Region-level photovoltaic output accurate prediction method and device based on subdivided grids
By combining fuzzy C-means clustering and a BP neural network optimized by a genetic algorithm with subdivided grid technology, the problem of inaccurate photovoltaic power generation prediction after distributed photovoltaic power stations are connected to the grid is solved, and accurate prediction of photovoltaic output and optimization of grid dispatch are achieved.
Patent Information
- Application Number
- CN202410979466.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-15
AI Technical Summary
After a high proportion of distributed photovoltaic power stations are connected to the regional power grid, the photovoltaic power generation is affected by complex external factors, and existing prediction methods are difficult to achieve accurate prediction, which leads to difficulties in power grid dispatch.
A BP neural network optimized by fuzzy C-means clustering algorithm and genetic algorithm, combined with subdivided grid technology, is used to construct a photovoltaic ultra-short-term output prediction model. The model can accurately predict photovoltaic output by using installation information, photovoltaic effective irradiance calculation and meteorological data classification.
It significantly improves the accuracy and generalizability of ultra-short-term photovoltaic power output forecasting, reduces grid dispatch risks, and enhances the grid connection and absorption capacity of solar energy.
Smart Images

Figure CN118970891A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and device for accurately predicting photovoltaic output at a subdivided grid level, and belongs to the technical field of photovoltaic power generation prediction. Background Art
[0002] With the rapid development of social economy in recent years, it is required to actively change the way of energy production and consumption, optimize the energy structure, increase the development of renewable energy, and promote the construction of distributed energy. Distributed photovoltaic power stations have the characteristics of small units and dispersed installed capacity; and some areas show characteristics such as high penetration rate and high density. The prediction accuracy of single distributed photovoltaic power generation output is not high.
[0003] When a high proportion of distributed photovoltaic power stations are connected to the regional power grid, the following impacts will be caused on the regional power grid: 1) Impact on the local voltage stability of the power grid; 2) Impact on the frequency stability of the power grid; 3) Impact on the quality of electric energy. Specifically, compared with traditional power sources, photovoltaic power generation is more complexly affected by external factors, and it is difficult to directly reflect its changing characteristics with only a few characteristics. Direct influencing factors include solar radiation intensity, atmospheric quality, cloud thickness and height, installation angle of photovoltaic panels, photovoltaic conversion efficiency of photovoltaic panels, etc., while indirect factors are various environmental factors such as wind speed, temperature, humidity, rainfall, etc. Sudden changes in meteorological conditions may lead to huge differences between the predicted value of photovoltaic power generation and the actual output, which poses new challenges to traditional optimization scheduling and control methods, making the optimal scheduling strategy based on deterministic optimization increasingly inadequate in economy. Summary of the invention
[0004] Aiming at the photovoltaic power generation prediction problem at the substation level, the present invention explores a distributed photovoltaic equivalent substitution method, establishes a photovoltaic ultra-short-term output prediction model based on a fuzzy C-means clustering algorithm and a BP neural network optimized by a genetic algorithm (GA), and corrects the photovoltaic output prediction error.
[0005] The present invention adopts the following technical means to achieve: a method for accurately predicting photovoltaic output at the subdivided grid level, comprising the following steps:
[0006] Step 1: construct a distributed photovoltaic power generation simulation model based on installation information;
[0007] Step 2: Based on the photovoltaic power generation simulation model, the fuzzy C-means clustering algorithm is applied to classify the historical weather data of the forecast area according to meteorological conditions, screen similar day sets, and divide the normalized similar day sets into training sets and test sets;
[0008] Step 3: Divide the forecast area into grids to determine the forecast location, and construct a high-dimensional input matrix based on the meteorological data of each grid;
[0009] Step 4: Use the high-dimensional input matrix as the input of the GA-BP neural network photovoltaic ultra-short-term output prediction model, establish the GA-BP neural network photovoltaic ultra-short-term output prediction model, use the training set of similar day sets for training, and use the genetic algorithm to optimize the parameters to improve the generalization and accuracy of the model prediction;
[0010] Step 5: Use the optimized GA-BP neural network photovoltaic ultra-short-term output prediction model to predict the photovoltaic ultra-short-term output of the test set of similar days, and make dynamic corrections based on the error between the actual output and the predicted output.
[0011] In step 1, the distributed photovoltaic power generation simulation model based on installation information specifically includes the following three parts:
[0012] The first part is the calculation of installation parameter information, which is the position of the sun relative to the observation device on the earth's surface and is a necessary input required by the distributed photovoltaic power generation simulation model;
[0013] The second part is the photovoltaic effective irradiance calculation, which is used to calculate the global horizontal irradiance, direct normal irradiance and diffuse horizontal irradiance, which is used as the input of the distributed photovoltaic power generation simulation model;
[0014] The third part is the calculation of photovoltaic power generation. The present invention uses the De Soto model to analyze and simulate the performance of photovoltaic modules.
