CN113487093A - Ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination - Google Patents

Ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination Download PDF

Info

Publication number
CN113487093A
CN113487093A CN202110778120.4A CN202110778120A CN113487093A CN 113487093 A CN113487093 A CN 113487093A CN 202110778120 A CN202110778120 A CN 202110778120A CN 113487093 A CN113487093 A CN 113487093A
Authority
CN
China
Prior art keywords
neighborhood
site
optimal
photovoltaic power
sites
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110778120.4A
Other languages
Chinese (zh)
Other versions
CN113487093B (en
Inventor
唐雅洁
龚迪阳
倪筹帷
李志浩
林达
汪湘晋
张雪松
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Electric Power Research Institute of State Grid Zhejiang Electric Power Co Ltd
Original Assignee
Electric Power Research Institute of State Grid Zhejiang Electric Power Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Electric Power Research Institute of State Grid Zhejiang Electric Power Co Ltd filed Critical Electric Power Research Institute of State Grid Zhejiang Electric Power Co Ltd
Priority to CN202110778120.4A priority Critical patent/CN113487093B/en
Publication of CN113487093A publication Critical patent/CN113487093A/en
Application granted granted Critical
Publication of CN113487093B publication Critical patent/CN113487093B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/067—Enterprise or organisation modelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • General Physics & Mathematics (AREA)
  • Tourism & Hospitality (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Theoretical Computer Science (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Health & Medical Sciences (AREA)
  • Educational Administration (AREA)
  • Public Health (AREA)
  • Water Supply & Treatment (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an ultrashort-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination. The technical scheme adopted by the invention is as follows: firstly, calculating a map linear distance according to the longitude and latitude of a site to be predicted, and establishing a neighborhood site set; secondly, establishing an ultra-short-term photovoltaic power prediction model based on the selected forward time sequence optimal combination feature set of the neighborhood sites according to the neighborhood set, sequentially checking the prediction effect, expanding the optimal combination set, and completing construction of the ultra-short-term photovoltaic power prediction model of the sites to be predicted until forward checking is finished. According to the method, the accuracy of the super-short-term station prediction model under weather fluctuation can be improved when ground meteorological observation data of the station lack.

