Photovoltaic power generation power prediction method based on physical mechanism and considering similar period machine learning

Abstract

The invention provides a photovoltaic generation power prediction method based on a physical mechanism and considering similar period machine learning, and the method comprises the following steps: employing a physical mechanism model, taking the meteorological data of a historical year as the input, and obtaining a photovoltaic generation power calculation value of a research region in a period t of the historical year; and training the machine learning model, and determining an optimal division mode of similar time periods in the year based on a sliding window method so as to divide the year into a plurality of continuous similar time periods, and determining the machine learning model corresponding to each similar time period for predicting the photovoltaic power generation power of the year. The invention provides a photovoltaic power generation power prediction method based on a physical mechanism and considering similar time period machine learning, in particular to a photovoltaic power generation power prediction method which takes grid meteorological data as input, is based on the physical mechanism and considers similar time period machine learning, and can effectively improve the accuracy of photovoltaic power generation power prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation prediction, and in particular relates to a photovoltaic power generation prediction method based on physical mechanism and machine learning considering similar time periods. Background Art

[0002] As the dominant clean and low-carbon energy, new energy such as wind power and photovoltaic power is an important measure to build a new power system and achieve emission reduction and carbon reduction. As of July 2024, the installed capacity of new energy will account for more than 38% of the total installed capacity of power generation, with a high penetration rate in the power system. Among them, the installed capacity of photovoltaic power generation has exceeded 730 million kilowatts, ranking first, and is the main force in the development of new energy. Photovoltaic power generation is characterized by randomness, volatility, and intermittency, which has a great impact on the safe and stable operation of the power system. At the same time, the formation of a "duck curve" directly affects the value and price of photovoltaic power generation in the power market. Therefore, more accurate photovoltaic power generation power prediction is of great practical significance. In practical applications, the accuracy of existing photovoltaic power generation power prediction methods still needs to be further improved. Summary of the invention

[0003] In view of the defects of the prior art, the present invention provides a photovoltaic power prediction method based on physical mechanism and machine learning considering similar time periods, which can effectively solve the above problems.

[0004] The technical solution adopted by the present invention is as follows:

[0005] The present invention provides a photovoltaic power generation prediction method based on physical mechanism and considering machine learning of similar time periods, comprising the following steps:

[0006] Step S1, obtaining the measured value of photovoltaic power generation and the meteorological data of each time period t in the study area in the historical year;

[0007] Step S2, using the physical mechanism model and taking the meteorological data of the historical year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the historical year;

[0008] Step S3, using the calculated value of photovoltaic power generation at each time period t of the historical year, combined with the meteorological data at each time period t as the feature quantity, and taking the measured value of photovoltaic power generation at each time period t of the historical year as the model target, the machine learning model is trained, and the best division method of similar time periods within the year is determined based on the sliding window method, so as to divide the year into multiple continuous similar time periods, and determine the machine learning model corresponding to each similar time period;

[0009] Step S4, using the physical mechanism model and taking the meteorological data of the forecast year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the forecast year;

[0010] Step S5, using the best division method of similar time periods within the year determined in step S3, dividing the forecast year into a plurality of continuous similar time periods;

[0011] For each similar time period, the corresponding photovoltaic power calculation value obtained in step S4 is input into the corresponding machine learning model determined in step S3 in combination with the meteorological data. The input photovoltaic power calculation value is corrected using the machine learning model, and the photovoltaic power prediction value corresponding to the similar time period is output; thereby completing the prediction of the photovoltaic power generation power for the predicted year.

[0012] Preferably, step S2 specifically includes:

[0013] Step S2.1, the physical mechanism model includes a photovoltaic module oblique radiation model and a photoelectric conversion model;

[0014] Step S2.2, the study area is in grid form, with multiple grids; obtain the meteorological data of each grid in the study area in each period t of the historical year, including: the total radiation received by the horizontal plane I, the direct radiation received by the horizontal plane I b and surface air temperature T a ;

[0015] Step S2.3, the horizontal plane receives the total radiation I and the horizontal plane receives the direct radiation I b As input, the photovoltaic module slope radiation model is run to output the total irradiance of the photovoltaic module slope I T ;

[0016] The slope radiation model of photovoltaic modules is:

[0017] I T =I b R b +I d R d +I T,g

[0018] in:

[0019] I b —The horizontal plane receives direct radiation;

[0020] R b —Correction factor for direct radiation from the horizontal plane; δ—declination; δ=23.45sin[(360×(284+n0) / 365×(π / 180)]×(π / 180); n0—the serial number of the day in the year; —latitude; β—inclination angle of the photovoltaic module, the optimal radiation angle for installing the photovoltaic module; γ—azimuth angle of the photovoltaic module; ω—hour angle; ω=[(λ-120) / 15+(0.258cosθ-7.416sinθ-3.648cos(2θ)-9.228sin(2θ)) / 60)+(t0–12)]×(π / 180); λ—longitude; θ—day angle; θ=2π(n0-1) / 365.242; t0—hour of the day at that time;

[0021] I d —The horizontal plane receives scattered radiation, I d =II b ;

[0022] R d —Correction factor for horizontal scattered radiation; R d =AR b +(1-A)[(1+cosβ) / 2)]; A—sky anisotropy index; A=I b / I0; I0—solar radiation at the top of the atmosphere; Correction coefficient for the distance between the sun and the earth, γ = 1 + 0.033cos[(360 × (284 + n0) / 365]; E sc —solar constant;

[0023] I T,g ——ground reflected radiation; I T,g =Iρ[(1-cosβ) / 2)]; I—total radiation received by the horizontal surface; ρ—surface albedo;

[0024] Step S2.4, taking the total irradiance of the photovoltaic module slope I T , surface air temperature T a As input, run the photoelectric conversion model and output the calculated photovoltaic power value P:

[0025] The photoelectric conversion model is:

[0026] P=I T ×η s ×[1-α(T c -25)]×S a ×K1×K2×K3×K4 / 1000

[0027] P—Calculated value of photovoltaic power generation;

[0028] η s —Photoelectric conversion efficiency under standard test conditions;

[0029] α—temperature coefficient;

[0030] T c =T a+[(T rated -20) / 800]×I T ; T a —Surface air temperature; T rated —Rated solar cell operating temperature;

[0031] T c —PV panel temperature;

[0032] S a —Effective area of ​​photovoltaic modules in power stations;

[0033] K1—PV module aging loss coefficient;

[0034] K1=1-k×y a ; k—the number of years the grid-connected photovoltaic power station has been put into use; y a —Annual attenuation rate of solar cell materials;

[0035] K2—PV module mismatch loss coefficient;

[0036] K3—dust shielding loss coefficient;

[0037] K4—DC circuit line loss coefficient.

