CN112507613A - Second-level ultra-short-term photovoltaic power prediction method - Google Patents

Second-level ultra-short-term photovoltaic power prediction method Download PDF

Info

Publication number
CN112507613A
CN112507613A CN202011385741.8A CN202011385741A CN112507613A CN 112507613 A CN112507613 A CN 112507613A CN 202011385741 A CN202011385741 A CN 202011385741A CN 112507613 A CN112507613 A CN 112507613A
Authority
CN
China
Prior art keywords
population
value
prediction
algorithm
optimization
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
CN202011385741.8A
Other languages
Chinese (zh)
Other versions
CN112507613B (en
Inventor
赵振兴
陈凯杰
戴瑜兴
陈颖
向道朴
赵葵银
宁勇
刘增
王环
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.)
Hunan Institute of Engineering
Original Assignee
Hunan Institute of Engineering
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 Hunan Institute of Engineering filed Critical Hunan Institute of Engineering
Priority to CN202011385741.8A priority Critical patent/CN112507613B/en
Publication of CN112507613A publication Critical patent/CN112507613A/en
Application granted granted Critical
Publication of CN112507613B publication Critical patent/CN112507613B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/08—Computing arrangements based on specific mathematical models using chaos models or non-linear system models
    • 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
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Electricity, gas or water supply

Abstract

The invention provides a second-level ultrashort-term photovoltaic power prediction method, and belongs to the technical field of photovoltaic power generation. According to the method, a prediction model is established by adopting an LSSVM algorithm, the influence of the super-parameters between the LSSVM algorithm on the prediction performance is large, chaotic numbers with more randomness are generated by adopting chaotic cube mapping, the initial population position calculation is optimized, and the population is initialized by combining the chaotic numbers; optimizing the hyper-parameters of the LSSVM for the first time by adopting a single iteration wolf optimization algorithm to obtain a better solution; the second optimization is to search a better solution by using an iterative local search with the improved Griewank function as a disturbance function to obtain a better solution; searching a more optimal solution by using a local adaptive differential algorithm LSaDE in the third optimization to obtain an optimal solution; and determining the hyper-parameters, and training the prediction model to obtain a predicted value. And determining the model precision according to the precision evaluation result, and carrying out parameter optimization again when the model precision does not reach the standard. The invention realizes the prediction effects of short prediction time and high prediction precision.

Description

Second-level ultra-short-term photovoltaic power prediction method
Technical Field
The invention relates to the technical field of photovoltaic power generation, in particular to a second-level ultra-short-term photovoltaic power prediction method.
Background
In recent years, the number of new photovoltaic installation machines is increased dramatically year by year, the permeability of photovoltaic power generation in the whole power grid system is higher and higher, however, in the large-scale photovoltaic grid connection process, as the output power of a photovoltaic power station is greatly influenced by the environment and has the characteristics of randomness and discontinuity, the photovoltaic output power generates large fluctuation, and large impact is caused to the power grid. Prediction and energy storage technology are key technologies for stabilizing photovoltaic output power fluctuation.
In the photovoltaic power generation process, power fluctuation is a normal phenomenon, however, the average change rate of the power fluctuation does not exceed 3%/s under the general condition, but under the complex weather condition, the instantaneous output power change rate can reach 75%/s at most due to the movement of cloud layers. With the development of hybrid energy storage technology equipped with super capacitors, it is possible to complete charging and discharging within a very short time. Correspondingly, it becomes the key to stabilize the instantaneous photovoltaic power fluctuation to accomplish prediction with high accuracy within an ultra-short time scale.
Chinese patent application document CN106372749A discloses an ultra-short-term photovoltaic power prediction method based on cloud change analysis, which predicts through meteorological conditions above a photovoltaic power station, and adds a predicted value as a correction parameter into a photovoltaic power generation model to further obtain a predicted value of the output power of the photovoltaic power station. Indirect prediction represented by the method usually only focuses on main factors influencing wide power generation, but ignores the influence of other secondary factors, and finally causes large deviation between the predicted value and the actual value of the output power. In addition, the indirect prediction method like this also has the problem of long prediction time.
Chinese patent application document CN106503828A discloses a photovoltaic output power ultra-short term chaos prediction method, which determines a central phase space point and an adjacent phase space point by reconstructing a photovoltaic output power time sequence phase space, calculates the weight values of the central phase space point and the adjacent phase space points, establishes a photovoltaic output weighted first-order local linear regression model, and calculates an optimal linear fitting coefficient matrix to obtain a photovoltaic output power prediction value. The method can obtain a predicted value with high precision in a short time, but the method is the same as other prediction algorithms, the prediction precision is sensitive to weather conditions, the prediction effect is good in sunny days, and the prediction effect is poor in rainy days. In complex weather conditions, the performance is poor.
The prior art has at least the following disadvantages:
1. the contradiction between prediction time and prediction accuracy. The prediction algorithm with shorter prediction time has more serious loss on prediction precision; the prediction algorithm with higher prediction accuracy has longer running time and cannot meet the requirement of second-level ultra-short-term power prediction.
2. The difficulty of setting the key parameters is high. In the prior art, manual experience is adopted, or an optimization algorithm with a long optimization period is adopted, so that the time consumption is long, and the effect is not obvious.
3. Sample data acquisition is difficult. In the prior art, comprehensive weather, cloud layer information and historical output power are mostly selected as input samples, the acquisition precision of the weather and cloud layer information is not high and expensive, and the weather and cloud layer information does not meet the economic operation requirement.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a second-level ultra-short-term photovoltaic power prediction method which is high in prediction precision, short in prediction time and high in universality, can achieve second-level prediction, and meanwhile, sample data is relatively easy to obtain. The method adopts an LSSVM algorithm to establish a prediction model, three times of optimization are adopted to optimize and optimize the hyper-parameters of the LSSVM prediction model, chaotic cube mapping is adopted to generate chaotic numbers with stronger randomness, a population position initialization calculation formula is optimized, the population position is initialized, a single iteration wolf algorithm is adopted to carry out first optimization on LSSVM algorithm parameters, optimization iteration time is shortened, improved iteration local search C-ILS is adopted for second optimization, a better adaptive value fixness obtained by the single iteration wolf optimization algorithm S-GWO is adopted for second optimization, an improved Griewank function is used as a disturbance function of the improved iteration local search C-ILS, the adaptive value is disturbed to obtain second optimization parameters, a local adaptive difference algorithm LSaDE is adopted to update the population position, and third optimization is carried out on the prediction model parameters. The invention provides a second-level ultrashort-term photovoltaic power prediction method aiming at the requirements of second-level ultrashort-term photovoltaic power prediction on time and precision, and the prediction effects of short prediction time and high prediction precision are realized.
The invention provides a second-level ultra-short-term photovoltaic power prediction method in three aspects of shortening prediction time, improving prediction precision and facilitating data sample acquisition.
The invention provides a second-level ultrashort-term photovoltaic power prediction method, which comprises the following steps:
s100: the method comprises the following steps of data acquisition, namely acquiring photovoltaic output power data from historical data of a photovoltaic power station before the predicted time at equal intervals, taking the photovoltaic output power data as sample data, and forming a one-dimensional array;
s200: building training data and test data, namely removing nighttime data from sample data, dividing the data into two parts, namely taking one part as the training data and taking the other part as the test data, and carrying out normalization processing on the data;
s300: constructing an LSSVM prediction model, wherein the LSSVM prediction model adopts an LSSVM algorithm, an RBF function is used as a kernel function, and a function estimate is selected as an algorithm type;
s400: the first optimization step of the hyperparameter of the LSSVM algorithm comprises the following steps:
s401: generation of chaotic numbers h for initialization using chaotic cube mappingiInitializing the population and locating the population
The initialization uses the following formula:
Position＝hi·(ub-lb) (1)
wherein the content of the first and second substances,
position is the initialized population Position;
hiis the chaos number generated by the chaos cubic mapping;
ubthe value of the hyper-parameter in the LSSVM algorithm is an upper bound;
lbthe value of the hyperparameter in the LSSVM algorithm is lower bound;
s402: carrying out first optimization on the hyper-parameters in the LSSVM algorithm by using a single iteration Husky optimization algorithm S-GWO to obtain a population initial adaptive value fitness and an initial population position Xpos；
S500: and a second optimization step of the hyperparameter of the LSSVM algorithm, wherein the second optimization is carried out on the preliminary population adaptive value fitness obtained in the step S400 by adopting improved iteration local search C-ILS (component-level-error-free) to obtain a second optimized population adaptive value betafitnessAnd optimal population position beterposThe improved iterative local search further reduces the search range of a better value on the basis of optimizing the search range for the first time;
s600: performing a third optimization step of the hyperparameter of the LSSVM algorithm, and adopting a local adaptive difference algorithm LSaDE to perform population adaptive value beter obtained after the second optimizationfitnessPerforming third optimization, updating the position of the population and obtaining the optimal adaptive value best of the populationfitnessAnd the corresponding optimal population position bestposThe local adaptive differential algorithm LSaDE further reduces the optimal value search range on the basis of optimizing the search range for the second time;
S700: determining the optimal value of the hyperparameter of the LSSVM algorithm according to the optimal adaptive value best of the populationfitnessDetermining an alpha wolf position, and acquiring optimal parameter information of the super-parameters of the LSSVM algorithm from the alpha wolf position information to serve as the optimal value of the super-parameters of the LSSVM algorithm;
s800: a prediction model training step, namely, bringing the obtained optimal value of the hyperparameter of the LSSVM algorithm into the LSSVM algorithm, inputting training data to train the LSSVM prediction model, and starting timing to obtain the trained LSSVM prediction model; s900: and a prediction model precision evaluation step, namely inputting test data, predicting the output power within T time, calculating the T time from the prediction moment, recording the prediction time T1 required by the prediction model from training to obtaining the prediction result, and calculating the error value of the prediction result and the test data, wherein the error value is used for evaluating the prediction precision of the prediction model.