Description

Ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination
Technical Field
The invention relates to the technical field of photovoltaic ultra-short-term power prediction, in particular to an ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination.
Background
Under the global carbon neutralization target, the vigorous development of clean zero-carbon new energy sources such as photovoltaic and the like becomes a key factor for carbon emission reduction. Because the photovoltaic output has stronger randomness and volatility, an accurate power prediction technology is the basis for participating in the dispatching operation of the power grid; meanwhile, under the mechanism of the power market, the important basis for consuming high-proportion photovoltaic is to correctly predict the ultra-short-term power of the generated power within the future 4-hour period.
The research on the single photovoltaic station power prediction technology in European and American countries starts earlier, and a neural network method is introduced in the 90 s of the 20 th century; the photovoltaic power prediction field starts relatively late in China, and the research speed is obviously increased in recent years. With the breakthrough of artificial intelligence algorithm, the current mainstream photovoltaic station ultra-short term power prediction methods at home and abroad are statistical prediction models which are based on machine learning algorithms such as neural networks and added with a real-time power feedback correction link and are combined and applied by aiming at a single-station multi-prediction modeling method. The prediction model does not consider complex photovoltaic equipment physical modeling, but mines correlation among multiple variables based on big data analysis, considers dynamic prediction error feedback, extracts real-time output characteristics corresponding to the photovoltaic under various meteorological and geographic states, and realizes ultra-short-term photovoltaic power prediction under multiple scenes. The research mainly focuses on enhancing the generalization of the modeling method by using the prediction model and enhancing the real-time performance by using the error correction method, so that the performance of the single-site ultra-short-term power prediction model is improved.
However, the input observation information of the ultra-short-term power prediction model proposed by the above research depends on the earth surface irradiation intensity observation device configured in the ground meteorological station, which is relatively high in cost, and is focused on a certain specific geographic orientation, and the capability of rapidly tracking real-time meteorological changes under weather fluctuation is still insufficient. Therefore, the improvement of the prediction accuracy of the distributed photovoltaic power can be realized according to the multi-site information. Currently, there is a method for increasing the predicted input feature dimension of the local station by using a neighboring station with higher power similarity. However, the geographical weather states of the sites with higher output similarity are generally similar, so that the problems of information reuse and less effective dimension increasing information of combination expansion easily occur in combination prediction, and the improvement of the prediction effect precision is limited.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provides an ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination, which enhances the mining of the incidence relation among the photovoltaic outputs of each neighborhood caused by weather change and increases the effective observation dimension of the geographic space of the model training characteristics so as to realize the improvement of the accuracy of the ultra-short-term prediction model of the site under the weather fluctuation when the ground meteorological observation data of the site is lacked.
Therefore, the invention adopts the following technical scheme: the ultrashort-term photovoltaic power prediction method based on the optimal combination of the neighborhood forward time sequences comprises the following steps:
firstly, calculating a map linear distance according to the longitude and latitude of a site to be predicted, and establishing a neighborhood site set;
secondly, establishing an ultra-short-term photovoltaic power prediction model based on the selected forward time sequence optimal combination feature set of the neighborhood sites according to the neighborhood set, sequentially checking the prediction effect, expanding the optimal combination set, and completing construction of the ultra-short-term photovoltaic power prediction model of the sites to be predicted until forward checking is finished.
Compared with the classical ultra-short-term photovoltaic prediction model based on power time sequence transition and site power similarity, the method provided by the invention can more accurately predict the change trend of the future photovoltaic power of the site according to the neighborhood space-time transition of weather under the condition of lacking meteorological observation information, and effectively improve the prediction precision.
Further, the specific content of establishing the neighborhood site set is as follows:
photovoltaic station A0Is set to be equal to A0All sites with a distance of no more than 240 km, assuming site a0Longitude and latitude of (x)0,y0) Another site a on the mapiLongitude and latitude of (x)i,yi) Then, the distance between two stations is calculated as follows:
Di＝arccos((siny0×sinyi)+(cosy0×cosyi×cos(x0-xi)))×r (1)
in the formula, DiAs station A0With another site AiThe linear distance therebetween; r is the earth mean radius; all variable units are kilometers;
calculating to obtain a neighborhood site set S which is a site set meeting the requirement of the influence distance, and comprises the following steps:
Figure BDA0003156577950000031
wherein M is the total number of neighborhood sites, PkNumbering the stations; station A0Contains A0Itself.
Further, let the station A to be predicted0The neighborhood site set obtained by the calculation in the step one is { A }0,A1,A2,...ANAnd selecting an optimal neighborhood site combination set as V, an optimal input feature set as X, a neighborhood site set to be verified as S and an optimal error verification value as e, and selecting a neighborhood forward time sequence optimal combination sequence by adopting an iterative extension verification method, wherein the method specifically comprises the following steps:
1) iterative initialization of
Figure BDA0003156577950000032
S(0)＝{A0,A1,A2,...AN}，e(0)＝+∞；
Wherein, V(0)Representing the initial optimal neighborhood site combination set, X(0)Representing initial optimal input feature sets which are all empty sets; s(0)Representing an initial neighborhood site set to be checked as a to-be-predicted sitePoint A0A set of all neighborhood sites (including themselves); e.g. of the type(0)Representing an initial optimal error check value which is infinite, and the initialization process can be regarded as 0 th iteration;