[0038] Preferably, step S3 specifically comprises:

[0039] Step S3.1, integrating the calculated photovoltaic power generation value at each time period t of the historical year and the meteorological data at each time period t, to obtain a photovoltaic power generation characteristic data sequence at each time period t of the historical year; wherein each set of photovoltaic power generation characteristic data includes the calculated photovoltaic power generation value and meteorological data of the corresponding time period;

[0040] The sample length Ts is determined, so that the photovoltaic power generation characteristic data sequence of each period t in the historical year is divided into m subsequences in sequence, each subsequence forms a sample, and its length is the sample length Ts, thereby obtaining a training sample set S including m samples = {s1, s2, s3, …, s m};

[0041] Step S3.2, determine the sliding window scale set U = {u1,u2,u3,…,u n}, where n represents the total number of sliding window scales;

[0042] Step S3.3, traverse the sliding window scale set U = {u1,u2,u3,…,u n}, at each sliding window scale, from the training sample set S = {s1, s2, s3, …, s m} to obtain m sample units, each sample unit being a sequence of multiple continuous samples;

[0043] Therefore, for n sliding window scales, a total of m*n sample units are obtained. Let N = m*n, and the N sample units are represented as: sample units d1, d2, d3, …, d N ;

[0044] Step S3.4, building a machine learning model XGBoost;

[0045] Take each sample unit d k , k = 1, 2, 3, ..., N, train the machine learning model XGBoost, take the photovoltaic power measured value sequence of the corresponding period as the model target, determine the model parameters, and thus obtain the sample unit d k The machine learning model XGBoost obtained after training k , and obtain the machine learning model XGBoost k The coefficient of determination

[0046] Step S3.5, according to each sample unit d k Corresponding machine learning model XGBoost k The coefficient of determination Construct a mathematical optimization model and solve it, so that in the sample units d1, d2, d3, …, d N In the process, the best sample unit combination is found Among them, M represents the number of sample units in the best sample unit combination, so that the best sample unit combination Satisfies the following conditions: Contains training sample set S = {s1, s2, s3, ..., s m}, and there are no duplicate samples between any two sample units, and the sum of the determination coefficients of the machine learning models of all sample units in the best sample unit combination reaches the maximum value;

[0047] Step S3.6, based on the best sample unit combination Determine the best way to divide similar time periods within a year, thereby dividing the year into multiple consecutive similar time periods, and determine the machine learning model corresponding to each similar time period;

[0048] Specifically:

[0049] Best sample unit combination In each sample unit The corresponding time period within the year is a similar time period Thus, the year is divided into M similar time periods, expressed as: Each similar time period The corresponding machine learning model is to use the corresponding sample unit The machine learning model XGBoost obtained after training r .

[0050] Preferably, in step S3.1, the sample length Ts is a month or a week.

[0051] Preferably, the meteorological data used in step S3.1 include irradiance, temperature and rainfall data on a monthly, daily, hourly and hourly scale.

[0052] Preferably, in step S3.3, at each sliding window scale, a training sample set S = {s1, s2, s3, ..., s m} to obtain m sample units, specifically:

[0053] Step S3.3.1, when the sliding window size u is used i When i = 1, 2, 3, ... n, for the training sample set S = {s1, s2, s3, ..., s m} to supplement and obtain the supplemented training sample set in,

[0054] Step S3.3.2, using a sliding window size of u i The sliding window, with a step size of 1, is in the supplemented training sample set Slide from sample s1 until you reach sample s m , thus obtaining m sample units corresponding to the sliding window of m positions; each sample unit has u i A continuous sample.

[0055] Preferably, in step S3.5, the mathematical optimization model is:

[0056] The optimization objective F of the mathematical optimization model is:

[0057]

[0058] The constraints of the mathematical optimization model are:

[0059]

[0060] in: The best sample unit combination Medium Sample Unit The machine learning model obtained after training; For machine learning models The coefficient of determination.

[0061] Preferably, the mathematical optimization model is solved using a 0-1 integer programming method.

[0062] The photovoltaic power generation prediction method provided by the present invention based on physical mechanism and considering machine learning of similar time periods has the following advantages:

[0063] The present invention provides a photovoltaic power prediction method based on physical mechanisms and machine learning considering similar time periods. Specifically, the present invention provides a photovoltaic power prediction method based on physical mechanisms and machine learning considering similar time periods, which can effectively improve the accuracy of photovoltaic power prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic flow chart of a photovoltaic power generation prediction method based on physical mechanism and machine learning considering similar time periods provided by the present invention;

[0065] Figure 2 A schematic diagram of the principle of the "similar time period" sliding optimization sample unit combination provided by the present invention;

[0066] Figure 3 A structural diagram of the preferred "similar time period" according to the proposed principle provided by the present invention;

[0067] Figure 4 The prediction effect of the test set of five "similar time periods" throughout the year provided by the present invention is demonstrated; among them, the left is the calculation result of the physical mechanism model, and the right is the result after machine learning optimization. DETAILED DESCRIPTION

[0068] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0069] The present invention provides a photovoltaic power prediction method based on physical mechanisms and machine learning considering similar time periods. Specifically, the present invention provides a photovoltaic power prediction method based on physical mechanisms and machine learning considering similar time periods, which can effectively improve the accuracy of photovoltaic power prediction.

