A machine learning-based distributed photovoltaic power prediction method and system

Abstract

The application belongs to the technical field of new energy prediction, and provides a distributed photovoltaic power prediction method and system based on machine learning, which comprises the following steps: a physical power prediction model of a photovoltaic power station is constructed, the photovoltaic power of the photovoltaic power station is predicted, and the prediction accuracy is evaluated; if the accuracy is low, a fusion statistical prediction method is triggered, a statistical power prediction model is constructed through a machine learning algorithm, the statistical power prediction model and the physical power prediction model are weighted and fused, a fusion power prediction model of the photovoltaic power station is constructed, and the prediction model adopted by the photovoltaic power station is determined according to the weight; the distributed photovoltaic system is divided into photovoltaic subsystems, whether the photovoltaic subsystems trigger a similar set prediction method is judged based on the proportion of different prediction models adopted by the photovoltaic power station; if the similar set prediction method is triggered, a similar prediction group is screened through a clustering algorithm, and the photovoltaic power of the photovoltaic power station is predicted in the similar prediction group through a core mapping method.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy prediction, and in particular to a distributed photovoltaic power prediction method and system based on machine learning. BACKGROUND

[0002] With the increasing demand for clean energy worldwide, distributed photovoltaic power generation systems have been widely used due to their flexibility and efficiency. Distributed photovoltaic systems are usually composed of multiple photovoltaic power stations, which are distributed in different geographical locations and are affected by local meteorological conditions, geographical environment and other factors. Accurate prediction of the power output of distributed photovoltaic systems is of great significance for the stable operation of power grids, power dispatching and energy trading.

[0003] In the prior art, photovoltaic power prediction mainly includes physical prediction methods and statistical prediction methods. Physical prediction methods are based on the physical characteristics of photovoltaic power stations and predict photovoltaic power by establishing physical models. Statistical prediction methods use historical meteorological data and photovoltaic power data to establish data-driven models such as neural networks, support vector machines and time series models through machine learning and statistical analysis to predict photovoltaic power. However, single prediction methods have obvious limitations. Physical models have certain errors in simulating meteorological conditions, and statistical prediction methods usually only consider statistical rules in historical data, making it difficult to meet the prediction needs of all photovoltaic power stations and unable to achieve high-precision power prediction.

[0004] On the other hand, with the continuous expansion of distributed photovoltaic systems, the operating conditions and meteorological conditions of photovoltaic power stations within the system show obvious regional characteristics and spatial and temporal differences. Traditional centralized prediction methods treat the entire distributed photovoltaic system as a whole for prediction, ignoring the differences between photovoltaic power stations in different regions and failing to fully utilize the similarity and complementarity between photovoltaic power stations, resulting in large deviations between predicted results and actual power. Moreover, existing prediction methods lack adaptive adjustment capability in response to sudden weather changes, making it difficult to quickly and accurately update the prediction model, further reducing the reliability and timeliness of the prediction.

[0005] To solve the above problems, the present application proposes a distributed photovoltaic power prediction method and system based on machine learning. SUMMARY

[0006] In order to make up for the shortcomings of the prior art and solve at least one technical problem in the background art.

[0007] The technical solution adopted by the application to solve the technical problem is: a distributed photovoltaic power prediction method based on machine learning, comprising:

[0008] Acquire the basic information of each photovoltaic power station in the distributed photovoltaic system, and construct a physical power prediction model of the photovoltaic power station;

[0009] The physical power prediction model of the photovoltaic power station is used to predict the photovoltaic power of the photovoltaic power station, and the prediction accuracy is evaluated through the actual meteorological data sequence and the actual photovoltaic power sequence obtained by real-time acquisition. If the accuracy is low, the fusion statistical prediction method is triggered;

[0010] The evaluation of the prediction accuracy is as follows:

[0011] A prediction period is set, and the interval period between adjacent prediction periods is a prediction analysis period;

[0012] In each prediction period, a plurality of prediction points are uniformly taken. At the beginning of the current prediction period, the prediction meteorological data sequence of the region where the photovoltaic power station is located in the prediction period is acquired, input into the physical power prediction model, and the prediction photovoltaic power sequence is obtained. Real-time acquisition of actual photovoltaic power, at the end of the current prediction period, the actual photovoltaic power sequence is obtained according to the time sequence integration;

[0013] The root mean square error between the prediction photovoltaic power sequence and the actual photovoltaic power sequence is calculated. If the root mean square error is greater than or equal to the preset error threshold, it is judged that the prediction accuracy of the physical power prediction model is low;

[0014] If the fusion statistical prediction method is triggered, a statistical power prediction model is constructed by combining the actual meteorological data sequence and the actual photovoltaic power sequence through a machine learning algorithm. The statistical power prediction model and the physical power prediction model are weighted and fused to construct a fusion power prediction model of the photovoltaic power station, and the prediction model used by the photovoltaic power station is determined according to the weight;

[0015] The judgment of the prediction model used by the photovoltaic power station is as follows:

[0016] The prediction model includes the fusion power prediction model and the physical power prediction model;

[0017] A weight evaluation flag value is set for each photovoltaic power station that triggers the fusion statistical prediction method, and the initial value is 0. The weight of the physical power prediction model in the fusion power prediction model is compared with the preset weight threshold. If the weight is greater than or equal to the weight threshold, the weight evaluation flag value is incremented by one. If the weight is less than the weight threshold, the weight evaluation flag value is reset to 0;