[0015] The specific processing flow of the fuzzy C-means clustering algorithm analysis in step 2 is as follows:
[0016] First, a sample data set X is established, and the fuzzy cluster center and the number of cluster centers are set. The membership matrix is updated by repeatedly adjusting the cluster center matrix. The optimal membership and cluster center are obtained based on the minimum sum of the Euclidean distances from each data sample to the cluster center. Finally, the sample points are clustered according to the membership. The specific formula is as follows:
[0017]
[0018] Among them, the data set X = {x 1 ,x 2 ,…,x n},v j (t) is the fuzzy cluster center; c is the number of cluster centers; n is the number of sample data; u ij (t) is the membership degree of the i-th sample to the j-th class, u ij (t)∈[0,1]; d ij(t) is the Euclidean distance from sample i to cluster center j, d ij (t)=||c i -c j ||;d ik (t) is the Euclidean distance from sample i to cluster center k, d ik (t)=||c i -c k ||; t represents the time of the data sample, and m is a weight parameter that describes the degree of fuzzification.
[0019] The similar day screening in step 2 specifically includes the following steps:
[0020] 4-1) According to the previous forecast day weather condition identification, determine which weather category the forecast day belongs to, and select all the data that match this weather condition category during the forecast period to obtain the similar day set A. Using the Chebyshev method, the important meteorological characteristic parameters of the forecast day, solar irradiance C(t) and precipitation G(t), are compared with the solar irradiance C(t) of the mth day in the similar day set A. m (t), precipitation G m (t) Calculate the corresponding Chebyshev distance, the distance d m The expression is as follows:
[0021]
[0022] Among them, ρ C,m C(t) and C m (t) correlation coefficient of meteorological characteristic parameters; ρ G,m G(t) and G m (t) is the correlation coefficient of the meteorological characteristic parameters; D(C(t)) is the variance of the meteorological characteristic parameters of C(t); D(G(t)) is the variance of the meteorological characteristic parameters of G(t).
[0023] 4-2) Using the Chebyshev method, we can find the Chebyshev distance between each day in the similar day set A and the predicted day, and get {d 1 ,d 2 ,…,d m}, set the ratio coefficient η to satisfy:
[0024]
[0025] The distance value sequence {d 1 ,d 2 ,…,d m} is less than or equal to (1+η)·min{d 1 ,d 2 ,…,d mAll feature diaries of} are subset B in similar day set A, and subset B is defined as the similar day set.
[0026] The regional grid described in step 3 is divided as follows:
[0027] Specifically, the distributed photovoltaic power station at the grid scale is taken as the research object, and the prediction area is divided into several independent small grids. Each grid is composed of several distributed photovoltaic power stations. According to the area setting of the distributed photovoltaic power station cluster, the size of each grid in the region is about 10km×10km, and each grid includes 10 or more 10kV lines.
[0028] In step 3, the input matrix uses the photovoltaic daily power curve within a single grid as the minimum analysis unit for research. The input of the matrix is the photovoltaic power generation power contained in the entire region, meteorological data such as solar irradiance, ambient temperature and wind speed. The output of the matrix is the photovoltaic power generation power of each grid on the day to be predicted. After the grid is divided, the photovoltaic power generation power curves of each grid with different meteorological characteristics are integrated together to predict the overall photovoltaic output under various weather conditions in the region.
[0029] The specific steps of photovoltaic output prediction are as follows:
[0030] Firstly, by screening the meteorological data of the entire prediction area, a similar day training set is obtained, and the corresponding relationship between the meteorological data and the grid output power is constructed and applied to the prediction of the day to be predicted; secondly, according to the prediction location and requirements, the overall regional prediction or the single grid prediction is selected for targeted prediction.
[0031] The output layer activation function of the GA-BP neural network photovoltaic ultra-short-term output prediction model in step 4 is:
[0032] E=SPη[1-0.005(t 0 +25)]
[0033] Where E is the value of photovoltaic output; S is the value of photovoltaic collection panel area; P is the effective irradiance; t 0 To monitor the temperature, η is the value of the energy conversion efficiency.
[0034] The present invention provides a subdivided grid-based photovoltaic output accurate prediction device, which includes the following modules:
[0035] A subdivided grid-based photovoltaic output accurate prediction device, the device includes the following modules:
[0036] Installation parameter information calculation module: used to observe the position of the sun relative to the earth's surface, which is a necessary input for the distributed photovoltaic power generation simulation model;
[0037] Photovoltaic effective irradiance calculation module: used to calculate global horizontal irradiance, direct normal irradiance and diffuse horizontal irradiance, which are used as input for distributed photovoltaic power generation simulation model;
[0038] Photovoltaic power calculation module: used to analyze and simulate the performance of photovoltaic modules;
[0039] Similar day clustering module: used to classify the historical weather data of the forecast area according to meteorological conditions, filter similar day sets, and divide the normalized similar day sets into training sets and test sets;
[0040] Grid division module: used to divide the forecast area into grids to determine the forecast location, and to form a high-dimensional input matrix based on the meteorological data of each grid;
[0041] GA-BP neural network photovoltaic ultra-short-term output prediction model module: used to accurately predict photovoltaic output, and make overall regional predictions or single grid predictions based on the prediction location and requirements.
[0042] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method for accurately predicting photovoltaic output at the subdivided grid level is implemented.
