, including:

[0008] Obtain relevant information about each site in the photovoltaic cluster within the preset area;

[0009] Snowstorm weather tags were assigned based on information from each station;

[0010] For each station in the cluster, an inversion model for the irradiance satellite remote sensing data of each power station is constructed separately;

[0011] Using the irradiance satellite remote sensing data inversion model, the historical measured surface irradiance of the area near each station in the cluster was retrieved;

[0012] Based on the inverted surface irradiance data of the region, a blizzard weather prediction model is constructed for each power station in the cluster.

[0013] Based on the blizzard weather prediction model, a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard weather conditions is constructed based on multi-dimensional observation information.

[0014] A comprehensive assessment is made based on the prediction results of the day-ahead ramp-up forecasting model for power generation, and a day-ahead ramp-up warning for photovoltaic cluster power generation is issued.

[0015] Furthermore, the relevant information for each station includes: the historical output of each station, the geographical location information of each station, and the meteorological data of each station;

[0016] The historical power output data for each station are as follows:

[0017] (0.1)

[0018] Wherein, P represents the set of historical power output data for all photovoltaic power plants within a preset time interval. Representing station i from time t0 to t 0+r The historical contribution of every moment Let represent the power output of power station i at time τ, n be the number of photovoltaic power stations in the photovoltaic cluster, and r be the length of the time interval, i.e., from the start time t0 to the end time t. 0+r The duration, where T is the set of sampling times, i.e. , Representing station number 1 at t 0+r Output value at any given moment Representing station n at t 0+r Output value at any given moment;

[0019] The geographical location information of each station is: L=(L1,L2,…,Ln), and the geographical location information Li of station i includes the longitude and latitude of its location;

[0020] The meteorological data from each station includes snowfall data and predicted irradiance data; the snowfall data is defined as the cumulative amount of snow that falls to the Earth's surface, with the accumulation period being the past hour, and the data is a time series with a time resolution of 1 hour; the predicted irradiance data is a time series with a time resolution of 15 minutes.

[0021] Furthermore, the blizzard weather labeling based on information from each station includes:

[0022] Temporal resolution for processing snowfall data;

[0023] A sliding window is used to calculate the cumulative snowfall over 24 hours.

[0024] Based on the time resolution and cumulative snowfall, determine whether it is blizzard weather and obtain a blizzard weather label.

[0025] Furthermore, for each station in the cluster, an irradiance satellite remote sensing data inversion model is constructed for each power station; including:

[0026] Use the clear sky index to classify weather patterns;

[0027] Based on the pattern segmentation results, feature data under different weather conditions are selected as input features for the model;

[0028] Based on the input characteristics, an inversion model for irradiance satellite remote sensing data under different weather patterns is established.

[0029] Furthermore, the method of using the clear sky index to classify weather conditions into patterns is specifically as follows:

[0030] The clear sky index is defined as:

[0031] (0.2)

[0032] Where: K t GHI is the clear sky index per hour, H0 is the ground horizontal irradiance, and H0 is the extraterrestrial solar irradiance.

[0033] The formula for calculating the extraterrestrial solar irradiance H0 is as follows:

[0034] (0.3)

[0035] Among them: I sc Let be the solar constant, which is 1367 W / m², and n be the nth day of the year. The solar zenith angle;

[0036] Calculate the hourly clear sky index K t This yields a daily clear sky index sequence of 24 points.

[0037] According to the clear sky index K t K-means clustering was performed to classify the weather patterns of adjacent regions into three categories: clear skies, cloudy skies, and rainy skies.

[0038] The clear sky index data is standardized as follows:

[0039] (0.4)

[0040] Among them, K t，norm The standardized clear sky index, This represents the average of the clear sky index. Standard deviation;

[0041] The K-means clustering algorithm was used to classify weather patterns into the following types: clear skies, cloudy, and rainy, as shown in the following formula:

[0042] (0.5)

[0043] Where k=3 represents three weather types: clear sky, cloudy, and rainy.

[0044] Based on the model segmentation results, feature data under different weather conditions are selected as input features for the model. Specifically, 14 channel data points for the power station location are acquired using the Fengyun-4A satellite, with a spatial range of 4km × 4km. For different weather patterns, the Pearson correlation coefficient is used to calculate the correlation between each channel data point and surface irradiance, as shown in the following formula:

[0045] (0.6)

[0046] in, The channel data value refers to the observation of a certain channel of the FY-4A satellite at the i-th sample time in the pixel where the field station is located. The mean of the channel data samples. This represents the surface irradiance value. is the average surface irradiance sample, n is the number of samples, i.e., the number of time samples participating in the correlation calculation under this weather pattern, and i is the sample index, with values ​​i=1,2,…,n, representing the time sample number;

[0047] Filtering correlation coefficients Channel data with a value greater than 0.5 were used as input features;

[0048] The step of establishing an irradiance satellite remote sensing data inversion model under different weather patterns based on input features is as follows: Selecting support vector machine, convolutional neural network, K-nearest neighbor regression and gradient boosting algorithm to establish an irradiance satellite remote sensing data inversion model to invert surface irradiance;

[0049] Support Vector Machines (SVMs) perform regression by constructing a hyperplane in a high-dimensional space. The regression objective of SVMs is solved by the following optimization problem:

[0050] (0.7)

[0051] Constraints:

[0052] (0.8)

[0053] (0.9)

[0054] (0.10)

[0055] Where W is the model weight vector. Here, C is the kernel function mapping for the input features, and C is the regularization parameter. Let n be the sample size and i be the sample index, where i = 1, 2, ..., n. ζ i is a slack variable, representing the deviation of sample i, and b is the bias / intercept term;

[0056] Using the Radial Basis Function (RBF) kernel function:

[0057] (0.11)

[0058] Where, x i、 x j Let be the input feature vector of the i-th and j-th samples. Let γ be the Euclidean distance between the two feature vectors, and γ be the kernel width parameter.

[0059] Convolutional Layer: Extracts local spatial features.

[0060] (0.12)

[0061] in, Let x be the response value of the k-th convolutional kernel at position (i,j) on the output feature map. i+m,j+n Let f(i,j) be the pixel value at offset (m,n) within the input window when the convolution window is aligned with the output position (i,j), where m=1,…,M and n=1,…,N are the row and column indices of the convolution kernel, respectively. Let be the weight of the k-th convolutional kernel at position (m,n). This is the bias term corresponding to the k-th convolution kernel;

[0062] Pooling Layer: Performs feature dimensionality reduction.

[0063] (0.13)

[0064] Among them, y i,j Let x be the pixel value at position (i,j) of the pooled feature map. 2i,2j x 2i+1,2j x 2i,2j+1 x 2i+1,2j+1 These are the four pixel values ​​within a 2×2 window with (2i,2j) as the top left corner in the feature map before pooling;

[0065] Fully Connected Layer: Maps the extracted features to predicted irradiance values;

[0066] K-nearest neighbor regression is a non-parametric method that predicts based on the average of the K nearest neighbors in the training set.

[0067] Distance metric: Euclidean distance is used as the distance between samples.

[0068] (0.14)

[0069] Where d is the distance metric function, d(x) i, x j ) is the sample x i With sample x j Euclidean distance in the feature space, x i,k Let x be the feature component of the i-th sample in the k-th dimension. jk Let n be the feature component of the j-th sample in the k-th dimension, n be the number of feature dimensions, and k be the dimension index.

[0070] Find the K nearest neighbor samples of the test sample and take the average of their surface irradiance:

[0071] (0.15)

[0072] in, The predicted surface irradiance for the test sample, K is the nearest neighbor number, y i Let be the true surface irradiance of the i-th nearest neighbor sample;

[0073] The goal of gradient boosting is to minimize the loss function iteratively.

[0074] (0.16)

[0075] Among them, F m (x) represents the prediction of the strong learner for sample x after the m-th iteration, F m-1 (x) represents the predictions of the cumulative model in the first m-1 rounds, η is the learning rate / shrinkage coefficient, which controls the contribution of the base learner in this round to the overall model, h m (x) is the fitting function of the m-th base learner for the current residual;

[0076] The model performance was evaluated using a validation set, with mean squared error (MSE) as the evaluation metric.

[0077] (0.17)

[0078] Where n is the number of samples in the evaluation set, y i Let i be the true surface irradiance of the i-th sample. Let be the predicted surface irradiance of the model for the i-th sample;

[0079] For each weather pattern, the model with the best performance on the validation set is selected as the final model.

[0080] Furthermore, the aforementioned irradiance satellite remote sensing data inversion model is used to invert the historical measured surface irradiance of the area near each station in the cluster; specifically, this includes:

[0081] (1) Calculate the typical distribution of each channel feature under various weather patterns of the power station;

[0082] The system acquires 14 channel characteristic data points for Power Station A under sunny, cloudy, and rainy weather conditions from the Fengyun-4 satellite and stores them separately. It then calculates the statistical measures for each weather type's characteristic data. If there are n days of data, each data point contains 14 characteristic dimensions, denoted as X. ij Where i represents the i-th day and j represents the j-th feature:

[0083] average value :

[0084] (0.18)

[0085] Where n is the sample size, i.e., the number of days used for statistics under this weather pattern, i is the sample index, taking i=1,2,…,n, corresponding to the i-th day, j is the feature / channel index, taking j=1,2,…,14, X ij Let j be the value of the j-th channel on the i-th day;

[0086] Standard deviation :

[0087] (0.19)

[0088] The above formula yields the typical distribution of each feature under the three weather patterns.

[0089] (2) For each day's data in the adjacent area, calculate its similarity to various weather type patterns and determine the weather type;

[0090] The typical distribution of each channel feature under three weather patterns for the power station was calculated. Specifically, feature data of 14 channels located in the four adjacent areas (north, south, east, and west) of each power station were obtained from the Fengyun-4 satellite. The data spatial range was 4 km × 4 km for each area, and the data of the adjacent areas on a certain day was set as a vector. =[y1,y2,...,y14], the average vector for a certain weather type =[μ1,μ2,...,μ14], and its Euclidean distance D with the adjacent region data is calculated as follows:

[0091] (0.20)

[0092] Repeat the above process for each weather type to obtain the distances between adjacent areas and the three weather types;

[0093] Based on the calculated Euclidean distance, the weather type corresponding to the minimum distance is selected as the prediction result; if the Euclidean distance for sunny weather is the minimum, then the weather in the adjacent area is predicted to be sunny on that day.

[0094] (3) Input the corresponding features of the adjacent area into the irradiance inversion model to obtain the measured surface irradiance data;

[0095] By using the channel data types for each weather pattern as input to the model, and the corresponding channel data for each day in the adjacent area as input to the irradiance inversion model, surface irradiance data for four adjacent 4km×4km areas of each power station are obtained.

[0096] Furthermore, based on the blizzard weather prediction model, a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard weather conditions is constructed based on multi-dimensional observation information. The model uses the GCN model, with each photovoltaic power station as a node. The node characteristics of each power station include the forecast irradiance data for the predicted day, the historical power data for 7 days, and the blizzard weather label for the predicted day.

[0097] The input characteristics of each photovoltaic power station node are represented as follows:

[0098] (0.21)

[0099] Among them: I tThe forecast irradiance data for the predicted day is in shape 96×1 with a time resolution of 15 minutes. This is historical photovoltaic power data for the past 7 days, in a 7×96 format; W t The label for the predicted blizzard weather is 96×1 and is a binary label indicating whether it is blizzard weather.

[0100] After concatenating the above features, the node feature matrix is ​​as follows:

[0101] (0.22)

[0102] Where: N is the number of nodes; F is the feature dimension of each node;

[0103] The adjacency matrix A between photovoltaic power plants is calculated based on the Pearson correlation coefficient of historical photovoltaic power output. For any two photovoltaic power plants i and j, the correlation coefficient is calculated using the following formula:

[0104] (0.23)

[0105] Where: i, j are the photovoltaic power station indices, representing any two power stations respectively; Let i be the power output of photovoltaic power station i at time step k; Let be the average power output of photovoltaic power plant i within the time window T, where T is the length of the historical time series and k is the time index, taking k=1,2,…,T;

[0106] Adjacency matrix A∈R N*N The element is defined as:

[0107] (0.24)

[0108] Among them, A ij The elements of the adjacency matrix represent the relevance edge weights between power station i and power station j, r. ij Let A be the Pearson correlation coefficient calculated from the historical power series, and let A be the adjacency matrix with dimension R. N*N Where N is the number of photovoltaic power stations, A∈R N*N This indicates that the adjacency matrix is ​​an N-row, N-column real matrix;

[0109] To adapt to the GCN model, the adjacency matrix is ​​normalized to... :

[0110] (0.25)

[0111] Where: I is the identity matrix, used to add self-loops; A is the adjacency matrix; GCN is a Graph Convolutional Network, which extracts features from graph-structured data through convolution operations between the adjacency matrix and the feature matrix; and D is the degree matrix, defined as:

[0112] (0.26)

[0113] Among them, D ii Let A be the diagonal element of the degree matrix, representing the degree of node i. ij For the elements of the adjacency matrix, I ij is an element of the identity matrix, and j is the node index;

[0114] The graph convolution operation used in GCN is represented as follows:

[0115] (0.27)

[0116] in: Let be the output feature matrix of the (l+1)th layer of the graph convolutional network. Let H(0) be the input feature matrix of the graph convolutional layer l, and let H(0) = X. Let be the trainable weight matrix of the graph convolutional layer l; For activation functions; This is the normalized adjacency matrix;

[0117] Based on the spatial features extracted by GCN, a time series model is used to model the temporal dimension of the node features; where LSTM is a Long Short-Term Memory network; it is assumed that after GCN computation, the output feature of each node is H∈R. N×T×F Where T is the number of time steps and F is the feature dimension;

[0118] The update formula for the time series model is:

[0119] (0.28)

[0120] Where: GRU is a gated loop unit, h t-1 This is the hidden state of the previous time step. For the input features at time step t, The hidden state at the current time step;

[0121] The hidden state is mapped to the power output using a fully connected layer:

[0122] (0.29)

[0123] in: W is the predicted power of power plant i at time step t.o and b o These are the weight matrix and bias vector of the fully connected layer, respectively, both of which are trainable parameters;

[0124] The output is the photovoltaic power of each power station for the next day, with a time resolution of 15 minutes.

[0125] Furthermore, the method involves comprehensively judging the prediction results of the day-ahead ramp-up prediction model for power generation to provide a day-ahead ramp-up warning for photovoltaic cluster power generation; specifically, this includes:

[0126] The corresponding slope recognition results were obtained using three different slope recognition methods, as follows:

[0127] Method 1. Short-time window ramp rate determination:

[0128] The ramp rate is the ratio of the absolute value of the power change of a photovoltaic power station within a certain time interval to the time interval itself, as shown in the following formula:

[0129] (0.30)

[0130] in: The gradeability is expressed as power change per time (MW / 15min). The observed power value at time t (in MW); The observation time interval is in minutes. =15 minutes);

[0131] At each time t, calculate the gradient rate. , where the threshold Selecting 5% of the power plant's installed capacity is considered a ramp-up event, and this moment is marked as 1; otherwise, it is marked as 0. This is suitable for capturing rapid changes in a short period of time.

[0132] (0.31)

[0133] Where: M 1(t) This is the hill-climbing event indicator for Method 1;

[0134] Method 2. Determining the ramp rate of a window sliding over a long period of time;

[0135] Set up a sliding window =4h, sliding step size is 15 minutes, traversing the time series Calculate the ramp rate for each time window:

[0136] (0.32)

[0137] in: The long-term ramp rate at time step t, The long window length is set to 4 hours, which corresponds to 16 sampling points, each with a 15-minute interval.

[0138] Determine if a hill-climbing event has occurred:

[0139] (0.33)

[0140] Where M2(t) is the hill-climbing event indicator of Method 2, and P2 is the threshold;

[0141] threshold Select 10% of the power plant's installed capacity. The sliding window logic is as follows: if a ramp event occurs within the window, mark all points in the current window as 1; if no ramp event occurs, mark the new points as 0.

[0142] Method 3. Determining the power difference during long-term window sliding;

[0143] Set up a sliding window =4h, sliding step size is 15 minutes; traverse the time series Calculate the power difference for each time window:

[0144] (0.34)

[0145] Where D3(t) is the power difference at time step t;

[0146] Determine if a climbing event has occurred:

[0147] (0.35)

[0148] Where M3(t) is the hill-climbing event indicator of method 3, and P3 is the threshold;

[0149] threshold Select 15% of the power plant's installed capacity; the sliding window logic is the same as in Method 2.

[0150] An expert voting method was used to comprehensively evaluate the three methods and obtain the climbing result. Specifically, for each time t, the evaluation results of the three methods were combined:

[0151] (0.36)

[0152] Where V(t) is the comprehensive judgment result, M1(t), M2(t), and M3(t) are the climbing event indicators of the three methods, respectively, and Majority is the majority voting function: if at least two of the three methods are judged as climbing events (value 1), then V(t) = 1, otherwise V(t) = 0.

[0153] If at least two of the three methods identify a climbing event as occurring, and this is marked as 1, then V(t) = 1; otherwise, V(t) = 0.

[0154] Output the hill-climbing event judgment result V(t) at each moment to realize photovoltaic power hill-climbing early warning.

[0155] A photovoltaic cluster day-ahead ramp prediction device based on multi-dimensional information under blizzard conditions, including:

[0156] The acquisition module is used to acquire relevant information about each site in the photovoltaic cluster of the preset area;

[0157] The tagging module is used to categorize blizzard weather tags based on information from each station.

[0158] The inversion model construction module is used to construct an inversion model of the irradiance satellite remote sensing data for each power station in the cluster.

[0159] The inversion module is used to invert the historical measured surface irradiance of the area near each station in the cluster using the irradiance satellite remote sensing data inversion model;

[0160] The blizzard weather prediction module for each power station area is used to construct a blizzard weather prediction model for each power station area in the cluster based on the inverted surface irradiance data of the area.

[0161] The module for constructing a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard conditions based on multi-dimensional observation information is used to construct a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard conditions based on multi-dimensional observation information.

[0162] The early warning module is used to make a comprehensive judgment on the prediction results of the day-ahead ramp-up prediction model for power generation and to issue early warnings for the day-ahead ramp-up of power generation in photovoltaic clusters.

[0163] Furthermore, the early warning module is used to comprehensively judge the prediction results of the day-ahead ramp-up prediction model for power generation and to issue a day-ahead ramp-up warning for photovoltaic cluster power generation; including:

[0164] The corresponding slope recognition results were obtained using three different slope recognition methods, as follows:

[0165] Method 1. Short-time window ramp rate determination:

[0166] The ramp rate is the ratio of the absolute value of the power change of a photovoltaic power station within a certain time interval to the time interval itself, as shown in the following formula:

[0167] (0.30)

[0168] in: The gradeability is expressed as power change per time (MW / 15min). The observed power value at time t (in MW); The observation time interval is in minutes. =15 minutes);

[0169] At each time t, calculate the gradient rate. , where the threshold Selecting 5% of the power plant's installed capacity is considered a ramp-up event, and this moment is marked as 1; otherwise, it is marked as 0. This is suitable for capturing rapid changes in a short period of time.

[0170] (0.31)

[0171] Where: M 1(t) This is the hill-climbing event indicator variable for Method 1;

[0172] Method 2. Determining the ramp rate of a window sliding over a long period of time;

[0173] Set up a sliding window =4h, sliding step size is 15 minutes, traversing the time series Calculate the ramp rate for each time window:

[0174] (0.32)

[0175] in: Let P(t) be the sliding ramp rate at time t, and P(t) be the power value at time t.

[0176] Determine if a climbing event has occurred:

[0177] (0.33)

[0178] Where M2(t) is the hill-climbing event indicator variable of Method 2, and P2 is the threshold;

[0179] threshold Select 10% of the power plant's installed capacity. The sliding window logic is as follows: if a ramp event occurs within the window, mark all points in the current window as 1; if no ramp event occurs, mark the new points as 0.

[0180] Method 3. Determining the power difference during long-term window sliding;

[0181] Set up a sliding window =4h, sliding step size is 15 minutes; traverse the time series Calculate the power difference for each time window:

[0182] (0.34)

[0183] Where D3(t) is the power difference at time step t;

[0184] Determine if a climbing event has occurred:

[0185] (0.35)

[0186] Where M3(t) is the hill-climbing event indicator variable of Method 3, and P3 is the threshold;

[0187] The threshold P3 is selected as 15% of the power plant's installed capacity, and the sliding window logic is the same as in method 2.

[0188] Step 72. Use expert voting to comprehensively evaluate the three methods and obtain the climbing result. Specifically, for each time t, combine the evaluation results of the three methods:

[0189] (0.36)

[0190] Where V(t) is the comprehensive judgment result, M1(t), M2(t), and M3(t) are the climbing event indicators of the three methods, respectively, and Majority is the majority voting function: if at least two of the three methods are judged as climbing events (value 1), then V(t) = 1, otherwise V(t) = 0.

[0191] If at least two of the three methods identify a climbing event as occurring, and this is marked as 1, then V(t) = 1; otherwise, V(t) = 0.

[0192] Output the hill-climbing event judgment result V(t) at each moment to realize photovoltaic power hill-climbing early warning.

[0193] A computer device includes a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the computer program to implement the steps of the photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in any one of the claims.

[0194] A computer storage medium storing a computer program, wherein when the computer program is executed by a processor, the steps of the photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in any one of the claims are implemented.

[0195] The present invention has the following beneficial effects and advantages:

[0196] Weather pattern decomposition and modeling: The clear sky index and its standardized clustering are introduced to divide the samples into three categories: clear sky, cloudy, and rainy. The models are trained separately under homogeneous meteorological conditions to reduce the bias caused by meteorological confounding and improve the stability and accuracy of irradiance inversion and power prediction.

[0197] Multi-source satellite features and optimal inversion model: FY-4A satellite channel features are integrated, channels are screened based on Pearson correlation coefficient, and validation sets of algorithms such as SVR, CNN, KNN, and GBDT are compared under each weather mode to select the best final irradiance satellite remote sensing data inversion model and improve inversion accuracy.

[0198] Enhanced neighborhood information: Satellite features and similarity discrimination of four neighborhoods around the power station have been added. The impact of regional cloud systems and cold wave paths on the power station has been incorporated into the modeling. Compared with only site location data, weather changes can be identified earlier, improving the lead time and accuracy of turning point identification.

[0199] Spatial-temporal joint modeling: A graph structure with historical power correlation as edge weights is constructed, and GCN+LSTM is used to jointly model spatial correlation and temporal dependence. Blizzard labels and NWP forecast irradiance are incorporated as node features into the model to improve the accuracy of power prediction.

[0200] Multi-criteria ramp identification and voting fusion: Three identification methods are proposed: "short window ramp rate, long window ramp rate, and long window power difference". The majority voting is used for comprehensive judgment, which effectively reduces false alarms / false negatives of the single threshold method and significantly improves the reliability of ramp early warning.

[0201] In summary, this invention improves the reliability of power ramp prediction and early warning during blizzards. Furthermore, by comparing the performance of these algorithms on the validation set, this invention selects the best-performing model as the final irradiance satellite remote sensing data inversion model, which is innovative.

[0202] This invention can not only effectively capture complex relationships in time series, but also reflect the priority of key features, making it very suitable for snowstorm weather forecasting tasks.

[0203] This invention can obtain more accurate power prediction results within the time range of blizzard weather, thereby improving the reliability of power ramp prediction and early warning under blizzard weather and providing effective support for power grid operation and scheduling. Attached Figure Description

[0204] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0205] Figure 1 This is an overall flowchart of the method of the present invention;

[0206] Figure 2 This is a flowchart of step 2 of the present invention, which defines blizzard weather labels based on snowfall data.

[0207] Figure 3 This is a flowchart of step 3 of the present invention for constructing the irradiance satellite remote sensing data inversion model for each power station.

[0208] Figure 4 This is a flowchart of step 4 of the present invention, which uses the irradiance satellite remote sensing data inversion model to invert the historical measured surface irradiance of the area near each photovoltaic power station in the cluster;

[0209] Figure 5 This is a flowchart of step 7 of the present invention, which involves comprehensively judging the prediction results of the day-ahead ramp-up prediction model for power generation and issuing a day-ahead ramp-up warning for photovoltaic cluster power generation. Detailed Implementation

[0210] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0211] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0212] The following reference Figures 1-5 The technical solutions of some embodiments of the present invention are described below.

[0213] Example 1

[0214] This invention provides an embodiment of a method for predicting the day-ahead ramp-up of photovoltaic (PV) cluster power generation under blizzard conditions based on multi-dimensional information. Specifically, it is a method for predicting and warning of the day-ahead ramp-up of PV cluster power generation under blizzard conditions based on multi-dimensional observation information. Figure 1 As shown, Figure 1 This is an overall flowchart of the method of the present invention.

[0215] The present invention specifically includes the following steps:

[0216] Step 1. Obtain the historical power output, geographical location information, and meteorological data of each station in the preset area photovoltaic cluster. The meteorological data of each station includes snowfall data and predicted irradiance data.

[0217] Step 2. Label the blizzard weather based on snowfall data.

[0218] Step 3. For each photovoltaic power station in the cluster, construct an inversion model for the irradiance satellite remote sensing data of each power station.

[0219] Step 4. Use the irradiance satellite remote sensing data inversion model to invert the historical measured surface irradiance of the area near each photovoltaic power station in the cluster.

[0220] Step 5. Based on the inverted surface irradiance data of the region, construct a blizzard weather prediction model for the region where each photovoltaic power station is located within the cluster.

[0221] Step 6. Based on the blizzard weather prediction model, construct a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard weather conditions based on multi-dimensional observation information.

[0222] Step 7. Based on the multi-dimensional observation information, the photovoltaic cluster power generation power day-ahead ramp-up prediction model under blizzard weather conditions is used to comprehensively judge the power time series of each power station obtained by the prediction using three ramp-up event identification methods combined with expert voting method, so as to realize the early warning of photovoltaic cluster power generation power day-ahead ramp-up.

[0223] In step 1, relevant information about each photovoltaic power station in the preset area photovoltaic cluster is obtained, wherein the historical power output data of n photovoltaic power stations are as follows:

[0224] (0.1)

[0225] Wherein, P represents the set of historical power output data for all photovoltaic power plants within a preset time interval. Representing station i from time t0 to t 0+r The historical contribution of every moment Let represent the power output of power station i at time τ, n be the number of photovoltaic power stations in the photovoltaic cluster, and r be the length of the time interval, i.e., from the start time t0 to the end time t. 0+r The duration, where T is the set of sampling times, i.e. , Representing station number 1 at t 0+r Output value at any given moment Representing station n at t 0+r The output value at any given moment.

[0226] In step 1, the geographical location information of each station in the photovoltaic cluster in the preset area is obtained. The geographical location information of n photovoltaic stations is: L=(L1,L2,…,Ln). The geographical location information Li of station i includes the longitude and latitude of its location.

[0227] In step 1, meteorological data from each station in the preset area photovoltaic cluster is acquired, including snowfall data and predicted irradiance data. Snowfall data is defined as the cumulative amount of snow that has fallen to the Earth's surface, with the accumulation period being the past hour, and the data is a time series with a time resolution of 1 hour. Predicted irradiance data is a time series with a time resolution of 15 minutes.

[0228] Step 2 describes the classification of blizzard weather tags based on snowfall data, such as... Figure 2 As shown, Figure 2 This is a flowchart of the process for classifying blizzard weather labels based on snowfall data in this invention, including:

[0229] Step 21. Process the temporal resolution of snowfall data;

[0230] Step 22. Calculate the cumulative snowfall over 24 hours using a sliding window;

[0231] Step 23. Based on the time resolution and cumulative snowfall, determine whether it is blizzard weather and obtain the blizzard weather label.

[0232] The time resolution for processing snowfall data in step 21 is as follows: for each hour of snowfall, it is evenly distributed into four corresponding 15-minute time intervals. For example, if the snowfall in a certain hour, such as 01:00 on February 1, 2025, is 0.8 mm, then that hour will be decomposed into the following four time points:

[0233] From 00:45 on February 1, 2025 to 01:00 on February 1, 2025: 0.2 mm

[0234] From 01:00 on February 1, 2025 to 01:15 on February 1, 2025: 0.2 mm

[0235] From 01:15 on February 1, 2025 to 01:30 on February 1, 2025: 0.2 mm

[0236] From 01:30 on February 1, 2025 to 01:45 on February 1, 2025: 0.2 mm.

[0237] Step 22, which involves calculating the 24-hour cumulative snowfall using a sliding window, specifically involves calculating the 24-hour cumulative snowfall of the sliding window based on data with a 15-minute resolution. This means calculating the cumulative snowfall of the previous 24 hours for each time point.

[0238] Step 23 describes determining whether it is blizzard weather based on time resolution and cumulative snowfall, and obtaining blizzard weather labels. Specifically, if the 24-hour cumulative snowfall at a certain time point is greater than 2.5 mm, it is marked as blizzard weather with a label of 1; otherwise, it is marked as non-blizzard weather with a label of 0, thus obtaining blizzard weather labels for each station in the cluster at a 15-minute resolution.

[0239] Step 3 describes constructing an irradiance satellite remote sensing data inversion model for each photovoltaic power station within the cluster, as follows: Figure 3 As shown, Figure 3 This is a flowchart of step 3 of the present invention for constructing the irradiance satellite remote sensing data inversion model for each power station, specifically including:

[0240] Step 31. Use the Clear Sky Index to classify weather conditions into patterns;

[0241] Step 32. Based on the pattern segmentation results, select feature data under different weather conditions as input features for the model;

[0242] Step 33. Based on the input characteristics, establish an inversion model for irradiance satellite remote sensing data under different weather patterns.

[0243] Step 31, which involves using the clear sky index to classify weather conditions into patterns, specifically includes:

[0244] The clear sky index is defined as:

[0245] (0.2)

[0246] Where: K t GHI is the clear sky index per hour, H0 is the ground horizontal irradiance, and H0 is the extraterrestrial solar irradiance.

[0247] The formula for calculating the extraterrestrial solar irradiance H0 is as follows:

[0248] (0.3)

[0249] Among them: I sc Let be the solar constant, which is 1367 W / m², and n be the nth day of the year. The solar zenith angle;

[0250] Calculate the hourly clear sky index K t This yields a daily clear sky index sequence of 24 points.

[0251] According to the clear sky index K tK-means clustering was performed to classify the weather patterns of adjacent areas into three categories: clear skies, cloudy skies, and rainy skies.

[0252] The clear sky index data is standardized as follows:

[0253] (0.4)

[0254] Among them, K t，norm The standardized clear sky index, This represents the average of the clear sky index. The standard deviation is denoted as .

[0255] Then, the K-means clustering algorithm is used to classify the weather patterns into three types: clear skies, cloudy, and rainy. The formula is as follows:

[0256] (0.5)

[0257] Where k=3 represents three weather types: clear sky, cloudy, and rainy.

[0258] Step 32, based on the pattern segmentation results, selects feature data under different weather conditions as input features for the model, specifically as follows:

[0259] Fourteen channels of data, covering a spatial area of ​​4km × 4km, were acquired using the Fengyun-4A satellite. For different weather patterns, the Pearson correlation coefficient was used to calculate the correlation between the data from each channel and surface irradiance, as shown in the following formula:

[0260] (0.6)

[0261] in, The channel data value refers to the observation of a certain channel of the FY-4A satellite at the i-th sample time in the pixel where the field station is located. The mean of the channel data samples. This represents the surface irradiance value. is the average surface irradiance sample, n is the number of samples, i.e., the number of time samples participating in the correlation calculation under this weather pattern, and i is the sample index, with values ​​i=1,2,…,n, representing the time sample number;

[0262] Filtering correlation coefficients Channel data with a value greater than 0.5 are used as input features.

[0263] In this embodiment, the FY-4A satellite scanning imaging radiometer primarily undertakes the task of acquiring cloud images, with a total of 14 channels. Based on the characteristic data of these 14 channels obtained from the Fengyun-4 satellite, the spatial range of the acquired satellite data is 4km × 4km. The Pearson correlation coefficient is used to calculate the correlation between the data from different satellite channels and the surface irradiance of the power station, and channel data with a correlation coefficient greater than 0.5 are selected as the input features for the model.

[0264] Step 33 describes establishing an irradiance satellite remote sensing data inversion model under different weather patterns based on input features. Specifically, in order to explore the most suitable machine learning inversion algorithm for each weather pattern, four machine learning algorithms were selected: support vector machine, convolutional neural network, K-nearest neighbor regression, and gradient boosting algorithm. An irradiance satellite remote sensing data inversion model was established to invert the surface irradiance.

[0265] Support Vector Machines (SVMs) perform regression by constructing a hyperplane in a high-dimensional space. The regression objective of SVM is solved by the following optimization problem:

[0266] (0.7)

[0267] Constraints:

[0268] (0.8)

[0269] (0.9)

[0270] (0.10)

[0271] Where W is the model weight vector. Here, C is the kernel function mapping for the input features, and C is the regularization parameter. Let n be the sample size and i be the sample index, where i = 1, 2, ..., n. ζ i is a slack variable, representing the deviation of sample i, and b is the bias / intercept term;

[0272] Using the Radial Basis Function (RBF) kernel function:

[0273] (0.11)

[0274] Where, x i、 x j Let i be the input feature vector of the i-th and j-th samples. Let γ be the Euclidean distance between the two feature vectors, and γ be the kernel width parameter.

[0275] Convolutional Layer: Extracts local spatial features.

[0276] (0.12)

[0277] in, Let x be the response value of the k-th convolutional kernel at position (i,j) on the output feature map. i+m,j+n Let f(i,j) be the pixel value at offset (m,n) within the input window when the convolution window is aligned with the output position (i,j), where m=1,…,M and n=1,…,N are the row and column indices of the convolution kernel, respectively. Let be the weight of the k-th convolutional kernel at position (m,n). This is the bias term corresponding to the k-th convolution kernel;

[0278] Pooling Layer: Performs feature dimensionality reduction.

[0279] (0.13)

[0280] Among them, y i,j Let x be the pixel value at position (i,j) of the pooled feature map. 2i,2j x 2i+1,2j x 2i,2j+1 x 2i+1,2j+1 These are the four pixel values ​​within a 2×2 window with (2i,2j) as the top left corner in the feature map before pooling;

[0281] Fully Connected Layer: Maps the extracted features to the predicted irradiance values.

[0282] K-Nearest Neighbors (KNN) regression is a simple non-parametric method that predicts based on the average of the K nearest neighbors in the training set.

[0283] Distance metric: Euclidean distance is used as the distance between samples.

[0284] (0.14)

[0285] Where d is the distance metric function, d(x i, x j ) is the sample x i With sample x j Euclidean distance in the feature space, x i,k Let x be the feature component of the i-th sample in the k-th dimension. jk Let n be the feature component of the j-th sample in the k-th dimension, n be the number of feature dimensions, and k be the dimension index.

[0286] Find the K nearest neighbor samples of the test sample and take the average of their surface irradiance:

[0287] (0.15)

[0288] in, The predicted surface irradiance for the test sample, K is the nearest neighbor number, y i Let be the true surface irradiance of the i-th nearest neighbor sample;

[0289] Gradient Boosting Regression (GBDT) is an ensemble learning method that obtains the final prediction result by weighted summation of multiple weak regression trees.

[0290] The goal of gradient boosting is to minimize the loss function iteratively.

[0291] (0.16)

[0292] Among them, F m (x) represents the prediction of the strong learner (cumulative model) for sample x after the m-th iteration, F m-1 (x) represents the predictions of the cumulative model in the first m-1 rounds, η is the learning rate / shrinkage coefficient, which controls the contribution of the base learner in this round to the overall model, h m (x) is the fitting function of the m-th base learner for the current residual.

[0293] The model performance was evaluated using a validation set, with mean squared error (MSE) as the evaluation metric.

[0294] (0.17)

[0295] Where n is the number of samples in the evaluation set, y i Let i be the true surface irradiance of the i-th sample. Let be the predicted surface irradiance of the model for the i-th sample.

[0296] For each weather pattern, the model with the best performance on the validation set is selected as the final model.

[0297] In this embodiment, given the significant differences in the relationship between irradiance and the 14 channel data under different weather patterns, the weather patterns were first meticulously categorized, and the most relevant high-similarity features were selected for each weather pattern. To determine the optimal machine learning inversion algorithm under various weather conditions, four advanced machine learning algorithms were evaluated: Support Vector Machine, Convolutional Neural Network, K-Nearest Neighbor Regression, and Gradient Boosting Decision Tree. By comparing the performance of these algorithms on the validation set, the best-performing model was selected as the final irradiance satellite remote sensing data inversion model, which is innovative.

[0298] Step 4 describes using the irradiance satellite remote sensing data inversion model to invert the historical measured surface irradiance of the area near each photovoltaic power station in the cluster, such as... Figure 4 As shown, Figure 4 To utilize satellite remote sensing data for irradiance inversion modeling, a flowchart is provided for retrieving historical measured surface irradiance in areas near each photovoltaic power station within the cluster. The flowchart includes:

[0299] Step 41. Calculate the typical distribution of each channel feature under the three weather patterns of the power station;

[0300] Step 42. For each day's data in the adjacent area, calculate its similarity to the three weather type patterns and determine the weather type;

[0301] Step 43. Input the corresponding features of the adjacent area into the irradiance inversion model to obtain the measured surface irradiance data.

[0302] Step 41, which calculates the typical distribution of each channel feature under three weather modes for the power station, specifically involves: acquiring 14 channel feature data for power station A under sunny, cloudy, and rainy weather conditions from the Fengyun-4 satellite and storing them separately. For each weather type, such as sunny, cloudy, and rainy, calculate the statistics of its feature data, including the mean and standard deviation. Assume there are n days of data, and each data point contains 14 feature dimensions, denoted as X. ij Where i represents the i-th day and j represents the j-th feature:

[0303] average value :

[0304] (0.18)

[0305] Where n is the number of samples, i.e., the number of days used for statistics under this weather pattern, i is the sample index, taking i=1,2,…,n, corresponding to the i-th day, and j is the feature / channel index, taking j=1,2,…,14, X ij Let be the value of the j-th channel on the i-th day.

[0306] Standard deviation :

[0307] (0.19)

[0308] The above formulas provide the typical distribution of each feature under the three weather patterns.

[0309] Step 42, which calculates the typical distribution of each channel feature under three weather patterns for the power station, specifically involves acquiring feature data for 14 channels located in the four adjacent areas (north, south, east, and west) of each power station from the Fengyun-4 satellite. The data spatial range is 4 km × 4 km for each area. Let the data for a particular day in the adjacent areas be a vector. =[y1,y2,...,y14], the average vector for a certain weather type =[μ1,μ2,...,μ14], and its Euclidean distance D with the adjacent region data is calculated as follows:

[0310] (0.20)

[0311] Repeat the above process for each weather type to obtain the distances between adjacent areas and the three weather types.

[0312] Based on the calculated Euclidean distance, the weather type corresponding to the smallest distance is selected as the prediction result. That is, if the Euclidean distance for sunny weather is the smallest, then the weather in that adjacent area is predicted to be sunny on that day.

[0313] Step 43 involves inputting the corresponding features of the adjacent areas into the irradiance inversion model to obtain measured surface irradiance data. Specifically, in step 3, the channel data types used as model inputs for each weather pattern have been identified. The corresponding channel data of the adjacent areas for each day are used as inputs to the irradiance inversion model to obtain surface irradiance data for four adjacent 4km×4km areas of each power station.

[0314] Step 5 describes the construction of a blizzard weather prediction model for each photovoltaic power station within the cluster, specifically for the region where each power station is located. This model employs a CNN-LSTM architecture incorporating an attention mechanism. The model inputs include the power station's historical power output for the past three days, inverted irradiance data from four surrounding regions for the past three days, and the predicted irradiance data (NWP) for the predicted date. The output is the blizzard weather label for the predicted date.

[0315] The model structure includes: a CNN module: performing 1D convolutions on each irradiance sequence to extract local features; an LSTM module: inputting the features extracted by the CNN into the LSTM to capture long-term dependencies in the time series; an attention mechanism: weighting the output sequences of the CNN-LSTM to learn the importance of different input sequences; and a fully connected layer: mapping the features output by the attention mechanism to binary classification results (blizzard weather labels).

[0316] This embodiment innovatively considers the irradiance of the area surrounding the power station, as the irradiance in these areas is highly correlated with the irradiance of the power station itself. In reality, only historical measured irradiance data of the power station's own measurement points are typically available, lacking data for its surrounding areas. Therefore, traditional methods fail to fully utilize meteorological observation information around the power station. Given that extreme weather events, especially snowfall, are often influenced by the propagation path of cold air waves, specifically selecting irradiance data from the area surrounding the power station is crucial for improving the accuracy of blizzard weather identification. Successfully acquiring this data through a resource inversion model and incorporating it into the blizzard weather prediction model is a significant innovation. The employed CNN-LSTM model combines the ability of LSTM to capture long-term dependencies in time series with the ability of CNN-enhanced models to perceive short-term changes. Simultaneously, it automatically learns the importance of different input features through an attention mechanism, thus highlighting the key role of predicting irradiance in the area where the power station is located. This model design not only effectively captures complex relationships in time series but also reflects the priority of key features, making it highly suitable for blizzard weather prediction tasks.

[0317] Step 6 describes the construction of a day-ahead power ramp-up prediction model for photovoltaic clusters under blizzard weather conditions, based on multi-dimensional observation information. The model uses the GCN model, with each photovoltaic power station as a node. The node characteristics of each power station include the forecast irradiance data for the predicted date, historical power data for the past 7 days, and the blizzard weather label for the predicted date.

[0318] The input characteristics of each photovoltaic power station (node) can be represented as follows:

[0319] (0.21)

[0320] Among them: I t The forecast irradiance data for the predicted day is in shape 96×1 with a time resolution of 15 minutes. This is historical photovoltaic power data for the past 7 days, in a 7×96 format; W t The label for the predicted blizzard weather is 96×1 and is a binary label indicating whether it is blizzard weather.

[0321] After concatenating these features, the node feature matrix is ​​as follows:

[0322] (0.22)

[0323] Where: N is the number of nodes; F is the feature dimension of each node.

[0324] The adjacency matrix A between photovoltaic power plants is calculated based on the Pearson correlation coefficient of historical photovoltaic power output. For any two photovoltaic power plants i and j, the correlation coefficient is calculated using the following formula:

[0325] (0.23)

[0326] Where: i, j are the photovoltaic power station indices, representing any two power stations respectively; Let i be the power output of the photovoltaic power station at time step k; Let be the average power output of the historical power sequence of photovoltaic power station i within the time window T; T is the length of the historical time sequence, and k is the time index, taking k=1,2,…,T.

[0327] Adjacency matrix A∈R N*N The element is defined as:

[0328] (0.24)

[0329] Among them, A ij The elements of the adjacency matrix represent the relevance edge weights between power station i and power station j, r. ij Let A be the Pearson correlation coefficient calculated from the historical power series, and let A be the adjacency matrix with dimension R. N*N Where N is the number of photovoltaic power stations, A∈R N*N This indicates that the adjacency matrix is ​​an N-row, N-column real matrix.

[0330] To adapt to the GCN model, the adjacency matrix needs to be normalized to... :

[0331] (0.25)

[0332] Where: I is the identity matrix, used to add self-loops; A is the adjacency matrix; GCN is a Graph Convolutional Network, which extracts features from graph-structured data through convolution operations between the adjacency matrix and the feature matrix; and D is the degree matrix, defined as:

[0333] (0.26)

[0334] Among them, D iiLet A be the diagonal element of the degree matrix, representing the degree of node i. ij For the elements of the adjacency matrix, I ij Let j be an element of the identity matrix, and j be the node index.

[0335] The graph convolution operation used in GCN can be represented as:

[0336] (0.27)

[0337] in: Let be the output feature matrix of the (l+1)th layer of the graph convolutional network. Let H(0) be the input feature matrix of the graph convolutional layer l, and let H(0) = X. Let be the trainable weight matrix of the graph convolutional layer l; For activation functions; This is the normalized adjacency matrix;

[0338] Based on the spatial features extracted by GCN, a time series model, such as LSTM, is used to model the temporal dimension of the node features. LSTM stands for Long Short-Term Memory. Assume that after GCN computation, the output feature of each node is H∈R. N×T×F , where T is the number of time steps and F is the feature dimension.

[0339] The update formula for the time series model is:

[0340] (0.28)

[0341] Where: GRU is a gated loop unit, h t-1 This is the hidden state of the previous time step. For the input features at time step t, The hidden state at the current time step;

[0342] The hidden state is mapped to the power output using a fully connected layer:

[0343] (0.29)

[0344] in: W is the predicted power of power plant i at time step t. o and b o These are the weight matrix and bias vector of the fully connected layer, respectively, both of which are trainable parameters;

[0345] The output is the photovoltaic power of each power station for the next day, with a time resolution of 15 minutes.

[0346] In this embodiment, the blizzard weather label from the above steps is innovatively input into the prediction model. The impact of blizzard weather is taken into account in the power prediction, thereby obtaining more accurate power prediction results within the time range of blizzard occurrence, thus improving the reliability of power ramp prediction and early warning under blizzard weather.

[0347] The photovoltaic cluster power generation ramp-up prediction model based on multi-dimensional observation information under blizzard weather conditions described in step 7 uses three ramp-up event identification methods combined with expert voting to make a comprehensive judgment on the predicted power time series of each power station, thereby realizing the early warning of photovoltaic cluster power generation ramp-up.

[0348] like Figure 5 As shown, Figure 5 The flowchart for constructing a blizzard weather prediction model for the region where each photovoltaic power station is located within the cluster includes the following:

[0349] Step 71. Obtain the corresponding ramp identification results using the three ramp identification methods respectively. Specifically, the three ramp event identification methods involved are as follows:

[0350] Method 1. Short-time window ramp rate determination:

[0351] The ramp rate is the ratio of the absolute value of the power change of a photovoltaic power station within a certain time interval to the time interval itself. The formula is:

[0352] (0.30)

[0353] in: The gradeability is expressed as power change per time (MW / 15min). The observed power value at time t (in MW); The observation time interval is in minutes. =15 minutes;

[0354] At each time t, calculate the gradient rate. , where the threshold A ramp-up event is considered to have occurred when 5% of the power plant's installed capacity is selected, and this moment is marked as 1; otherwise, it is marked as 0. This method is suitable for capturing rapid changes over a short period of time.

[0355] (0.31)

[0356] Where: M 1(t) This is the hill-climbing event indicator for Method 1.

[0357] Method 2. Determining the ramp rate of a long-term window sliding motion

[0358] Set up a sliding window =4h, sliding step size is 15 minutes, traversing the time series Calculate the ramp rate for each time window:

[0359] (0.32)

[0360] in: The long-term ramp rate at time step t, The long window length is set to 4 hours, which corresponds to 16 sampling points, each with a 15-minute interval.

[0361] Determine if a hill-climbing event has occurred:

[0362] (0.33)

[0363] Where M2(t) is the hill-climbing event indicator of Method 2, and P2 is the threshold.

[0364] The threshold P2 is selected as 10% of the installed capacity of the power plant. The sliding window logic is that if a ramping event occurs within the window, all points in the current window are marked as 1. If no ramping event occurs, only the newly added points are marked as 0.

[0365] Method 3. Determining the power difference during long-term window sliding

[0366] Set up a sliding window =4h, sliding step size is 15 minutes; traverse the time series Calculate the power difference for each time window:

[0367] (0.34)

[0368] Where D3(t) is the power difference at time step t.

[0369] Determine if a hill-climbing event has occurred:

[0370] (0.35)

[0371] Where M3(t) is the hill-climbing event indicator of method 3, and P3 is the threshold.

[0372] The threshold P3 is selected as 15% of the power plant's installed capacity, and the sliding window logic is the same as in method 2.

[0373] Step 72. Use expert voting to comprehensively evaluate the three methods and obtain the climbing result. Specifically: Use expert voting to comprehensively evaluate the three methods and obtain the climbing result. For each time t, combine the evaluation results of the three methods:

[0374] (0.36)

[0375] Where V(t) is the comprehensive judgment result, M1(t), M2(t), and M3(t) are the climbing event indicators of the three methods respectively, and Majority is the majority voting function: if at least two of the three methods are judged as climbing events with a value of 1, then V(t) = 1, otherwise V(t) = 0.

[0376] If at least two of the three methods identify a climbing event as occurring, and this is marked as 1, then V(t) = 1; otherwise, V(t) = 0.

[0377] Output the hill-climbing event judgment result V(t) at each moment to realize photovoltaic power hill-climbing early warning.

[0378] Example 2

[0379] This invention provides another embodiment of a photovoltaic cluster day-ahead ramp prediction device based on multi-dimensional information under blizzard weather, used to implement a photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather, including:

[0380] The acquisition module is used to acquire relevant information about each site in the photovoltaic cluster of the preset area;

[0381] The tagging module is used to categorize blizzard weather tags based on information from each station.

[0382] The inversion model construction module is used to construct an inversion model of the irradiance satellite remote sensing data for each photovoltaic power station in the cluster.

[0383] The inversion module is used to invert the historical measured surface irradiance of the area near each photovoltaic power station in the cluster using the irradiance satellite remote sensing data inversion model;

[0384] The blizzard weather prediction module for each power station's location is used to build blizzard weather prediction models for each photovoltaic power station in the cluster.

[0385] The module for constructing a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard conditions based on multi-dimensional observation information is used to construct a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard conditions based on multi-dimensional observation information.

[0386] The early warning module is used to make a comprehensive judgment on the prediction results of the day-ahead ramp-up prediction model for power generation and to issue early warnings for the day-ahead ramp-up of power generation in photovoltaic clusters.

[0387] Example 3

[0388] The present invention provides another embodiment of a photovoltaic cluster day-ahead ramp prediction device based on multi-dimensional information under blizzard weather, which is used to implement a photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather.

[0389] This includes comprehensively judging the prediction results of the day-ahead ramp-up prediction model for power generation and issuing a day-ahead ramp-up warning for photovoltaic cluster power generation; specifically including:

[0390] Step 71. Obtain the corresponding slope recognition results using the three slope recognition methods respectively, as follows:

[0391] Method 1. Short-time window ramp rate determination:

[0392] The ramp rate is the ratio of the absolute value of the power change of a photovoltaic power station within a certain time interval to the time interval itself, as shown in the following formula:

[0393] (0.30)

[0394] in: The gradeability is expressed as power change per time (MW / 15min). The observed power value at time t (in MW); The observation time interval is in minutes. =15 minutes;

[0395] At each time t, calculate the gradient rate. , where the threshold Selecting 5% of the power plant's installed capacity is considered a ramp-up event, and this moment is marked as 1; otherwise, it is marked as 0. This is suitable for capturing rapid changes in a short period of time.

[0396] (0.31)

[0397] Where: M 1(t) This is the hill-climbing event indicator variable for Method 1;

[0398] Method 2. Determining the ramp rate over a long window sliding period;

[0399] Set up a sliding window =4h, sliding step size is 15 minutes, traversing the time series Calculate the ramp rate for each time window:

[0400] (0.32)

[0401] in: Let P(t) be the sliding ramp rate at time t, and P(t) be the power value at time t.

[0402] Determine if a hill-climbing event has occurred:

[0403] (0.33)

[0404] Where M2(t) is the hill-climbing event indicator variable of Method 2, and P2 is the threshold;

[0405] The threshold P2 is selected as 10% of the installed capacity of the power plant. The sliding window logic is that if a ramping event occurs in the window, all points in the current window are marked as 1. If no ramping event occurs, the newly added points are marked as 0.

[0406] Method 3. Determining the power difference during long-term window sliding;

[0407] Set up a sliding window =4h, sliding step size is 15 minutes; traverse the time series Calculate the power difference for each time window:

[0408] (0.34)

[0409] Where D3(t) is the power difference at time step t;

[0410] Determine if a hill-climbing event has occurred:

[0411] (0.35)

[0412] Where M3(t) is the hill-climbing event indicator variable of Method 3, and P3 is the threshold;

[0413] The threshold P3 is selected as 15% of the power plant's installed capacity, and the sliding window logic is the same as in method 2.

[0414] Step 72. Use expert voting to comprehensively evaluate the three methods and obtain the climbing result. Specifically, for each time t, combine the evaluation results of the three methods:

[0415] (0.36)

[0416] Where V(t) is the comprehensive judgment result, M1(t), M2(t), and M3(t) are the climbing event indicators of the three methods respectively, and Majority is the majority voting function: if at least two of the three methods are judged as climbing events with a value of 1, then V(t) = 1, otherwise V(t) = 0.

[0417] If at least two of the three methods identify a climbing event as occurring, and this is marked as 1, then V(t) = 1; otherwise, V(t) = 0.

[0418] Output the hill-climbing event judgment result V(t) at each moment to realize photovoltaic power hill-climbing early warning.

[0419] Example 4

[0420] Based on the same inventive concept, embodiments of the present invention also provide a computer device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor. When the processor executes the computer program, it implements the steps of the photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather described in Embodiment 1 or 2.

[0421] Example 5

[0422] Based on the same inventive concept, this embodiment of the invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in embodiment 1 or 2.

[0423] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0424] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0425] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0426] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0427] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the day-ahead ramp-up of photovoltaic clusters under blizzard conditions based on multi-dimensional information, characterized by: include: Obtain relevant information about each site in the photovoltaic cluster within the preset area; Snowstorm weather tags were assigned based on information from each station; For each station in the cluster, an inversion model for the irradiance satellite remote sensing data of each power station is constructed separately; Using the irradiance satellite remote sensing data inversion model, the historical measured surface irradiance of the area near each station in the cluster was retrieved; Based on the inverted surface irradiance data of the region, a blizzard weather prediction model is constructed for each power station in the cluster. Based on the blizzard weather prediction model, a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard weather conditions is constructed based on multi-dimensional observation information. A comprehensive assessment is made based on the prediction results of the day-ahead ramp-up forecasting model for power generation, and a day-ahead ramp-up warning for photovoltaic cluster power generation is issued.