2) let us say that the current iteration has been to the ith time, i ∈ [1, N]Then after the last iteration, the selected neighborhood site combination set is V(i-1)The optimal input feature set is X(i-1)The iteration can select the neighborhood site set as S(i-1)Then, the combination mode of the current round of the to-be-verified preferred expandable neighborhood sites is as follows:
Figure BDA0003156577950000036
wherein, Vj(j∈[1,i-1]) Neighborhood sites selected for jth iteration
Figure BDA0003156577950000037
Subscripts;
photovoltaic power at predicted time T is used as output YTThe length of the period L of the optimal common translation input feature set for all the sites in the enumeration set ViIncorporating time-shift features of the current predicted instant
Figure BDA0003156577950000035
The method comprises the following steps of { minute, hour, day, month, quarter, year }, and the obtained multidimensional input feature vector set to be verified at the prediction moment is shown as a formula (4):
Figure BDA0003156577950000033
wherein the content of the first and second substances,
Figure BDA0003156577950000034
for the kth preceding current prediction time (k e [1, L ]i]) Historical predicted time tkEach neighborhood site in V
Figure BDA00031565779500000410
(j∈[1,i-1]) At tkReal photovoltaic power at a moment
Figure BDA0003156577950000041
Is a combined vector of
Figure BDA0003156577950000042
When a training check model at all time F (X) Y is formed, transversely expanding variable values contained in all vectors in the vectors to form a total input vector of a training set;
3) and performing K-fold cross validation on the F (X) -Y model training sample to be verified by adopting a machine learning algorithm.
Furthermore, the K-fold cross-validation method divides the training sample into K partitions with the same size, sequentially selects different partitions as a test set, and takes the remaining K-1 partitions as a training set; the error of each training is the average percentage error between the predicted value and the true value and is marked as MAPEkThe final error is the average of all training errors
Figure BDA00031565779500000411
The formula is as follows:
Figure BDA0003156577950000043
Figure BDA0003156577950000044
wherein n is the number of samples in the test area; y isiThe actual value of the photovoltaic power is obtained;
Figure BDA0003156577950000045
and (4) the photovoltaic power predicted value of the k-th cross test.
Still further, under a K-fold cross validation method, enumerating the period length L 'of each combination V' to be verified, and calculating the error mean value of the X (V ') model of the input feature set X (V') to be verified formed each time
Figure BDA0003156577950000046
Let the error optimal mean be
Figure BDA0003156577950000047
The neighborhood site set corresponding to the optimal error is V'minFeature set X'minThe newly added optimal neighborhood site in V' of the iteration is
Figure BDA0003156577950000048
If e(i)≤e(i-1)Then order V(i)＝V′min，X(i)＝X'min，
Figure BDA0003156577950000049
i is i + 1; otherwise, the algorithm is terminated; meanwhile, when i is N +1, the algorithm terminates; v obtained finally(i-1)Namely the site A to be predicted0Optimal set sequence after neighborhood selection, corresponding X(i-1)I.e. the model input feature set.
The method comprehensively considers the multi-meteorological-state coupling scene, expands effective information dimension, performs efficient training modeling under the condition of not depending on meteorological observation devices such as earth surface irradiation intensity and the like, improves the ultra-short-term photovoltaic power prediction precision under the scene, has the characteristics of economy, light weight and easiness in deployment, and is particularly suitable for the ultra-short-term power prediction of small distributed photovoltaic stations which are generally not matched with ground meteorological stations. The result shows that under the condition that only photovoltaic power is used as an effective input feature set of a prediction model, compared with a classical ultra-short-term photovoltaic prediction model based on power time sequence transition and site power similarity, the method provided by the invention can more accurately predict the change trend of the photovoltaic power of a site in a future ultra-short-term period by means of the associated coupling feature dimension expanded by the neighborhood space-time transition of weather, and effectively improve the prediction precision.
Drawings
FIG. 1 is a schematic diagram of a typical single site ultra short term power timing prediction model;
FIG. 2 is a schematic diagram of an ultra-short term power prediction model based on a neighborhood forward timing optimal combination according to the present invention;
FIG. 3 is a schematic diagram of a distributed photovoltaic system site in a certain area of Ningbo in Zhejiang in an application example of the present invention;
fig. 4 is a comparison graph of the ultra-short term power prediction results in the scenario 1 in the application example of the present invention (fig. 4a is a graph of the ultra-short term power prediction results of multiple models in cloudy days of 5 months and 8 days, and fig. 4b is a graph of the ultra-short term power prediction results of multiple models in sunny days of 5 months and 9 days);
fig. 5 is a comparison diagram of the ultra-short term power prediction results in the scenario 2 in the application example of the present invention (fig. 5a is a diagram of the ultra-short term power prediction results of the multi-model in a sunny day of 5 months and 8 days, and fig. 5b is a diagram of the ultra-short term power prediction results of the multi-model in a sunny day of 5 months and 9 days).
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Examples
The embodiment provides an ultrashort-term photovoltaic power prediction method based on the optimal combination of neighborhood forward time sequences, which comprises the following steps:
step one, establishing a neighborhood site set
For photovoltaic site A0According to the fast moving speed of the strong convection cloud cluster, which is generally 60 km/h, and 4 hours of ultra-short-term prediction, the cloud cluster within 240 km (the altitude is ignored during the distance calculation, the following same below) from the station may move to the station position within the prediction time period, so that the continuity relevance of the meteorological change in the area is brought. Photovoltaic station a0Can be set to be equal to A0All sites with a distance of no more than 240 km. Suppose A0The station latitude and longitude is (x)0,y0) Another site a on the mapiLongitude and latitude of (x)i,yi) Then, the distance between two stations is calculated as follows:
Di＝arccos((siny0×sinyi)+(cosy0×cosyi×cos(x0-xi)))×r (1)
