A method for predicting the power generation of photovoltaic power generation

Abstract

The present invention relates to the technical field of photovoltaic power generation power prediction, and discloses a method for predicting the power generation power of photovoltaic power generation, including the following steps: Step S101, collecting the characteristic parameters of a photovoltaic power generation unit and dividing them into M clusters through cluster analysis; Step S102, constructing the characteristic sequences of the M clusters; Step S103, constructing time series graph structure data as the sample data of the training samples; Step S104, collecting the total value of the output power of the photovoltaic power generation unit in the second preset time period T2 as the sample label of the training samples; Step S105, repeatedly executing Step S101 to Step S104 until U training samples are obtained, and training the power generation power prediction model through the training samples; The present invention constructs time series graph structure data from the characteristic parameters of multiple photovoltaic power generation units, and performs information aggregation on the time series graph structure data through the power generation power prediction model, thereby improving the prediction accuracy of the power generation power of the entire photovoltaic power station.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation prediction, and more specifically, it relates to a method for predicting the power generation of photovoltaic power generation. Background Technique

[0002] With the increasing global demand for renewable energy, photovoltaic power generation, as a clean and sustainable energy form, has been widely applied. However, the output power of photovoltaic power generation is greatly affected by weather conditions (such as solar radiation, temperature, cloud cover, etc.), with strong volatility and uncertainty, which brings great challenges to power grid dispatching and energy management. Therefore, accurately predicting the power generation of photovoltaic power stations is of great significance for optimizing power dispatching and improving the utilization rate of photovoltaics.

[0003] Photovoltaic power stations usually consist of multiple photovoltaic generator sets, which directly convert solar energy into electrical energy using the photovoltaic effect. Currently, the historical operation data of photovoltaic power stations are classified by grey relational analysis to obtain similar-day data sets under different weather types, and time series prediction models (such as LSTM, GRU, etc.) are trained through the similar-day data sets to eliminate the influence of operation data fluctuations caused by weather factors. Then, the parameters of the time series prediction model are updated through optimization algorithms (such as particle swarm optimization algorithm, whale optimization algorithm, etc.) to avoid it falling into local optimal solutions, thereby improving the prediction accuracy of the time series prediction model.

[0004] However, the above scheme does not consider the influence of individual differences between different photovoltaic generator sets on the power generation of the entire photovoltaic power station. For example, the geographical location, azimuth angle, shading conditions, and aging degree of each photovoltaic generator set are different, resulting in differences in the response of different photovoltaic generator sets to meteorological conditions at the same time, thus leading to low prediction accuracy of the power generation of the entire photovoltaic power station. Summary of the Invention

[0005] The present invention provides a method for predicting the power generation of photovoltaic power generation to solve the technical problems in the above background technique.

[0006] The present invention provides a method for predicting the power generation of photovoltaic power generation, including the following steps:

[0007] Step S101, within the first preset time period T1, collect the characteristic parameters of K photovoltaic generator sets in the photovoltaic power station at a preset time interval t, and divide the K photovoltaic generator sets into M clusters through cluster analysis;

[0008] The characteristic parameters include: rated power, photoelectric conversion efficiency, installation inclination angle, azimuth angle, output power, solar radiation intensity, temperature, rainfall, effective light area, cumulative operation duration, number of fault occurrences, and aging degree;

[0009] Step S102, construct the feature sequences of M clusters through a sliding window;

[0010] The feature sequence includes N sequence units, each sequence unit is represented by a feature vector, and the feature vector is obtained by preprocessing the feature parameters of all photovoltaic generator sets in the cluster through a sliding window;

[0011] Step S103, construct the time series graph structure data based on the feature sequences of M clusters as the sample data of the training samples;

[0012] The time series graph structure data includes: M nodes, the features of the nodes, and the edges between the nodes;

[0013] The M nodes are respectively mapped to the M clusters;

[0014] The features of the nodes are represented by the feature sequences of the clusters mapped to the nodes;

[0015] Step S104, after the first preset time period T1, collect the sum value of the output powers of K photovoltaic generator sets in the photovoltaic power station in the second preset time period T2 as the sample label of the training sample;

[0016] Step S105, repeatedly execute Step S101 to Step S104 until U training samples are obtained, and train the power prediction model through the training samples;

[0017] The input of the power prediction model corresponds to the sample data of the training sample;

[0018] The output of the power prediction model corresponds to the sample label of the training sample;

[0019] Step S106, execute Step S101 to Step S103 to obtain the time series graph structure data and input it into the trained power prediction model, and output the predicted value of the power generation of the photovoltaic power station in the second preset time period T2.