[0018] If the weight evaluation flag value is less than the stable flag value, the photovoltaic power station uses the fusion power prediction model. Otherwise, the photovoltaic power station ends the fusion statistical prediction method and uses the physical power prediction model;

[0019] The fusion power prediction model is obtained as follows:

[0020] The statistical power prediction model and the physical power prediction model are fused by weighting to construct a fusion power prediction model, wherein the sum of the weights of the statistical power prediction model and the physical power prediction model is 1;

[0021] A population of weights of the physical power prediction model is randomly generated by a genetic algorithm, the total actual meteorological data sequence is input into the fusion power prediction model, the root mean square error between the prediction result and the corresponding actual photovoltaic power is calculated, the population is continuously evolved by selection, crossover and mutation genetic operations, and the weight of the physical power prediction model that minimizes the root mean square error is gradually screened out, so as to determine the weight of the physical power prediction model and obtain the fusion power prediction model;

[0022] The statistical power prediction model is obtained in the following manner:

[0023] It is judged whether the photovoltaic power station has a statistical power prediction model;

[0024] If not, all actual photovoltaic power sequences and actual meteorological data sequences of the photovoltaic power station up to the current time are obtained, the total photovoltaic power sequence and the total actual meteorological data sequence are obtained by time sequence integration respectively, a training data set is obtained, a random forest model is constructed based on a machine learning algorithm and trained by using the training data set, and the trained random forest model is obtained as the statistical power prediction model of the photovoltaic power station;

[0025] If yes, actual photovoltaic power sequences and actual meteorological data sequences that are not included in the training data set are obtained to supplement the training data set, the statistical power prediction model is supplemented and trained, and the statistical power prediction model of the photovoltaic power station is updated;

[0026] The distributed photovoltaic system is divided into photovoltaic subsystems, it is judged whether the photovoltaic subsystem triggers a similar set prediction method, if not, the photovoltaic power of the photovoltaic power station is predicted according to the prediction model used, if yes, a similar prediction group is screened by a clustering algorithm, and the photovoltaic power of the photovoltaic power station is predicted by using a core mapping method in the similar prediction group;

[0027] The judgment manner of whether the similar set prediction method is triggered is as follows:

[0028] The distributed photovoltaic system is divided into a plurality of photovoltaic subsystems according to the jurisdiction range of the dispatching department, the grid connection mode and the geographical distance, for each photovoltaic subsystem, the number of photovoltaic power stations in the photovoltaic subsystem that use the fusion power prediction model is obtained, the ratio of the number of photovoltaic power stations to the total number of photovoltaic power stations in the photovoltaic subsystem is calculated, and if the ratio is greater than a preset proportion standard, the similar set prediction method is triggered in the photovoltaic subsystem;

[0029] The similar prediction group is obtained in the following manner:

[0030] All photovoltaic power stations in the photovoltaic subsystem are divided into two clusters, the centers of the two clusters are calculated, and the Euclidean distance between the cluster centers is compared with the set judgment threshold value;

[0031] If greater than or equal to the set judgment threshold value, each cluster is further divided into two clusters, and the Euclidean distance between the centers of the two clusters is calculated and compared again, and if the number of clusters is greater than the preset number standard, the re-clustering and comparison are stopped;

[0032] If less than the set judgment threshold value, the two clusters are marked as similar clusters, the normalized photovoltaic power sequence of the photovoltaic power stations in the cluster and the Euclidean distance between the cluster centers are obtained and calculated, and the photovoltaic power stations less than the judgment threshold value are classified into the same similar prediction group;

[0033] The two clusters are divided in the following manner:

[0034] The actual photovoltaic power sequence of the photovoltaic power stations in the previous prediction period is obtained and normalized to obtain a normalized photovoltaic power sequence, and a K-means clustering algorithm is used to cluster all photovoltaic power stations in the photovoltaic subsystem based on the normalized photovoltaic power sequence, and all photovoltaic power stations are divided into two clusters;

[0035] The core mapping method includes:

[0036] The Euclidean distance between the normalized photovoltaic power sequence of the photovoltaic power stations in the similar prediction group and the cluster center is obtained;

[0037] The photovoltaic power station with the smallest Euclidean distance is taken as the core prediction station of the similar prediction group, and the photovoltaic power stations other than the core prediction station are taken as the mapping prediction stations of the similar prediction group;

[0038] For the core prediction station, the prediction meteorological data sequence is obtained, which is input into the fusion power prediction model of the core prediction station to obtain the prediction photovoltaic power sequence of the core prediction station, and based on the proportional relationship between the core prediction station and the mapping prediction station in the actual photovoltaic power sequence in the previous prediction period, the prediction photovoltaic power sequence of the mapping prediction station is obtained according to the prediction photovoltaic power sequence of the core prediction station.

[0039] A distributed photovoltaic power prediction system based on machine learning, which is used to implement the distributed photovoltaic power prediction method described above, includes the following modules:

[0040] Physical prediction module: obtain the basic information of each photovoltaic power station in the distributed photovoltaic system, and construct a physical power prediction model of the photovoltaic power station;

[0041] Accurate analysis module: the physical power prediction model of the photovoltaic power station is used to predict the photovoltaic power of the photovoltaic power station, and the prediction accuracy is evaluated through the actual meteorological data sequence and the actual photovoltaic power sequence obtained by real-time collection, if the accuracy is low, the fusion statistical prediction method is triggered;

[0042] Fusion prediction module: if the fusion statistical prediction method is triggered, a statistical power prediction model is constructed through a machine learning algorithm combined with the actual meteorological data sequence and the actual photovoltaic power sequence, the statistical power prediction model and the physical power prediction model are weighted and fused to construct a fusion power prediction model of the photovoltaic power station, and the prediction model adopted by the photovoltaic power station is determined according to the weight;

[0043] Set prediction module: the distributed photovoltaic system is divided into photovoltaic subsystems, it is judged whether the similar set prediction method is triggered for the photovoltaic subsystems, if not, the photovoltaic power of the photovoltaic power station is predicted according to the prediction model adopted, if yes, the photovoltaic power of the photovoltaic power station is predicted by using a core mapping method in a similar prediction group selected through a clustering algorithm.