[0070] See also Figure 1 The present invention provides a photovoltaic power prediction method based on physical mechanism and considering machine learning of similar time periods, comprising the following steps:

[0071] Step S1, obtaining the measured value of photovoltaic power generation and the meteorological data of each time period t in the study area in the historical year;

[0072] Step S2, using the physical mechanism model and taking the meteorological data of the historical year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the historical year;

[0073] Step S3, using the calculated value of photovoltaic power generation at each time period t of the historical year, combined with the meteorological data at each time period t as the feature quantity, and taking the measured value of photovoltaic power generation at each time period t of the historical year as the model target, the machine learning model is trained, and the best division method of similar time periods within the year is determined based on the sliding window method, so as to divide the year into multiple continuous similar time periods, and determine the machine learning model corresponding to each similar time period;

[0074] Step S4, using the physical mechanism model and taking the meteorological data of the forecast year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the forecast year;

[0075] Step S5, using the best division method of similar time periods within the year determined in step S3, dividing the forecast year into a plurality of continuous similar time periods;

[0076] For each similar time period, the corresponding photovoltaic power calculation value obtained in step S4 is input into the corresponding machine learning model determined in step S3 in combination with the meteorological data. The input photovoltaic power calculation value is corrected using the machine learning model, and the photovoltaic power prediction value corresponding to the similar time period is output; thereby completing the prediction of the photovoltaic power generation power for the predicted year.

[0077] The following is a detailed description of steps S1 to S5:

[0078] Step S1, obtaining the measured value of photovoltaic power generation and the meteorological data of each time period t in the study area in the historical year;

[0079] In the specific implementation, time-space matching and data preprocessing are first performed:

[0080] (1) Spatial matching:

[0081] Taking the historical year 2022 as an example, the actual measured values ​​of photovoltaic power generation power of photovoltaic power stations every 15 minutes throughout 2022 are collected. For the photovoltaic power station site, the hourly meteorological data of the meteorological grid at the corresponding location in the historical year are obtained, including the surface temperature T a , total radiation received by the horizontal plane I, direct radiation received by the horizontal plane I b and precipitation data.

[0082] (2) Time matching:

[0083] Generally, the gridded meteorological data is in hourly steps, and the power station provides the measured photovoltaic power generation every 15 minutes. Therefore, the photovoltaic power generation with small steps is weighted and counted to large steps, that is, the measured photovoltaic power generation value for each hour of the historical year is obtained by weighted average.

[0084] (3) Delete all hourly data on the day with missing measurements.

[0085] Step S2, using the physical mechanism model and taking the meteorological data of the historical year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the historical year;

[0086] Step S2 is specifically as follows:

[0087] Step S2.1, the physical mechanism model includes a photovoltaic module oblique radiation model and a photoelectric conversion model;

[0088] Step S2.2, the study area is in grid form; obtain the meteorological data of the grid where the station site in the study area is located for each time period t in the historical year, including: the total radiation received by the horizontal plane I, the direct radiation received by the horizontal plane I b and surface air temperature T a ;

[0089] Step S2.3, the horizontal plane receives the total radiation I and the horizontal plane receives the direct radiation I b As input, the photovoltaic module slope radiation model is run to output the total irradiance of the photovoltaic module slope I T ;

[0090] The slope radiation model of photovoltaic modules is:

[0091] I T =I b R b +I d R d +I T,g

[0092] in:

[0093] I b —Horizontal plane receives direct radiation, unit: W / m 2 ;

[0094] R b —Correction factor for direct radiation from the horizontal plane; δ—declination, unit: rad; δ=23.45sin[(360×(284+n0) / 365×(π / 180)]×(π / 180); n0—the serial number of the day in the year; —latitude, unit: rad; β—inclination angle of photovoltaic module, the best radiation angle for photovoltaic module installation, unit: rad; γ—azimuth angle of photovoltaic module, unit: rad; ω—hour angle, unit: rad; ω=[(λ-120) / 15+(0.258cosθ-7.416sinθ-3.648cos(2θ)-9.228sin(2θ)) / 60)+(t0–12)]×(π / 180); λ—longitude, unit: °; θ—day angle, unit: rad; θ=2π(n0-1) / 365.242; t0—hour of the day at that time;

[0095] I d —The horizontal plane receives scattered radiation, I d =II b , unit: W / m 2 ;

[0096] R d —Correction factor for horizontal scattered radiation; R d =AR b +(1-A)[(1+cosβ) / 2)]; A—sky anisotropy index; A=I b / I0; I0—solar radiation at the top of the atmosphere, unit: W / m 2 ; γ—Sun-Earth distance correction coefficient, γ=1+0.033cos[(360×(284+n0) / 365]; E sc —Solar constant, E sc =1367W / m 2 ;

[0097] I T,g ——Ground reflected radiation, unit: W / m 2 ;I T,g =Iρ[(1-cosβ) / 2)]; I—total radiation received by the horizontal plane, unit: W / m 2 ; ρ—surface albedo, can be 0.2-0.25;

[0098] Therefore, the slope radiation model of photovoltaic modules is used to comprehensively consider the horizontal surface receiving direct radiation I b And the horizontal plane direct radiation correction factor R b , Horizontal plane receives scattered radiation I d and horizontal scattered radiation correction factor R d 、Ground reflected radiation I T,g , the total irradiance of the photovoltaic module slope is corrected to obtain a more accurate total irradiance of the photovoltaic module slope I T .

[0099] Step S2.4, taking the total irradiance of the photovoltaic module slope I T , surface air temperature T a As input, run the photoelectric conversion model and output the calculated photovoltaic power value P:

[0100] The photoelectric conversion model is:

[0101] P=I T ×η s ×[1-α(T c -25)]×S a ×K1×K2×K3×K4 / 1000

[0102] P—Calculated value of photovoltaic power generation, unit: kW;

[0103] η s —Photoelectric conversion efficiency under standard test conditions;

[0104] α—temperature coefficient, unit: ℃ -1 ;

[0105] T c =T a +[(T rated -20) / 800]×I T ; T a —Surface air temperature, unit: °C; T rated —Rated solar cell operating temperature, unit: °C;

[0106] T c —PV module temperature, unit: °C;

[0107] S a —Effective area of ​​photovoltaic modules in power station, unit: m 2 ;

[0108] K1—PV module aging loss coefficient;

[0109] K1=1-k×y a ; k—the number of years the grid-connected photovoltaic power station has been put into use; y a —Annual attenuation rate of solar cell materials;

[0110] K2—PV module mismatch loss coefficient;

[0111] K3—dust shielding loss coefficient;

[0112] K4—DC circuit line loss coefficient.