in the formula, DiAs station A0And site AiThe linear distance therebetween; r is the earth mean radius; all variables are in kilometers.
Calculating to obtain a neighborhood site set S which is a site set meeting the requirement of the influence distance, and comprises the following steps:
Figure BDA0003156577950000061
wherein M is the total number of neighborhood sites, PkNumbering the stations; in particular, site A0Contains A0Itself.
Step two, ultra-short-term photovoltaic power prediction model based on neighborhood forward time sequence optimal combination
A time sequence passing algorithm based on data mining is a basic method for ultra-short-term photovoltaic prediction, and mainly aims at learning trends of historical photovoltaic data. Since photovoltaic output is univariate data, the characteristics mainly include two parts: a translation feature and a time feature. The translation characteristic is that the current moment is taken as a time node, and the previous historical photovoltaic output data is taken as a characteristic; the time characteristic generally refers to a calendar characteristic of the predicted time of day, such as hours, minutes, and the like. The typical single-photovoltaic-station ultra-short-term power timing model is modeled by combining the relationship between the two and the photovoltaic output value, as shown in fig. 1.
Sequence Forward Selection (SFS) is a commonly used feature Selection method. The SFS algorithm is started from an empty set, an optimal feature k is selected from unselected features each time and added into a feature set X, so that an evaluation function J (X) trained by a learner is optimal; when J (X) fails to obtain better, the forward selection ends. SFS is a chain greedy algorithm. On the basis of a typical time sequence prediction model, a modeling method combined with neighborhood site translation features is provided, an SFS framework is used for performing neighborhood site optimization screening on a sequence S, observation dimensions are expanded, the correlation coupling of meteorological space-time transition reflected on neighborhood photovoltaic output is further deeply mined, and the site prediction precision is further improved. The specific process is as follows:
order station A0The neighborhood site set obtained by the calculation in the step one is { A }0,A1,A2,...ANAnd e, selecting an optimal neighborhood site combination set as V, an optimal input feature set as X, a neighborhood site set to be verified as S and an optimal error verification value as e, and selecting a neighborhood forward time sequence optimal combination sequence by adopting an iterative extension verification method. The method comprises the following specific steps:
2) iterative initialization, initialization
Figure BDA0003156577950000071
S(0)＝{A0,A1,A2,...AN}，e(0)＝+∞；
3) Let us now iterate to the ith (i ∈ [1, N)]) Then, the neighbor site combination set selected in the last iteration is V(i -1)The input feature set is X(i-1)The set of neighborhood sites can be selected as S(i-1)Then, the combination mode of the current round of the to-be-verified preferred expandable neighborhood sites is as follows:
Figure BDA0003156577950000077
wherein, Vj(j∈[1,i-1]) Neighborhood sites selected for jth iteration
Figure BDA0003156577950000078
Subscripts.
Photovoltaic power at predicted time T is used as output YTThe length of the period L of the optimal common translation input feature set for all the sites in the enumeration set ViIncorporating time-shift features of the current predicted instant
Figure BDA0003156577950000076
Including { minute, hour, day, month, quarter, year }, and the obtained multidimensional input feature vector set to be verified at the predicted time is shown in formula (4)。
Figure BDA0003156577950000072
Wherein the content of the first and second substances,
Figure BDA0003156577950000073
for the kth preceding current prediction time (k e [1, L ]i]) Historical predicted time tkEach neighborhood site in V
Figure BDA0003156577950000079
(j∈[1,i-1]) At tkReal photovoltaic power at a moment
Figure BDA0003156577950000074
Is a combined vector of
Figure BDA0003156577950000075
When a training check model at all time F (X) and Y in the whole period is formed, variable values contained in all vectors in the vectors are transversely expanded to form a total input vector of a training set. The model is shown in fig. 2.
3) And performing K-fold cross validation on the F (X) -Y model training sample to be verified by adopting a machine learning algorithm. The K-fold cross verification method divides a training sample into K partitions with the same size, sequentially selects different partitions as a test set, and takes the remaining K-1 partitions as a training set. The error of each training is the average percentage error between the predicted value and the true value and is marked as MAPEkThe final error is the average of all training errors
Figure BDA0003156577950000081
The formula is as follows:
Figure BDA0003156577950000082
Figure BDA0003156577950000083
wherein n is the number of samples in the test area; y isiThe actual value of the photovoltaic power is obtained;
Figure BDA0003156577950000084
and (4) the photovoltaic power predicted value of the k-th cross test.
Under a K-fold cross validation method, enumerating the period length L 'of each combination V' to be verified, and calculating the error mean value of the X (V ') model of the input feature set X (V') to be verified formed each time
Figure BDA0003156577950000085
Let the error optimal mean be
Figure BDA0003156577950000086
The neighborhood site set corresponding to the optimal error is V'minFeature set X'minThe newly added optimal neighborhood site in V' of the iteration is
Figure BDA0003156577950000087
If e(i)≤e(i-1)Then order V(i)＝V′min，X(i)＝X'min，
Figure BDA0003156577950000088
i is i + 1; otherwise the algorithm terminates. Meanwhile, when i is N +1, the algorithm terminates; v obtained finally(i-1)Namely the site A to be predicted0Optimal set sequence after neighborhood selection, corresponding X(i-1)I.e. the model input feature set.
The optimal period length calculated in the last effective iteration process is set to be L, and an ultra-short-term photovoltaic power prediction model based on the neighborhood forward time sequence optimal combination is shown in table 1.
TABLE 1 ultrashort term photovoltaic power prediction based on neighborhood forward timing optimal combination
Figure BDA0003156577950000089
Figure BDA0003156577950000091
And the model prediction is carried out by adopting a rolling calculation method, and the predicted time power is used as the power translation characteristic of the next predicted time until the power prediction of the 4 th hour of the ultra-short term prediction is completed.
Application example