[0043] The present invention also provides a computer-readable storage medium, which includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the method for accurately predicting photovoltaic output at the subdivided grid level.
[0044] The beneficial effects of the present invention after adopting the above technical solution are:
[0045] In the case of days with similar weather, photovoltaic data is affected by multivariate meteorological factors, with large data volume and high data dimension. The present invention applies the fuzzy C-means cluster analysis method to achieve flexible fuzzy division, which can maximize the rationality of the classification results. Compared with the prior art, the technical solution of the present invention can significantly improve the accuracy and generalization of photovoltaic ultra-short-term output forecasts, reduce the grid dispatch risks caused by photovoltaic power generation power fluctuations, and help the power dispatching department to reasonably arrange the operation mode and system dispatching plan of photovoltaic power sources, improve the grid connection and absorption capacity of solar energy, and has good practicality and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a diagram of a distributed photovoltaic power generation system;
[0047] Figure 2 It is a distributed photovoltaic power generation simulation flow chart;
[0048] Figure 3 This is a schematic diagram of a similar day dataset for sunny scenes;
[0049] Figure 4 This is a schematic diagram of a dataset of similar days with cloudy scenes;
[0050] Figure 5 This is a schematic diagram of a similar day dataset for rainy scenes;
[0051] Figure 6 It is the structure diagram of BP neural network;
[0052] Figure 7 It is the training flow chart of GA-BP neural network;
[0053] Figure 8 This is the basic idea diagram of the prediction model;
[0054] Fig. 9 It is a short-term prediction diagram of photovoltaic power generation on a typical similar day on a sunny day;
[0055] Fig.10 It is a short-term forecast diagram of photovoltaic power generation on a typical similar day on a cloudy day;
[0056] Fig.11 It is a short-term prediction diagram of photovoltaic power generation on a typical similar day in rainy days. DETAILED DESCRIPTION
[0057] The technical solution of the present invention is described clearly and completely below in conjunction with the accompanying drawings.
[0058] First, distributed photovoltaic power generation simulation modeling based on installation information
[0059] A typical distributed photovoltaic power generation system mainly includes: solar cell arrays (modules), cables (modules), power electronic converters, energy storage devices (batteries), loads (users), such as Figure 1 As shown, photovoltaic components and energy storage devices constitute the power supply system, the controller and power electronic converter constitute the control and protection system, and the terminal of the system is the load.
[0060] The present invention will establish a distributed photovoltaic power generation simulation model based on the Python PVILB library to calculate the solar irradiance. The specific process is as follows: Figure 2 shown.
[0061] The specific parts of the simulation model include: calculation of installation parameter information, calculation of photovoltaic effective irradiance and calculation of photovoltaic power generation.
[0062] The first step is to calculate the installation parameter information. The installation parameter information is the position of the sun relative to the observation device on the earth's surface, which is a necessary input for the distributed photovoltaic power generation simulation model. Common terms for describing the position of the sun include: solar altitude angle (θ h ), solar zenith angle (θ Z ), and the solar azimuth (θ A ).θ h Defined as the angle between the horizon and the sun:
[0063]
[0064] where δ is the declination angle of the sun, is the latitude of the observer (positive in the northern hemisphere), and ω is the hour angle. δ is defined as the angular position of the sun at noon relative to the equatorial plane and is given by:
[0065]
[0066] ω is defined as the angular displacement of the sun east or west of the local meridian (zero for due south, negative for east, positive for west) and can be calculated from local time (hours, minutes, seconds):
[0067]
[0068] The local time must be entered in the format of 0-24 hours. Z is θ h Additional angles:
[0069] θ Z =90°-θ h
[0070] θ A Defined as the angular displacement of the projection of the beam radiation on the horizontal plane from the south. Generally speaking, A = 0 is due south, A < 0 is east, and A > 0 is west. θ A The calculation method is:
[0071]
[0072] Next is the calculation of photovoltaic effective irradiance. Irradiance data, including global horizontal irradiance (GHI), direct normal irradiance (DNI) and diffuse horizontal irradiance (DHI), are used as inputs to the distributed photovoltaic power generation simulation model. Given the hourly GHI and DHI, DNI is determined according to the following formula:
[0073] GHI=DHI+DNI·cos(θ Z )
[0074] Next, the irradiance on the plane of the photovoltaic panel (E POA), the calculation formula is:
[0075] E POA =E b +E g +E d
[0076] Among them, E b The DNI is calculated by adjusting the angle of incidence (AOI) between the sun's rays and the PV array as follows:
[0077] E b =DNI×cos(AOI)
[0078] The calculation formula of AOI is:
[0079] AOI=cos -1 [cos(θ Z )cos(θ T )+sin(θ Z )sin(θ T )cos(θ A -θ A,a )]