[0113] Therefore, through the photoelectric conversion model, the total irradiance of the photovoltaic module slope I T , surface air temperature T aAs input, the photovoltaic conversion model considering the effects of photovoltaic conversion efficiency, dust shielding, DC circuit line loss, etc. can obtain a more accurate calculation value of photovoltaic power generation.

[0114] Step S3, using the calculated value of photovoltaic power generation at each time period t of the historical year, combined with the meteorological data at each time period t as the feature quantity, and taking the measured value of photovoltaic power generation at each time period t of the historical year as the model target, the machine learning model is trained, and the best division method of similar time periods within the year is determined based on the sliding window method, so as to divide the year into multiple continuous similar time periods, and determine the machine learning model corresponding to each similar time period;

[0115] Step S3 is specifically as follows:

[0116] Step S3.1, merge the calculated photovoltaic power generation power at each time period t of the historical year and the meteorological data at each time period t to obtain a photovoltaic power generation characteristic data sequence at each time period t of the historical year; wherein each set of photovoltaic power generation characteristic data includes the calculated photovoltaic power generation power and meteorological data of the corresponding time period; the meteorological data here include irradiance, temperature and rainfall data at monthly, daily, hourly and hourly scales.

[0117] Determine the sample length Ts, for example, the sample length Ts is a month or a week, so that the photovoltaic power generation characteristic data sequence of each period t of the historical year is sequentially divided into m subsequences, each subsequence forms a sample, and its length is the sample length Ts, thereby obtaining a training sample set S including m samples = {s1, s2, s3, ..., s m};

[0118] Step S3.2, determine the sliding window scale set U = {u1,u2,u3,…,u n}, where n represents the total number of sliding window scales;

[0119] For example, two sliding window scales are set, namely, 2-month and 3-month scales. The present invention does not limit the set sliding window scale and the type of sliding window scale, which can be flexibly set according to actual needs.

[0120] Step S3.3, traverse the sliding window scale set U = {u1,u2,u3,…,u n}, at each sliding window scale, from the training sample set S = {s1, s2, s3, …, s m} to obtain m sample units, each sample unit being a sequence of multiple continuous samples;

[0121] In this step, at each sliding window scale, the training sample set S = {s1, s2, s3, ..., s m} to obtain m sample units, specifically:

[0122] Step S3.3.1, when the sliding window size u is used i When i = 1, 2, 3, ... n, for the training sample set S = {s1, s2, s3, ..., s m} to supplement and obtain the supplemented training sample set in,

[0123] Step S3.3.2, using a sliding window size of u i The sliding window, with a step size of 1, is in the supplemented training sample set Slide from sample s1 until you reach sample s m , thus obtaining m sample units corresponding to the sliding window of m positions; each sample unit has u i A continuous sample.

[0124] Therefore, for n sliding window scales, a total of m*n sample units are obtained. Let N = m*n, and the N sample units are represented as: sample units d1, d2, d3, …, d N ;

[0125] Step S3.4, constructing a machine learning model XGBoost; XGBoost is an extreme gradient boosting machine learning model;

[0126] Take each sample unit d k , k = 1, 2, 3, ..., N, train the machine learning model XGBoost, take the photovoltaic power measured value sequence of the corresponding period as the model target, determine the model parameters, and thus obtain the sample unit d k The machine learning model XGBoost obtained after training k , and obtain the machine learning model XGBoost k The coefficient of determination

[0127] In this step, the sample unit d k When training the machine learning model XGBoost, the sample unit d k The calculated photovoltaic power generation value and meteorological data of the corresponding period are included in the data. The meteorological data here are different from the meteorological data input into the physical mechanism model. In addition to time factors such as month, day, and hour, and dominant factors such as total radiation and direct radiation, it also includes key factors such as hourly rainfall as feature quantities, which can more effectively train the machine learning model XGBoost. By training the machine learning model XGBoost and optimizing the parameters, the sample unit d is obtained. kXGBoost machine learning model after training k The measured value sequence of photovoltaic power generation is used as the model target to determine the machine learning model XGBoost. k The coefficient of determination Used to evaluate the machine learning model XGBoost k The accuracy of The larger the value, the higher the accuracy.

[0128] for Figure 2 In the illustrated embodiment, after obtaining 24 sample units consisting of two sliding window scales, each sample unit obtains a corresponding machine learning model and a determination coefficient of the machine learning model.

[0129] Step S3.5, according to each sample unit d k Corresponding machine learning model XGBoost k The coefficient of determination Construct a mathematical optimization model and solve it, so that in the sample units d1, d2, d3, …, d N In the process, the best sample unit combination is found Among them, M represents the number of sample units in the best sample unit combination, so that the best sample unit combination Satisfies the following conditions: Contains training sample set S = {s1, s2, s3, ..., s m}, and there are no duplicate samples between any two sample units, and the sum of the determination coefficients of the machine learning models of all sample units in the best sample unit combination reaches the maximum value;

[0130] In this step, the mathematical optimization model constructed is:

[0131] The optimization objective F of the mathematical optimization model is:

[0132]

[0133] The constraints of the mathematical optimization model are:

[0134]

[0135] in: The best sample unit combination Medium Sample Unit The machine learning model obtained after training; For machine learning models The coefficient of determination.

[0136] The mathematical optimization model is solved by using 0-1 integer programming method.

[0137] Step S3.6, based on the best sample unit combination Determine the best way to divide similar time periods within a year, thereby dividing the year into multiple consecutive similar time periods, and determine the machine learning model corresponding to each similar time period;

[0138] Specifically:

[0139] Best sample unit combination In each sample unit The corresponding time period within the year is a similar time period Thus, the year is divided into M similar time periods, expressed as: Each similar time period The corresponding machine learning model is to use the corresponding sample unit The machine learning model obtained after training

[0140] Step S4, using the physical mechanism model and taking the meteorological data of the forecast year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the forecast year;

[0141] The physical mechanism model of this step is the same as the physical mechanism model used in step S2, with the only difference being that: in step S2, the meteorological data of the historical year is input into the physical mechanism model, and the calculated value of the photovoltaic power generation at each time period t of the historical year is output; whereas in this step S4, the meteorological data of the predicted year is input into the physical mechanism model, and the calculated value of the photovoltaic power generation at each time period t of the predicted year is output.

[0142] Step S5, using the best division method of similar time periods within the year determined in step S3, dividing the forecast year into a plurality of continuous similar time periods;

[0143] For each similar time period, the corresponding photovoltaic power calculation value obtained in step S4 is input into the corresponding machine learning model determined in step S3 in combination with the meteorological data. The input photovoltaic power calculation value is corrected using the machine learning model, and the photovoltaic power prediction value corresponding to the similar time period is output; thereby completing the prediction of the photovoltaic power generation power for the predicted year.

[0144] In this step, the input meteorological data is still the predicted irradiance, temperature and rainfall data on a monthly, daily, hourly and hourly scale for the corresponding period.

[0145] The present invention has the following characteristics:

[0146] 1. First, the use of distributed grid meteorological data input can reflect the impact of meteorological factors such as solar radiation and temperature at the location of the photovoltaic station.

[0147] 2. Second, the physical mechanism model fully considers the direct radiation, scattered radiation, reflected radiation actually received by the slope of the photovoltaic module, as well as the inverter efficiency, photovoltaic cell conversion rate and various attenuations and losses in the photoelectric conversion. Therefore, the calculated photovoltaic power generation value is more accurate, with an accuracy of R 2 Can reach above 0.8.

[0148] 3. Third, after the introduction of machine learning XGBoost, in addition to time series elements such as month, day, and time, and dominant factors of photovoltaic power generation such as radiation, hourly precipitation is also used as a feature quantity, which can reflect the key factor of local weather influence, which is conducive to further improving the calculation results of the physical mechanism model.

[0149] 4. Fourth, the proposed “similar time period” division method can effectively improve the applicability of the machine learning XGBoost model. This is mainly done by using machine learning training and testing for 24 groups of “sample units” from January to December, 2 months at a time and 3 months at a time, and taking into account the parameter optimization in machine learning, and traversing 24 groups of calculation results evaluation indicators R 2 , and the optimal grouping method, namely “similar time periods” and its machine learning model, is obtained.

[0150] 5. Fifth, for the intended prediction period, the physical mechanism model and the corresponding machine learning model of the "similar time period" are used for prediction. Since the "similar time period" is a period that is identified to have similar clustering characteristics, the use of targeted machine learning models for different "similar time periods" is more adaptable. Therefore, the prediction effect will be further improved.

[0151] Combine the following Figure 2 , introduce a specific embodiment:

[0152] The implementation case selected a photovoltaic power station in Qinghai Province with an installed capacity of 8MW. The method flow is as follows:

[0153] Step S1, obtaining and statistically calculating the hourly measured photovoltaic power generation power and hourly meteorological data of the study area in the historical year 2022;

[0154] Step S2, using the physical mechanism model and taking the meteorological data of the historical year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the historical year;

[0155] Step S3, using the calculated value of photovoltaic power generation at each time period t of the historical year, combined with the meteorological data at each time period t as the feature quantity, and taking the measured value of photovoltaic power generation at each time period t of the historical year as the model target, the machine learning model is trained, and the best division method of similar time periods within the year is determined based on the sliding window method, so as to divide the year into multiple continuous similar time periods, and determine the machine learning model corresponding to each similar time period;

[0156] Step S3.1, integrating the calculated photovoltaic power generation value and the hourly meteorological data of the historical year to obtain the photovoltaic power generation characteristic data sequence of the historical year;

[0157] The sample length Ts is determined to be months. Therefore, for the collected hourly photovoltaic power generation characteristic data sequence of the photovoltaic power station throughout 2022, the month is used as the sample length Ts, so that the hourly photovoltaic power generation characteristic data sequence of the photovoltaic power station throughout 2022 is divided into 12 subsequences, namely the subsequence corresponding to January, the subsequence corresponding to February, …, the subsequence corresponding to December. The training sample set S thus obtained has 12 samples, so S = {s1, s2, s3, …, s 12}, respectively, samples in January, samples in February, …, samples in December;

[0158] Step S3.2, setting two sliding window scales, 2 months and 3 months respectively;

[0159] Step S3.3, such as Figure 2 As shown, for the monthly samples processed by the photovoltaic power station in 2022, the training sample set S = {s1, s2, s3, ..., s 12}; When the sliding window size is 2 months, 12 sample units are obtained, namely: a sample unit formed by sample s1 and sample s2, a sample unit formed by sample s2 and sample s3, and so on, until sample s 12 and sample s1 form a sample unit;

[0160] When the sliding window size is 3 months, 12 sample units are obtained, namely: a sample unit formed by sample s1, sample s2 and sample s3, a sample unit formed by sample s2, sample s3 and sample s4, and so on, until sample s 12 , sample s1 and sample s2 form a sample unit.

[0161] Thus, a total of 24 sample units consisting of two sliding window scales are obtained, which are represented as sample units d1, d2, d3, …, d 24 .

[0162] Step S3.4, using each sample unit d k , k=1,2,3,…,24, train the machine learning model XGBoost, take the measured value sequence of photovoltaic power generation in the corresponding period as the model target, determine the model parameters, and obtain the sample unit d k The machine learning model XGBoost obtained after training k , and obtain the machine learning model XGBoost k The coefficient of determination

[0163] Step S3.5, construct a mathematical optimization model and solve it.