The method provided by the invention is applied to a distributed photovoltaic system in a certain area of Ningbo in Zhejiang as shown in FIG. 3, and the distributed photovoltaic system comprises 24 photovoltaic stations which are numbered from 1 to 24. The longitude and latitude of each station and photovoltaic power data are known, and the power observation time and the time resolution of the training set test data set are completely consistent, namely, the data are respectively from 4/1/2020/1 to 5/9/2021/15, and each data point is 15 minutes.
In order to verify the effect of the neighborhood combination prediction, the ultra-short term photovoltaic power prediction precision under two different scenes is considered.
Scene 1 ultra-short-term photovoltaic power prediction of sites 4 with denser neighborhood sites.
And (2) ultra-short-term photovoltaic power prediction of a site 3 with sparse neighborhood sites.
The method selects the 4 th hour prediction point average percentage error MAPE of all test set ultra-short-term prediction time periods as an evaluation index of the ultra-short-term photovoltaic prediction model performance. The calculation formula is as follows:
Figure BDA0003156577950000092
in the formula, ytActual photovoltaic power generation data of the 4 th hour predicted for each ultra-short period;
Figure BDA0003156577950000093
predicting photovoltaic power generation data for the 4 th hour of each ultra-short-term prediction; t is the total amount of data.
In order to verify the prediction effect of the invention under different meteorological conditions, the data set is used as a training sample from 2020, 4 and 1 to 2021, 5 and 7 days, and the data set is used as a testing sample from 2021, 5 and 8 days to 2021, 5 and 9 days, and the invention is compared with an ultra-short-term power prediction model based on single-site characteristic time sequence transition and based on site power similarity under different scenes. In order to realize reasonable comparison of prediction effects, all the adopted machine learning algorithms are XGboost (gradient spanning Tree) methods, all the model parameters are kept consistent, and all the model parameters are tree depth 7, learning rate 0.01, tree number 1000, maximum leaf node number 128 and characteristic sampling rate 1.0.
The following detailed description of the principles of the invention is provided in connection with the accompanying drawings.
And D, calculating according to the first step to obtain that all the sites in the area are in the neighborhood range. The process of constructing the ultra-short-term photovoltaic power prediction model extension based on the neighborhood forward time sequence optimal combination and the calculation result pair are shown in table 2. The table shows that each test scene gradually improves the prediction precision according to the feature dimension added by the neighborhood site set, and all sites to be predicted enter the optimal combination queue first. In addition, not all neighborhood sites have an effect of improving the prediction effect of the site, and the following two reasons are mainly adopted: firstly, the influence of real geographic factors, such as movement of mountains blocking clouds and the like, causes weak weather continuity of two sites, so that the relevance of photovoltaic output is weak; secondly, influence factors of partial neighborhood sites on the photovoltaic output of the site are reflected in the selected site set, and the expanded characteristic dimension is redundant, so that the generalization of the model is reduced.
TABLE 2 comparison of ultra-short term photovoltaic power prediction model calculation results
Figure BDA0003156577950000101
Figure BDA0003156577950000111
Table 2 also shows the error comparison of the prediction effect of the method of the present invention with the ultra-short term power prediction model based on the power feature time sequence transition and the power similarity, respectively. The method comprises the following steps that a site power time sequence based transition model is a neighborhood forward time sequence optimal combination prediction model when only the site features exist in a set; and based on the station power similarity model, the daily curve average Euclidean distance is used as a standard, the most similar stations with the same number as the optimal combined stations of the model are selected for modeling, and the comparison of the station combination selection method is realized. Similarity S between site number i and site number jijThe calculation formula is as follows:
Figure BDA0003156577950000112
in the formula (I), the compound is shown in the specification,
Figure BDA0003156577950000113
the kth power data of site i and site j on day d respectively; n is the number of daily power curve data; d is the number of days of the data.
As can be seen from Table 2, the prediction accuracy of the model under the scene 1 is respectively improved by 1.06% and 1.01%; the prediction accuracy of the model under the scene 2 is improved by 1.58% and 1.04% respectively, and the comparison graphs of the prediction curves are shown in fig. 4 and 5. As can be seen from the figure, the model prediction curve constructed by the method has stronger follow-up performance on the actual output trend when the weather of the cloudy days changes frequently, has less burrs when the actual output is stable in the sunny days, and can obtain the best prediction effect under two typical weather scenes.
In summary, without relying on a meteorological station observation device, the ultra-short-term prediction accuracy of the stations can be improved from the perspective of the integrity and continuity of meteorological changes by efficiently integrating the photovoltaic power monitoring information provided by all the stations in the area.
The invention provides an ultrashort-term photovoltaic power prediction method based on neighborhood forward time sequence combination, and the method analyzes and contrasts the effect of a prediction model under a multi-site scene. The results show that:
1) under the condition of not depending on an earth surface irradiation intensity observation device, a prediction model established by considering multi-dimensional coupling correlation among photovoltaic output of adjacent area sites can improve the ultra-short-term photovoltaic prediction precision simply and efficiently.
2) The method has the characteristics of economy, light weight and easiness in deployment, and is particularly suitable for ultra-short-term power prediction of small distributed photovoltaic stations which are not matched with weather stations generally.
While the invention has been described in further detail with reference to preferred embodiments thereof, it should be emphasized that the above-described embodiments are not to be considered as limiting. After reading the above description, it will be apparent to those skilled in the art that various modifications, substitutions and alterations can be made without departing from the spirit of the invention, and all such modifications, substitutions and alterations should be considered as the scope of the invention. Accordingly, the scope of the invention should be determined from the following claims.