[0080] where θ T and θ A,a are the tilt angle and azimuth of the photovoltaic panel respectively. The irradiance E reflected by the ground on the inclined surface g It can be expressed as a function of the GHI, the reflectivity of the ground surface (called albedo, which defaults to 0.2), and the surface tilt angle:
[0081]
[0082] Diffuse irradiance E of the sky d An empirical model developed by Sandia National Laboratories can be used to determine:
[0083]
[0084] Finally, the photovoltaic power generation calculation, the present invention uses the De Soto model (also known as the five-parameter model) to analyze and simulate the performance of photovoltaic modules. The three main points on the module current-voltage (IV) curve, namely the short-circuit current (I sc ), open circuit voltage (V oc ) and the maximum power point (I mp ,V mp ), is given by the following formula:
[0085]
[0086]
[0087]
[0088]
[0089] f 1 The air mass modifier is called the air mass modifier, which is the absolute air mass (AM a ) is a fourth-order polynomial function:
[0090] f 1 =a 0 +a 1 AM a +a 2 (AM a ) 2 +a 3 (AM a ) 3 +a 4 (AM a ) 4
[0091] where a is the coefficient vector determined by the modularity test. a Provides a relative measure of the path length the Sun must travel through the atmosphere (the sea level AM when the Sun is directly overhead). a =1), and can be quickly calculated from the solar zenith angle and altitude:
[0092]
[0093] f 2 It is called the incident angle correction function and is a fifth-order polynomial function of AOI:
[0094] f 2 =b 0 +b 1 AOI+b 2 (AOI) 2 +b 3 (AOI) 3 +b 4 (AOI) 4 +b 5 (AOI) 5
[0095] where b is the coefficient vector determined by the modularity test. δ is T c The function is defined as:
[0096]
[0097] where n is the empirically determined diode coefficient, k is the Boltzmann constant (1.381×10-23 J / K), and q is the fundamental charge constant (1.6022×10-19 Coulombs).
[0098] The power at the maximum power point (I mp ,V mp ) is determined as:
[0099] P mp =I mp ×V mp
[0100] In most cases, it is assumed that the inverter of the standard array (with maximum power point tracking capability) can perfectly track the maximum power point. Based on the above distributed photovoltaic power generation simulation model, the longitude and latitude information of the power station is input to obtain the solar irradiance of the area.
[0101] Second, similar day weather state recognition based on fuzzy C-means clustering
[0102] Based on the simulation of distributed photovoltaic power generation, the present invention will further explore the changes in the effective irradiance of centralized photovoltaics under different weather conditions, pair the distributed photovoltaics and photovoltaic clusters with the highest similarity under various conditions, and establish an empirical method for converting distributed photovoltaics into equivalent photovoltaic clusters.
[0103] Under similar weather conditions, photovoltaic power generation data is affected by multiple meteorological factors, and has the characteristics of large data volume and high dimension. The present invention will apply the fuzzy C-means (FCM) clustering analysis method to explore the changes in effective photovoltaic irradiance under similar weather conditions. The fuzzy C-means clustering algorithm is improved from the K-means algorithm. It is a fuzzy clustering algorithm based on distance and membership as the objective function, which realizes flexible fuzzy partitioning and can maximize the rationality of the classification results.
[0104] This paper takes a certain area in Zhenjiang City, Jiangsu Province as the research area, and uses the meteorological data collected by the environmental detector and the 10kW distributed photovoltaic measurement point data, with a total of 720 samples, including four dimensions: temperature, wind speed, precipitation, and effective light irradiance. On this basis, the fuzzy C-means clustering method is used to perform overall clustering of meteorological elements to obtain the optimal number of clusters. The specific processing flow is as follows:
[0105] First, a sample data set X is established, and the fuzzy cluster center and the number of cluster centers are set. The membership matrix is updated by repeatedly adjusting the cluster center matrix. The optimal membership and cluster center are obtained based on the minimum sum of the Euclidean distances from each data sample to the cluster center. Finally, the sample points are clustered according to the membership. The specific formula is as follows:
[0106]
[0107] Among them, the data set X = {x 1 ,x2 ,…,x n},v j (t) is the fuzzy cluster center; c is the number of cluster centers; n is the number of sample data; u ij (t) is the membership degree of the i-th sample to the j-th class, u ij (t)∈[0,1]; d ij (t) is the Euclidean distance from sample i to cluster center j, d ij (t)=||c i -c j ||;d ik (t) is the Euclidean distance from sample i to cluster center k, d ik (t)=||c i -c k ||; t represents the time of the data sample, and m is a weight parameter that describes the degree of fuzzification.
[0108] According to the results of the discrimination, when the number of clusters is 3, the membership degree and comprehensive index value are both optimal, so the number of cluster layers is set to 3, and the meteorological scenarios are divided into sunny, cloudy and rainy. There are 720 data for the whole month. After cluster analysis, there are 360 clusters for sunny days, 144 clusters for cloudy days, and 216 clusters for rainy days. The clustering results are as follows: Figure 3-Figure 5 shown.
[0109] Figure 3-Figure 5 In the case of clear weather, the effective irradiance per unit area is relatively smooth, with an amplitude of about 1000W. In this case, due to the periodicity, trend, and seasonality of the data, the prediction accuracy of the scheme is relatively high; in the case of gloomy weather, due to the influence of cloudy weather during the day, the effective irradiance per unit area fluctuates greatly, with the maximum fluctuation of only about 900W, and the trend of this change is very obvious, and there is no significant data characteristic. In this scenario, it is difficult for conventional forecast models to accurately grasp the mutual influence of weather and photovoltaic power generation, and it is difficult to make effective predictions; in the case of rainy weather, the effective irradiance data received by solar panels fluctuates greatly due to changes in wind speed and lower temperatures, and the peak effective irradiance per unit area is only 700W, with the worst trend; at the same time, the weather data under rainy weather conditions has a strong randomness, and it is difficult to find its long-term correlation with photovoltaic power generation.
[0110] In the three clustering analyses, the illumination period and duration change with the change of seasons. On this basis, the distributed photovoltaic power generation simulation model will use the self-attention mechanism to automatically capture the photoelectric response generated by the change of illumination time. The present invention intends to study the long-term correlation between the effective irradiance of light and the photovoltaic power generation, pre-classify it in the distributed photovoltaic power generation model, and equate it to a photovoltaic cluster, so as to achieve high-precision prediction of the effective irradiance of light and the photovoltaic power generation.
[0111] According to the previous forecast day meteorological condition identification, it is determined which meteorological category the forecast day belongs to, and all data that match this meteorological condition category within the forecast period are screened to obtain the similar day set A. Using the Chebyshev method, the important meteorological characteristic parameters of the forecast day, solar irradiance C(t) and precipitation G(t), are compared with the solar irradiance C(t) of the mth day in the similar day set A. m (t), precipitation G m (t) Calculate the corresponding Chebyshev distance, the distance d m The expression is as follows:
[0112]
[0113] Among them, ρ C,m C(t) and C m (t) correlation coefficient of meteorological characteristic parameters; ρ G,m G(t) and G m (t) is the correlation coefficient of the meteorological characteristic parameters; D(C(t)) is the variance of the meteorological characteristic parameters of C(t); D(G(t)) is the variance of the meteorological characteristic parameters of G(t).
[0114] Using the Chebyshev method, we can find the Chebyshev distance between each day in the similar day set A and the predicted day, and get {d 1 ,d 2 ,…,d m}, set the ratio coefficient η to satisfy:
[0115]
[0116] The distance value sequence {d 1 ,d 2 ,…,d m} is less than or equal to (1+η)·min{d 1 ,d 2 ,...,d m All feature diaries of} are subset B in similar day set A, and subset B is defined as the similar day set.
[0117] Third, the establishment of a subdivided grid-based photovoltaic ultra-short-term output prediction model
[0118] In view of the fact that the current photovoltaic output prediction method based on numerical weather forecast lacks the extraction of time series data, it is easily interfered by meteorological factors, and has poor anti-interference performance for extreme climate events such as heavy precipitation, resulting in low prediction accuracy. In addition, the use of conventional numerical weather forecast prediction methods has problems such as long time consumption and high cost. Based on the above characteristics, the present invention will establish a photovoltaic ultra-short-term output prediction model based on a GA-BP multi-layer forward neural network. The neural network consists of three layers: input layer, hidden layer, and output layer, each of which is composed of a number of neurons. The basic model framework is as follows: Figure 6 shown.
[0119] Figure 6 In the paper, the main idea of the multi-layer feedforward neural network is: forward training and backward training in two stages; the first link is forward transmission, starting from the input layer, through the training of the hidden layer, reaching the output layer, completing the forward transmission; secondly, due to the reverse transmission of the error, when the first-level signal reaches the output layer, its initial data is often not accurate enough and has a large deviation from the real data, so it must be learned. The correction is performed in the second step of the neural network. By comparing the data of the output layer with the real data, the error is transmitted to the input layer through the hidden layer, and the accuracy of the model is improved by adjusting the weights and biases of the input-hidden and hidden-output layers, thereby achieving an accurate prediction effect.
[0120] The input data of the model are the irradiance, temperature, wind speed and power at the previous moment, and the output data is the predicted value of photovoltaic power generation. The present invention calculates the number of hidden layer nodes through an empirical formula, and the calculation method is as follows:
[0121]
[0122] In the formula, m is the number of input layer nodes, h is the number of hidden layer nodes, p is the number of output layer nodes, and a is any integer between 0 and 10. From the above formula, it can be determined that the optimal range of hidden layer nodes is 3 to 12.
[0123] The traditional BP neural network is based on the gradient descent method. Its optimal solution is relatively cumbersome, and there are problems such as slow convergence speed and easy to fall into local extreme values. In view of the above shortcomings, the present invention applies genetic algorithm optimization to the learning process of BP neural network and optimizes the model. Its training flow chart is as follows Figure 7 shown.
[0124] According to the main factors affecting photovoltaic output that have been screened out before, the temperature, effective irradiance and wind speed collected by the environmental monitoring system in a certain area of Zhenjiang City, Jiangsu Province are extracted, as well as the historical photovoltaic power generation power. In order to ensure that the number of samples is sufficient and to provide sufficient training sets for the neural network structure to improve the accuracy of the prediction, the present invention collects the annual meteorological data of the environmental monitoring system in 2023, with a time interval of 5 minutes. The 60 working days of the data collected throughout the year are used as the test set, and the remaining data are used as the training set. This sample is used as the training sample, and a high-precision neural network is obtained through training. The established prediction model is used to predict the typical daily hourly photovoltaic power generation power data of each similar day set. The output layer activation function of the photovoltaic output prediction model used in the present invention is:
[0125] E=SPη[1-0.005(t 0 +25)]
[0126] Where E is the value of photovoltaic output; S is the value of photovoltaic collection panel area; P is the effective irradiance; t 0 To monitor the temperature, η is the value of the energy conversion efficiency.
[0127] The present invention will use a subdivided grid prediction method to process the overall data of the region, specifically taking the distributed photovoltaic power station at the grid scale as the research object, dividing the prediction area into several independent small grids, each grid is composed of several distributed photovoltaic power stations, and is set according to the area of the distributed photovoltaic power station cluster. The size of each grid in the region is about 10km×10km, and each grid includes 10 or more 10kV lines. The input matrix studies the photovoltaic daily power curve in a single grid as the minimum analysis unit. The input of the matrix is the photovoltaic power generation power, meteorological data such as solar irradiance, ambient temperature and wind speed contained in the entire region. The output of the matrix is the photovoltaic power generation power of each grid to be predicted on the day. After the grid is divided, the photovoltaic power generation power curves with different meteorological characteristics of each grid are integrated together to predict the overall photovoltaic output under various weather conditions in the region.
[0128] First, by screening the entire grid to obtain a similar day training set, the corresponding relationship between meteorological data and grid output power is constructed and applied to the prediction of the day to be predicted; secondly, according to the prediction location and requirements, the overall regional prediction or single grid prediction is selected for targeted prediction. The present invention intends to use the method with the maximum global similarity to screen similar days. The prediction process is as follows Figure 8 As shown, the specific steps are as follows:
[0129] 1) The fuzzy C-means clustering method is used to identify the meteorological conditions of the overall photovoltaic power generation of the grid, and it is divided into three types: sunny, cloudy and rainy. The credibility of the screening is improved by screening the reliable data of the environmental monitoring station.
[0130] 2) By dividing the overall area into grids, statistical analysis is performed on the daily power curve in each grid based on the historical records of photovoltaic power generation changes, so as to identify similar days in different time periods and improve accuracy.
[0131] 3) A high-dimensional input matrix is constructed based on the temperature, wind speed, effective irradiance and the power curve of the distributed photovoltaic power station in each grid, and this is used as the input to build a photovoltaic ultra-short-term output prediction model based on the GA-BP neural network. By training each grid unit, the photovoltaic output daily prediction curve of each grid unit is obtained. Compared with a single area, the subdivided grid prediction improves the accuracy of the prediction by introducing distributed photovoltaic power generation data in neighboring areas and spatial distribution information between grids.
[0132] 4) Output the predicted daily photovoltaic power generation curve of the target grid.
[0133] Fourth, prediction results and analysis
[0134] To verify the effectiveness of the present invention, it is illustrated by specific examples below. The present invention utilizes the historical data of photovoltaic power generation of environmental detectors in a certain area of Zhenjiang City, Jiangsu Province and the corresponding weather data (temperature, effective irradiance, wind speed), and uses the fuzzy C-means clustering method and the GA-BP neural network photovoltaic ultra-short-term output prediction model for processing. The photovoltaic power generation power is sampled every 5 minutes, and the data normalized with the temperature, effective irradiance and wind speed are used as the input of the photovoltaic ultra-short-term output prediction model, and the predicted power value is output after network training. The present invention will study the situations of three types of typical similar days respectively, and select December 28, 2023 as a typical sunny day, December 30, 2023 as a typical cloudy day, and December 18, 2023 as a typical rainy day for prediction and comparison with the measured values. The prediction results are as follows Fig. 9 , Fig.10 , Fig.11 shown.
[0135] Figures 9 to 11 The following are comparison diagrams of the predicted values (GA-BP neural network, BP neural network) and the measured data under three typical similar day conditions: sunny day, cloudy day, and rainy day. The error curve is the change curve of the absolute value of the difference between the predicted value and the actual value. Fig. 9 , Fig.10It can be seen that the prediction model has a high accuracy for sunny and cloudy days, and the maximum errors compared with the measured power are 3.9MW and 1.4MW respectively. The error in the prediction of rainy weather is relatively large, which is mainly related to the rapid movement and thickness change of clouds, but the overall prediction trend of photovoltaic power generation is still accurate. The average prediction accuracy of three typical similar days is 93.4%, so the GA-BP neural network photovoltaic ultra-short-term output prediction model has a high prediction accuracy and can meet the actual needs of the project. In addition, adding satellite monitoring cloud meteorological data to the input of the model will further improve the accuracy of the model's prediction of various meteorological types.
[0136] In order to further measure the prediction error of the prediction model for different categories of similar day sets, the mean absolute percentage error MAPE (Mean Absolute Percentile Error), root mean square error RMSE (Root Mean Square Error) and absolute coefficient R 2 The prediction results are analyzed, where MAPE is used to analyze the quality of the model prediction results, RMSE is used to measure the deviation between the predicted value and the true value, and R 2 It is used to measure the degree to which the data fits the regression model. It is defined as:
[0137]
[0138] Among them, x i represents the actual value of photovoltaic power generation at time i, x i, , represents the predicted value of photovoltaic power generation at time i; N represents the total number of prediction samples; x mean Represents the average value of the true value of the test sample.
[0139] The smaller the MAPE and RMSE values are, the higher the prediction accuracy of the prediction model is. 2 The closer the value of is to 1, the better the fitting degree of the prediction model is. The specific prediction error index analysis results are shown in Table 1.
[0140] Table 1 Analysis of prediction error indicators
[0141]
[0142] As shown in Table 1, the GA-BP neural network prediction model has the best prediction performance for sunny similar days, with MAPE, RMSE, R 2 The corresponding values are 0.524, 0.552 and 0.992 respectively, which shows that the prediction effect is good and the prediction goal can be achieved well.
[0143] The present invention provides a subdivided grid-based photovoltaic output accurate prediction device, which includes the following modules:
[0144] Installation parameter information calculation module: The installation parameter information is the position of the sun relative to the observation device on the earth's surface, which is a necessary input required by the distributed photovoltaic power generation simulation model;
[0145] Photovoltaic effective irradiance calculation module: used to calculate global horizontal irradiance, direct normal irradiance and diffuse horizontal irradiance, which are used as input for distributed photovoltaic power generation simulation model;
[0146] Photovoltaic power generation calculation module: The present invention uses the De Soto model to analyze and simulate the performance of photovoltaic modules.
[0147] Similar day clustering module: Apply the fuzzy C-means clustering algorithm to classify the historical weather data of the forecast area according to meteorological conditions, screen similar day sets, and divide the normalized similar day sets into training sets and test sets;
[0148] Grid division module: divide the forecast area into grids to determine the forecast location, and form a high-dimensional input matrix based on the meteorological data of each grid;
[0149] GA-BP neural network photovoltaic ultra-short-term output prediction model module: used for accurate photovoltaic output prediction, which can be used for overall regional prediction or single grid prediction according to the prediction location and requirements.
[0150] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method for accurately predicting photovoltaic output at the subdivided grid level is implemented.
[0151] The present invention also provides a computer-readable storage medium, which includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the method for accurately predicting photovoltaic output at the subdivided grid level.
[0152] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the protection scope required by the present invention.
Claims
1. A method for accurately predicting photovoltaic output at the subdivided grid level, characterized in that: The following steps are involved: Step 1: construct a distributed photovoltaic power generation simulation model based on installation information; Step 2: Based on the photovoltaic power generation simulation model, the fuzzy C-means clustering algorithm is applied to classify the historical weather data of the forecast area according to meteorological conditions, screen similar day sets, and divide the normalized similar day sets into training sets and test sets; Step 3: Divide the forecast area into grids to determine the forecast location, and construct a high-dimensional input matrix based on the meteorological data of each grid; Step 4: Use the high-dimensional input matrix as the input of the GA-BP neural network photovoltaic ultra-short-term output prediction model, establish the GA-BP neural network photovoltaic ultra-short-term output prediction model, use the training set of similar day sets for training, and use the genetic algorithm to optimize the parameters to improve the generalization and accuracy of the model prediction; Step 5: Use the optimized GA-BP neural network photovoltaic ultra-short-term output prediction model to predict the photovoltaic ultra-short-term output of the test set of similar days, and make dynamic corrections based on the error between the actual output and the predicted output.
2. According to claim 1, a method for accurately predicting photovoltaic output at the subdivided grid level, characterized in that: The distributed photovoltaic power generation simulation model based on installation information in step 1 specifically includes the following three parts: The first part is the calculation of installation parameter information, which is the position of the sun relative to the observation device on the earth's surface and is a necessary input required by the distributed photovoltaic power generation simulation model; The second part is the photovoltaic effective irradiance calculation, which is used to calculate the global horizontal irradiance, direct normal irradiance and diffuse horizontal irradiance, which is used as the input of the distributed photovoltaic power generation simulation model; The third part is the calculation of photovoltaic power generation, using the De Soto model to analyze and simulate the performance of photovoltaic modules.
3. According to claim 1, a method for accurately predicting photovoltaic output at the subdivided grid level, characterized in that: The specific processing flow of the fuzzy C-means clustering algorithm analysis in step 2 is as follows: First, a sample data set X is established, and the fuzzy cluster center and the number of cluster centers are set. The membership matrix is updated by repeatedly adjusting the cluster center matrix. The optimal membership and cluster center are obtained based on the minimum sum of the Euclidean distances from each data sample to the cluster center. Finally, the sample points are clustered according to the membership. The specific formula is as follows: Among them, the data set X={x1,x2,…,x n },v j (t) is the fuzzy cluster center; c is the number of cluster centers; n is the number of sample data; u ij (t) is the membership degree of the i-th sample to the j-th class, u ij (t)∈[0,1]; d ij (t) is the Euclidean distance from sample i to cluster center j, d ij (t)=||c i -c j ||;d ik (t) is the Euclidean distance from sample i to cluster center k, d ik (t)=||c i -c k ||; t represents the time of the data sample, and m is a weight parameter that describes the degree of fuzzification.
4. According to the method of accurate prediction of photovoltaic output at the subdivided grid level according to claim 1, it is characterized in that: The similar day screening in step 2 specifically includes the following steps: 4-1) According to the previous forecast day weather condition identification, determine which weather category the forecast day belongs to, and select all the data that match this weather condition category during the forecast period to obtain the similar day set A. Using the Chebyshev method, the important meteorological characteristic parameters of the forecast day, solar irradiance C(t) and precipitation G(t), are compared with the solar irradiance C(t) of the mth day in the similar day set A. m (t), precipitation G m (t) Calculate the corresponding Chebyshev distance, the distance d m The expression is as follows: Among them, ρ C,m C(t) and C m (t) correlation coefficient of meteorological characteristic parameters; ρ G,m G(t) and G m (t) is the correlation coefficient of the meteorological characteristic parameters; D(C(t)) is the variance of the meteorological characteristic parameters of C(t); D(G(t)) is the variance of the meteorological characteristic parameters of G(t). 4-2) Using the Chebyshev method, we can find the Chebyshev distance between each day in the similar day set A and the predicted day, and get {d1, d2, …, d m }, set the ratio coefficient η to satisfy: The distance value sequence {d1, d2, …, d m } is less than or equal to (1+η)·min{d1,d2,…,d m All feature diaries of} are subset B in similar day set A, and subset B is defined as the similar day set.
5. The method for accurately predicting photovoltaic output at the subdivided grid level according to claim 1 is characterized in that: In step 3, the regional grid is divided as follows: Taking the distributed photovoltaic power station at the grid scale as the research object, the prediction area is divided into several independent small grids, each grid is composed of several distributed photovoltaic power stations. According to the area setting of the distributed photovoltaic power station cluster, the size of each grid in the area is about 10km×10km, and each grid includes 10 or more 10kV lines.
6. The method for accurately predicting photovoltaic output at the subdivided grid level according to claim 1 is characterized in that: In step 3, the input matrix uses the photovoltaic daily power curve within a single grid as the minimum analysis unit for research. The input of the matrix is the photovoltaic power generation power contained in the entire region, meteorological data such as solar irradiance, ambient temperature and wind speed. The output of the matrix is the photovoltaic power generation power of each grid on the day to be predicted. After the grid is divided, the photovoltaic power generation power curves of each grid with different meteorological characteristics are integrated together to perform ultra-short-term photovoltaic output forecasts under various weather conditions for the overall region.
7. The method for accurately predicting photovoltaic output at the subdivided grid level according to claim 6 is characterized in that: The specific steps of photovoltaic output prediction are as follows: Firstly, by screening the meteorological data of the entire prediction area, a similar day training set is obtained, and the corresponding relationship between the meteorological data and the grid output power is constructed and applied to the prediction of the day to be predicted; secondly, according to the prediction location and requirements, the overall regional prediction or the single grid prediction is selected for targeted prediction.
8. The method for accurately predicting photovoltaic output at the subdivided grid level according to claim 1 is characterized in that: The output layer activation function of the GA-BP neural network photovoltaic ultra-short-term output prediction model in step 4 is: E=SPη[1-0.005(t0+25)] Among them, E is the value of photovoltaic output; S is the value of photovoltaic collection panel area; P is the effective irradiance; t0 is the monitoring temperature, and η is the value of energy conversion efficiency.
9. A photovoltaic output accurate prediction device at the subdivided grid level, characterized in that: The device includes the following modules: Installation parameter information calculation module: used to observe the position of the sun relative to the earth's surface, which is a necessary input for the distributed photovoltaic power generation simulation model; Photovoltaic effective irradiance calculation module: used to calculate global horizontal irradiance, direct normal irradiance and diffuse horizontal irradiance, which are used as input for distributed photovoltaic power generation simulation model; Photovoltaic power calculation module: used to analyze and simulate the performance of photovoltaic modules; Similar day clustering module: used to classify the historical weather data of the forecast area according to meteorological conditions, filter similar day sets, and divide the normalized similar day sets into training sets and test sets; Grid division module: used to divide the forecast area into grids to determine the forecast location, and to form a high-dimensional input matrix based on the meteorological data of each grid; GA-BP neural network photovoltaic ultra-short-term output prediction model module: used to accurately predict photovoltaic output, and make overall regional predictions or single grid predictions based on the prediction location and requirements.
10. An electronic device, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, a method for accurately predicting photovoltaic output at the subdivided grid level according to any one of claims 1 to 8 is implemented.
11. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed, the device where the storage medium is located is controlled to execute a method for accurately predicting photovoltaic output at the subdivided grid level according to any one of claims 1 to 8.
Citation Information
Cited By
Intelligent key management system and method for transformer substation
CN119541092A
Photovoltaic module and system
CN120301298A
A photovoltaic module and system
CN120301298B
Photovoltaic power station irradiance prediction method and device based on multi-type data content identification, and medium
CN120430455A
Irradiance ultra-short-term forecasting method based on wind cloud 4B satellite
CN121211112A