[0164] Figure 2 In the example, when 24 sample units are obtained, corresponding to 24 machine learning models and determination coefficients, the mathematical optimization model of the present invention is used and solved, and finally the best sample unit combination is found from the 24 sample units: There are 5 sample units in total, namely:

[0165] Sample unit formed by samples s1 and s2

[0166] Sample unit formed by samples s3, s4 and s5

[0167] Sample unit formed by samples s6, s7 and s8

[0168] Sample s9, sample s 10 The sample unit formed

[0169] Samples 11 , samples 12 The sample unit formed

[0170] Specifically, when determining each sample unit in the optimal sample unit combination, each sample in each sample unit is considered to have a higher clustering characteristic, and from the physical gradient law, it is a sample of consecutive months; using the same machine learning model in each sample unit can obtain better results than other sample unit combinations (R 2 According to the following principles, determine the best sample unit combination:

[0171] Principle 1: The timing of each sample unit must be continuous;

[0172] Principle 2: Evaluation index R of the annual “sample unit” 2 The sum should be as large as possible;

[0173] Principle 3 “Sample Unit” Evaluation Index R 2 It is better than not dividing the "sample unit", that is, the average index R when the 12 months of the year are calculated uniformly. 2 .

[0174] Step S3.6, based on the best sample unit combination Determine the best way to divide similar time periods within a year, thereby dividing the year into multiple consecutive similar time periods, and determine the machine learning model corresponding to each similar time period;

[0175] Therefore, the whole year of 2022 is divided into five consecutive similar time periods, namely: January and February are the first similar time period, March to May are the second similar time period, June to August are the third similar time period, September and October are the fourth similar time period, and November and December are the fifth similar time period.

[0176] As a control, when the whole year of 2022 is not divided into multiple similar time periods, the samples of the whole year of 2022 are divided into training set and test set; only one machine learning model XGBoost is used for training to obtain the trained machine learning model XGBoost; the test set is used to test the trained machine learning model XGBoost, and the results are as follows: Figure 3 As shown, the solid line is the index R of the trained machine learning model XGBoost 2 .

[0177] When the whole year of 2022 is divided into 5 consecutive similar time periods according to the method of the present invention, the machine learning model XGBoost is trained respectively, and its index R 2 , the results are as follows Figure 3 As shown, the dotted line is the index R of the machine learning model XGBoost in each similar time period 2 .

[0178] right Figure 3 The results are analyzed and the following conclusions can be drawn: According to the division method of the present invention, the R values ​​of the machine learning model XGBoost corresponding to the five similar time periods are 2 The average value is 0.9185, and the annual R is not divided into similar time periods throughout the year. 2 The R obtained by the method of the present invention is 0.9073. 2 The average value is higher, which shows that the overall performance of the machine learning model obtained by the method of the present invention is better than the performance of the machine learning model that is not divided into similar time periods.

[0179] Step S4, using the physical mechanism model and taking the meteorological data of the forecast year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the forecast year;

[0180] Step S5, using the best division method of similar time periods within the year determined in step S3, dividing the forecast year into five consecutive similar time periods;

[0181] For each similar time period, the corresponding photovoltaic power calculation value obtained in step S4 is input into the corresponding machine learning model determined in step S3 in combination with the meteorological data. The input photovoltaic power calculation value is corrected using the machine learning model, and the photovoltaic power prediction value corresponding to the similar time period is output; thereby completing the prediction of the photovoltaic power generation power for the predicted year.

[0182] In step S3, the best sample unit combination is: There are 5 sample units in total, namely: and The corresponding trained machine learning model is expressed as:

[0183] Assuming that it is necessary to predict the photovoltaic power generation in 2025, after obtaining the hourly photovoltaic power generation calculation value in 2025 using the physical mechanism model, the whole year of 2025 is divided into five similar time periods, namely: January and February are the first similar time period, March to May are the second similar time period, June to August are the third similar time period, September and October are the fourth similar time period, and November and December are the fifth similar time period;

[0184] The calculated photovoltaic power generation values ​​for January and February 2025 are combined with meteorological data and input into the machine learning model In the calculation, the calculated values ​​of photovoltaic power generation in January and February 2025 are corrected, and the predicted values ​​of photovoltaic power generation in January and February 2025 are output;

[0185] Similarly, the calculated photovoltaic power generation values ​​for March, April and May 2025 are combined with meteorological data and input into the machine learning model. In the calculation, the calculated values ​​of photovoltaic power generation in March, April and May 2025 are corrected, and the predicted values ​​of photovoltaic power generation in March, April and May 2025 are output;

[0186] The calculated photovoltaic power generation values ​​for June, July and August 2025 are combined with meteorological data and input into the machine learning model In the calculation, the calculated values ​​of photovoltaic power generation in June, July and August 2025 are corrected, and the predicted values ​​of photovoltaic power generation in June, July and August 2025 are output;

[0187] The calculated photovoltaic power generation values ​​for September and October 2025 are combined with meteorological data and input into the machine learning model In the calculation, the calculated values ​​of photovoltaic power generation in September and October 2025 are corrected, and the predicted values ​​of photovoltaic power generation in September and October 2025 are output;

[0188] The calculated photovoltaic power generation values ​​for November and December 2025 are combined with meteorological data and input into the machine learning model In the calculation, the calculated values ​​of photovoltaic power generation in November and December 2025 are corrected, and the predicted values ​​of photovoltaic power generation in November and December 2025 are output.

[0189] This results in hourly forecasts of photovoltaic power generation throughout 2025.

[0190] As a verification, the 2022 meteorological data is used as input. Figure 4 The prediction results of the five "similar time periods" test sets throughout the year provided for this embodiment are shown in the figure; the left one is the calculation result of the physical mechanism model, and the right one is the result after machine learning optimization. Figure 4 It can be seen that the photovoltaic power generation value after machine learning optimization is closer to the actual value of photovoltaic power generation, thereby realizing the correction of the calculation results of the physical mechanism model and improving the accuracy of photovoltaic power generation prediction.

[0191] The photovoltaic power generation prediction method provided by the present invention based on physical mechanism and considering machine learning of similar time periods has the following characteristics:

[0192] (1) The present invention proposes a physical mechanism model that uses meteorological data as input to obtain a relatively accurate photovoltaic power calculation value;

[0193] Specifically, grid meteorological data, such as the total horizontal irradiance and direct horizontal irradiance of the reanalysis grid (31km×31km) as input, can reflect the influence of meteorological factors such as solar radiation and temperature at the location of the photovoltaic station. The slope radiation received by the photovoltaic module takes into account the direct horizontal radiation and slope correction, scattered radiation and slope correction, and ground reflected radiation affected by the slope. Photovoltaic efficiency conversion takes into account factors such as the difference between the air temperature and the rated operating temperature of the solar cell, the temperature of the photovoltaic module, the photoelectric conversion efficiency, the dust shielding loss, and the line loss. Therefore, a more accurate calculation value of photovoltaic power generation can be obtained.