[0020] Furthermore, , where length represents the size of the sliding window, step represents the step size of the sliding window, floor represents rounding down, the size and step size of the sliding window are both user-defined parameters, and the first preset time period T1, the preset time interval t, the number K of photovoltaic generator sets, the number M of clusters, the second preset time period T2, and the number U of training samples are all user-defined parameters.

[0021] Furthermore, the effective light receiving area of the photovoltaic generator set The calculation formula is as follows:

[0022] ;

[0023] Among them represents the total area of the photovoltaic panels, represents the output power of the photovoltaic power generation unit, represents the photoelectric conversion efficiency of the photovoltaic power generation unit, represents the solar radiation intensity;

[0024] The calculation formula for the aging degree Aging of the photovoltaic power generation unit is as follows:

[0025] ;

[0026] ;

[0027] Among them and respectively represent the output power and rated power of the photovoltaic power generation unit, e represents the natural constant, represents the time decay coefficient, and its value range is between 0.01 and 0.03. Time represents the cumulative operation duration, expressed in years, represents the fault influence factor, and its value range is between 0.9 and 1. Freq represents the number of fault occurrences, represents the environmental adaptation coefficient, represents the rainfall influence coefficient, and its value range is between 0.01 and 0.02. Rain represents the rainfall.

[0028] Furthermore, K photovoltaic power generation units are divided into M clusters through cluster analysis, including the following steps:

[0029] Step S201, randomly select 1 photovoltaic power generation unit from the K photovoltaic power generation units as the central node, and then select M - 1 photovoltaic power generation units with the highest similarity to the characteristic parameters of this central node as the central nodes, and initialize the current iteration number to 1;

[0030] The calculation formula for the similarity Sim is as follows:

[0031] ;

[0032] where T = T1 / t, 1 ≤ a ≤ K, 1 ≤ b ≤ K, a ≠ b, and respectively represent the characteristic parameters of the a-th and b-th photovoltaic power generation units at the i-th time point. Dis represents the Euclidean distance, and cosine represents the cosine similarity, and respectively represent the first weight coefficient and the second weight coefficient, both of which are user-defined parameters, and + = 1.

[0033] Step S202: Calculate the similarity between the characteristic parameters of the remaining photovoltaic power generation units and those of the central node, and add the photovoltaic power generation unit with the maximum similarity to the central node to form a cluster;

[0034] The number of photovoltaic power generation units in each cluster must be greater than or equal to K / 4M;

[0035] Step S203: Calculate the average value of the characteristic parameters of all photovoltaic power generation units in each cluster respectively, and select the photovoltaic power generation unit with the maximum similarity between each cluster and this average value as the central node, and increment the current iteration count by 1;

[0036] Step S204: Repeat steps S202 to S203 until M clusters are obtained when the iteration termination condition is met;

[0037] The iteration termination conditions include: the current iteration count is greater than or equal to the maximum iteration count, or the similarity between the characteristic parameters of the central node at the current iteration count and those of the central node at the previous iteration count is less than the similarity threshold, where both the maximum iteration count and the similarity threshold are user-defined parameters.

[0038] Furthermore, preprocess the characteristic parameters of all photovoltaic power generation units within the cluster through a sliding window to obtain a feature vector, including the following steps:

[0039] Step S301: Fill it by taking the average value of the characteristic parameters at the adjacent two time points of the missing value;

[0040] Step S302: Obtain the first vector by taking the average value of each characteristic parameter of all photovoltaic power generation units within the sliding window;

[0041] Step S303: Perform normalization processing on the first vector through the Min-Max normalization method to obtain the feature vector.

[0042] Furthermore, if the correlation degree between any two clusters is greater than or equal to the correlation degree threshold, then build an edge between the nodes mapped to the clusters, where the correlation degree threshold is a user-defined parameter;

[0043] The calculation formula for the correlation degree Rel between two clusters is as follows:

[0044] ;

[0045] where 1 ≤ c ≤ M, 1 ≤ d ≤ M, c ≠ d, and respectively represent the nth sequence unit of the characteristic sequences of the cth and dth clusters, Dis represents the Euclidean distance, min represents taking the minimum value, max represents taking the maximum value, Denote the resolution coefficient, which is assigned a value of 0.5.

[0046] Furthermore, the power generation prediction model includes: M first hidden layers, M extraction layers, 1 second hidden layer, 1 third hidden layer, and 1 first classifier;

[0047] The M first hidden layers share weight parameters and bias parameters;

[0048] Each first hidden layer includes N hidden units. The nth hidden unit of the mth first hidden layer inputs the nth sequence unit of the feature sequence corresponding to the mth node of the time series graph structure data and outputs a first update vector, where 1 ≤ m ≤ M and 1 ≤ n ≤ N;

[0049] The mth extraction layer is used to extract the first update vector output by the Nth hidden unit of the mth first hidden layer as the feature of the mth node of the time series graph structure data;

[0050] The second hidden layer inputs the time series graph structure data and outputs a graph update matrix;

[0051] The graph update matrix includes M row vectors, and each row vector corresponds to a second update vector of a node;

[0052] The third hidden layer inputs the graph update matrix and outputs a third update vector;

[0053] Input the third update vector into the first classifier, and the classification space of the first classifier represents the power generation power of the photovoltaic power station in the second preset time period T2.

[0054] Furthermore, the first hidden layer is constructed based on a gated recurrent neural network.

[0055] Furthermore, the calculation formula of the second hidden layer includes:

[0056] ;

[0057] ;

[0058] ;

[0059] where 1 ≤ i ≤ M, 1 ≤ j ≤ M, i ≠ j, matrix represents the graph update matrix output by the second hidden layer, represents the second update vector of the ith node, represents the set of nodes that have an edge connection with the ith node, and respectively represent the first update vectors of the ith and jth nodes, represents the attention score between the ith node and the jth node, and is a real value between 0 and 1, and respectively represent the first weight parameter and the second weight parameter, MLP represents a multi-layer perceptron, and concat represents a concatenation operation. represents a stacking operation on the second update vectors of M nodes, and Swish represents the Swish activation function;

[0060] The calculation formula for the third hidden layer is as follows:

[0061] ;

[0062] where represents the third update vector output by the third hidden layer, represents the third weight parameter, and b represents the bias parameter.

[0063] Furthermore, before training the power generation prediction model, it is pre-trained first. During the pre-training of the power generation prediction model, the first update vector output by the Nth hidden unit of the first hidden layer is input into the second classifier. The classification space of the second classifier represents the total value of the output powers of all photovoltaic generator sets within the corresponding cluster in the next time interval. During the training process of the power generation prediction model, the maximum number of iterations of the power generation prediction model is set, and the difference between the predicted value output by the power generation prediction model and the sample label of the training sample in each iteration is specified as the loss function.

[0064] The beneficial effects of the present invention are as follows: The present invention constructs the characteristic parameters of multiple photovoltaic generator sets into time-series graph structure data, and aggregates the information of the time-series graph structure data through the power generation prediction model, thereby improving the prediction accuracy of the power generation power of the entire photovoltaic power station. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a flowchart of a method for predicting the power generation power of a photovoltaic power generation of the present invention;

[0066] Figure 2 is a flowchart of dividing K photovoltaic generator sets into M clusters through cluster analysis of the present invention;

[0067] Figure 3 is a flowchart of preprocessing the characteristic parameters of all photovoltaic generator sets within a cluster through a sliding window to obtain feature vectors of the present invention;

[0068] Figure 4 is a comparison diagram of verification results of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0069] Reference will now be made to exemplary embodiments to discuss the subject matter described herein. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the content of this specification. Each example may omit, substitute, or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0070] It should be noted that unless otherwise defined, the technical terms or scientific terms used in one or more embodiments of the present invention should be the ordinary meanings understood by those of ordinary skill in the art to which the present invention pertains. The "first", "second" and similar terms used in one or more embodiments of the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. Words such as "including" or "comprising" mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Upper", "lower", "left", "right", etc. are only used to represent relative position relationships, and when the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0071] As Figures 1 to 4 shown, a method for predicting the power generation of photovoltaic power generation includes the following steps:

[0072] Step S101, within a first preset time period T1, collect the characteristic parameters of K photovoltaic generator sets in a photovoltaic power station at a preset time interval t, and divide the K photovoltaic generator sets into M clusters through cluster analysis;

[0073] The characteristic parameters include: rated power, photoelectric conversion efficiency, installation inclination angle, azimuth angle, output power, solar radiation intensity, temperature, rainfall, effective light area, cumulative operation duration, number of fault occurrences, and degree of aging;

[0074] Step S102, construct characteristic sequences of the M clusters through a sliding window;

[0075] The characteristic sequence includes N sequence units, each sequence unit is represented by a characteristic vector, and the characteristic vector is obtained by preprocessing the characteristic parameters of all photovoltaic generator sets in the cluster through a sliding window;

[0076] Step S103, construct time-series graph structure data based on the characteristic sequences of the M clusters as the sample data of the training sample;

[0077] The time series graph structure data includes: M nodes, the features of the nodes, and the edges between the nodes;

[0078] The M nodes are respectively mapped to M clusters;

[0079] The features of the nodes are represented by the feature sequences of the clusters that are mapped to the nodes;

[0080] Step S104, after the first preset time period T1, collect the total value of the output powers of K photovoltaic generator sets in the photovoltaic power station during the second preset time period T2 as the sample label of the training sample;

[0081] Step S105, repeatedly execute Step S101 to Step S104 until U training samples are obtained, and train the power prediction model with the training samples;

[0082] The input of the power prediction model corresponds to the sample data of the training sample;

[0083] The output of the power prediction model corresponds to the sample label of the training sample;

[0084] Step S106, execute Step S101 to Step S103 to obtain the time series graph structure data and input it into the trained power prediction model, and output the predicted value of the power generation of the photovoltaic power station during the second preset time period T2.

[0085] It should be noted that for medium and large-scale photovoltaic power stations, which may consist of thousands of photovoltaic generator sets, if each photovoltaic generator set is used as a node of the time series graph structure data, then the scale of the time series graph structure data is relatively large, resulting in a slow operation speed of the power prediction model. Therefore, in the present invention, through clustering analysis, the photovoltaic generator sets with similar features are grouped into the same cluster, greatly reducing the scale of the time series graph structure data. Thus, while retaining the data features, the training speed and calculation efficiency of the power prediction model are greatly improved.

[0086] In an embodiment of the present invention, the first preset time period T1, the preset time interval t, the number K of photovoltaic generator sets, the number M of clusters, the second preset time period T2, and the number U of training samples are all user-defined parameters. Preferably, the default value of T1 is 24 hours, the default value of t is 15 minutes, K is set according to the actual number of photovoltaic generator sets in the photovoltaic power station, M should be much smaller than K, the default value of M is 20, the default value of T2 is 24 hours, and the default value of U is 2000. , where length represents the size of the sliding window, step represents the step size of the sliding window, floor represents rounding down. Both the size and step size of the sliding window are user-defined parameters. Preferably, length = 4×t and step = length / 2. Then, according to the default values of T1 and t above, N = 47 can be calculated.

[0087] It should be noted that the first preset time period T1 and the second preset time period T2 can be set according to the requirements of the prediction duration. For example, in short-term prediction, T1 and T2 can be set to 1 hour, and in long-term prediction, T1 and T2 can be set to 30 days. In addition, the sliding window can reduce the number of time points for collecting characteristic parameters, that is, reduce the length of the characteristic sequence. Similarly, while retaining the data characteristics, it can improve the training speed and calculation efficiency of the power generation prediction model, which will not be elaborated here.

[0088] In an embodiment of the present invention, the effective illumination area of the photovoltaic power generation unit The calculation formula is as follows:

[0089] ;

[0090] Where represents the total area of the photovoltaic panels, represents the output power of the photovoltaic power generation unit, represents the photoelectric conversion efficiency of the photovoltaic power generation unit, represents the solar radiation intensity;

[0091] The calculation formula for the aging degree Aging of the photovoltaic power generation unit is as follows:

[0092] ;

[0093] ;

[0094] Where and respectively represent the output power and rated power of the photovoltaic power generation unit, e represents the natural constant, represents the time decay coefficient, and its value range is between 0.01 and 0.03. time represents the cumulative operation duration, expressed in years, represents the fault influence factor, and its value range is between 0.9 and 1. Freq represents the number of fault occurrences, represents the environmental adaptation coefficient, represents the rainfall influence coefficient, and its value range is between 0.01 and 0.02. Rain represents the rainfall amount.

[0095] It should be noted that the photoelectric conversion efficiency refers to the efficiency of the photovoltaic power generation unit in converting the received solar radiation energy into electrical energy, which is related to the type of photovoltaic module (such as monocrystalline silicon, polycrystalline silicon, thin film, etc.) and is provided by the photovoltaic panel manufacturer; the installation inclination angle represents the inclination angle of the photovoltaic panel relative to the horizontal plane; the azimuth angle represents the orientation angle of the photovoltaic panel; the solar radiation intensity, temperature, and rainfall are all expressed by the solar radiation intensity, temperature, and rainfall per unit time, such as the preset time interval t; the effective illumination area represents the area on the photovoltaic panel that can effectively absorb sunlight.

[0096] In an embodiment of the present invention, the output power of the photovoltaic power generation unit can be directly measured by monitoring devices such as inverters or power meters installed in the photovoltaic power generation unit, or can be obtained by the following calculation formula:

[0097] ;

[0098] Where represents the output power of the photovoltaic power generation unit under standard test conditions, that is, the output power of the photovoltaic power generation unit under a solar radiation intensity of 1000 W / ㎡ and a temperature of 25 °C, represents the solar radiation intensity, T represents the actual temperature of the photovoltaic panel, represents the temperature coefficient, which is provided by the manufacturer and represents the degree of change in the output power of the photovoltaic panel with temperature, The default value of is 0.3%. The above calculation formula will have a certain error compared with the output power directly measured by the monitoring device, and is applicable to the situation where the monitoring device fails or the cost is limited.

[0099] In an embodiment of the present invention, as Figure 2 shown, clustering analysis is used to divide K photovoltaic power generation units into M clusters, including the following steps:

[0100] Step S201, randomly select 1 photovoltaic power generation unit from the K photovoltaic power generation units as the central node, and then select M - 1 photovoltaic power generation units with the highest similarity of characteristic parameters to this central node as the central nodes, and initialize the current iteration number to 1;

[0101] The calculation formula of the similarity Sim is as follows:

[0102] ;

[0103] Where T = T1 / t, 1 ≤ a ≤ K, 1 ≤ b ≤ K, a ≠ b, and respectively represent the characteristic parameters of the a-th and b-th photovoltaic power generation units at the i-th time point, Dis represents the Euclidean distance, and cosine represents the cosine similarity, and respectively represent the first weight coefficient and the second weight coefficient, both of which are user-defined parameters, and + = 1. Preferably, and are respectively set to 0.3 and 0.7.

[0104] Step S202: Calculate the similarity between the characteristic parameters of the remaining photovoltaic generator sets and the characteristic parameters of the central node, and add the photovoltaic generator set with the maximum similarity to the central node to form a cluster;

[0105] where the number of photovoltaic generator sets in each cluster must be greater than or equal to K / 4M;

[0106] Step S203: Calculate the average value of the characteristic parameters of all photovoltaic generator sets in each cluster respectively, and select the photovoltaic generator set with the maximum similarity between each cluster and the average value as the central node, and increment the current iteration count by 1;

[0107] Step S204: Repeat steps S202 to S203 until M clusters are obtained when the iteration termination condition is met;

[0108] The iteration termination conditions include: the current iteration count is greater than or equal to the maximum iteration count, or the similarity between the characteristic parameters of the central node of the current iteration count and the characteristic parameters of the central node of the previous iteration count is less than the similarity threshold. Both the maximum iteration count and the similarity threshold are user-defined parameters. Preferably, the maximum iteration count is set to 10, and the similarity threshold is set to 0.05.

[0109] It should be noted that in the present invention, by selecting M central nodes according to the maximum distance principle, randomness can be reduced, the separation degree of clusters can be improved, or M photovoltaic generator sets can also be directly randomly selected as the central nodes.

[0110] In an embodiment of the present invention, as Figure 3 shown, the characteristic vectors are obtained by preprocessing the characteristic parameters of all photovoltaic generator sets in the cluster through a sliding window, including the following steps:

[0111] Step S301: Fill it by taking the average value of the characteristic parameters of the adjacent two time points of the missing value;

[0112] Step S302: Obtain the first vector by taking the average value of each characteristic parameter of all photovoltaic generator sets within the sliding window;

[0113] That is, the size of the first vector is 1×12, and the 12 dimensions respectively represent the average values of the rated power, photoelectric conversion efficiency, installation inclination angle, azimuth angle, output power, solar radiation intensity, temperature, rainfall, effective lighting area, cumulative operation duration, number of fault occurrences, and aging degree of all photovoltaic power generation units;

[0114] Step S303, perform normalization processing on the first vector through the Min - Max normalization method to obtain a feature vector.

[0115] In an embodiment of the present invention, if the correlation degree between any two clusters is greater than or equal to the correlation degree threshold, an edge is constructed between the nodes mapped to the clusters, where the correlation degree threshold is a user - defined parameter. Preferably, the correlation degree threshold is set to 0.5;

[0116] The calculation formula for the correlation degree Rel between two clusters is as follows:

[0117] ;

[0118] where 1≤c≤M, 1≤d≤M, c≠d, and respectively represent the nth sequence unit of the feature sequences of the c - th and d - th clusters, Dis represents the Euclidean distance, min represents taking the minimum value, max represents taking the maximum value, represents the discrimination coefficient, and is assigned a value of 0.5.

[0119] In an embodiment of the present invention, the power generation prediction model includes: M first hidden layers, M extraction layers, 1 second hidden layer, 1 third hidden layer, and 1 first classifier;

[0120] The M first hidden layers share weight parameters and bias parameters;

[0121] Each first hidden layer includes N hidden units. The nth hidden unit of the m - th first hidden layer inputs the nth sequence unit of the feature sequence corresponding to the m - th node of the time - series graph structure data and outputs a first updated vector, where 1≤m≤M, 1≤n≤N;

[0122] The m - th extraction layer is used to extract the first updated vector output by the N - th hidden unit of the m - th first hidden layer as the feature of the m - th node of the time - series graph structure data;

[0123] The second hidden layer inputs the time - series graph structure data and outputs a graph update matrix;

[0124] The graph update matrix includes M row vectors, and each row vector corresponds to a second updated vector of a node;

[0125] The third hidden layer inputs the graph update matrix and outputs a third updated vector;

[0126] Input the third update vector into the first classifier, and the classification space of the first classifier represents the power generation power of the photovoltaic power station in the second preset time period T2.

[0127] In an embodiment of the present invention, the first hidden layer is constructed based on GRU (Gated Recurrent Unit), and can also be constructed based on LSTM (Long Short-Term Memory). In addition, the number of the first hidden layers is determined by the number M of clusters. This parallel design can greatly improve the calculation speed of the power generation power prediction model.

[0128] In an embodiment of the present invention, the calculation formula of the second hidden layer includes:

[0129] ;

[0130] ;

[0131] ;

[0132] where 1 ≤ i ≤ M, 1 ≤ j ≤ M, i ≠ j, matrix represents the graph update matrix output by the second hidden layer, represents the second update vector of the i-th node, represents the set of nodes having an edge connection with the i-th node, and respectively represent the first update vectors of the i-th and j-th nodes, represents the attention score between the i-th node and the j-th node, and is a real value between 0 and 1, and respectively represent the first weight parameter and the second weight parameter, MLP represents a multi-layer perceptron, concat represents a concatenation operation, represents a stacking operation on the second update vectors of M nodes, Swish represents the Swish activation function;

[0133] The calculation formula of the third hidden layer is as follows:

[0134] ;

[0135] where represents the third update vector output by the third hidden layer, represents the third weight parameter, and b represents the bias parameter.

[0136] It should be noted that both the weight parameters and bias parameters in the power generation prediction model are learnable parameters. For example, if the size of the first update vector is 1×Q1, the first weight parameter and the second weight parameter can be designed as a matrix of size Q1×Q2, then the size of the second update vector is 1×Q2, and the size of the graph update matrix is M×Q2. The third weight parameter can be designed as a vector of size Q2×1, and the size of the third update vector is M×1, where both Q1 and Q2 are user-defined parameters. For example, Q1 is set to 16 and Q2 is set to 8.

[0137] In an embodiment of the present invention, before training the power generation prediction model, it is pre-trained first. During the pre-training of the power generation prediction model, the first update vector output by the Nth hidden unit of the first hidden layer is input into the second classifier. The classification space of the second classifier represents the total value of the output powers of all photovoltaic generator sets within the corresponding cluster in the next time interval. During the training of the power generation prediction model, the maximum number of iterations of the power generation prediction model is set, and the difference between the predicted value output by the power generation prediction model and the sample label of the training sample in each iteration is specified as the loss function.

[0138] It should be noted that pre-training can accelerate the training speed of the power generation prediction model, reduce the number of training samples during training, and accelerate the convergence speed of the model. During the training process, the weight parameters and bias parameters in the power generation prediction model are updated through the gradient descent algorithm and the chain rule. The training of the neural network model belongs to conventional technical means and will not be elaborated here.

[0139] As Figure 4 shown, the power generation prediction model provided by the present invention and the time series prediction model (constructed based on LSTM) of the prior art are respectively used to test 10 verification data. The verification data is the power generation of 1000 photovoltaic generator sets in one day. It is found through observation that the prediction accuracy of the power generation prediction model provided by the present invention is higher.

[0140] The above describes the embodiments of this embodiment, but this embodiment is not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make many forms, all of which fall within the protection scope of this embodiment.

Claims

1. A method for predicting the power generation of photovoltaic power generation, characterized in that, It includes the following steps: Step S101, within the first preset time period T1, collect the characteristic parameters of K photovoltaic generator sets in the photovoltaic power station at a preset time interval t, and divide the K photovoltaic generator sets into M clusters through cluster analysis; The characteristic parameters include: rated power, photoelectric conversion efficiency, installation inclination angle, azimuth angle, output power, solar radiation intensity, temperature, rainfall, effective light area, cumulative operation duration, number of fault occurrences, and degree of aging; Step S102, construct the characteristic sequences of the M clusters through a sliding window; The characteristic sequence includes N sequence units, each sequence unit is represented by a feature vector, and the feature vector is obtained by preprocessing the characteristic parameters of all photovoltaic generator sets in the cluster through a sliding window; Step S103, construct time series graph structure data based on the characteristic sequences of the M clusters as the sample data of the training samples; The time series graph structure data includes: M nodes, the characteristics of the nodes, and the edges between the nodes; The M nodes are respectively mapped to the M clusters; The characteristics of the nodes are represented by the characteristic sequences of the clusters mapped to the nodes; Step S104, after the first preset time period T1, collect the total value of the output powers of the K photovoltaic generator sets in the photovoltaic power station in the second preset time period T2 as the sample label of the training sample; Step S105, repeatedly execute Step S101 to Step S104 until U training samples are obtained, and train the power prediction model through the training samples; The input of the power prediction model corresponds to the sample data of the training sample; The output of the power prediction model corresponds to the sample label of the training sample; Step S106, execute Step S101 to Step S103 to obtain the time series graph structure data and input it into the trained power prediction model, and output the predicted value of the power generation of the photovoltaic power station in the second preset time period T2.

2. The power generation prediction method for photovoltaic power generation according to claim 1, characterized in that, , where length represents the size of the sliding window, step represents the step size of the sliding window, floor represents rounding down. Both the size and step size of the sliding window are user-defined parameters. Moreover, the first preset time period T1, the preset time interval t, the number K of photovoltaic generator sets, the number M of clusters, the second preset time period T2, and the number U of training samples are all user-defined parameters.

3. A method for predicting the power generation of a photovoltaic power generation according to claim 1, characterized in that Effective illumination area of a photovoltaic power generation unit The calculation formula is as follows: ； Among them represents the total area of the photovoltaic panel represents the output power of the photovoltaic power generation unit represents the photoelectric conversion efficiency of the photovoltaic power generation unit represents the solar radiation intensity The calculation formula for the degree of aging Aging of the photovoltaic generator set is as follows: ； ； Among them and represent the output power and rated power of the photovoltaic power generation unit respectively, e represents the natural constant, represents the time decay coefficient, with a value range between 0.01 and 0.03, time represents the cumulative operation duration, expressed in years, represents the fault impact factor, with a value range between 0.9 and 1, Freq represents the number of fault occurrences, represents the environmental adaptation coefficient, represents the rainfall impact coefficient, with a value range between 0.01 and 0.02, Rain represents the rainfall amount.

4. A method for predicting the power generation of a photovoltaic power generation according to claim 1, characterized in that, Dividing the K photovoltaic generator sets into M clusters through cluster analysis includes the following steps: Step S201, randomly select 1 photovoltaic generator set from the K photovoltaic generator sets as the central node, and then select M - 1 photovoltaic generator sets with the highest similarity to the characteristic parameters of this central node as the central nodes, and initialize the current iteration number to 1; The calculation formula for the similarity Sim is as follows: ； where \(T = T1 / t\), \(t\) represents the preset time interval, \(1\leq a\leq K\), \(1\leq b\leq K\), \(a\neq b\), and respectively represent the characteristic parameters of the \(a\)-th and \(b\)-th photovoltaic power generation units at the \(i\)-th time point, \(Dis\) represents the Euclidean distance, and \(cosine\) represents the cosine similarity, and respectively represent the first weight coefficient and the second weight coefficient, both of which are user-defined parameters, and + = 1; Step S202, calculate the similarity between the characteristic parameters of the remaining photovoltaic generator sets and the characteristic parameters of the central nodes, and add the photovoltaic generator set with the highest similarity to the central nodes to form a cluster; where the number of photovoltaic generator sets in each cluster must be greater than or equal to K / 4M; Step S203, calculate the average value of the characteristic parameters of all photovoltaic generator sets in each cluster respectively, and select the photovoltaic generator set with the highest similarity to this average value in each cluster as the central node, and add 1 to the current iteration number; Step S204, repeat Step S202 to S203 until M clusters are obtained when the iteration termination condition is met; The iteration termination conditions include: the current iteration number is greater than or equal to the maximum iteration number, or the similarity between the characteristic parameters of the central node at the current iteration number and the characteristic parameters of the central node at the previous iteration number is less than the similarity threshold, where both the maximum iteration number and the similarity threshold are user-defined parameters.

5. A method for predicting the power generation of a photovoltaic power generation according to claim 1, characterized in that, Preprocessing the characteristic parameters of all photovoltaic power generation units in the cluster through a sliding window to obtain a feature vector, including the following steps: Step S301, filling the missing value by taking the average of the characteristic parameters at two adjacent time points. Step S302, obtaining a first vector by taking the average of each characteristic parameter of all photovoltaic power generation units within the sliding window. Step S303, performing normalization processing on the first vector through the Min-Max normalization method to obtain the feature vector.

6. A method for predicting the power generation of a photovoltaic power generation according to claim 1, wherein If the correlation degree between any two clusters is greater than or equal to the correlation degree threshold, an edge is constructed between the nodes mapped to the clusters, where the correlation degree threshold is a user-defined parameter; The calculation formula for the correlation degree Rel between two clusters is as follows: ； where 1 ≤ c ≤ M, 1 ≤ d ≤ M, and c ≠ d, and respectively represent the n-th sequence element of the feature sequences of the c-th and d-th clusters, Dis represents the Euclidean distance, min represents taking the minimum value, and max represents taking the maximum value. represents the discrimination coefficient, which is assigned a value of 0.

5.

7. A method for predicting the power generation of a photovoltaic power generation according to claim 1, characterized in that, The power generation prediction model includes: M first hidden layers, M extraction layers, 1 second hidden layer, 1 third hidden layer, and 1 first classifier; The M first hidden layers share weight parameters and bias parameters; Each first hidden layer includes N hidden units. The nth hidden unit of the mth first hidden layer inputs the nth sequence unit of the feature sequence corresponding to the mth node of the time series graph structure data and outputs a first updated vector, where 1 ≤ m ≤ M and 1 ≤ n ≤ N; The mth extraction layer is used to extract the first updated vector output by the Nth hidden unit of the mth first hidden layer as the feature of the mth node of the time series graph structure data; The second hidden layer inputs the time series graph structure data and outputs a graph update matrix; The graph update matrix includes M row vectors, and each row vector corresponds to a second updated vector of a node; The third hidden layer inputs the graph update matrix and outputs a third updated vector; Input the third updated vector into the first classifier, and the classification space of the first classifier represents the power generation power of the photovoltaic power station in the second preset time period T2.

8. A method for predicting the power generation of a photovoltaic power generation according to claim 7, characterized in that, The first hidden layer is constructed based on a gated recurrent neural network.

9. A method for predicting the power generation of a photovoltaic power generation according to claim 7, wherein The calculation formula of the second hidden layer includes: ； ； ； where \(1\leq i\leq M\), \(1\leq j\leq M\), \(i\neq j\), matrix represents the graph update matrix output by the second hidden layer, represents the second update vector of the \(i\)-th node, represents the set of nodes that have edge connections with the \(i\)-th node, and represent the first update vectors of the \(i\)-th and \(j\)-th nodes respectively, represents the attention score between the \(i\)-th node and the \(j\)-th node, and is a real value between 0 and 1, and represent the first weight parameter and the second weight parameter respectively, MLP represents the multi-layer perceptron, and concat represents the concatenation operation, represents the stacking operation of the second update vectors of \(M\) nodes, and Swish represents the Swish activation function; The calculation formula of the third hidden layer is as follows: ； Among them represents the third updated vector output by the third hidden layer represents the third weight parameter, and b represents the bias parameter 10. A method for predicting the power generation of a photovoltaic power generation according to claim 1, characterized in that, Before training the power generation prediction model, pre-train it first. During the pre-training of the power generation prediction model, input the first updated vector output by the Nth hidden unit of the first hidden layer into the second classifier. The classification space of the second classifier represents the total value of the output powers of all photovoltaic power generation units within the corresponding cluster in the next time interval. During the training process of the power generation prediction model, set the maximum iteration number of the power generation prediction model, and specify the difference between the predicted value output by the power generation prediction model at each iteration number and the sample label of the training sample as the loss function.

Images