Claims (5)

1. An ultrashort-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination is characterized in that,
firstly, calculating a map linear distance according to the longitude and latitude of a site to be predicted, and establishing a neighborhood site set;
secondly, establishing an ultra-short-term photovoltaic power prediction model based on the selected forward time sequence optimal combination feature set of the neighborhood sites according to the neighborhood set, sequentially checking the prediction effect, expanding the optimal combination set, and completing construction of the ultra-short-term photovoltaic power prediction model of the sites to be predicted until forward checking is finished.
2. The ultrashort-term photovoltaic power prediction method based on the optimal combination of the neighborhood forward time sequences as claimed in claim 1, wherein the specific content of establishing the neighborhood site set is as follows:
photovoltaic station A0Is set to be equal to A0All sites with a distance of no more than 240 km, assuming site a0Longitude and latitude of (x)0,y0) Another site a on the mapiLongitude and latitude of (x)i,yi) Then, the distance between two stations is calculated as follows:
Di＝arccos((siny0×sinyi)+(cosy0×cosyi×cos(x0-xi)))×r (1)
in the formula, DiAs station A0With another site AiThe linear distance therebetween; r is the earth mean radius; all variable units are kilometers;
calculating to obtain a neighborhood site set S which is a site set meeting the requirement of the influence distance, and comprises the following steps:
Figure FDA0003156577940000011
wherein M is the total number of neighborhood sites, PkNumbering the stations; station A0Contains A0Itself.
3. The ultrashort-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination as claimed in claim 1, wherein let site to be predicted A0The neighborhood site set obtained by the calculation in the step one is { A }0,A1,A2,...ANAnd selecting an optimal neighborhood site combination set as V, an optimal input feature set as X, a neighborhood site set to be verified as S and an optimal error verification value as e, and selecting a neighborhood forward time sequence optimal combination sequence by adopting an iterative extension verification method, wherein the method specifically comprises the following steps:
1) iterative initialization of
Figure FDA0003156577940000012
S(0)＝{A0,A1,A2,...AN}，e(0)＝+∞；
Wherein, V(0)Representing the initial optimal neighborhood site combination set, X(0)Representing initial optimal input feature sets which are all empty sets; s(0)Representing an initial neighborhood site set to be verified asPredicting site A0A set of all neighborhood sites; e.g. of the type(0)Representing an initial optimal error check value which is infinite, and taking the initialization process as the 0 th iteration;
2) let us say that the current iteration has been to the ith time, i ∈ [1, N]Then after the last iteration, the selected neighborhood site combination set is V(i-1)The optimal input feature set is X(i-1)The iteration selects a neighborhood site set as S(i-1)Then, the combination mode of the current round of the to-be-verified preferred expandable neighborhood sites is as follows:
Figure FDA0003156577940000028
wherein, Vj(j∈[1,i-1]) Neighborhood sites selected for jth iteration
Figure FDA0003156577940000029
Subscripts;
photovoltaic power at predicted time T is used as output YTThe length of the period L of the optimal common translation input feature set for all the sites in the enumeration set ViCombined with the current predicted time t0Time shift feature of
Figure FDA0003156577940000027
The method comprises the following steps of { minute, hour, day, month, quarter, year }, and the obtained multidimensional input feature vector set to be verified at the prediction moment is shown as a formula (4):
Figure FDA0003156577940000021
Figure FDA0003156577940000022
predicting the k-th historical time t before the current prediction timekEach neighborhood site in V
Figure FDA0003156577940000023
At tkReal photovoltaic power at a moment
Figure FDA0003156577940000024
Is a combined vector of
Figure FDA0003156577940000025
k∈[1,Li]，j∈[1,i-1](ii) a When a training check model at all time F (X) Y is formed, transversely expanding variable values contained in all vectors in the vectors to form a total input vector of a training set;
3) and performing K-fold cross validation on the F (X) -Y model training sample to be verified by adopting a machine learning algorithm.
4. The ultrashort-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination as claimed in claim 3, wherein the K-fold cross-validation method divides a training sample into K partitions with the same size, sequentially selects different partitions as a test set, and takes the remaining K-1 partitions as a training set; the error of each training is the average percentage error between the predicted value and the true value and is marked as MAPEkThe final error is the average of all training errors
Figure FDA0003156577940000026
The formula is as follows:
Figure FDA0003156577940000031
Figure FDA0003156577940000032
wherein n is the number of samples in the test area; y isiThe actual value of the photovoltaic power is obtained;
Figure FDA0003156577940000033
and (4) the photovoltaic power predicted value of the k-th cross test.
5. The ultrashort-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination as claimed in claim 4, wherein under K-fold cross validation, the period length L 'under each combination V' to be verified is enumerated, and the error mean value of the X (V ') model of the input feature set X (V') to be verified formed each time is calculated
Figure FDA0003156577940000034
Let the error optimal mean be
Figure FDA0003156577940000035
The neighborhood site set corresponding to the optimal error is V'minFeature set X'minThe newly added optimal neighborhood site in V' of the iteration is
Figure FDA0003156577940000036
If e(i)≤e(i-1)Then order V(i)＝V′min，X(i)＝X′min，
Figure FDA0003156577940000037
i is i + 1; otherwise, the algorithm is terminated; meanwhile, when i is N +1, the algorithm terminates; v obtained finally(i-1)Namely the site A to be predicted0Optimal set sequence after neighborhood selection, corresponding X(i-1)I.e. the model input feature set.
CN202110778120.4A 2021-07-09 2021-07-09 Ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination Active CN113487093B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110778120.4A CN113487093B (en) 2021-07-09 2021-07-09 Ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110778120.4A CN113487093B (en) 2021-07-09 2021-07-09 Ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination

Publications (2)

Publication Number Publication Date
CN113487093A true CN113487093A (en) 2021-10-08
CN113487093B CN113487093B (en) 2024-06-14

Family

ID=77937843

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110778120.4A Active CN113487093B (en) 2021-07-09 2021-07-09 Ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination

Country Status (1)

Country Link
CN (1) CN113487093B (en)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103559561A (en) * 2013-11-13 2014-02-05 上海电气集团股份有限公司 Super-short-term prediction method of photovoltaic power station irradiance
JP2014063372A (en) * 2012-09-21 2014-04-10 Toshiba Corp Apparatus and method for predicting power generation capacity
CN108053048A (en) * 2017-11-06 2018-05-18 中国电力科学研究院有限公司 A kind of gradual photovoltaic plant ultra-short term power forecasting method of single step and system
CN109299821A (en) * 2018-09-21 2019-02-01 华北电力大学（保定） A kind of distributed photovoltaic power generation predicting platform and method
CN111967652A (en) * 2020-07-22 2020-11-20 国网浙江省电力有限公司电力科学研究院 Double-layer cooperative real-time correction photovoltaic prediction method
CN112132328A (en) * 2020-09-04 2020-12-25 国网上海市电力公司 Photovoltaic output power ultra-short-term local emotion reconstruction neural network prediction method
CN112381282A (en) * 2020-11-09 2021-02-19 上海交通大学 Photovoltaic power generation power prediction method based on width learning system