[0194] (2) The present invention proposes an optimal method for dividing similar time periods within a year based on a sliding window method. When the machine learning model XGBoost is trained using sample units corresponding to each similar time period divided in this way, multiple machine learning models XGBoost with the best overall determination coefficient can be obtained, so that the machine learning model XGBoost can more accurately correct the calculated value of photovoltaic power generation output by the physical mechanism model, thereby obtaining a more accurate photovoltaic power generation prediction value.

[0195] Specifically, in order to improve the prediction accuracy of the machine learning model XGBoost, the "similar time period" division is very important for the classification construction of the machine learning model XGBoost. The present invention proposes that, for example, by sliding monthly, by 2 months and by 3 months, 24 groups of "sample units" are formed, and on the basis of the machine learning XGBoost, 24 evaluation indicators of the machine learning model XGBoost, namely, the determination coefficient R 2 , and then select the annual comprehensive R 2 The most effective combination is to divide the whole year into "similar time periods". For each "similar time period", an adaptive machine learning model is trained to predict the same period of the year.

[0196] In addition, when using sample units to train the machine learning model XGBoost, the sample units not only have the calculated value of photovoltaic power generation, but also have meteorological characteristics. The meteorological characteristics take into account time factors such as month, day, and hour, dominant factors such as total radiation and direct radiation, and key factors such as hourly rainfall, thereby further improving the training effect of the machine learning model XGBoost. In particular, the introduction of hourly rainfall characteristics fully considers the impact of local instantaneous weather on the basis of dominant factors such as radiation changes over time, and then corrects the radiation model results, which is very important for photovoltaic power generation.

[0197] Therefore, the present invention uses machine learning XGBoost (extreme gradient boosting method) to further correct the results based on the use of physical mechanism models to calculate photovoltaic power generation. Using the XGBoost method, through the optimization of the "sample unit" combination, a "similar time period" is proposed. According to the division of the "similar time period", different sample unit combinations are respectively machine-learned and optimized to obtain the machine learning model XGBoost of each "similar time period". For the time period that needs to be predicted, the machine learning model XGBoost of the corresponding time period is used for correction to improve the prediction accuracy.

[0198] The present invention aims at the limited basic data of photovoltaic power generation simulation prediction and the need to improve accuracy. It provides a photovoltaic power prediction method based on gridded reanalysis meteorological data or numerical weather forecast data as input, and based on physical mechanism model and machine learning XGBoost correction in "similar time period". Not only can gridded meteorological data be used to make up for the lack of measured meteorological data at the site, but also photovoltaic power can be directly calculated through horizontal plane radiation, photovoltaic module inclined plane radiation correction, and photoelectric conversion series model, and "similar time period" is selected through sliding optimization, and the physical mechanism model results are corrected by machine learning algorithm in "similar time period" to improve the prediction accuracy. The construction of the method of the present invention is not limited by basic data conditions, including physical mechanism and machine learning mode, and the results of photovoltaic power prediction are calculated and improved. It has strong adaptability, low application threshold, and can be widely promoted and applied.

[0199] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A photovoltaic power generation prediction method based on physical mechanism and machine learning considering similar time periods, characterized in that: The following steps are involved: Step S1, obtaining the measured value of photovoltaic power generation and the meteorological data of each time period t in the study area in the historical year; Step S2, using the physical mechanism model and taking the meteorological data of the historical year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the historical year; Step S3, using the calculated value of photovoltaic power generation at each time period t of the historical year, combined with the meteorological data at each time period t as the feature quantity, and taking the measured value of photovoltaic power generation at each time period t of the historical year as the model target, the machine learning model is trained, and the best division method of similar time periods within the year is determined based on the sliding window method, so as to divide the year into multiple continuous similar time periods, and determine the machine learning model corresponding to each similar time period; Step S4, using the physical mechanism model and taking the meteorological data of the forecast year as input, obtaining the calculated value of photovoltaic power generation in each time period t of the study area in the forecast year; Step S5, using the best division method of similar time periods within the year determined in step S3, dividing the forecast year into a plurality of continuous similar time periods; For each similar time period, the corresponding photovoltaic power calculation value obtained in step S4 is input into the corresponding machine learning model determined in step S3 in combination with the meteorological data. The input photovoltaic power calculation value is corrected using the machine learning model, and the photovoltaic power prediction value corresponding to the similar time period is output; thereby completing the prediction of the photovoltaic power generation power for the predicted year.

2. A photovoltaic power prediction method based on physical mechanism and considering machine learning of similar time periods according to claim 1, characterized in that: Step S2 specifically includes: Step S2.1, the physical mechanism model includes a photovoltaic module oblique radiation model and a photoelectric conversion model; Step S2.2, the study area is in grid form; obtain the meteorological data of the grid where the station site in the study area is located for each time period t in the historical year, including: the total radiation received by the horizontal plane I, the direct radiation received by the horizontal plane I b and surface air temperature T a ; Step S2.3, the horizontal plane receives the total radiation I and the horizontal plane receives the direct radiation I b As input, the photovoltaic module slope radiation model is run to output the total irradiance of the photovoltaic module slope I T ; The slope radiation model of photovoltaic modules is: I T =I b R b +I d R d +I T,g in: I b —The horizontal plane receives direct radiation; R b —Correction factor for direct radiation from the horizontal plane; δ—declination; δ=23.45sin[(360×(284+n0) / 365×(π / 180)]×(π / 180); n0—the serial number of the day in the year; —latitude; β—inclination angle of the photovoltaic module, the optimal radiation angle for installing the photovoltaic module; γ—azimuth angle of the photovoltaic module; ω—hour angle; ω=[(λ-120) / 15+(0.258cosθ-7.416sinθ-3.648cos(2θ)-9.228sin(2θ)) / 60)+(t0–12)]×(π / 180); λ—longitude; θ—day angle; θ=2π(n0-1) / 365.242; t0—hour of the day at that time; I d —The horizontal plane receives scattered radiation, I d =II b ; R d —Correction factor for horizontal scattered radiation; R d =AR b +(1-A)[(1+cosβ) / 2)]; A—sky anisotropy index; A=I b / I0; I0—solar radiation at the top of the atmosphere; γ—Sun-Earth distance correction coefficient, γ=1+0.033cos[(360×(284+n0) / 365]; E sc —solar constant; I T,g ——ground reflected radiation; I T,g =Iρ[(1-cosβ) / 2)]; I—total radiation received by the horizontal surface; ρ—surface albedo; Step S2.4, taking the total irradiance of the photovoltaic module slope I T , surface air temperature T a As input, run the photoelectric conversion model and output the calculated photovoltaic power value P: The photoelectric conversion model is: P=I T ×η s ×[1-α(T c -25)]×S a ×K1×K2×K3×K4 / 1000 P—Calculated value of photovoltaic power generation; η s —Photoelectric conversion efficiency under standard test conditions; α—temperature coefficient; T c =T a +[(T rated -20) / 800]×I T ; T a —Surface air temperature; T rated —Rated solar cell operating temperature; T c —PV panel temperature; S a —Effective area of ​​photovoltaic modules in power stations; K1—PV module aging loss coefficient; K1=1-k×y a ; k—the number of years the grid-connected photovoltaic power station has been put into use; y a —Annual attenuation rate of solar cell materials; K2—PV module mismatch loss coefficient; K3—dust shielding loss coefficient; K4—DC circuit line loss coefficient.

3. The photovoltaic power generation prediction method based on physical mechanism and machine learning considering similar time periods according to claim 1 is characterized in that: Step S3 is specifically as follows: Step S3.1, integrating the calculated photovoltaic power generation value at each time period t of the historical year and the meteorological data at each time period t, to obtain a photovoltaic power generation characteristic data sequence at each time period t of the historical year; wherein each set of photovoltaic power generation characteristic data includes the calculated photovoltaic power generation value and meteorological data of the corresponding time period; The sample length Ts is determined, so that the photovoltaic power generation characteristic data sequence of each period t in the historical year is divided into m subsequences in sequence, each subsequence forms a sample, and its length is the sample length Ts, thereby obtaining a training sample set S including m samples = {s1, s2, s3, …, s m }; Step S3.2, determine the sliding window scale set U = {u1,u2,u3,…,u n }, where n represents the total number of sliding window scales; Step S3.3, traverse the sliding window scale set U = {u1,u2,u3,…,u n }, at each sliding window scale, from the training sample set S = {s1, s2, s3, …, s m } to obtain m sample units, each sample unit being a sequence of multiple continuous samples; Therefore, for n sliding window scales, a total of m*n sample units are obtained. Let N = m*n, and the N sample units are represented as: sample units d1, d2, d3, …, d N ; Step S3.4, building a machine learning model XGBoost; Take each sample unit d k , k = 1, 2, 3, ..., N, train the machine learning model XGBoost, take the photovoltaic power measured value sequence of the corresponding period as the model target, determine the model parameters, and thus obtain the sample unit d k The machine learning model XGBoost obtained after training k , and obtain the machine learning model XGBoost k The coefficient of determination Step S3.5, according to each sample unit d k Corresponding machine learning model XGBoost k The coefficient of determination Construct a mathematical optimization model and solve it, so that in the sample units d1, d2, d3, …, d N In the process, find the best sample unit combination Among them, M represents the number of sample units in the best sample unit combination, so that the best sample unit combination Satisfies the following conditions: Contains training sample set S = {s1, s2, s3, ..., s m }, and there are no duplicate samples between any two sample units, and the sum of the determination coefficients of the machine learning models of all sample units in the best sample unit combination reaches the maximum value; Step S3.6, based on the best sample unit combination Determine the best way to divide similar time periods within a year, thereby dividing the year into multiple consecutive similar time periods, and determine the machine learning model corresponding to each similar time period; Specifically: Best sample unit combination In each sample unit The corresponding time period within the year is a similar time period Thus, the year is divided into M similar time periods, expressed as: Each similar time period The corresponding machine learning model is to use the corresponding sample unit The machine learning model XGBoost obtained after training r .

4. A photovoltaic power generation prediction method based on physical mechanism and considering machine learning of similar time periods according to claim 3, characterized in that: In step S3.1, the sample length Ts is a month or a week.

5. The photovoltaic power generation prediction method based on physical mechanism and considering machine learning of similar time periods according to claim 3 is characterized in that: The meteorological data used in step S3.1 include irradiance, temperature and rainfall data at monthly, daily, hourly and hourly scales.

6. The photovoltaic power generation prediction method based on physical mechanism and machine learning considering similar time periods according to claim 3 is characterized in that: In step S3.3, at each sliding window scale, a sample is selected from the training sample set S = {s1, s2, s3, ..., s m } to obtain m sample units, specifically: Step S3.3.1, when the sliding window size u is used i When i = 1, 2, 3, ... n, for the training sample set S = {s1, s2, s3, ..., s m } to supplement and obtain the supplemented training sample set in, Step S3.3.2, using a sliding window size of u i The sliding window, with a step size of 1, is in the supplemented training sample set Slide from sample s1 until you reach sample s m , thus obtaining m sample units corresponding to the sliding window of m positions; each sample unit has u i A continuous sample.

7. The photovoltaic power generation prediction method based on physical mechanism and machine learning considering similar time periods according to claim 3 is characterized in that: In step S3.5, the mathematical optimization model is: The optimization objective F of the mathematical optimization model is: The constraints of the mathematical optimization model are: in: The best sample unit combination Medium Sample Unit The machine learning model obtained after training; For machine learning models The coefficient of determination.

8. The photovoltaic power generation prediction method based on physical mechanism and machine learning considering similar time periods according to claim 3 is characterized in that: The mathematical optimization model is solved by using 0-1 integer programming method.

Images