[0044] The beneficial effects of the present application are as follows:

[0045] 1、The present application realizes intelligent switching and fusion of prediction models by constructing a physical power prediction model and a statistical power prediction model, and dynamically triggering a fusion statistical prediction method according to real-time meteorological and power data. This flexible model selection mechanism can fully utilize the advantages of physical models and statistical models, effectively improve the accuracy of photovoltaic power prediction under different meteorological conditions and power station operating states, and provide more reliable data support for power grid dispatching and energy management.

[0046] 2、The present application divides the distributed photovoltaic system into photovoltaic subsystems and judges whether the similar set prediction method is triggered. For the subsystems that trigger the similar set prediction, a similar prediction group is selected by using a clustering algorithm, and a core mapping method is used for prediction. This innovative strategy fully utilizes the correlation information between similar power stations, not only improves the prediction efficiency of similar power stations, but also further improves the prediction accuracy of the entire distributed photovoltaic system, which helps to optimize the operation and management of the distributed photovoltaic system and improve the energy utilization efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0047] The present application will be further described below with reference to the accompanying drawings.

[0048] Figure 1 is a step flow chart of a distributed photovoltaic power prediction method based on machine learning according to an embodiment of the present application;

[0049] Figure 2It is a system module architecture diagram of a distributed photovoltaic power prediction system based on machine learning. DETAILED DESCRIPTION

[0050] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application will be further described below in conjunction with specific embodiments.

[0051] Embodiment 1

[0052] Please refer to Figure 1 The photovoltaic power prediction method and system based on machine learning comprises the following steps:

[0053] S1: Obtain the basic information of each photovoltaic power station in the distributed photovoltaic system, and construct a physical power prediction model of the photovoltaic power station;

[0054] For the photovoltaic power station in the distributed photovoltaic system, the basic information of the photovoltaic power station is obtained, including the latitude and longitude coordinates of the photovoltaic power station and the physical characteristic parameters of the photovoltaic module;

[0055] The physical characteristic parameters of the photovoltaic module include the short-circuit current, open-circuit voltage, maximum power point current, maximum power point voltage and module conversion efficiency of the photovoltaic module, which are obtained from the factory technical document of the photovoltaic module;

[0056] Among them, the short-circuit current represents the current value when the photovoltaic module is short-circuited under standard test (STC) conditions, and the open-circuit voltage represents the voltage value when the photovoltaic module is in an open circuit state without connecting a load under standard test conditions;

[0057] Among them, the maximum power point current and the maximum power point voltage respectively represent the current value and the voltage value corresponding to the maximum power output of the photovoltaic module under standard test conditions, and the product of the maximum power point current and the maximum power point voltage is the maximum photovoltaic power of the module under standard test conditions ;

[0058] Among them, the module conversion efficiency represents the efficiency of the photovoltaic module in converting the received solar radiation energy into electrical energy under standard test conditions, and the calculation formula is:

[0059] ;

[0060] Among them, A represents the effective area of the photovoltaic module, represents the solar irradiance under standard test conditions;

[0061] Based on the photoelectric conversion principle, combined with the physical characteristic parameters of photovoltaic modules of photovoltaic power stations, a physical power prediction model is constructed for photovoltaic power stations with meteorological data as input and photovoltaic power as output. The meteorological data includes solar radiation intensity, ambient temperature, ambient humidity and wind speed. The solar radiation intensity directly determines the generation rate of photo-generated carriers of photovoltaic modules, and has an effect on the photovoltaic efficiency of photovoltaic modules, and is a key determinant of photovoltaic power. The ambient temperature affects the conversion efficiency of photovoltaic modules by affecting the band gap and carrier mobility of semiconductor materials, and thus affects the photovoltaic efficiency of photovoltaic modules, and is a main influencing factor of photovoltaic power.

[0062] For each photovoltaic power station in the distributed photovoltaic system, a physical power prediction model of each photovoltaic power station is constructed ;

[0063] S2: The physical power prediction model of the photovoltaic power station is used to predict the photovoltaic power of the photovoltaic power station, and the prediction accuracy is evaluated by the actual meteorological data sequence and the actual photovoltaic power sequence obtained by real-time collection. If the accuracy is low, the fusion statistical prediction method is triggered.

[0064] For each photovoltaic power station, the photovoltaic power of the photovoltaic power station is predicted by the physical power prediction model of the photovoltaic power station.

[0065] Specifically, a prediction period is set, the length of the prediction period is the same, and the interval period between adjacent prediction periods is marked as a prediction analysis period, and the length of the prediction analysis period is the same.

[0066] The starting point of the first prediction period of the photovoltaic power station is the starting time of the initial use of the photovoltaic power station. A plurality of prediction points are taken in each prediction period. The starting point of each prediction period is the first prediction point in the prediction period, and the end point of each prediction period is the final prediction point in the prediction period. The interval length between adjacent prediction points is the same.

[0067] At the starting point of the current prediction period, the prediction value of the meteorological data of the photovoltaic power station in the prediction period is obtained by a professional meteorological forecasting agency. The prediction meteorological data sequence is obtained by integrating the time sequence. The prediction meteorological data is input into the physical power prediction model of the photovoltaic power station, and the prediction photovoltaic power sequence of the photovoltaic power station is output :

[0068] ;

[0069] Wherein, , Pn represents the prediction value of the photovoltaic power of the photovoltaic power station at the nth prediction point in the current prediction period, and N represents the total number of prediction points in the current prediction period.

[0070] At each prediction point in the current prediction period, real-time actual meteorological data is collected and recorded by meteorological monitoring equipment deployed at the photovoltaic power station, including a solar radiation sensor, a temperature sensor, a humidity sensor, and a wind speed sensor, and real-time actual photovoltaic power is collected and recorded by high-precision power monitoring equipment installed at the photovoltaic power station.

[0071] At the end of the current prediction period, the actual meteorological data collected in real time in the current prediction period is integrated according to the time sequence to obtain an actual meteorological data sequence, and the actual photovoltaic power collected in real time in the current prediction period is integrated according to the time sequence to obtain an actual photovoltaic power sequence. :

[0072] ;

[0073] wherein, represents the actual photovoltaic power of the photovoltaic power station at the nth prediction point in the current prediction period;

[0074] The root mean square error (RMSE) is used as an index to evaluate the accuracy of the physical power prediction model, and the root mean square error RMSE between the predicted photovoltaic power sequence and the actual photovoltaic power sequence is calculated, and the calculation formula is:

[0075] ;

[0076] The calculated root mean square error is compared with a preset error threshold;

[0077] If the calculated root mean square error is less than the preset error threshold, it is judged that the prediction accuracy of the physical power prediction model is high, and the model does not need to be changed;

[0078] If the calculated root mean square error is greater than or equal to the preset error threshold, it is judged that the prediction accuracy of the physical power prediction model is low, and the fusion statistical prediction method is triggered;

[0079] It should be noted that the physical power prediction model is based on mathematical formula derivation and calculation, and has clear logic, low calculation complexity, low data dependency, high calculation efficiency, and low performance requirement for the calculation device. Therefore, when the accuracy of the physical power prediction model can meet the real-time power dispatching demand, the fusion statistical prediction method does not need to be used, and the prediction cost is saved;

[0080] S3: If the fusion statistical prediction method is triggered, a statistical power prediction model is constructed by combining the actual meteorological data sequence and the actual photovoltaic power sequence through a machine learning algorithm, the statistical power prediction model and the physical power prediction model are weighted and fused to construct a fusion power prediction model of the photovoltaic power station, and the prediction model adopted by the photovoltaic power station is determined according to the weight;

[0081] If the fusion statistical prediction method is triggered, the prediction analysis period is entered, and for the photovoltaic power station triggering the fusion statistical prediction method, it is determined whether the photovoltaic power station has a statistical power prediction model;

[0082] If there is no statistical power prediction model in the photovoltaic power station, all actual photovoltaic power sequences of the photovoltaic power station up to the current time are obtained, the sequences are integrated into a total photovoltaic power sequence according to time sequence and normalized to obtain a total normalized photovoltaic power sequence, and the elements in the total normalized photovoltaic power sequence are marked as normalized photovoltaic power;

[0083] All actual meteorological data sequences of the photovoltaic power station up to the current time are obtained, the sequences are integrated into a total actual meteorological data sequence, the total normalized photovoltaic power sequence and the total actual meteorological data sequence are one-to-one corresponding according to the prediction point, and a training data set is obtained by integration;

[0084] Based on the machine learning algorithm, a random forest model is constructed as a statistical power prediction model of the photovoltaic power station;

[0085] Specifically, a plurality of decision trees are constructed in the random forest model and integrated, the total actual meteorological data sequence is taken as an input feature, the total normalized photovoltaic power sequence is taken as an output label, and the input is input into the random forest model. The random forest model learns the influence weight of different input features on the normalized photovoltaic power by feature division and decision tree construction on the input data. When constructing the decision tree, the machine learning algorithm selects the input feature with the highest degree of discrimination according to the information gain and gini index of the input feature to divide the input data, and gradually constructs a decision tree that can accurately predict the normalized photovoltaic power;

[0086] Through the cross-validation method, the training data set is divided into a plurality of subsets, one subset is taken as a validation set and the remaining subsets are taken as training sets in turn, the random forest model is trained and evaluated on the validation set multiple times, the parameters of the random forest model are adjusted, including the number of decision trees and the depth of the tree, and the mean square error is taken as the optimization target. By continuously optimizing the parameters of the random forest model, the random forest model reaches the best fitting effect on the training data set, and a trained random forest model is obtained. The output data of the trained random forest model is de-normalized based on the corresponding total photovoltaic power sequence to obtain a statistical power prediction model of the photovoltaic power station in the next prediction period ;

[0087] If there is a statistical power prediction model in the photovoltaic power station, the actual photovoltaic power sequence and the actual meteorological data sequence not belonging to the training data set are obtained, the actual photovoltaic power sequence and the actual meteorological data sequence not belonging to the training data set are classified into the training data set, the statistical power prediction model is supplemented and trained, and the statistical power prediction model of the photovoltaic power station in the next prediction period is obtained ;

[0088] The statistical power prediction model is fused in a weighted average manner and the physical power prediction model to obtain a fused power prediction model The weighted average formula is:

[0089] ;

[0090] wherein, is the weight of the physical power prediction model, and the value range is between [0, 1], is the weight of the physical power prediction model, determined by a genetic algorithm;

[0091] Specifically, a group of initial values (population) of are randomly generated, each value represents an individual, for each individual, the total actual meteorological data sequence is input into the fused power prediction model, the root mean square error RMSE between the prediction result of the fused power prediction model and the corresponding actual photovoltaic power is calculated, and the population is continuously evolved through selection, crossover and mutation genetic operations, and the individual with the smallest root mean square error is gradually screened out After multiple generations of evolution, the optimal is determined, and the fused power prediction model is obtained;

[0092] A weight evaluation flag value is set for each photovoltaic power station triggering the fused statistical prediction method, and the initial value is 0, the weight of the physical power prediction model is compared with the preset weight threshold value, if the weight is greater than or equal to the weight threshold value, the weight evaluation flag value is added by one, and if the weight is less than the weight threshold value, the weight evaluation flag value is 0;

[0093] The prediction model adopted by the photovoltaic power station is determined based on the weight evaluation flag value, and the prediction model includes the fused power prediction model and the physical power prediction model;

[0094] Specifically, the weight evaluation flag value is compared with the preset stability flag value;

[0095] If the weight evaluation flag value is less than the stability flag value, the photovoltaic power station adopts the fusion power prediction model;

[0096] If the weight evaluation flag value is greater than or equal to the stability flag value, it is judged that the accuracy of the physical power prediction model returns to the normal standard, the fusion statistical prediction method is ended, and the photovoltaic power station adopts the physical power prediction model;

[0097] It should be noted that the role of this step is to introduce a machine learning algorithm when the physical model prediction accuracy is low, fully excavate the complex nonlinear relationship in the actual meteorological data and power data, construct a statistical power prediction model, and fuse the statistical power prediction model and the physical power prediction model through weighted averaging. The theoretical basis of the physical power prediction model and the data fitting ability of the statistical power prediction model are used to make up for each other, significantly improve the accuracy and adaptability of the prediction model, and at the same time, through the weight evaluation flag value, the prediction model adopted by the photovoltaic power station is judged, realizing the dynamic selection and optimization of the prediction model. Under the premise of ensuring the prediction accuracy, the prediction efficiency is improved and the calculation resource consumption is reduced;

[0098] S4: The distributed photovoltaic system is divided into photovoltaic subsystems, and based on the proportion of different prediction models adopted by the photovoltaic power station, it is judged whether the photovoltaic subsystem triggers the similar set prediction method. If not, the photovoltaic power of the photovoltaic power station is predicted according to the adopted prediction model, and if so, the similar prediction group is screened through a clustering algorithm, and the photovoltaic power of the photovoltaic power station is predicted in the similar prediction group using a core mapping method.

[0099] The distributed photovoltaic system is divided into a plurality of photovoltaic subsystems according to the jurisdiction range of the dispatching department, the grid connection mode and the geographical distance;

[0100] Specifically, according to the jurisdiction range of different dispatching departments, the coverage area of the distributed photovoltaic system is divided into a plurality of jurisdiction sub-areas, the nodes of the photovoltaic power station connected to the power grid are obtained through the power grid planning data, and the photovoltaic power stations in the jurisdiction sub-area are classified according to the voltage level of the nodes connected to the power grid in each jurisdiction sub-area. The voltage level includes high voltage, medium voltage and low voltage, and the photovoltaic power stations of different voltage levels in the jurisdiction sub-area are obtained.

[0101] The jurisdiction sub-area is evenly divided into a plurality of grid units, and the size of the grid unit is determined based on the size of the jurisdiction sub-area. The photovoltaic power stations with the same voltage level in each grid unit are divided into a photovoltaic subsystem.

[0102] For each photovoltaic subsystem, the number of photovoltaic power stations in the photovoltaic subsystem that adopt the fusion power prediction model is obtained, the ratio of the number of photovoltaic power stations to the total number of photovoltaic power stations in the photovoltaic subsystem is calculated, and if the ratio is greater than a preset proportion standard, the similar set prediction method is triggered in the photovoltaic subsystem.

[0103] If the similar set prediction method is not triggered, in the photovoltaic subsystem, the predicted meteorological data sequence of each photovoltaic power station in the next prediction period is obtained, which is input into the fusion power prediction model or the physical power model adopted by each photovoltaic power station, to obtain the predicted photovoltaic power sequence of each photovoltaic power station in the next prediction period;

[0104] If the similar set prediction method is triggered, for any photovoltaic power station, the actual photovoltaic power sequence of the photovoltaic power station in the previous prediction period before the current prediction analysis period is obtained, and the actual photovoltaic power sequence is normalized to obtain the normalized photovoltaic power sequence of the i th photovoltaic power station in the photovoltaic subsystem ;

[0105] ;

[0106] Wherein, , represents the normalized actual photovoltaic power at the j th prediction point in the normalized photovoltaic power sequence, and J represents the total number of prediction points in the normalized photovoltaic power sequence;

[0107] The K-means clustering algorithm is used to cluster all photovoltaic power stations in the photovoltaic subsystem based on the normalized photovoltaic power sequence, and all photovoltaic power stations are divided into two clusters;

[0108] Specifically, in the normalized photovoltaic power sequence of all photovoltaic power stations in the photovoltaic subsystem, two normalized photovoltaic power sequences are randomly selected as initial cluster centers, for each photovoltaic power station in the photovoltaic subsystem, the Euclidean distance between the normalized photovoltaic power sequence of the photovoltaic power station and the two cluster centers is calculated, and each photovoltaic power station is assigned to the cluster with a closer distance according to the calculated Euclidean distance, when all photovoltaic power stations complete the cluster assignment, the average value of the normalized photovoltaic power sequence of all photovoltaic power stations in each cluster is recalculated, and the average value is taken as the new cluster center point;

[0109] The iteration optimization is repeated until the difference between the cluster center points calculated by adjacent two iterations is less than the preset difference threshold, or the total number of iterations reaches the preset maximum number of iterations, at this time, the stable two clusters are obtained, and the cluster centers are marked as ;

[0110] ;

[0111] Wherein, k is the serial number of the cluster, represents the j th value in the cluster center;

[0112] The Euclidean distance between the two cluster centers is calculated , and the calculation formula is:

[0113] ;

[0114] Among them, k1 and k2 represent the serial numbers of the two clusters respectively;

[0115] If the Euclidean distance between two cluster centers If the value is greater than or equal to the set judgment threshold, the photovoltaic power fluctuations of the photovoltaic power stations in the photovoltaic subsystem are judged to have low similarity. Re-clustering is performed within the two clusters, and each cluster is further divided into two clusters. The Euclidean distance between the centers of the two clusters is calculated and compared with the set judgment threshold again. If the number of clusters is greater than the preset number standard, re-clustering and comparison are stopped.

[0116] If the Euclidean distance between the centers of two clusters in a cluster is If the value is less than the set judgment threshold, it is judged that the photovoltaic power fluctuations of the photovoltaic power stations in the two clusters are highly similar, and the two clusters are marked as similar clusters;

[0117] For similar clusters, the Euclidean distance between the normalized photovoltaic power series of the photovoltaic power stations in the cluster and the cluster center is calculated, and the photovoltaic power stations with a Euclidean distance less than the judgment threshold are classified into the same similar prediction group;

[0118] In the similar prediction group, the core mapping method is used to predict the photovoltaic power of the photovoltaic power stations in the similar prediction group;

[0119] Specifically, the photovoltaic power station with the smallest Euclidean distance in the similar prediction group is used as the core prediction power station of the similar prediction group, and the photovoltaic power stations that are not core prediction power stations in the similar prediction group are used as mapping prediction power stations of the similar prediction group;

[0120] Within the similar prediction group, for the core prediction power station, the predicted meteorological data sequence for the next prediction period is obtained and input into the fusion power prediction model of the core prediction power station to obtain the predicted photovoltaic power sequence of the core prediction power station in the next prediction period. Based on the proportional relationship between the actual photovoltaic power sequences of the core prediction power station and the mapped prediction power station in the previous prediction period, the predicted photovoltaic power sequences of the core prediction power station are mapped to obtain the predicted photovoltaic power sequence of the mapped prediction power station in the next prediction period.

[0121] For PV power stations in the PV subsystem that do not belong to any similar prediction group, obtain the predicted meteorological data sequence for the next prediction period and input it into the fusion power prediction model or physical power model used by each PV power station to obtain the predicted PV power sequence for each PV power station in the next prediction period;

[0122] It needs to be explained that the role of this step is to divide the distributed photovoltaic system into a photovoltaic subsystem according to the jurisdiction of the dispatching department, the connection mode of the power grid and the geographical distance, and to flexibly judge whether the similar set prediction method is triggered based on the proportion of different prediction models adopted by the photovoltaic power station in the photovoltaic subsystem. When the method is triggered, the photovoltaic power stations with similar power fluctuation characteristics are screened out by the K-means clustering algorithm to form a similar prediction group, and the core mapping method is used for prediction. This hierarchical zoning and similarity-based prediction strategy fully considers the spatial distribution characteristics of the distributed photovoltaic system and the correlation between photovoltaic power stations, reduces repeated calculation, improves prediction efficiency, and further improves the accuracy of prediction, and is especially suitable for the power prediction scene of large-scale distributed photovoltaic systems.

[0123] The technical scheme of the embodiment of the application is: acquiring the basic information of each photovoltaic power station in the distributed photovoltaic system, constructing a physical power prediction model of the photovoltaic power station, predicting the photovoltaic power of the photovoltaic power station by using the physical power prediction model of the photovoltaic power station, and evaluating the prediction accuracy by using the actual meteorological data sequence and the actual photovoltaic power sequence obtained by real-time acquisition. If the accuracy is low, the fusion statistical prediction method is triggered. If the fusion statistical prediction method is triggered, a statistical power prediction model is constructed by using a machine learning algorithm in combination with the actual meteorological data sequence and the actual photovoltaic power sequence. The statistical power prediction model and the physical power prediction model are weighted and fused to construct a fusion power prediction model of the photovoltaic power station. The prediction model adopted by the photovoltaic power station is determined according to the weight. The distributed photovoltaic system is divided into a photovoltaic subsystem. Whether the similar set prediction method is triggered for the photovoltaic subsystem is determined based on the proportion of different prediction models adopted by the photovoltaic power station. If the method is not triggered, the photovoltaic power of the photovoltaic power station is predicted according to the adopted prediction model. If the method is triggered, the photovoltaic power of the photovoltaic power station is predicted by using the core mapping method in the similar prediction group.

[0124] Embodiment 2:

[0125] As shown in Figure 2 A distributed photovoltaic power prediction system based on machine learning according to the embodiment of the application comprises the following modules:

[0126] The physical prediction module acquires the basic information of each photovoltaic power station in the distributed photovoltaic system and constructs a physical power prediction model of the photovoltaic power station.

[0127] The accurate analysis module predicts the photovoltaic power of the photovoltaic power station by using the physical power prediction model of the photovoltaic power station, and evaluates the prediction accuracy by using the actual meteorological data sequence and the actual photovoltaic power sequence obtained by real-time acquisition. If the accuracy is low, the fusion statistical prediction method is triggered.

[0128] Fusion prediction module: if the fusion statistical prediction method is triggered, a statistical power prediction model is constructed by combining the actual meteorological data sequence and the actual photovoltaic power sequence through a machine learning algorithm, the statistical power prediction model and the physical power prediction model are weighted fused, the fusion power prediction model of the photovoltaic power station is constructed, and the prediction model adopted by the photovoltaic power station is judged according to the weight;

[0129] Set prediction module: the distributed photovoltaic system is divided into photovoltaic subsystems, it is judged whether the similar set prediction method is triggered for the photovoltaic subsystems, if not, the photovoltaic power of the photovoltaic power station is predicted according to the adopted prediction model, if yes, the similar prediction groups are screened through a clustering algorithm, and the photovoltaic power of the photovoltaic power station is predicted in the similar prediction groups by using a core mapping method.

[0130] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application, and various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A machine learning based distributed photovoltaic power prediction method, characterized in that: The method comprises the following steps: acquiring basic information of each photovoltaic power station in a distributed photovoltaic system, and constructing a physical power prediction model of the photovoltaic power station; predicting photovoltaic power of the photovoltaic power station by using the physical power prediction model of the photovoltaic power station, and evaluating the prediction accuracy by using actual meteorological data sequence and actual photovoltaic power sequence obtained by real-time acquisition, and triggering a fusion statistical prediction method if the accuracy is low; if the fusion statistical prediction method is triggered, constructing a statistical power prediction model by using a machine learning algorithm and combining the actual meteorological data sequence and the actual photovoltaic power sequence, weighting the statistical power prediction model and the physical power prediction model, constructing a fusion power prediction model of the photovoltaic power station, and judging the prediction model used by the photovoltaic power station according to the weight; the judgment mode of the prediction model used by the photovoltaic power station is as follows: the prediction model comprises the fusion power prediction model and the physical power prediction model; a weight evaluation flag value is set for each photovoltaic power station triggering the fusion statistical prediction method, and the initial value is 0, the weight of the physical power prediction model in the fusion power prediction model is compared with a preset weight threshold value, the weight evaluation flag value is increased by one if the weight is greater than or equal to the weight threshold value, and the weight evaluation flag value is reset to 0 if the weight is less than the weight threshold value; if the weight evaluation flag value is less than a stable flag value, the photovoltaic power station uses the fusion power prediction model, otherwise, the photovoltaic power station ends the fusion statistical prediction method and uses the physical power prediction model; the distributed photovoltaic system is divided into photovoltaic subsystems, whether the photovoltaic subsystem triggers a similar set prediction method is judged, if not, the photovoltaic power of the photovoltaic power station is predicted according to the prediction model used, if yes, similar prediction groups are screened by using a clustering algorithm, and the photovoltaic power of the photovoltaic power station is predicted by using a core mapping method in the similar prediction groups; the core mapping method comprises: obtaining the Euclidean distance between the normalized photovoltaic power sequence of the photovoltaic power station in the similar prediction group and the cluster center; the photovoltaic power station with the minimum Euclidean distance is taken as the core prediction station of the similar prediction group, and the photovoltaic power stations other than the core prediction station are taken as the mapping prediction stations of the similar prediction group; for the core prediction station, the prediction meteorological data sequence is obtained, input into the fusion power prediction model of the core prediction station, the prediction photovoltaic power sequence of the core prediction station is obtained, and the prediction photovoltaic power sequence of the mapping prediction station is obtained according to the proportion relationship between the core prediction station and the mapping prediction station in the actual photovoltaic power sequence in the previous prediction period.

2. The method of claim 1, wherein: the evaluation mode of the prediction accuracy is as follows: a prediction period is set, and the interval period between adjacent prediction periods is a prediction analysis period; in each prediction period, a plurality of prediction points are uniformly taken, at the beginning of the current prediction period, the prediction meteorological data sequence of the region where the photovoltaic power station is located in the prediction period is obtained, input into the physical power prediction model, the prediction photovoltaic power sequence is obtained, the actual photovoltaic power is acquired in real time, at the end of the current prediction period, the actual photovoltaic power sequence is obtained according to time sequence integration; Calculate the root mean square error between the predicted photovoltaic power sequence and the actual photovoltaic power sequence, and if the root mean square error is greater than or equal to the preset error threshold, determine that the prediction accuracy of the physical power prediction model is low.

3. The method of claim 1, wherein: The fusion power prediction model is obtained in the following manner: The statistical power prediction model and the physical power prediction model are weighted and fused to obtain a fusion power prediction model, wherein the sum of the weights of the statistical power prediction model and the physical power prediction model is 1. A population of weights of the physical power prediction model is randomly generated by a genetic algorithm, the total actual meteorological data sequence is input into the fusion power prediction model, the root mean square error between the prediction result and the corresponding actual photovoltaic power is calculated, and the population is continuously evolved by selection, crossover and mutation genetic operations to gradually screen out the weight of the physical power prediction model that minimizes the root mean square error, thereby determining the weight of the physical power prediction model and obtaining the fusion power prediction model.