2. The photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in claim 1, characterized in that: The relevant information for each station includes: historical power output of each station, geographical location information of each station, and meteorological data of each station; The historical power output data for each station are as follows: ； Wherein, P represents the set of historical power output data for all photovoltaic power plants within a preset time interval. Representing station i from time t0 to t 0+r The historical contribution of every moment Let represent the power output of power station i at time τ, n be the number of photovoltaic power stations in the photovoltaic cluster, and r be the length of the time interval, i.e., from the start time t0 to the end time t. 0+r The duration, where T is the set of sampling times, i.e. , This represents the output value of stadium 1 at time t0+r. Representing station n at t 0+r Output value at any given moment; The geographical location information of each station is: L=(L1,L2,…,Ln), and the geographical location information Li of station i includes the longitude and latitude of its location; The meteorological data from each station includes snowfall data and predicted irradiance data; the snowfall data is defined as the cumulative amount of snow that falls to the Earth's surface, with the accumulation period being the past hour, and the data is a time series with a time resolution of 1 hour; the predicted irradiance data is a time series with a time resolution of 15 minutes.

3. The photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in claim 1, characterized in that: The blizzard weather tags are defined based on information from each station; including: Temporal resolution for processing snowfall data; Calculate 24-hour cumulative snowfall using a sliding window; Based on the time resolution and cumulative snowfall, determine whether it is blizzard weather and obtain a blizzard weather label.

4. The photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in claim 1, characterized in that: The method involves constructing an irradiance satellite remote sensing data inversion model for each power station within the cluster; including: Use the clear sky index to classify weather patterns; Based on the pattern segmentation results, feature data under different weather conditions are selected as input features for the model; Based on the input characteristics, an inversion model for irradiance satellite remote sensing data under different weather patterns is established.

5. The photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in claim 4, characterized in that: The method of classifying weather conditions using the clear sky index is as follows: The clear sky index is defined as: ； Where: K t GHI is the hourly clear sky index, GHI is the ground horizontal irradiance, and H0 is the extraterrestrial solar irradiance. The formula for calculating the extraterrestrial solar irradiance H0 is as follows: ； Among them: I sc Here, is the solar constant, which is 1367 W / m², and n is the nth day of the year. The solar zenith angle; Calculate the hourly clear sky index K t This yields a daily clear sky index sequence of 24 points. According to the clear sky index K t K-means clustering was performed to classify the weather patterns of adjacent regions into three categories: clear skies, cloudy skies, and rainy skies. The clear sky index data is standardized as follows: ； Among them, K t，norm The standardized clear sky index, This represents the average of the clear sky index. Standard deviation; The K-means clustering algorithm was used to classify weather patterns into the following types: clear skies, cloudy, and rainy, as shown in the following formula: ； Where k=3 represents three weather types: clear sky, cloudy, and rainy. Based on the model segmentation results, feature data under different weather conditions are selected as input features for the model. Specifically, 14 channel data points for the power station location are acquired using the Fengyun-4A satellite, with a spatial range of 4km × 4km. For different weather patterns, the Pearson correlation coefficient is used to calculate the correlation between each channel data point and surface irradiance, as shown in the following formula: ； in, The channel data value refers to the observation of a certain channel of the FY-4A satellite at the i-th sample time in the pixel where the field station is located. The mean of the channel data samples. This represents the surface irradiance value. is the average surface irradiance sample, n is the number of samples, i.e., the number of time samples participating in the correlation calculation under this weather pattern, and i is the sample index, with values ​​i=1,2,…,n, representing the time sample number; Filtering correlation coefficients Channel data with a value greater than 0.5 were used as input features; The step of establishing an irradiance satellite remote sensing data inversion model under different weather patterns based on input features is as follows: Selecting support vector machine, convolutional neural network, K-nearest neighbor regression and gradient boosting algorithm to establish an irradiance satellite remote sensing data inversion model to invert surface irradiance; Support Vector Machines (SVMs) perform regression by constructing a hyperplane in a high-dimensional space. The regression objective of SVMs is solved by the following optimization problem: ； Constraints: ； ； ； Where W is the model weight vector. Here, C is the kernel function mapping for the input features, and C is the regularization parameter. Let n be the sample size and i be the sample index, where i = 1, 2, ..., n. ζ i is a slack variable, representing the deviation of sample i, and b is the bias / intercept term; Using the Radial Basis Function (RBF) kernel function: ； Where, x i、 x j Let be the input feature vector of the i-th and j-th samples. Let γ be the Euclidean distance between the two feature vectors, and γ be the kernel width parameter. Convolutional Layer: Extracts local spatial features. ； in, Let x be the response value of the k-th convolutional kernel at position (i,j) on the output feature map. i+m,j+n Let f(i,j) be the pixel value at offset (m,n) within the input window when the convolution window is aligned with the output position (i,j), where m=1,…,M and n=1,…,N are the row and column indices of the convolution kernel, respectively. Let be the weight of the k-th convolutional kernel at position (m,n). This is the bias term corresponding to the k-th convolution kernel; Pooling Layer: Performs feature dimensionality reduction. ； Among them, y i,j Let x be the pixel value at position (i,j) of the pooled feature map. 2i,2j x 2i+1,2j x 2i,2j+1 x 2i+1,2j+1 These are the four pixel values ​​within a 2×2 window with (2i,2j) as the top left corner in the feature map before pooling; Fully Connected Layer: Maps the extracted features to predicted irradiance values; K-nearest neighbor regression is a non-parametric method that predicts based on the average of the K nearest neighbors in the training set. Distance metric: Euclidean distance is used as the distance between samples. ； Where d is the distance metric function, d(x i, x j ) is the sample x i With sample x j Euclidean distance in the feature space, x i,k Let x be the feature component of the i-th sample in the k-th dimension. jk Let n be the feature component of the j-th sample in the k-th dimension, n be the number of feature dimensions, and k be the dimension index. Find the K nearest neighbor samples of the test sample and take the average of their surface irradiance: ； in, The predicted surface irradiance for the test sample, K is the nearest neighbor number, y i Let be the true surface irradiance of the i-th nearest neighbor sample; The goal of gradient boosting is to minimize the loss function iteratively. ； Among them, F m (x) represents the prediction of the strong learner for sample x after the m-th iteration, F m-1 (x) represents the predictions of the cumulative model in the first m-1 rounds, η is the learning rate / shrinkage coefficient, which controls the contribution of the base learner in this round to the overall model, h m (x) is the fitting function of the m-th base learner for the current residual; The model performance was evaluated using a validation set, with mean squared error (MSE) as the evaluation metric. ； Where n is the number of samples in the evaluation set, y i Let i be the true surface irradiance of the i-th sample. Let be the predicted surface irradiance of the model for the i-th sample; For each weather pattern, the model with the best performance on the validation set is selected as the final model.

6. The photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in claim 1, characterized in that: The aforementioned inversion model using satellite remote sensing data of irradiance is used to invert historical measured surface irradiance in the area near each station in the cluster; specifically, it includes: (1) Calculate the typical distribution of each channel feature under various weather patterns of the power station; The system acquires 14 channel characteristic data points for Power Station A under sunny, cloudy, and rainy weather conditions from the Fengyun-4 satellite and stores them separately; it calculates the statistical measures of the characteristic data for each weather type; if there are n days of data, each data point contains 14 characteristic dimensions, denoted as X. ij Where i represents the i-th day and j represents the j-th feature: average value : ； Where n is the sample size, i.e., the number of days used for statistics under this weather pattern, i is the sample index, taking i=1,2,…,n, corresponding to the i-th day, j is the feature / channel index, taking j=1,2,…,14, X ij Let j be the value of the j-th channel on the i-th day; Standard deviation : ； The above formula yields the typical distribution of each feature under the three weather patterns. (2) For each day's data in the adjacent area, calculate its similarity to various weather type patterns and determine the weather type; The typical distribution of each channel feature under three weather patterns for the power station was calculated. Specifically, feature data of 14 channels located in the four adjacent areas (north, south, east, and west) of each power station were obtained from the Fengyun-4 satellite. The data spatial range was 4 km × 4 km for each area, and the data of the adjacent areas on a certain day was set as a vector. =[y1,y2,...,y14], the average vector for a certain weather type =[μ1,μ2,...,μ14], and its Euclidean distance D with the adjacent region data is calculated as follows: ； Repeat the above process for each weather type to obtain the distances between adjacent areas and the three weather types; Based on the calculated Euclidean distance, the weather type corresponding to the minimum distance is selected as the prediction result; if the Euclidean distance for sunny weather is the minimum, then the weather in the adjacent area is predicted to be sunny on that day. (3) Input the corresponding features of the adjacent area into the irradiance inversion model to obtain the measured surface irradiance data; By using the channel data types that serve as input to the model under each weather pattern, and the corresponding channel data of the adjacent areas for each day as input to the irradiance inversion model, surface irradiance data for four adjacent 4km×4km areas of each power station are obtained.

7. The photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in claim 1, characterized in that: The aforementioned blizzard weather prediction model constructs a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard weather conditions based on multi-dimensional observation information. The model uses the GCN model, with each photovoltaic power station as a node. The node characteristics of each power station include the forecast irradiance data for the predicted day, the historical power data for 7 days, and the blizzard weather label for the predicted day. The input characteristics of each photovoltaic power station node are represented as follows: ； Among them: I t The forecast irradiance data for the predicted day is in shape 96×1 with a time resolution of 15 minutes. This is historical photovoltaic power data for the past 7 days, in a 7×96 format; W t The label for the predicted blizzard weather is 96×1 and is a binary label indicating whether it is blizzard weather. After concatenating the above features, the node feature matrix is ​​as follows: ； Where: N is the number of nodes; F is the feature dimension of each node; The adjacency matrix A between photovoltaic power plants is calculated based on the Pearson correlation coefficient of historical photovoltaic power output. For any two photovoltaic power plants i and j, the correlation coefficient is calculated using the following formula: ； Where: i, j are the photovoltaic power station indices, representing any two power stations respectively; Let i be the power output of photovoltaic power station i at time step k; Let be the average power output of photovoltaic power plant i within the time window T, where T is the length of the historical time series and k is the time index, taking k=1,2,…,T; Adjacency matrix A∈R N*N The element is defined as: ； Among them, A ij The elements of the adjacency matrix represent the relevance edge weights between power station i and power station j, r. ij Let A be the Pearson correlation coefficient calculated from the historical power series, and let A be the adjacency matrix with dimension R. N*N Where N is the number of photovoltaic power stations, A∈R N*N This indicates that the adjacency matrix is ​​an N-row, N-column real matrix; To adapt to the GCN model, the adjacency matrix is ​​normalized to... : ； Where: I is the identity matrix, used to add self-loops; A is the adjacency matrix; GCN is a Graph Convolutional Network, which extracts features from graph-structured data through convolution operations between the adjacency matrix and the feature matrix; and D is the degree matrix, defined as: ； Among them, D ii Let A be the diagonal element of the degree matrix, representing the degree of node i. ij For the elements of the adjacency matrix, I ij is an element of the identity matrix, and j is the node index; The graph convolution operation used in GCN is represented as follows: ； in: Let be the output feature matrix of the (l+1)th layer of the graph convolutional network. Let H(0) be the input feature matrix of the graph convolutional layer l, and let H(0) = X. Let be the trainable weight matrix of the graph convolutional layer l; For activation functions; This is the normalized adjacency matrix; Based on the spatial features extracted by GCN, a time series model is used to model the temporal dimension of the node features; where LSTM is a Long Short-Term Memory network; it is assumed that after GCN computation, the output feature of each node is H∈R. N×T×F Where T is the number of time steps and F is the feature dimension; The update formula for the time series model is: ； Where: GRU is a gated loop unit, h t-1 This is the hidden state of the previous time step. For the input features at time step t, The hidden state at the current time step; The hidden state is mapped to the power output using a fully connected layer: ； in: W is the predicted power of power plant i at time step t. o and b o These are the weight matrix and bias vector of the fully connected layer, respectively, both of which are trainable parameters; The output is the photovoltaic power of each power station for the next day, with a time resolution of 15 minutes.

8. The photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in claim 1, characterized in that: The method involves comprehensively judging the prediction results of the day-ahead ramp-up prediction model for power generation and issuing a day-ahead ramp-up warning for photovoltaic cluster power generation; specifically, it includes: The corresponding slope recognition results were obtained using three different slope recognition methods, as follows: Method 1. Short-time window ramp rate determination: The ramp rate is the ratio of the absolute value of the power change of a photovoltaic power station within a certain time interval to the time interval itself, as shown in the following formula: ； in: The gradeability is expressed as power change per time (MW / 15min). The observed power value at time t (in MW); The observation time interval is in minutes. =15 minutes); At each time t, calculate the gradient rate. , where the threshold Selecting 5% of the power plant's installed capacity is considered a ramp-up event, and this moment is marked as 1; otherwise, it is marked as 0. This is suitable for capturing rapid changes in a short period of time. ； Where: M 1(t) This is the hill-climbing event indicator for Method 1; Method 2. Determining the ramp rate of a window sliding over a long period of time; Set up a sliding window =4h, sliding step size is 15 minutes, traversing the time series Calculate the ramp rate for each time window: ； in: The long-term ramp rate at time step t, The long window length is set to 4 hours, which corresponds to 16 sampling points, each with a 15-minute interval. Determine if a hill-climbing event has occurred: ； Where M2(t) is the hill-climbing event indicator of Method 2, and P2 is the threshold; threshold Select 10% of the power plant's installed capacity. The sliding window logic is as follows: if a ramp event occurs within the window, mark all points in the current window as 1; if no ramp event occurs, mark the new points as 0. Method 3. Determining the power difference during long-term window sliding; Set up a sliding window =4h, sliding step size is 15 minutes; traverse the time series Calculate the power difference for each time window: ； Where D3(t) is the power difference at time step t; Determine if a hill-climbing event has occurred: ； Where M3(t) is the hill-climbing event indicator of method 3, and P3 is the threshold; threshold Select 15% of the power plant's installed capacity; the sliding window logic is the same as in Method 2. An expert voting method was used to comprehensively evaluate the three methods and obtain the climbing result. Specifically, for each time t, the evaluation results of the three methods were combined: ； Where V(t) is the comprehensive judgment result, M1(t), M2(t), and M3(t) are the climbing event indicators of the three methods, respectively, and Majority is the majority voting function: if at least two of the three methods are judged as climbing events (value 1), then V(t) = 1, otherwise V(t) = 0. If at least two of the three methods identify a climbing event as occurring, mark it as 1. =1, otherwise =0; Output the hill-climbing event judgment result at each time step. This enables early warning of photovoltaic power ramp-up.

9. A photovoltaic cluster day-ahead ramp prediction device based on multi-dimensional information under blizzard weather, characterized by: include: The acquisition module is used to acquire relevant information about each site in the photovoltaic cluster of the preset area; The tagging module is used to categorize blizzard weather tags based on information from each station. The inversion model construction module is used to construct an inversion model of the irradiance satellite remote sensing data for each power station in the cluster. The inversion module is used to invert the historical measured surface irradiance of the area near each station in the cluster using the irradiance satellite remote sensing data inversion model; The blizzard weather prediction module for each power station area is used to construct a blizzard weather prediction model for each power station area in the cluster based on the inverted surface irradiance data of the area. The module for constructing a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard conditions based on multi-dimensional observation information is used to construct a day-ahead ramp-up prediction model for photovoltaic cluster power generation under blizzard conditions based on multi-dimensional observation information. The early warning module is used to make a comprehensive judgment on the prediction results of the day-ahead ramp-up prediction model for power generation and to issue early warnings for the day-ahead ramp-up of power generation in photovoltaic clusters.

10. The photovoltaic cluster day-ahead ramp prediction device based on multi-dimensional information under blizzard weather as described in claim 9, characterized in that: The early warning module is used to make a comprehensive judgment on the prediction results of the day-ahead ramp-up prediction model for power generation and to issue a day-ahead ramp-up warning for the power generation of the photovoltaic cluster. These include: The corresponding slope recognition results were obtained using three different slope recognition methods, as follows: Method 1. Short-time window ramp rate determination: The ramp rate is the ratio of the absolute value of the power change of a photovoltaic power station within a certain time interval to the time interval itself, as shown in the following formula: ； in: The gradeability is expressed as power change per time (MW / 15min). The observed power value at time t (in MW); The observation time interval is in minutes. =15 minutes); At each time t, calculate the gradient rate. , where the threshold Selecting 5% of the power plant's installed capacity is considered a ramp-up event, and this moment is marked as 1; otherwise, it is marked as 0. This is suitable for capturing rapid changes in a short period of time. ； Where: M 1(t) This is the hill-climbing event indicator variable for Method 1; Method 2. Determining the ramp rate of a window sliding over a long period of time; Set up a sliding window =4h, sliding step size is 15 minutes, traversing the time series Calculate the ramp rate for each time window: ； in: Let P(t) be the sliding ramp rate at time t, and P(t) be the power value at time t. Determine if a hill-climbing event has occurred: ； Where M2(t) is the hill-climbing event indicator variable of Method 2, and P2 is the threshold; threshold Select 10% of the power plant's installed capacity. The sliding window logic is as follows: if a ramp event occurs within the window, mark all points in the current window as 1; if no ramp event occurs, mark the new points as 0. Method 3. Determining the power difference during long-term window sliding; Set up a sliding window =4h, sliding step size is 15 minutes; traverse the time series Calculate the power difference for each time window: ； Where D3(t) is the power difference at time step t; Determine if a hill-climbing event has occurred: ； Where M3(t) is the hill-climbing event indicator variable of Method 3, and P3 is the threshold; The threshold P3 is selected as 15% of the power plant's installed capacity, and the sliding window logic is the same as in method 2. Step 72. Use expert voting to comprehensively evaluate the three methods and obtain the climbing result. Specifically, for each time t, combine the evaluation results of the three methods: ； Where V(t) is the comprehensive judgment result, M1(t), M2(t), and M3(t) are the climbing event indicators of the three methods, respectively, and Majority is the majority voting function: if at least two of the three methods are judged as climbing events (value 1), then V(t) = 1, otherwise V(t) = 0. If at least two of the three methods identify a climbing event as occurring, mark it as 1. =1, otherwise =0; Output the hill-climbing event judgment result at each time step. This enables early warning of photovoltaic power ramp-up.

11. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in any one of claims 1-8.

12. A computer storage medium, characterized in that: The computer storage medium contains a computer program, which, when executed by a processor, implements the steps of the photovoltaic cluster day-ahead ramp prediction method based on multi-dimensional information under blizzard weather as described in any one of claims 1-8.

Images