Preferably, in step S500, an improved Griewank function is used as a perturbation function of the improved iterative local search C-ILS, a population preliminary adaptive value is perturbed, and a second optimization is performed, where an expression of the improved Griewank function is as follows:
xi＝Xpos*rand() (2)
Figure BDA0002809638900000031
wherein the content of the first and second substances,
i is a wolf population number, and the value is 0,1,2,., n;
n is the total population of wolfsbane;
rand () is a systematic random function, generating random numbers between [0,1 ];
xiis the randomized position of the ith wolf population;
g is a disturbance value of the population position adaptive value and is a one-dimensional array with the length of n.
Preferably, in step S500, the process of performing the second optimization by using the disturbance value includes the following steps:
randomizing the population position obtained by the S-GWO algorithm;
substituting the randomized population position into an improved Griewank function, and taking an output value as a disturbance adaptive value;
comparing the disturbance adaptive value with the population adaptive value obtained by the S-GWO algorithm, and selecting a smaller value as the betafitness；
Updating the position of the population to obtain the better position of the population betapos。
Preferably, the mathematical expression of the variant contraction factor of the locally adaptive differential algorithm LSaDE in step S600 is as follows:
Figure BDA0002809638900000041
Figure BDA0002809638900000042
Figure BDA0002809638900000043
wherein the content of the first and second substances,
dd is a self-defining parameter set for narrowing the search range;
max _ iteration is the maximum number of iterations;
f is the contraction factor of the current population;
F0is the initial shrinkage factor;
i is the current population number, and the numeric area is [1, sizepop ], wherein sizepop is a self-defined parameter called population number or population scale;
Figure BDA0002809638900000044
and
Figure BDA0002809638900000045
is a random vector of the current population i, i is a random vector of the current population i in the variation process2And i3Is [1, sizepop ]]Internal random inequality and not being an integer of i, sizepop is the population size;
betterpos: is the better population position obtained by the second optimization;
bestpos: and the optimal population position obtained by the third optimization.
Preferably, in step S600, a population adaptive value beta obtained by the second optimization is obtainedfitnessPerforming variation and cross operation, and updating the population position beterposThen comparing the current adaptive value with the betafitnessTo obtain the optimal population adaptive value bestfitnessAnd the corresponding optimal population position bestpos。
Preferably, in step S700, the obtained optimal population adaptive value best is determinedfitnessAnd sequencing, wherein the minimum adaptive value is used as the adaptive value of the alpha wolf of the optimal population, and the optimal value of the hyperparameter of the LSSVM algorithm is obtained from the position information of the alpha wolf of the optimal population.
Preferably, in S900, MSE, MAE and RMSE are simultaneously used for error evaluation, and the accuracy of the prediction model is evaluated, where the root mean square error RMSE is a measure of the deviation between the predicted value and the true value; the Mean Square Error (MSE) is the square of the difference value between the real value and the predicted value and then is summed and averaged; the mean absolute error MAE is an average of absolute errors.
Preferably, in S900, the MSE, MAE and MAPE are used together for error estimation to evaluate the accuracy of the prediction model.
Preferably, in step S100, the normalizing the data specifically includes:
the min-max normalization, also called dispersion normalization, is chosen to perform a linear transformation on the raw data, so that the resulting values are mapped between [0,1], and the transfer function is as follows:
Figure BDA0002809638900000051
wherein the content of the first and second substances,
xminis the minimum value of the sample data;
xmaxis the maximum value of the sample data;
x*is the normalized value.
Preferably, in S900, the effectiveness of the prediction model is also evaluated through the recorded prediction time t1, and when the prediction time t1 is smaller than a preset prediction time threshold t _ thred, the prediction is considered to meet the effectiveness requirement; and when the prediction time t1 is greater than or equal to a preset prediction time threshold t _ thred, the prediction model prediction is considered not to meet the effectiveness requirement, and the prediction model parameter optimization is carried out again.
Compared with the prior art, the invention has the following beneficial effects:
(1) the invention optimizes the repeated iteration process which consumes a large amount of time in the traditional wolf optimization algorithm (GWO), finds out relatively better population position and population adaptive value by adopting single iteration, simplifies the iteration process, shortens the optimization time and realizes the requirement of ultra-short-term power prediction on time;
(2) in order to make up for the loss of precision of single iteration in the wolf optimization algorithm, the second optimization adopts an iterative local search algorithm to further approach the searched LSSVM hyperparameter to the optimal value, so that the requirement on power prediction precision is realized;
(3) the improved iteration local search algorithm is used in the second optimization, disturbance is added on the basis of a better population position adaptive value obtained by a single iteration Hui wolf optimization algorithm (S-GWO), and an improved Griewank function is used as a disturbance function of the adaptive value, so that a better adaptive value is obtained by local search, and the requirement on power prediction precision is met;
(4) before the first optimization of the LSSVM algorithm parameters, the population is initialized, an initialized population position calculation formula is optimized, a brand new calculation formula is used on the basis that chaotic numbers with randomness and uniformity are generated by using chaotic cube mapping, a better initialized population position is obtained, and the requirement on power prediction precision is met;
(5) the invention reduces the search factor to enhance the local search capability in order to realize the local search in a smaller range on the basis of obtaining a better seed group position by iterative local search. In the third optimization, a local adaptive differential algorithm (LSaDE) is selected, precision re-optimization is carried out, and the optimal population position and the population adaptive value are obtained.
Drawings
FIG. 1 is a graph of an adaptation search plot according to an embodiment of the present invention;
FIG. 2 is a graph of actual data fit to predicted data for one embodiment of the present invention;
FIG. 3 is a prediction flow diagram for one embodiment of the present invention;
FIG. 4 is a flowchart of the prediction model hyper-parameter optimization process according to an embodiment of the present invention, which is a specific process of the LLSVM algorithm hyper-parameter optimization in the prediction process;
FIG. 5 is a prediction flow diagram for one embodiment of the present invention;
FIG. 6 is pseudo code for a predictive model hyper-parameter optimization process implementation according to an embodiment of the invention.
Detailed Description
The following detailed description of the embodiments of the present invention is provided in conjunction with the accompanying drawings of fig. 1-5.
The invention provides a second-level ultrashort-term photovoltaic power prediction method, which comprises the following steps:
s100: the method comprises the following steps of data acquisition, namely acquiring photovoltaic output power data from historical data of a photovoltaic power station before the predicted time at equal intervals, taking the photovoltaic output power data as sample data, and forming a one-dimensional array;
because the numerical weather forecast and the solar irradiance of the third-party organization are reported according to regions, certain errors exist between the numerical weather forecast and the solar irradiance of the specific photovoltaic power station location, and sample data with high precision is difficult to obtain when the meteorological data is in the real-time change process, the historical output power of the photovoltaic power station, which is relatively easy to obtain, is selected as the sample data.
S200: building training data and test data, namely removing nighttime data from sample data, dividing the data into two parts, namely taking one part as the training data and taking the other part as the test data, and carrying out normalization processing on the data;
the photovoltaic power generation needs enough illumination conditions, so the photovoltaic power station can not generate power at night, and the output power of the photovoltaic power station has day and night alternation. According to the characteristics of photovoltaic power generation, output power data sampled at equal intervals in the time period from sunrise to sunset are taken as sample data.
S300: constructing an LSSVM prediction model, wherein the LSSVM prediction model adopts an LSSVM algorithm, an RBF function is used as a kernel function, and a function estimate is selected as an algorithm type; and selecting a regularization parameter C and a kernel parameter sigma by subsequent super-parameter optimization of the LSSVM algorithm. The regularization parameter C and the kernel parameter σ of the LSSVM algorithm are contained in the population optimal position information of the GWO algorithm, the initial parameter definition of the GWO algorithm has a dimension dim value, and there are two parameters that the LSSVM needs to optimize, so that dim is defined to be 2.
The kernel function can effectively process the nonlinear problem, original data are mapped to a Hilbert space and are learned in a high-dimensional space, the kernel function is used for calculating a high-dimensional dot product in a low dimension, and the space is generated by mapping sample data from the low dimension to the high dimension. The invention selects a radial basis function RBF as a kernel function of the invention. The expression of the RBF function is:
Figure BDA0002809638900000071
wherein the content of the first and second substances,
K(x,xc) Is the output of the RBF function;
x is any point in space;
xcis the kernel function center;
σ is a nuclear parameter;
the LSSVM algorithm has two function models, and the function estimation is a regression model which is mostly used for regression fitting; classification is a classification model which is used for classification, and the classification model adopts function estimation as an algorithm type.
S400: the first optimization step of the hyperparameter of the LSSVM algorithm comprises the following steps:
in the group intelligent algorithm, the initialization processing is carried out on the group, the initial group can be prevented from being excessively concentrated, the group convergence is accelerated, and the accuracy of the final solution is improved. The method realizes the rapid relative optimization of the hyper-parameter regularization parameter C and the kernel parameter sigma influencing the prediction precision in the LSSVM algorithm by using a single iteration Hui wolf optimization algorithm (S-GWO), and finds out a relatively better population position and a relatively better adaptive value thereof.
In the traditional grey wolf optimization algorithm (GWO), the characteristics of initialization randomness and uniformity of the population are not obviously characterized, and the risk that the algorithm falls into local optimization is increased. The gray wolf optimization algorithm (GWO) has the disadvantages of random initialization, non-ideal population distribution and non-ideal optimization result. The better the randomness of the initial population, the faster the convergence speed, and the higher the final solution precision, the less likely to fall into local optimality. The invention improves the original population position formula, and replaces the pseudo-random number in the original initialization with the chaotic number generated by chaotic cubic mapping, thereby obtaining a more random initial population position.
S401: generation of chaotic numbers h for initialization using chaotic cube mappingiInitializing a population, wherein the initialization of the population position adopts the following formula:
Position＝hi·(ub-lb) (1)
wherein the content of the first and second substances,
position is the initialized population Position;
hiis the chaos number generated by the chaos cubic mapping;
ubthe value of the hyper-parameter in the LSSVM algorithm is an upper bound;
lbthe value of the hyperparameter in the LSSVM algorithm is lower bound;
ubthe value of the hyper-parameter in the LSSVM algorithm is an upper bound, and in specific application, the following values can be obtained: u. ofb＝[1000,1000]The definition of the upper bound parameter is not unique, and the limit can be larger;
lbthe value of the hyper-parameter in the LSSVM algorithm is the lower bound, and in specific application, the following values can be obtained: lb＝[0.01,0.01]The definition of the lower bound parameter is not unique, and the limit can be larger;
sizepop is the population number;
dim is the number of the optimized parameters; since we need to optimize both the regularization parameter C and the kernel parameter σ, dim is taken to be 2, with an upper and lower bound for each parameter.
S402: carrying out first optimization on the hyper-parameters in the LSSVM algorithm by using a single iteration Husky optimization algorithm S-GWO to obtain a population initial adaptive value fitness and an initial population position Xpos；
The chaotic mapping can also be put into the first optimization, the chaotic mapping is the optimization of the initial wolf population, and the optimization result of the single iteration wolf optimization algorithm S-GWO is the population initial adaptive value fitness and the initial population position Xpos。
In the traditional gray wolf optimization algorithm (GWO), the whole optimization process needs repeated iterative search of the gray wolf population, and a large amount of time is consumed in the iterative process. In order to meet the requirement of second-level ultra-short-term power prediction on time, the invention only carries out one-time search on the gray wolf optimization algorithm (GWO), and provides a single-iteration gray wolf optimization algorithm (S-GWO), thereby greatly shortening the search time of the super-parameter of the LSSVM algorithm.
S500: and a second optimization step of the hyperparameter of the LSSVM algorithm, wherein the second optimization is carried out on the preliminary population adaptive value fitness obtained in the step S400 by adopting improved iteration local search C-ILS (component-level-error-free) to obtain a second optimized population adaptive value betafitnessAnd optimal population position beterposThe improved iterative local search further reduces the search range of a better value on the basis of optimizing the search range for the first time;
in order to compensate for the prediction precision loss caused by single iteration in the step S400, the invention adopts improved iteration local search (C-ILS) to re-optimize the better adaptive value obtained by the single iteration gray wolf optimization algorithm (S-GWO) in the second optimization, and takes the improved Griewank function as the iteration local search (I)LS) and the adaptive value of the disturbance, so that the local optimum is skipped, and the optimal adaptive value beta is foundfitnessAnd better population position beterpos。
S600: performing a third optimization step of the hyperparameter of the LSSVM algorithm, and adopting a local adaptive difference algorithm LSaDE to perform population adaptive value beter obtained after the second optimizationfitnessPerforming third optimization, updating the position of the population and obtaining the optimal adaptive value best of the populationfitnessAnd the corresponding optimal population position bestposThe local adaptive differential algorithm LSaDE further reduces the optimal value search range on the basis of optimizing the search range for the second time, and performs optimal value search in a very small local range;
in order to further improve the searching speed, the invention adopts a local self-adaptive differential algorithm LSaDE for reducing the optimal value searching range to carry out the third optimization, and the local self-adaptive differential algorithm LSaDE is an improvement of the traditional self-adaptive differential algorithm, thereby greatly reducing the searching range and providing the optimizing speed.
In the third optimization process, each optimization is based on the previous optimization, an improved algorithm is adopted, the search range is further narrowed based on the previous optimization search range, and the search range is narrowed to a minimum range in the third optimization, so that the search speed is greatly improved.
The optimal parameter is obtained by comparing population adaptive values, the population position and the population adaptive value are obtained through multiple times of optimization, the adaptive value is compared every time, the value with the minimum population adaptive value is selected, and the value corresponding to the minimum population adaptive value is the optimal population position.
S700: determining the optimal value of the hyperparameter of the LSSVM algorithm according to the optimal adaptive value best of the populationfitnessDetermining an alpha wolf position, and acquiring optimal parameter information of the super-parameters of the LSSVM algorithm from the alpha wolf position information to serve as the optimal value of the super-parameters of the LSSVM algorithm;
from the definition of the grey wolf optimization algorithm (GWO), it can be determined that the best parameter information of the LSSVM algorithm is hidden in the position information of the alpha wolf of the best population, so that the position information of the alpha wolf is determined, and the optimal value of the hyper-parameter of the LSSVM algorithm is also determined.
S800: a prediction model training step, namely, bringing the obtained optimal value of the hyperparameter of the LSSVM algorithm into the LSSVM algorithm, inputting training data to train the LSSVM prediction model, and starting timing to obtain the trained LSSVM prediction model;
the method divides the sample data into two groups of training data and testing data, trains the training data through a prediction algorithm to further obtain prediction data, limits the number of output values of the prediction data to be equal to that of the testing data, and facilitates comparison of prediction effects. The degree of deviation of the prediction data decreases as the time from the prediction time point increases.
S900: and a prediction model precision evaluation step, namely inputting test data, predicting the output power within T time, calculating the T time from the prediction moment, recording the prediction time T1 required by the prediction model from training to obtaining the prediction result, and calculating the error value of the prediction result and the test data, wherein the error value is used for evaluating the prediction precision of the prediction model.
As a preferred embodiment, in step S500, an improved Griewank function is used as a perturbation function of an improved iterative local search C-ILS, a population preliminary adaptive value is perturbed, and a second optimization is performed, where an expression of the improved Griewank function is as follows:
xi＝Xpos*rand() (2)
Figure BDA0002809638900000101
wherein the content of the first and second substances,
i is a wolf population number, and the value is 0,1,2,., n;
n is the total population of wolfsbane;
rand () is a systematic random function, generating random numbers between [0,1 ];
xiis the randomized position of the ith wolf population;
g is a disturbance value of the population position adaptive value and is a one-dimensional array with the length of n.
The Griewank function is a common adaptive perturbation function, and the original formula is as follows:
Figure BDA0002809638900000102
the invention improves the original Griewank function, and aims to jump out local optimum and search an optimum value in a larger range. (3) The right most positive sign of the formula in the equation is to unify the adaptation value to the negative half-axis of the y-axis. 4000 is defined by the original function, the value can also change, but the coefficient change is too large, so that the iterative local search is easy to jump out of the search range. Therefore, the invention only increases 1.1 times of multiplication in the formula (3), wherein 1.1 is the better result obtained according to the test, and of course, according to the result of the actual test, other values can be used to replace 1.1 to obtain better effect.
Besides the improved Griewank function adopted by the invention, disturbance functions such as a Rastrigin function, a Schafer function, an Ackley function, a Rosenbrock function and the like and improved functions of the functions can be used for disturbance.
And adding disturbance on the basis of the global better solution obtained by single global search, and carrying out local re-optimization on the global better solution. The basic idea of Iterative Local Search (ILS) is that there is some commonality between the preferred solutions. The iterative local search is the initial optimization of the LSSVM algorithm with super-parameter precision, and a better adaptive value is searched in a relatively larger range. The purpose of the improved Griewank function here is to focus the search for the optimal solution to a smaller extent.
As a preferred embodiment, in step S500, the process of performing the second optimization by using the disturbance value includes the following steps:
randomizing the population position obtained by the S-GWO algorithm;
substituting the randomized population position into an improved Griewank function, and taking an output value as a disturbance adaptive value;
comparing the disturbance adaptive value with the population adaptive value obtained by the S-GWO algorithm, and selecting a smaller value as the betafitness；
Updating the position of the population to obtain the better position of the population betapos。
As a preferred embodiment, the mathematical expression of the variant contraction factor of the locally adaptive differential algorithm LSaDE described in step S600 is as follows:
Figure BDA0002809638900000111
Figure BDA0002809638900000112
Figure BDA0002809638900000113
wherein the content of the first and second substances,
dd is a self-defining parameter set for narrowing the search range;
max _ iteration is the maximum number of iterations;
f is the contraction factor of the current population;
F0is the initial shrinkage factor;
i is the current population number, and the numeric area is [1, sizepop ], wherein sizepop is a self-defined parameter called population number or population scale;
Figure BDA0002809638900000114
and
Figure BDA0002809638900000115
is a random vector of the current population i, i is a random vector of the current population i in the variation process2And i3Is [1, sizepop ]]The internal random is not equal and is not an integer of i, and sizepop is the size of the population;
betterpos: is the better population position obtained by the second optimization;
bestpos: and the optimal population position obtained by the third optimization.
The adaptive value beter obtained after the initial optimization of the local adaptive difference algorithm LSaDE pairfitnessAnd (4) re-optimizing, and performing local search again on the basis of improved Iterative Local Search (ILS), so that the method is an optimal solution approximation process in a small range.
The differential evolution algorithm (DE) has the characteristics of strong local mining capability, high convergence speed and good stability, but the traditional differential evolution algorithm (DE) is easy to fall into local optimization and generates premature convergence. Although premature convergence can be adequately addressed by increasing the population, this increases the processor computation time, contrary to the time requirements for ultra-short-term power prediction in seconds.
Aiming at the advantages of the differential evolution algorithm (DE) in the local mining capability and the problem of premature convergence, the invention adopts the local self-adaptive differential algorithm (LSaDE) to update the population position to obtain the optimal population position and the corresponding optimal adaptive value. The method further reduces the search range and improves the search speed.
As a preferred embodiment, in step S600, the population adaptive value better obtained by the second optimizationfitnessPerforming variation and cross operation, and updating the population position beterposThen comparing the current adaptive value with the betafitnessTo obtain the optimal population adaptive value bestfitnessAnd the corresponding optimal population position bestpos。
In a preferred embodiment, in step S700, the obtained optimal population adaptive value best is comparedfitnessAnd sequencing, wherein the minimum adaptive value is used as the adaptive value of the alpha wolf of the optimal population, and the optimal value of the hyperparameter of the LSSVM algorithm is obtained from the position information of the alpha wolf of the optimal population.
As a preferred embodiment, in S900, MSE, MAE and RMSE are simultaneously used for error evaluation, and the accuracy of the prediction model is evaluated, where the root-mean-square error RMSE is a measure of the deviation between the predicted value and the true value; the Mean Square Error (MSE) is the square of the difference value between the real value and the predicted value and then is summed and averaged; the mean absolute error MAE is an average of absolute errors.
In S900, the MSE, MAE and MAPE are used to evaluate the accuracy of the prediction model.
As a preferred embodiment, in step S100, the normalizing the data specifically includes:
the min-max normalization, also called dispersion normalization, is chosen to perform a linear transformation on the raw data, so that the resulting values are mapped between [0,1], and the transfer function is as follows:
Figure BDA0002809638900000121
wherein the content of the first and second substances,
xminis the minimum value of the sample data;
xmaxis the maximum value of the sample data;
x*is the normalized value.
In S900, the effectiveness of the prediction model is also evaluated according to the recorded prediction time t1, and when the prediction time t1 is less than the preset prediction time threshold t _ thred, the prediction is considered to meet the effectiveness requirement; and when the prediction time t1 is greater than or equal to a preset prediction time threshold t _ thred, the prediction model prediction is considered not to meet the effectiveness requirement, and the prediction model parameter optimization is carried out again.
Example 1
The second-level ultra-short-term photovoltaic power prediction method of the present invention is described in detail below according to an embodiment of the present invention.
The invention provides a second-level ultrashort-term photovoltaic power prediction method, which comprises the following steps:
s100: the method comprises the following steps of data acquisition, namely acquiring photovoltaic output power data from historical data of a photovoltaic power station before the predicted time at equal intervals, taking the photovoltaic output power data as sample data, and forming a one-dimensional array;
in step S100, the normalization processing of the data specifically includes:
the min-max normalization, also called dispersion normalization, is chosen to perform a linear transformation on the raw data, so that the resulting values are mapped between [0,1], and the transfer function is as follows:
Figure BDA0002809638900000131
wherein the content of the first and second substances,
xminis the minimum value of the sample data;
xmaxis the maximum value of the sample data;
x*is the normalized value.
S200: building training data and test data, namely removing nighttime data from sample data, dividing the data into two parts, namely taking one part as the training data and taking the other part as the test data, and carrying out normalization processing on the data;
s300: constructing an LSSVM prediction model, wherein the LSSVM prediction model adopts an LSSVM algorithm, an RBF function is used as a kernel function, and a function estimate is selected as an algorithm type;
s400: the first optimization step of the hyperparameter of the LSSVM algorithm comprises the following steps:
s401: generation of chaotic numbers h for initialization using chaotic cube mappingiInitializing a population, wherein the initialization of the population position adopts the following formula:
Position＝hi·(ub-lb) (1)
wherein the content of the first and second substances,
position is the initialized population Position;
hiis the chaos number generated by the chaos cubic mapping;
ubthe value of the hyper-parameter in the LSSVM algorithm is an upper bound;
lbthe value of the hyperparameter in the LSSVM algorithm is lower bound;
s402: carrying out first optimization on the hyper-parameters in the LSSVM algorithm by using a single iteration Husky optimization algorithm S-GWO to obtain a population initial adaptive value fitness and an initial population position Xpos；
S500: second optimization step of hyperparameter of LSSVM algorithmStep two, adopting improved iteration local search C-ILS to carry out second optimization on the preliminary population adaptive value fitness obtained in the step S400 to obtain a second optimized population adaptive value betafitnessAnd optimal population position beterposThe improved iterative local search further reduces the search range of a better value on the basis of optimizing the search range for the first time;
in step S500, an improved Griewank function is used as a perturbation function of the improved iterative local search C-ILS, a population preliminary adaptive value is perturbed, and a second optimization is performed, where an expression of the improved Griewank function is as follows:
xi＝Xpos*rand() (2)
Figure BDA0002809638900000141
wherein the content of the first and second substances,
i is a wolf population number, and the value is 0,1,2,., n;
n is the total population of wolfsbane;
rand () is a systematic random function, generating random numbers between [0,1 ];
xiis the randomized position of the ith wolf population;
g is a disturbance value of the population position adaptive value and is a one-dimensional array with the length of n.
In step S500, the process of performing the second optimization using the perturbation value includes the following steps:
randomizing the population position obtained by the S-GWO algorithm;
substituting the randomized population position into an improved Griewank function, and taking an output value as a disturbance adaptive value;
comparing the disturbance adaptive value with the population adaptive value obtained by the S-GWO algorithm, and selecting a smaller value as the betafitness；
Updating the position of the population to obtain the better position of the population betapos。
S600: a third optimization step of the hyperparameter of the LSSVM algorithm, which adopts local selfAdapting to the population adaptive value beta obtained after the second optimization of the differential algorithm LSaDEfitnessPerforming third optimization, updating the position of the population and obtaining the optimal adaptive value best of the populationfitnessAnd the corresponding optimal population position, the local self-adaptive difference algorithm LSaDE further reduces the optimal value search range on the basis of the second time of optimizing the search range;
the mathematical expression of the variant contraction factor of the local adaptive difference algorithm LSaDE in the step S600 is as follows:
Figure BDA0002809638900000151
Figure BDA0002809638900000152
Figure BDA0002809638900000153
wherein the content of the first and second substances,
dd is a self-defining parameter set for narrowing the search range;
max _ iteration is the maximum number of iterations;
f is the contraction factor of the current population;
F0is the initial shrinkage factor;
i is the current population number, and the numeric area is [1, sizepop ], wherein sizepop is a self-defined parameter called population number or population scale;
Figure BDA0002809638900000154
and
Figure BDA0002809638900000155
is a random vector of the current population i, i is a random vector of the current population i in the variation process2And i3Is [1, sizepop ]]The internal random is not equal and is not an integer of i, and sizepop is the size of the population;
betterpos: is the better population position obtained by the second optimization;
bestpos: and the optimal population position obtained by the third optimization.
In step S600, a population adaptive value beta obtained by the second optimizationfitnessPerforming variation and cross operation, and updating the population position beterposThen comparing the current adaptive value with the betafitnessTo obtain the optimal population adaptive value bestfitnessAnd the corresponding optimal population position bestpos。
S700: determining the optimal value of the hyperparameter of the LSSVM algorithm according to the optimal adaptive value best of the populationfitnessDetermining an alpha wolf position, and acquiring optimal parameter information of the super-parameters of the LSSVM algorithm from the alpha wolf position information to serve as the optimal value of the super-parameters of the LSSVM algorithm;
in step S700, the obtained optimal population adaptive value best is subjected tofitnessAnd sequencing, wherein the minimum adaptive value is used as the adaptive value of the alpha wolf of the optimal population, and the optimal value of the hyperparameter of the LSSVM algorithm is obtained from the position information of the alpha wolf of the optimal population.
S800: a prediction model training step, namely, bringing the obtained optimal value of the hyperparameter of the LSSVM algorithm into the LSSVM algorithm, inputting training data to train the LSSVM prediction model, and starting timing to obtain the trained LSSVM prediction model; s900: and a prediction model precision evaluation step, namely inputting test data, predicting the output power within T time, calculating the T time from the prediction moment, recording the prediction time T1 required by the prediction model from training to obtaining the prediction result, and calculating the error value of the prediction result and the test data, wherein the error value is used for evaluating the prediction precision of the prediction model.
In S900, MSE, MAE and RMSE are adopted to carry out error evaluation, the precision of the prediction model is evaluated, and the root mean square error RMSE is the deviation between the predicted value and the true value; the Mean Square Error (MSE) is the square of the difference value between the real value and the predicted value and then is summed and averaged; the mean absolute error MAE is an average of absolute errors.
Or, if necessary, in S900, MSE, MAE and MAPE are simultaneously used for error evaluation to evaluate the accuracy of the prediction model.
In S900, evaluating the effectiveness of the prediction model through the recorded prediction time t1, and when the prediction time t1 is smaller than a preset prediction time threshold t _ thred, considering that the prediction meets the effectiveness requirement; and when the prediction time t1 is greater than or equal to a preset prediction time threshold t _ thred, the prediction model prediction is considered not to meet the effectiveness requirement, and the prediction model parameter optimization is carried out again.
Example 2
The effect of the second-level ultra-short-term photovoltaic power prediction method according to the present invention is described in detail below according to an embodiment of the present invention.
Fig. 1 shows a variation curve of the adaptive value in the process of the parameter optimization process of the LLSVM algorithm of the present invention, and it can be seen from the figure that since a large number of iteration processes are cancelled, the population adaptive value does not show the characteristic of increasing and decreasing in a single iteration, and the global search and the two local searches of the single iteration are adopted, the adaptive value of the population is changed greatly, and the optimal solution is searched in different search ranges.
Example 3
The effect of the second-level ultra-short-term photovoltaic power prediction method according to the present invention is described in detail below according to an embodiment of the present invention.
Fig. 2 is a fitting graph of the prediction result and the actual data by using the prediction model in this embodiment. In this embodiment, the sample data is divided into training data and test data, and the test data is not involved in the prediction algorithm training and is only used for comparison with the prediction data. The test data uses 48 samples and five minutes as time nodes, and in practical application, the sampling interval can be shortened to 15-60 seconds by referring to the prediction time.
Example 4
The effect of the second-level ultra-short-term photovoltaic power prediction method according to the present invention is described in detail below according to an embodiment of the present invention.
The basic weather information of the 9 th 6 th to 10 th day is given in table 1.
TABLE 1 basic weather information from 9 months, 6 days to 10 days in certain places
Figure BDA0002809638900000171
The following table 2 shows sample data collected under the above conditions, and the evaluation indexes and the prediction running times of prediction precisions of different prediction models are adopted to embody the effects brought by a series of methods for quickly and accurately determining the optimal population adaptive value, such as a local adaptive differential algorithm LSaDE (local adaptive differential algorithm) and the like, in a Huilles algorithm adopted by the invention, a single iteration is adopted, a chaotic number generated by chaotic cubic mapping is initialized for a population position and is used as a random number, a second optimization is carried out by an iterative local search algorithm, and the like.
TABLE 2 prediction accuracy evaluation index and prediction time of each prediction model
Figure BDA0002809638900000172
As can be seen from Table 2, the GWO-LSSVM algorithm which adopts the traditional GWO algorithm to carry out parameter optimization on the LSSVM algorithm has poor prediction accuracy, is easy to fall into local optimization and has long prediction time.
Aiming at the problem of poor prediction accuracy, the SaDE algorithm is selected to optimize the GWO-LSSVM algorithm which optimizes the LSSVM algorithm by the traditional GWO algorithm, so that the SaDE-GWO-LSSVM algorithm is obtained, the prediction accuracy is high, but the prediction time is long due to repeated iteration of the traditional GWO algorithm.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention shall fall within the protection scope of the present invention.

Claims (10)

1. A second-level ultra-short-term photovoltaic power prediction method is characterized by comprising the following steps:
s100: the method comprises the following steps of data acquisition, namely acquiring photovoltaic output power data from historical data of a photovoltaic power station before the predicted time at equal intervals, taking the photovoltaic output power data as sample data, and forming a one-dimensional array;
s200: building training data and test data, namely removing nighttime data from sample data, dividing the data into two parts, namely taking one part as the training data and taking the other part as the test data, and carrying out normalization processing on the data;
s300: constructing an LSSVM prediction model, wherein the LSSVM prediction model adopts an LSSVM algorithm, an RBF function is used as a kernel function, and a function estimate is selected as an algorithm type;
s400: the first optimization step of the hyperparameter of the LSSVM algorithm comprises the following steps:
s401: generation of chaotic numbers h for initialization using chaotic cube mappingiInitializing a population, wherein the initialization of the population position adopts the following formula:
Position＝hi·(ub-lb) (1)
wherein the content of the first and second substances,
position is the initialized population Position;
hiis the chaos number generated by the chaos cubic mapping;
ubthe value of the hyper-parameter in the LSSVM algorithm is an upper bound;
lbthe value of the hyperparameter in the LSSVM algorithm is lower bound;
s402: carrying out first optimization on the hyper-parameters in the LSSVM algorithm by using a single iteration Husky optimization algorithm S-GWO to obtain a population initial adaptive value fitness and an initial population position Xpos；
S500: and a second optimization step of the hyperparameter of the LSSVM algorithm, wherein the second optimization is carried out on the preliminary population adaptive value fitness obtained in the step S400 by adopting improved iteration local search C-ILS (component-level-error-free) to obtain a second optimized population adaptive value betafitnessAnd better population position beterposThe improved iterative local search is further based on the first optimized search rangeThe more optimal value searching range is narrowed;
s600: performing a third optimization step of the hyperparameter of the LSSVM algorithm, and adopting a local adaptive difference algorithm LSaDE to perform population adaptive value beter obtained after the second optimizationfitnessPerforming third optimization, updating the position of the population and obtaining the optimal adaptive value best of the populationfitnessAnd the corresponding optimal population position bestposThe local adaptive differential algorithm LSaDE further reduces the optimal value search range on the basis of optimizing the search range for the second time;
s700: determining the optimal value of the hyperparameter of the LSSVM algorithm according to the optimal adaptive value best of the populationfitnessSequencing and determining alpha wolf positions, and acquiring optimal parameter information of the super parameters of the LSSVM algorithm from the alpha wolf position information to serve as optimal values of the super parameters of the LSSVM algorithm;
s800: a prediction model training step, namely, bringing the obtained optimal value of the hyperparameter of the LSSVM algorithm into an LSSVM prediction model, inputting training data to train the LSSVM prediction model, and starting timing to obtain a trained LSSVM prediction model;
s900: and a prediction model precision evaluation step, namely inputting test data, predicting the output power within T time, calculating the T time from the prediction moment, recording the prediction time T1 required by the prediction model from training to obtaining the prediction result, and calculating the error value of the prediction result and the test data, wherein the error value is used for evaluating the prediction precision of the prediction model.
2. The method for predicting the second-level ultrashort-term photovoltaic power as claimed in claim 1, wherein in step S500, an improved Griewank function is adopted as a perturbation function of an improved iterative local search C-ILS, a population preliminary adaptive value is perturbed, and a second optimization is performed, wherein an expression of the improved Griewank function is as follows:
xi＝Xpos*rand() (2)
Figure FDA0002809638890000021
wherein the content of the first and second substances,
i is a wolf population number, and the value is 0,1,2,., n;
n is the total population of wolfsbane;
rand () is a systematic random function, generating random numbers between [0,1 ];
xiis the randomized position of the ith wolf population;
g is a disturbance value of the population position adaptive value and is a one-dimensional array with the length of n.
3. The method for predicting the second-level ultrashort-term photovoltaic power as claimed in claim 2, wherein the step S500 of performing the second optimization by using the perturbation value comprises the following steps:
randomizing the population position obtained by the S-GWO algorithm;
substituting the randomized population position into an improved Griewank function, and taking an output value as a disturbance adaptive value;
comparing the disturbance adaptive value with the population adaptive value obtained by the S-GWO algorithm, and selecting a smaller value as the betafitness；
Updating the position of the population to obtain the better position of the population betapos。
4. The method for predicting the second-level ultrashort-term photovoltaic power as claimed in claim 1, wherein the mathematical expression of the variation contraction factor of the locally adaptive differential algorithm LSaDE in the step S600 is as follows:
Figure FDA0002809638890000031
Figure FDA0002809638890000032
Figure FDA0002809638890000033
wherein the content of the first and second substances,
dd is a self-defining parameter set for narrowing the search range;
max _ iteration is the maximum number of iterations;
f is the contraction factor of the current population;
F0is the initial shrinkage factor;
i is the current population number, and the numeric area is [1, sizepop ], wherein sizepop is a self-defined parameter called population number or population scale;
Figure FDA0002809638890000034
and
Figure FDA0002809638890000035
is a random vector of the current population i, i is a random vector of the current population i in the variation process2And i3Is [1, sizepop ]]The internal random is not equal and is not an integer of i, and sizepop is the size of the population;
betterpos: is the better population position obtained by the second optimization;
bestpos: and the optimal population position obtained by the third optimization.
5. The method for predicting second-level ultrashort-term photovoltaic power as claimed in claim 4, wherein in step S600, the population adaptive value beta obtained by the second optimizationfitnessPerforming variation and cross operation, and updating the population position beterposThen comparing the current adaptive value with the betafitnessTo obtain the optimal population adaptive value bestfitnessAnd the corresponding optimal population position bestpos。
6. The method for predicting the second-level ultrashort-term photovoltaic power as claimed in claim 1, wherein in step S700, the best population adaptive value best is obtainedfitnessSorting is performed, wherein the minimum adaptation valueAnd the adaptive value of the alpha wolf of the optimal population is used, and the optimal value of the hyperparameter of the LSSVM algorithm is obtained from the position information of the alpha wolf of the optimal population.
7. The second-level ultrashort-term photovoltaic power prediction method of claim 1, wherein in S900, MSE, MAE and RMSE are simultaneously used for error evaluation to evaluate the accuracy of the prediction model, and the root-mean-square error RMSE is a measure of the deviation between the predicted value and the true value; the Mean Square Error (MSE) is the square of the difference value between the real value and the predicted value and then is summed and averaged; the mean absolute error MAE is an average of absolute errors.
8. The method for predicting the second-level ultrashort-term photovoltaic power as claimed in claim 1, wherein in S900, the MSE, MAE and MAPE are simultaneously used for error estimation to evaluate the accuracy of the prediction model.
9. The method for predicting the second-level ultrashort-term photovoltaic power as claimed in claim 1, wherein in the step S100, the normalization processing of the data specifically comprises:
the min-max normalization, also called dispersion normalization, is chosen to perform a linear transformation on the raw data, so that the resulting values are mapped between [0,1], and the transfer function is as follows:
Figure FDA0002809638890000041
wherein the content of the first and second substances,
xminis the minimum value of the sample data;
xmaxis the maximum value of the sample data;
x*is the normalized value.
10. The second-level ultrashort-term photovoltaic power prediction method of claim 1, wherein in S900, effectiveness of the prediction model is further evaluated through the recorded prediction time t1, and when the prediction time t1 is smaller than a preset prediction time threshold t _ thred, the prediction is considered to meet the requirement of effectiveness; and when the prediction time t1 is greater than or equal to a preset prediction time threshold t _ thred, the prediction model prediction is considered not to meet the effectiveness requirement, and the prediction model parameter optimization is carried out again.
CN202011385741.8A 2020-12-01 2020-12-01 Second-level ultra-short-term photovoltaic power prediction method Active CN112507613B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011385741.8A CN112507613B (en) 2020-12-01 2020-12-01 Second-level ultra-short-term photovoltaic power prediction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011385741.8A CN112507613B (en) 2020-12-01 2020-12-01 Second-level ultra-short-term photovoltaic power prediction method

Publications (2)

Publication Number Publication Date
CN112507613A true CN112507613A (en) 2021-03-16
CN112507613B CN112507613B (en) 2022-05-10

Family

ID=74969784

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011385741.8A Active CN112507613B (en) 2020-12-01 2020-12-01 Second-level ultra-short-term photovoltaic power prediction method

Country Status (1)

Country Link
CN (1) CN112507613B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114662393A (en) * 2022-03-25 2022-06-24 湖南工程学院 Photovoltaic power prediction error compensation method based on weather conditions
CN114706445A (en) * 2022-03-10 2022-07-05 三峡大学 Photovoltaic maximum power point tracking method based on DE-GWO algorithm

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104978613A (en) * 2015-07-28 2015-10-14 国家电网公司 Short-period forecasting method for photovoltaic output in consideration of assembly temperature
US20180046924A1 (en) * 2015-08-31 2018-02-15 Guangzhou Institute Of Energy Conversion, Chinese Academy Of Sciences Whole-life-cycle power output classification prediction system for photovoltaic systems
CN107992645A (en) * 2017-10-30 2018-05-04 嘉兴学院 Sewage disposal process soft-measuring modeling method based on chaos-fireworks hybrid algorithm
CN108510074A (en) * 2018-05-30 2018-09-07 江苏理工学院 A kind of implementation method for improving GWO algorithms
CN109635506A (en) * 2019-01-10 2019-04-16 浙江大学 The photovoltaic cell model parameter identification method of adaptive chaos tree and seed algorithm
KR20190123040A (en) * 2018-04-23 2019-10-31 한국전자통신연구원 Solar power generation prediction model management APPARATUS AND METHOD for setting the renewal cycle of the deep run-based solar power generation prediction model
CN110942205A (en) * 2019-12-05 2020-03-31 国网安徽省电力有限公司 Short-term photovoltaic power generation power prediction method based on HIMVO-SVM
CN111709524A (en) * 2020-07-03 2020-09-25 江苏科技大学 RBF neural network optimization method based on improved GWO algorithm

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104978613A (en) * 2015-07-28 2015-10-14 国家电网公司 Short-period forecasting method for photovoltaic output in consideration of assembly temperature
US20180046924A1 (en) * 2015-08-31 2018-02-15 Guangzhou Institute Of Energy Conversion, Chinese Academy Of Sciences Whole-life-cycle power output classification prediction system for photovoltaic systems
CN107992645A (en) * 2017-10-30 2018-05-04 嘉兴学院 Sewage disposal process soft-measuring modeling method based on chaos-fireworks hybrid algorithm
KR20190123040A (en) * 2018-04-23 2019-10-31 한국전자통신연구원 Solar power generation prediction model management APPARATUS AND METHOD for setting the renewal cycle of the deep run-based solar power generation prediction model
CN108510074A (en) * 2018-05-30 2018-09-07 江苏理工学院 A kind of implementation method for improving GWO algorithms
CN109635506A (en) * 2019-01-10 2019-04-16 浙江大学 The photovoltaic cell model parameter identification method of adaptive chaos tree and seed algorithm
CN110942205A (en) * 2019-12-05 2020-03-31 国网安徽省电力有限公司 Short-term photovoltaic power generation power prediction method based on HIMVO-SVM
CN111709524A (en) * 2020-07-03 2020-09-25 江苏科技大学 RBF neural network optimization method based on improved GWO algorithm

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZHAO ZHENXING,ET AL.: "Analysis,Design,and Implementation of Improved LLC Resonant Transformer for Efficiency Enhancement", 《ENERGIES》 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114706445A (en) * 2022-03-10 2022-07-05 三峡大学 Photovoltaic maximum power point tracking method based on DE-GWO algorithm
CN114706445B (en) * 2022-03-10 2023-08-01 三峡大学 DE-GWO algorithm-based photovoltaic maximum power point tracking method
CN114662393A (en) * 2022-03-25 2022-06-24 湖南工程学院 Photovoltaic power prediction error compensation method based on weather conditions
CN114662393B (en) * 2022-03-25 2023-06-30 湖南工程学院 Photovoltaic power prediction error compensation method based on weather conditions

Also Published As

Publication number Publication date
CN112507613B (en) 2022-05-10

Similar Documents

Publication Publication Date Title
CN112487702B (en) Method for predicting residual service life of lithium ion battery
CN110705743B (en) New energy consumption electric quantity prediction method based on long-term and short-term memory neural network
CN111260136A (en) Building short-term load prediction method based on ARIMA-LSTM combined model
CN109871977A (en) Load forecasting method based on wavelet transform and the minimum support vector machines of optimization
CN111667098B (en) Wind power station output power prediction method based on multi-model combination optimization
Yang et al. Hybrid prediction method for wind speed combining ensemble empirical mode decomposition and Bayesian ridge regression
CN112507613B (en) Second-level ultra-short-term photovoltaic power prediction method
CN110059867B (en) Wind speed prediction method combining SWLSTM and GPR
Zhang et al. Wind speed prediction research considering wind speed ramp and residual distribution
CN110555548B (en) ICS-ELM ultra-short-term wind power prediction method based on data mining original error correction
Senthil Kumar Improved prediction of wind speed using machine learning.
CN115511177A (en) Ultra-short-term wind speed prediction method based on INGO-SWGMN hybrid model
CN114462718A (en) CNN-GRU wind power prediction method based on time sliding window
CN113554466A (en) Short-term power consumption prediction model construction method, prediction method and device
Niu et al. Short-term wind speed hybrid forecasting model based on bias correcting study and its application
CN115310669A (en) Ultra-short-term wind power prediction method based on secondary decomposition and IWOA-LSSVM
CN116578551A (en) GRU-GAN-based power grid data restoration method
CN110276478B (en) Short-term wind power prediction method based on segmented ant colony algorithm optimization SVM
Kumar Improved Prediction of Wind Speed using Machine Learning.
CN114648147A (en) IPSO-LSTM-based wind power prediction method
CN116681154A (en) Photovoltaic power calculation method based on EMD-AO-DELM
CN116316542A (en) Regional distributed photovoltaic power prediction method, regional distributed photovoltaic power prediction device, computer equipment and storage medium
CN115796327A (en) Wind power interval prediction method based on VMD (vertical vector decomposition) and IWOA-F-GRU (empirical mode decomposition) -based models
Cheng et al. A hybrid probabilistic estimation method for photovoltaic power generation forecasting
CN113762591A (en) Short-term electric quantity prediction method and system based on GRU and multi-core SVM counterstudy

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