Short-term photovoltaic output prediction model construction method and prediction method

Abstract

The present application relates to the field of photovoltaic power generation prediction, and discloses a short-term photovoltaic output prediction model construction method and a prediction method, which are used for improving the prediction accuracy of photovoltaic output. The short-term photovoltaic output prediction model construction method comprises collecting historical data of a photovoltaic power generation system, the historical data of the photovoltaic power generation system comprising meteorological environment data of multiple regions and corresponding photovoltaic output data; using an RBF-MeanShift model to cluster the meteorological environment data to obtain multiple meteorological clusters; optimizing an initial ResNet model based on a DGBCO algorithm to obtain a feature extraction model, and using the feature extraction model to extract features of each meteorological cluster to generate multiple training samples; inputting the multiple training samples into a DeepAR model for training to obtain a source domain short-term photovoltaic output prediction model; obtaining a target domain small sample, and inputting the target domain small sample into the source domain short-term photovoltaic output prediction model for transfer learning to obtain a target domain short-term photovoltaic output prediction model.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, and in particular to a method for constructing a short-term photovoltaic power output prediction model and a prediction method. Background Technology

[0002] With the widespread application of photovoltaic power generation systems, accurate prediction of photovoltaic output is of great significance for power grid planning and energy dispatch. Research on the prediction of photovoltaic output in microgrids is particularly challenging due to the strong periodicity of the data, especially for short-term forecasts. Although many prediction models have been developed, the development of simple yet highly accurate models remains a key research direction for researchers in both academia and business due to the volatility of short-term data.

[0003] Furthermore, the photovoltaic power plant data collection database suffers from issues such as insufficient data samples and data loss in some areas, failing to meet training requirements. Additionally, a large amount of data is needed for pre-training during the cold start of a network in a new region. If historical data storage for that region is insufficient, network training will struggle to converge, resulting in unsatisfactory performance and low model prediction accuracy.

[0004] Therefore, existing technologies still need improvement and development. Summary of the Invention

[0005] This invention provides a method for constructing a short-term photovoltaic power output prediction model and a prediction method, which can be used to improve the accuracy of photovoltaic power output prediction.

[0006] The first aspect of this invention provides a method for constructing a short-term photovoltaic power output prediction model, the method comprising: Collect historical data of photovoltaic power generation systems, including meteorological and environmental data of multiple regions and corresponding photovoltaic power output data; The meteorological and environmental data were clustered using the RBF-MeanShift model to obtain multiple meteorological clusters; The initial ResNet model is optimized based on the DGBCO algorithm to obtain a feature extraction model. The feature extraction model is then used to extract features of each meteorological cluster to generate multiple training samples. The training samples include the features of each meteorological cluster and the corresponding photovoltaic power output data. Multiple training samples are input into the DeepAR model for training to obtain a source domain short-term photovoltaic power output prediction model. A small sample of the target domain is obtained and input into the short-term photovoltaic power output prediction model of the source domain for transfer learning to obtain the short-term photovoltaic power output prediction model of the target domain.

[0007] Preferably, the step of using the RBF-MeanShift model to cluster meteorological environmental data to obtain multiple meteorological clusters includes, prior to, performing dimensionless processing on the historical data of the power generation system.

[0008] Preferably, when clustering meteorological and environmental data using the RBF-MeanShift model, RBF is used as the kernel function, and the Bhattacharyya coefficient is defined as the similarity measure function of the meteorological and environmental data. The relational expression is as follows: (1) In the formula, The value of is between 0 and 1, p(y) represents the candidate target model, q represents the target model, and u is a natural number. The larger the value of p, the stronger the candidate target model. u (y) and target model q u The more similar they are.

[0009] Preferably, the formula (1) is Taylor expanded at the initial position y0, and the coefficients of Bhattacharyya are... To find the maximum value, we can obtain the Meanshift vector, which is calculated using the following formula:

[0010] In the formula, Let h represent the Meanshift vector. The scale, y1 represents the new iteration position, x i (i=1,2,…n) represents The position, g(·) is a unit kernel function, w i This represents the weight of each point, where w is the weight. i The relational expression is as follows:

[0011] In the formula, δ represents the Kronecker function, and b(x) i ) indicates that in x i The index value of the location.

[0012] Preferably, the network structure of the initial ResNet model is represented as follows: (4) In the formula, Let S represent a shallow unit, N represent a deep unit, i = 0, 1, ..., N-1, and w represent the weights. Let w be the optimization target S. i and set , R represents the set of real numbers, d represents the problem dimension, and the optimal vector is found by the DGBCO algorithm.

[0013] Preferably, said finding the optimal vector by the DGBCO algorithm comprises: S31, initialize the population size Sizepop, the maximum number of iterations , mutation rate mRate, the percentage epct of the population accounted for by the exploration group and the decline index k, and set the upper and lower limits of the search space for each solution as [S i,min , S i,max ; S32, calculate the fitness value of each solution S i to find the optimal solution with the best fitness value; S33, divide the individuals of the population into an exploration group and an exploitation group according to the set epct value, dynamically manage the number of individuals in each group at the beginning of each iteration, and increase the number of solutions in the exploration group if the optimal fitness does not change compared with the previous two iterations; S34, for the exploration group, obtain the crossover probability p', where p' is a random number between [0, 1]. If p'≥mRate, the individual searches around the current solution, which is achieved by iteratively searching for a better solution in its adjacent possible solutions, and the implemented calculation function is: (5) (6) In the formulas, and respectively represent coefficient vectors in the intervals [0, 2] and [0, 1], t represents the current iteration, represents the current solution vector, represents the diameter of the circle in the expected region where the solution will be searched; If p'<mRate, a mutant solution is generated, and local random perturbation is performed based on the crossover probability p', and the perturbation formula is expressed as: (7) In the formula, R represents a neighborhood function; S35, for the exploitation group, if p'≥mRate, the individual moves towards the optimal solution, and the implemented calculation function is: (8) (9) In the formulas, represents a random vector in the interval [0, 2], represents the optimal solution; If p' < 0.5, then the strategy of an individual searching for the optimal solution can be calculated using the following formula: (10) (11) (12) In the formula, and Both represent the random vector interval [0, 1]; k represents the control. A constant, k iterates from 2 to 0; S36. Modify individuals that are outside the search space and randomly change the order of individuals to swap the roles of exploration and development group members; S37. Determine whether the current iteration count has reached the set maximum iteration count; if yes, proceed to step S38; if no, repeat steps S32 to S36. S38. Optimization complete, optimal vector obtained.

[0014] Preferably, the DeepAR model network structure is represented as follows: (13) In the formula, This represents the predicted photovoltaic power output from time t0 to time T. This represents the actual photovoltaic power output from historical time 1 to t0-1. h represents the time-series characteristic parameter. j,t This represents the output of the j-th parameter of the hidden layer in the autoregressive recurrent network at time t, where j = 1, 2, ..., z, and z is the number of input feature parameters. j,t The relational expression is as follows: (14) In the formula, Given parameters for the LSTM cell states used in the autoregressive recurrent network, θ represents the neural network model being trained, and H represents the autoregressive recurrent network, with a likelihood function added to the autoregressive recurrent network. This makes the output a probability distribution prediction, and the likelihood function uses a Gaussian distribution function.

[0015] Preferably, the step of obtaining a small sample from the target domain and inputting the small sample from the target domain into the short-term photovoltaic power output prediction model of the source domain for transfer learning to obtain the short-term photovoltaic power output prediction model of the target domain includes: Obtain the latent features of photovoltaic power output data from a small sample in the target domain. , Represented as: ; Obtain the class labels of small samples in the target domain, and denote the feature set of the dataset of photovoltaic power curves of the k-th class in the target domain as . , Represented as: ,in This represents the size of the k-th class dataset, representing a small sample in the target domain. calculate Chinese characteristics and cosine similarity And based on cosine similarity get Fusion characteristics , and They are represented as follows: (15) (16) Based on the fusion features, small samples in the target domain The short-term photovoltaic power output prediction model in the input domain is transferred to the target domain to obtain the short-term photovoltaic power output prediction model.

[0017] A second aspect of the present invention provides a short-term photovoltaic power output prediction method, the short-term photovoltaic power output prediction method comprising: A short-term photovoltaic power output prediction model and a photovoltaic power generation digital twin model are constructed for the target domain. The method for constructing the short-term photovoltaic power output prediction model for the target domain adopts the short-term photovoltaic power output prediction model construction method described above. Meteorological environmental data within the forecast period is acquired, and features of the acquired meteorological environmental data are extracted. The extracted features are then input into the short-term photovoltaic power output prediction model of the target domain to obtain short-term photovoltaic power output prediction data. The acquired meteorological environmental data and the short-term photovoltaic output prediction data of the target domain are input into the photovoltaic power generation digital twin model, and then displayed through the photovoltaic power generation digital twin model; Obtain the actual photovoltaic power output data within the prediction period, calculate the difference between the actual photovoltaic power output data within the prediction period and the short-term photovoltaic power output prediction data, and optimize the short-term photovoltaic power output prediction model for the target domain based on the difference.

[0018] Preferably, the photovoltaic power generation digital twin model includes a real physical sub-model and a virtual sub-model. The real physical sub-model is used to model the meteorological characteristics, equipment information status characteristics, and output characteristics of the photovoltaic power generation system. The virtual sub-model includes simulating the operating status of the photovoltaic power generation system.

[0019] The technical solution provided by this invention introduces a MeanShift clustering algorithm with a radial basis function (RBF) neural network as the kernel function to adjust the position of the centroid of the candidate boxes, formulate similarity measurement criteria, and improve the performance of the model, enabling it to find and accurately cluster multiple meteorological clusters. Furthermore, the DGBCO algorithm is used to optimize the ResNet network structure, realizing the adaptive adjustment function of the neural network parameters, which can accurately extract the features of each meteorological cluster, further improving the accuracy of photovoltaic power output prediction. Moreover, by training the DeepAR model, a source domain short-term photovoltaic power output prediction model is obtained. The DeepAR model is trained on a large number of time series using an autoregressive model. Recursive network models can effectively learn global models from relevant time series data and are capable of learning complex patterns. This reduces the impact of uncertainties in data such as photovoltaic output and meteorology, enabling the prediction of probability distributions for data with time series characteristics. Finally, small samples from the target domain are obtained and input into the source domain short-term photovoltaic output prediction model for transfer learning. This yields the target domain short-term photovoltaic output prediction model. Transfer learning methods can be used to adjust sample weights, feature mappings, or model training between domains, reducing domain differences between the source and target domains, deeply mining data features between information, and achieving information sharing. Ultimately, this enables an adaptive adjustment mechanism. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the method for constructing a short-term photovoltaic power output prediction model provided in this embodiment of the invention; Figure 2 A flowchart of a short-term photovoltaic power output prediction method provided in an embodiment of the present invention. Detailed Implementation

[0021] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 As shown, an embodiment of the method for constructing a short-term photovoltaic power output prediction model in this invention includes: S101. Collect historical data of the photovoltaic power generation system. The historical data of the photovoltaic power generation system includes meteorological and environmental data of multiple regions and corresponding photovoltaic power output data.

[0023] S102. Use the RBF-MeanShift model to cluster meteorological and environmental data to obtain multiple meteorological clusters.

[0024] S103. Optimize the initial ResNet model based on the DGBCO algorithm to obtain the feature extraction model, and use the feature extraction model to extract the features of each meteorological cluster to generate multiple training samples. The training samples include the features of each meteorological cluster and the corresponding photovoltaic power output data.

[0025] S104. Input multiple training samples into the DeepAR model for training to obtain the source domain short-term photovoltaic power output prediction model.

[0026] S105. Obtain small samples from the target domain and input them into the source domain short-term photovoltaic power output prediction model for transfer learning to obtain the target domain short-term photovoltaic power output prediction model.

[0027] Understandably, the executing entity of this invention can be a short-term photovoltaic power output prediction model building device, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0028] In this embodiment, meteorological environmental data includes temperature, humidity, total solar irradiance, direct solar irradiance, diffuse solar irradiance, wind speed, wind direction, air pressure, and particulate matter index.

[0029] In this embodiment, the RBF-MeanShift model is used to cluster meteorological environmental data to obtain multiple meteorological clusters. Prior to this, the historical data of the power generation system is processed to be dimensionless.

[0030] Generally, dimensionless scaling can include min-max scaling, standardization, and decimal scaling. Min-max scaling refers to linearly scaling the data to a specified range, typically [0, 1] or [-1, 1]. Standardization transforms the data into a standard normal distribution with a mean of 0 and a variance of 1. Decimal scaling involves shifting the decimal point to make the data values ​​fall within the range [-1, 1].

[0031] It should be noted that the RBF-MeanShift model is a clustering model based on the Radial Basis Function (RBF) and the MeanShift algorithm.

[0032] RBF is a kernel function used to map data points to a high-dimensional feature space. In the RBF-MeanShift model, RBF is first used to map the original data points to a high-dimensional feature space, and then the MeanShift algorithm is applied to cluster the mapped data.

[0033] The MeanShift algorithm is a density-based clustering algorithm. Its core idea is to continuously update the position of data points, causing them to move towards the region with the highest local density, and eventually converge to the cluster center. The algorithm works by continuously calculating the mean shift vector of each data point and updating the position of the data point based on this vector until the convergence condition is met.

[0034] Preferably, in step S102, when clustering meteorological and environmental data using the RBF-MeanShift model, RBF is used as the kernel function, and the Bhattacharyya coefficient is defined as the similarity measure function of the meteorological and environmental data. The relational expression is as follows: (1) In the formula, The value of is between 0 and 1, p(y) represents the candidate target model, q represents the target model, and u is a natural number.

[0035] Furthermore, by Taylor expanding equation (1) at the initial position y0, and considering the Bhattacharyya coefficients... To find the maximum value, we can obtain the Meanshift vector, which is calculated using the following formula:

[0036] In the formula, Let h represent the Meanshift vector. The scale, y1 represents the new iteration position, x i (i=1,2,…n) represents The position, g(·) is a unit kernel function, w i This represents the weight of each point, where w is the weight. i The relational expression is as follows:

[0037] In the formula, δ represents the Kronecker function, and b(x) i ) indicates that in x i The index value of the location.

[0038] In this embodiment, in step S102, the network structure of the initial ResNet model is represented as follows: (4) In the formula, Let S represent a shallow unit, N represent a deep unit, i = 0, 1, ..., N-1, and w represent the weights. Let w be the optimization target S. i and set R represents the set of real numbers, d represents the problem dimension, and the optimal vector is found using the DGBCO algorithm.

[0039] It should be noted that the DGBCO (Differential Grey Wolf Binomial Crossover Optimization) algorithm is an optimization algorithm that combines grey wolf optimization with a binomial crossover operator. This algorithm is based on the binary crossover operation of grey wolf behavior and probabilities to optimize the parameters of a neural network.

[0040] Preferably, finding the optimal vector using the DGBCO algorithm includes: S31. Initialize population size Sizepop, maximum number of iterations. The mutation rate mRate, the percentage of the exploration group in the population epct, and the glide slope index k are used, and the upper and lower limits of the search space for each solution are set to [S]. i,min ,S i,max ].

[0041] Specifically, the population size is Sizepop=30, and the maximum number of iterations is... =500, mutation rate mRate=0.5, exploration group percentage of population epct=70, glide index k=2, and the upper and lower limits of the search space for each solution are [-10,10].

[0042] S32. Calculate each solution S i Find the optimal solution with the best fitness value.

[0043] S33. Divide the individuals in the population into two groups, exploration and development, according to the set epct value. At the beginning of each iteration, dynamically manage the number of individuals in each group. If the optimal fitness has not changed compared with the previous two iterations, increase the number of solutions in the exploration group.

[0044] S34. For the exploration group, obtain the crossover probability p', where p' is a random number between [0,1]. If p' ≥ mRate, then the individual searches around the current solution by iteratively finding a better solution among its neighboring possible solutions. The calculation function is as follows: (5) (6) In the formula, and respectively represent coefficient vectors in the intervals [0, 2] and [0, 1], t represents the current iteration, represents the current solution vector, represents the diameter of the circle of the expected region where the solution will be searched; if p' < mRate, a mutation solution is generated, and local random perturbation is performed based on the crossover probability p', and the perturbation formula is expressed as: (7) In the formula, R represents a neighborhood function.

[0045] S35, for the development group, if p' ≥ mRate, the individual moves towards the optimal solution, and the implemented calculation function is: (8) (9) In the formula, represents a random vector in the interval [0, 2], represents the optimal solution; if p' < 0.5, the individual adopts a strategy of searching around for the optimal solution, and the implemented calculation formula is: Formula (10) Formula (11) Formula (12) In the formula, and both represent random vectors in the interval [0, 1]; k represents a constant for controlling , and k iterates from 2 to 0.

[0046] S36, modifying individuals exceeding the search space, and randomly changing the order of individuals to exchange the roles of exploration and development group members.

[0047] S37, determining whether the current number of iterations reaches the set maximum number of iterations; if yes, executing step S38; if not, repeatedly executing steps S32 to S36.

[0048] S38, ending the optimization, and obtaining the optimal vector.

[0049] It's important to note that DeepAR (Deep AutoRegressive) is a deep learning model used for time series forecasting. Based on a recurrent neural network (RNN) structure, DeepAR can model and predict time series data with dynamic and complex correlations. Compared to traditional time series models, DeepAR can automatically learn seasonality, trends, and periodicity characteristics in time series data, making it suitable for multi-step forecasting and uncertainty modeling.

[0050] In this embodiment, in step S104, the DeepAR model network structure is represented as follows: (13) In the formula, This represents the predicted photovoltaic power output from time t0 to time T. This represents the actual photovoltaic power output from historical time 1 to t0-1. h represents the characteristic parameters in the time series. j,t This represents the output of the j-th parameter of the hidden layer in the autoregressive recurrent network at time t, where j = 1, 2, ..., z, and z is the number of input feature parameters. j,t The relational expression is as follows: (14) In the formula, Given parameters for the LSTM cell states used in the autoregressive recurrent network, θ represents the neural network model being trained, and H represents the autoregressive recurrent network, with a likelihood function added to the autoregressive recurrent network. This makes the output a probability distribution prediction, and the likelihood function uses a Gaussian distribution function.

[0051] In this embodiment, step S105 involves obtaining a small sample of the target domain and inputting it into the source domain short-term photovoltaic power output prediction model for transfer learning to obtain the target domain short-term photovoltaic power output prediction model, including: Obtain the latent features of photovoltaic power output data from a small sample in the target domain. , Represented as: ; Obtain the class labels of small samples in the target domain, and denote the feature set of the dataset of photovoltaic power curves of the k-th class in the target domain as . , Represented as: ,in This represents the size of the k-th class dataset, representing a small sample in the target domain. calculate Chinese characteristics and cosine similarity And based on cosine similarity get Fusion characteristics , and They are represented as follows: (15) (16) Based on the fusion features, small samples in the target domain The short-term photovoltaic power output prediction model in the input domain is transferred to the target domain to obtain the short-term photovoltaic power output prediction model.

[0053] This embodiment provides a method for constructing a short-term photovoltaic (PV) output prediction model. It introduces a MeanShift clustering algorithm with a radial basis function (RBF) neural network as the kernel function, adjusts the position of the centroid of the candidate boxes, and formulates a similarity metric to improve model performance, enabling it to find and accurately cluster multiple meteorological clusters. Furthermore, it uses the DGBCO algorithm to optimize the ResNet network structure, achieving adaptive adjustment of neural network parameters and accurately extracting features of each meteorological cluster, further improving the accuracy of PV output prediction. Additionally, by training a DeepAR model, a source domain short-term PV output prediction model is obtained. The DeepAR model, through extensive time series... Training an autoregressive recurrent network model on the column can effectively learn a global model from relevant time series data and can learn complex patterns, reducing the impact of uncertainties in data such as photovoltaic output and meteorology, and achieving the prediction probability distribution of data with time series characteristics. Finally, a small sample of the target domain is obtained and input into the source domain short-term photovoltaic output prediction model for transfer learning, resulting in a short-term photovoltaic output prediction model for the target domain. Using the transfer learning method, sample weights, feature mappings, or model training between domains can be adjusted to reduce the domain differences between the source and target domains, deeply explore the data features between information, and achieve information sharing, ultimately realizing an adaptive adjustment mechanism.

[0054] Please see Figure 2 As shown, embodiments of the present invention also provide a short-term photovoltaic power output prediction method, including: S201. Construct a short-term photovoltaic power output prediction model and a photovoltaic power generation digital twin model for the target domain.

[0055] In this embodiment, the method for constructing the short-term photovoltaic power output prediction model for the target domain adopts the short-term photovoltaic power output prediction model construction method described above.

[0056] In this embodiment, the photovoltaic power generation digital twin model includes a real physical sub-model and a virtual sub-model. The real physical sub-model is used to model the meteorological characteristics, equipment information status characteristics, and output characteristics of the photovoltaic power generation system; the virtual sub-model includes simulating the operating status of the photovoltaic power generation system.

[0057] S202. Obtain meteorological and environmental data within the forecast period, extract the features of the obtained meteorological and environmental data, and input the extracted features into the short-term photovoltaic power output prediction model of the target domain to obtain short-term photovoltaic power output prediction data.

[0058] When extracting features from acquired meteorological and environmental data, the features can be extracted using the aforementioned feature extraction model, or other methods can be used.

[0059] S203. Input the acquired meteorological environment data and short-term photovoltaic output forecast data of the target domain into the photovoltaic power generation digital twin model, and display it through the photovoltaic power generation digital twin model.

[0060] S204. Obtain the actual photovoltaic output data within the prediction period, calculate the difference between the actual photovoltaic output data within the prediction period and the short-term photovoltaic output prediction data, and optimize the short-term photovoltaic output prediction model in the target domain based on the difference.

[0061] This embodiment provides a short-term photovoltaic power output prediction method, which predicts short-term photovoltaic power output data through a target domain short-term photovoltaic power output prediction model, thereby improving prediction accuracy. Furthermore, by constructing a photovoltaic power generation digital twin model, it is possible to monitor, simulate, evaluate, and accurately predict the photovoltaic power generation system of a microgrid. The method utilizes continuously accumulated new data to improve the accuracy and adaptability of the digital twin model to better support the operation and management of the microgrid.

[0062] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a short-term photovoltaic power output prediction model, characterized in that, The construction method of the short-term photovoltaic output prediction model comprises: collecting historical data of the photovoltaic power generation system, wherein the historical data of the photovoltaic power generation system includes meteorological and environmental data of multiple regions and corresponding photovoltaic output data; clustering the meteorological and environmental data by using the RBF-MeanShift model to obtain a plurality of meteorological clusters; optimizing an initial ResNet model based on the DGBCO algorithm to obtain a feature extraction model, extracting features of each meteorological cluster by using the feature extraction model, and generating a plurality of training samples, wherein the training samples include the features of each meteorological cluster and the corresponding photovoltaic output data; inputting the plurality of training samples into a DeepAR model for training to obtain a source-domain short-term photovoltaic output prediction model; obtaining target-domain small samples, inputting the target-domain small samples into the source-domain short-term photovoltaic output prediction model for transfer learning, and obtaining a target-domain short-term photovoltaic output prediction model; the network structure of the initial ResNet model is expressed as: （4） In the formula, Let S represent a shallow unit, N represent a deep unit, i = 0, 1, ..., N-1, and w represent the weights. Let w be the optimization target S. i and set R represents the set of real numbers, d represents the dimension of the problem, and the optimal vector is found using the DGBCO algorithm. the step of finding the optimal vector by the DGBCO algorithm comprises: S31. Initialize population size Sizepop, maximum number of iterations. The mutation rate mRate, the percentage of the exploration group in the population epct, and the glide slope index k are used, and the upper and lower limits of the search space for each solution are set to [S]. i,min ,S i,max ]; S32. Calculate each solution S i Find the optimal solution with the best fitness value; S33, dividing individuals of a population into an exploration group and an exploitation group according to a set epct value, dynamically managing the number of individuals in each group at the beginning of each iteration, and increasing the number of solutions in the exploration group if the optimal fitness does not change compared with the previous two iterations; S34, for the exploration group, obtaining a crossover probability p', wherein p' is a random number between [0,1], if p'≥mRate, the individual searches around the current solution, which is realized by iteratively finding a better solution in adjacent possible solutions thereof, and the realized calculation function is: (5) (6) In the formula, and Let represent the coefficient vectors in the intervals [0, 2] and [0, 1], respectively, and t represent the current iteration. Represents the current solution vector. This means the solution will find the diameter of the circle containing the desired region; if p'<mRate, a mutation solution is generated, and based on local random perturbation of the crossover probability p', the perturbation formula is expressed as: (7) In the formula, Represents the neighborhood function; S35, for the exploitation group, if p'≥mRate, the individual moves towards the optimal solution, and the realized calculation function is: (8) (9) In the formula, Let [a] be a random vector interval [0, 2]. This indicates the best solution; if p'<0.5, the individual adopts a strategy of searching around for the optimal solution, and the realized calculation formula is: (10) (11) (12) In the formula, and Both represent the random vector interval [0, 1]; k represents the control. A constant, k iterates from 2 to 0; S36, modifying individuals exceeding the search space, and randomly changing the order of individuals to exchange the roles of members of the exploration group and the exploitation group; S37, judging whether the current number of iterations reaches the set maximum number of iterations; if yes, executing step S38; if not, repeatedly executing steps S32 to S36; S38, ending the optimization and obtaining the optimal vector.

2. The method for constructing a short-term photovoltaic power output prediction model according to claim 1, characterized in that, before the step of clustering the meteorological and environmental data by using the RBF-MeanShift model to obtain a plurality of meteorological clusters, the method further comprises: performing dimensionless processing on the historical data of the power generation system.

3. The method for constructing a short-term photovoltaic power output prediction model according to claim 1, characterized in that, When using the RBF-MeanShift model to cluster meteorological and environmental data, RBF is used as the kernel function, and the Bhattacharyya coefficient is defined as the similarity measure function of the meteorological and environmental data. The relational expression is as follows: （1） In the formula, The value of is between 0 and 1, p(y) represents the candidate target model, q represents the target model, and u is a natural number.

4. The method for constructing a short-term photovoltaic power output prediction model according to claim 3, characterized in that, Expanding equation (1) using Taylor at the initial position y0, and considering the Bhattacharyya coefficients... To find the maximum value, we can obtain the Meanshift vector, which is calculated using the following formula: In the formula, Let h represent the Meanshift vector. The scale, y1 represents the new iteration position, x i (i=1,2,…n) represents The position, g(·) is a unit kernel function, w i This represents the weight of each point, where w is the weight. i The relational expression is as follows: wherein δ represents the Kronecker delta, and b(xi) represents an index value at position xi.

5. The method for constructing a short-term photovoltaic power output prediction model according to claim 1, characterized in that, the network structure of the DeepAR model is expressed as: (13) In the formula, This represents the predicted photovoltaic power output from time t0 to time T. This represents the actual photovoltaic power output from historical time 1 to t0-1. h represents the time-series characteristic parameter. j,t This represents the output of the j-th parameter of the hidden layer in the autoregressive recurrent network at time t, where j = 1, 2, ..., z, and z is the number of input feature parameters. j,t The relational expression is as follows: (14) In the formula, Given parameters for the LSTM cell states used in the autoregressive recurrent network, θ represents the neural network model being trained, and H represents the autoregressive recurrent network, with a likelihood function added to the autoregressive recurrent network. This makes the output a probability distribution prediction, and the likelihood function uses a Gaussian distribution function.

6. The method for constructing a short-term photovoltaic power output prediction model according to claim 1, characterized in that, the step of obtaining target-domain small samples, inputting the target-domain small samples into the source-domain short-term photovoltaic output prediction model for transfer learning, and obtaining the target-domain short-term photovoltaic output prediction model comprises: Obtain the latent features of photovoltaic power output data from a small sample in the target domain. , Represented as: ； Obtain the class labels of small samples in the target domain, and denote the feature set of the dataset of photovoltaic power curves of the k-th class in the target domain as . , Represented as: ,in This represents the size of the k-th class dataset, representing a small sample in the target domain. calculate Chinese characteristics and cosine similarity And based on cosine similarity get Fusion characteristics , and They are represented as follows: (15) (16) Based on the fusion features, small samples in the target domain The short-term photovoltaic power output prediction model in the input domain is transferred to the target domain to obtain the short-term photovoltaic power output prediction model.

7. A method for predicting short-term photovoltaic power output, characterized in that, the short-term photovoltaic output prediction method comprises: A short-term photovoltaic power output prediction model and a photovoltaic power generation digital twin model are constructed for the target domain. The method for constructing the short-term photovoltaic power output prediction model for the target domain adopts the short-term photovoltaic power output prediction model construction method as described in any one of claims 1-6. Meteorological environmental data within the forecast period is acquired, and features of the acquired meteorological environmental data are extracted. The extracted features are then input into the short-term photovoltaic power output prediction model of the target domain to obtain short-term photovoltaic power output prediction data. The acquired meteorological environmental data and the short-term photovoltaic output prediction data of the target domain are input into the photovoltaic power generation digital twin model, and then displayed through the photovoltaic power generation digital twin model; Obtain the actual photovoltaic power output data within the prediction period, calculate the difference between the actual photovoltaic power output data within the prediction period and the short-term photovoltaic power output prediction data, and optimize the short-term photovoltaic power output prediction model for the target domain based on the difference.

8. The short-term photovoltaic power output prediction method according to claim 7, characterized in that, The photovoltaic power generation digital twin model includes a real physical sub-model and a virtual sub-model. The real physical sub-model is used to model the meteorological characteristics, equipment information status characteristics, and output characteristics of the photovoltaic power generation system. The virtual sub-model includes the simulation of the operating status of the photovoltaic power generation system.

Images