4. The method of claim 3, wherein: The statistical power prediction model is obtained in the following manner: Determine whether the photovoltaic power station has a statistical power prediction model; If not, obtain all actual photovoltaic power sequences and actual meteorological data sequences of the photovoltaic power station up to the current time, respectively integrate the total photovoltaic power sequence and the total actual meteorological data sequence based on time sequence to obtain a training data set, construct a random forest model based on a machine learning algorithm and train the model using the training data set, and obtain the trained random forest model as the statistical power prediction model of the photovoltaic power station; If so, obtain actual photovoltaic power sequences and actual meteorological data sequences that are not included in the training data set to supplement the training data set, supplementally train the statistical power prediction model, and update the statistical power prediction model of the photovoltaic power station.

5. The method of claim 2, wherein: The determination method of whether to trigger the similar set prediction method is as follows: Divide the distributed photovoltaic system into a plurality of photovoltaic subsystems according to the jurisdiction of the dispatching department, the grid connection mode and the geographical distance, obtain the number of photovoltaic power stations in each photovoltaic subsystem that use the fusion power prediction model, calculate the ratio of the number of photovoltaic power stations to the total number of photovoltaic power stations in the photovoltaic subsystem, and if the ratio is greater than a preset proportion standard, trigger the similar set prediction method in the photovoltaic subsystem.

6. The method of claim 5, wherein: The acquisition method of the similar prediction group is as follows: Divide all photovoltaic power stations in the photovoltaic subsystem into two clusters, calculate the Euclidean distance between the centers of the two clusters and compare it with a set judgment threshold; If greater than or equal to the set judgment threshold, divide each cluster into two clusters again, calculate the Euclidean distance between the centers of the two clusters again, and compare them again, and if the number of clusters is greater than a preset number standard, stop re-clustering and comparison; If less than the set judgment threshold, mark the two clusters as similar clusters, obtain and calculate the normalized photovoltaic power sequence of the photovoltaic power stations in the clusters and the Euclidean distance between the cluster centers, and the photovoltaic power stations less than the judgment threshold are included in the same similar prediction group.

7. The method of claim 6, wherein: The method of dividing into two clusters is as follows: The actual photovoltaic power sequence of the photovoltaic power station in a previous prediction period is acquired for normalization processing to obtain a normalized photovoltaic power sequence, and the K-means clustering algorithm is used to cluster all photovoltaic power stations in the photovoltaic subsystem based on the normalized photovoltaic power sequence, so as to divide all photovoltaic power stations into two clusters.

8. A machine learning based distributed photovoltaic power prediction system for implementing the prediction method according to any one of claims 1-7, characterized in that: Comprise the following modules: A physical prediction module: acquires the basic information of each photovoltaic power station in the distributed photovoltaic system, and constructs a physical power prediction model of the photovoltaic power station; An accurate analysis module: uses the physical power prediction model of the photovoltaic power station to predict the photovoltaic power of the photovoltaic power station, and evaluates the prediction accuracy through the actual meteorological data sequence and the actual photovoltaic power sequence collected in real time, and if the accuracy is low, the fusion statistical prediction method is triggered; A fusion prediction module: if the fusion statistical prediction method is triggered, a statistical power prediction model is constructed by combining the actual meteorological data sequence and the actual photovoltaic power sequence through a machine learning algorithm, the statistical power prediction model and the physical power prediction model are weighted and fused to construct a fusion power prediction model of the photovoltaic power station, and the prediction model used by the photovoltaic power station is determined according to the weight; The determination mode of the prediction model used by the photovoltaic power station is: The prediction model comprises a fusion power prediction model and a physical power prediction model; A weight evaluation flag value is set for each photovoltaic power station triggering the fusion statistical prediction method, and the initial value is 0, the weight of the physical power prediction model in the fusion power prediction model is compared with a preset weight threshold value, if the weight is greater than or equal to the weight threshold value, the weight evaluation flag value is increased by one, and if the weight is less than the weight threshold value, the weight evaluation flag value is reset to 0; If the weight evaluation flag value is less than a stable flag value, the photovoltaic power station uses the fusion power prediction model, otherwise, the photovoltaic power station ends the fusion statistical prediction method and uses the physical power prediction model; A set prediction module: the distributed photovoltaic system is divided into photovoltaic subsystems, it is judged whether the photovoltaic subsystem triggers a similar set prediction method, if not, the photovoltaic power of the photovoltaic power station is predicted according to the prediction model used, if yes, a similar prediction group is screened through a clustering algorithm, and the photovoltaic power of the photovoltaic power station in the similar prediction group is predicted by using a core mapping method; The core mapping method comprises: The Euclidean distance between the normalized photovoltaic power sequence of the photovoltaic power station in the similar prediction group and the cluster center is acquired; The photovoltaic power station with the smallest Euclidean distance is taken as the core prediction station of the similar prediction group, and the photovoltaic power stations other than the core prediction station are taken as the mapping prediction stations of the similar prediction group; For the core prediction station, the prediction meteorological data sequence is acquired and input into the fusion power prediction model of the core prediction station to obtain the prediction photovoltaic power sequence of the core prediction station, and the prediction photovoltaic power sequence of the mapping prediction station is obtained by mapping the prediction photovoltaic power sequence of the core prediction station according to the proportional relationship between the core prediction station and the mapping prediction station in the actual photovoltaic power sequence in a previous prediction period.

Images