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2014063372A (en) * 2012-09-21 2014-04-10 Toshiba Corp Apparatus and method for predicting power generation capacity
CN103559561A (en) * 2013-11-13 2014-02-05 上海电气集团股份有限公司 Super-short-term prediction method of photovoltaic power station irradiance
CN108053048A (en) * 2017-11-06 2018-05-18 中国电力科学研究院有限公司 A kind of gradual photovoltaic plant ultra-short term power forecasting method of single step and system
CN109299821A (en) * 2018-09-21 2019-02-01 华北电力大学（保定） A kind of distributed photovoltaic power generation predicting platform and method
CN111967652A (en) * 2020-07-22 2020-11-20 国网浙江省电力有限公司电力科学研究院 Double-layer cooperative real-time correction photovoltaic prediction method
CN112132328A (en) * 2020-09-04 2020-12-25 国网上海市电力公司 Photovoltaic output power ultra-short-term local emotion reconstruction neural network prediction method
CN112381282A (en) * 2020-11-09 2021-02-19 上海交通大学 Photovoltaic power generation power prediction method based on width learning system

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
NATTAWAT SODSONG 等: "Short-Term Solar PV Forecasting Using Gated Recurrent Unit with a Cascade Model", 《2019 INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE IN INFORMATION AND COMMUNICATION (ICAIIC)》, 21 March 2019 (2019-03-21), pages 292 - 297 *
唐雅洁 等: "基于邻域前向时序最优组合的分布式光伏超短期功率预测", 《浙江电力》, vol. 40, no. 10, 25 October 2021 (2021-10-25), pages 95 - 101 *
赵康宁: "基于贝叶斯神经网络的分布式光伏预测及应用", 中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑, no. 2021, 15 April 2021 (2021-04-15), pages 042 - 261 *

Also Published As

Publication number Publication date
CN113487093B (en) 2024-06-14

Similar Documents

Publication Publication Date Title
US20220373984A1 (en) Hybrid photovoltaic power prediction method and system based on multi-source data fusion
CN111860982A (en) Wind power plant short-term wind power prediction method based on VMD-FCM-GRU
US20230035108A1 (en) Photovoltaic cell parameter identification method based on improved equilibrium optimizer algorithm
CN110909924A (en) City expansion multi-scenario simulation cellular automata method based on cross entropy optimizer
CN110826244B (en) Conjugated gradient cellular automaton method for simulating influence of rail transit on urban growth
CN110942205A (en) Short-term photovoltaic power generation power prediction method based on HIMVO-SVM
CN115374995A (en) Distributed photovoltaic and small wind power station power prediction method
CN112712169A (en) Model building method and application of full residual depth network based on graph convolution
CN109583588A (en) A kind of short-term wind speed forecasting method and system
CN116796168A (en) CNN-BiLSTM high-altitude multi-factor power transmission line audible noise prediction method based on multi-head attention mechanism
Aliberti et al. Forecasting Short-term Solar Radiation for Photovoltaic Energy Predictions.
CN117613883A (en) Method and device for predicting generated power, computer equipment and storage medium
CN115936177A (en) Photovoltaic output power prediction method and system based on neural network
CN116894504A (en) Wind power cluster power ultra-short-term prediction model establishment method
CN115759389A (en) Day-ahead photovoltaic power prediction method based on weather type similar day combination strategy
CN116345555A (en) CNN-ISCA-LSTM model-based short-term photovoltaic power generation power prediction method
CN114004152B (en) Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network
CN114862023A (en) Distributed photovoltaic power prediction method and system based on four-dimensional point-by-point meteorological forecast
CN114882373A (en) Multi-feature fusion sandstorm prediction method based on deep neural network
CN113240219A (en) Land utilization simulation and prediction method
CN117293809A (en) Multi-time space scale new energy generation power prediction method based on large model
CN115587664B (en) Solar radiation prediction method and system based on graph convolution long-short-term memory network
CN115907131B (en) Method and system for constructing electric heating load prediction model in northern area
CN113487093B (en) Ultra-short-term photovoltaic power prediction method based on neighborhood forward time sequence optimal combination
CN114386666A (en) Wind power plant short-term wind speed prediction method based on space-time correlation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant