CN116167465A - Solar irradiance prediction method based on multivariate time series ensemble learning - Google Patents
Solar irradiance prediction method based on multivariate time series ensemble learning Download PDFInfo
- Publication number
- CN116167465A CN116167465A CN202310441767.7A CN202310441767A CN116167465A CN 116167465 A CN116167465 A CN 116167465A CN 202310441767 A CN202310441767 A CN 202310441767A CN 116167465 A CN116167465 A CN 116167465A
- Authority
- CN
- China
- Prior art keywords
- representing
- solar irradiance
- time series
- prediction
- data
- 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.)
- Withdrawn
Links
- 238000000034 method Methods 0.000 title claims abstract description 65
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 19
- 230000006870 function Effects 0.000 claims description 40
- 230000008569 process Effects 0.000 claims description 17
- 230000004913 activation Effects 0.000 claims description 9
- 230000002441 reversible effect Effects 0.000 claims description 5
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000003213 activating effect Effects 0.000 claims description 3
- 238000012986 modification Methods 0.000 claims description 3
- 230000004048 modification Effects 0.000 claims description 3
- 238000012549 training Methods 0.000 abstract description 14
- 238000010801 machine learning Methods 0.000 abstract description 8
- 238000012360 testing method Methods 0.000 abstract description 5
- 230000000694 effects Effects 0.000 abstract description 4
- 238000005516 engineering process Methods 0.000 abstract description 4
- 238000013136 deep learning model Methods 0.000 abstract description 3
- 238000009826 distribution Methods 0.000 description 12
- 238000011156 evaluation Methods 0.000 description 8
- 238000012545 processing Methods 0.000 description 7
- 238000010248 power generation Methods 0.000 description 5
- 238000010606 normalization Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 125000003275 alpha amino acid group Chemical group 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000012854 evaluation process Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000011002 quantification Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 230000001932 seasonal effect Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 230000000087 stabilizing effect Effects 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- 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
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a solar irradiance prediction method based on multivariate time series ensemble learning, which combines a CEEMDAN decomposition model with a WGAN and LSTM prediction model, and the provided CEEMDAN-WGAN-LSTM model uses a data decomposition technology and an advanced Machine Learning (ML) and Deep Learning (DL) model to identify the dependency relationship and network topology among data in the solar irradiance time series. CEEMDAN decomposed the original univariate solar irradiance dataset. The single column of GHI data is converted into a plurality of sub-sequence signals and a residual signal. Next, the obtained sub-sequences are divided into high and low frequencies, and each sub-sequence is divided into a training set and a test set for a subsequent prediction model. The invention delivers high frequency classes through WGAN and low frequency classes through LSTM. Finally, the prediction results for each sub-sequence are accumulated to produce a final prediction result. Experimental results show that the prediction effect and the prediction stability of the method are obviously improved compared with the existing solar irradiance prediction method.
Description
Technical Field
The invention relates to the technical field of photovoltaic data processing, in particular to a solar irradiance prediction method based on multivariate time series ensemble learning.
Background
Photovoltaic energy has become one of the most promising sources of electricity generation in residential, commercial and industrial applications. Because solar energy has the advantages of abundant resources, no pollution, free use and no transportation, in recent years, the global photovoltaic industry has exponentially increased, and the global photovoltaic installation amount reaches 260GW by the end of 2022. However, solar power generation has volatility and intermittence, which are mainly due to different weather conditions, and the integration of the photovoltaic power generation system with the power grid is limited by the volatility and intermittence of solar power generation. Therefore, the method for accurately predicting the photovoltaic power generation capacity in a short period has important significance for promoting the reasonable power dispatching of energy companies, improving the operation coordination between solar energy and other energy sources (such as wind energy, thermal power and the like) and micro-grids or large-scale grids, optimizing power resource dispatching and improving power economic benefit and environmental benefit. Photovoltaic power generation is most directly and significantly affected by solar irradiance on the earth's surface, and therefore, it is important for accurate prediction of solar irradiance.
According to the prior studies, scholars have proposed a number of data-driven prediction methods to predict solar irradiance, which can be roughly divided into two directions: a single or hybrid model is used. The single model prediction method mainly comprises a traditional statistical method, a classical machine learning method and a deep learning method. Research shows that a single model has a hysteresis problem in a prediction task, and the randomness of photovoltaic fluctuation cannot be reflected well. The hybrid model prediction method usually combines a plurality of models to solve the limitation of independent models, utilizes the characteristics of the plurality of models to improve the prediction performance, or combines the methods with characteristic engineering to solve the problem of poor single model prediction effect. Among them, the most widely used hybrid model is the decomposition-integration model.
At present, although a hybrid model based on a decomposition technology can be proved to improve prediction accuracy, there is still room for improvement, for example, only one model is used for predicting decomposed subsequences after data is decomposed at present, spectrum differences among the decomposed sequences are not considered, and diversity and suitability of matching high-frequency data and low-frequency data with a prediction model are ignored, so that the existing prediction method for solar irradiance by applying the hybrid model has great room for improvement, and the prediction accuracy is further improved.
Disclosure of Invention
The invention provides a solar irradiance prediction method based on multivariate time series integrated learning, which is characterized in that a decomposition-integration technology is continuously applied, a CEEMDAN algorithm, a WGAN model and an LSTM long-short-term memory network are combined, a high-frequency subsequence and a low-frequency subsequence with obvious frequency difference in solar irradiance data are decomposed by the CEEMDAN algorithm, then the WGAN model is utilized to predict the high-frequency subsequence, the LSTM is utilized to predict the low-frequency subsequence, finally the predicted values of all components are added, and finally the obtained solar irradiance prediction result has higher accuracy.
To achieve the purpose, the invention adopts the following technical scheme:
the solar irradiance prediction method based on multivariate time series ensemble learning comprises the following steps:
s1, decomposing solar irradiance time series data by using CEEMDAN algorithm
Obtaining a high-frequency subsequence and a low-frequency subsequence;
s2, predicting each high-frequency subsequence by using an improved WGAN model, predicting each low-frequency subsequence by using a stacked LSTM network, and adding the prediction results of each subsequence to obtain a final solar irradiance prediction result.
Preferably, in step S1, the time-series data is decomposed
The method of (1) comprises the steps of:
s11, for the time sequence data
White noise +.>
Obtain->
, wherein ,
Representing the time series data after adding white noise +.>
， 
Representing a noise figure;
s12, decomposing each by using an EMD modal decomposition algorithm
And averaging the components obtained by the decomposition to obtain the final eigenmode function +.>
And residual->
。
Preferably, each is decomposed
The method of (1) comprises the steps of:
s121, defining the EMD modal decomposition algorithm to decompose the kth component as an operator
，
Is that
First order modal component sequence obtained via EMD, < >>
，
Representing time series data +.>
Adding the total number of Gaussian white noise with the mean value of 0, and then decomposing each +.>
Extracting first order eigenmode function +.>
；
S122, calculating a first residual error r 1 (t)
；
S123, decomposing residual error
Obtain->
；
S124, for the rest
，
Decomposing the residual error by using the method of step S122-S123, and finally calculating to obtain the final residual error +.>
。
Preferably, the eigenmode functions and residuals of the front K/2 are the high-frequency subsequences and the eigenmode functions and residuals of the remaining K/2 are the low-frequency subsequences, ordered from high to low frequency.
Preferably, the method for predicting each of the high frequency subsequences using the modified WGAN model specifically includes the steps of:
a1, will be defined as
Is input into a BiGRU layer of a generator G of the WGAN model, the BiGRU layer being populated with data from pairs of forward and reverse GRU layers>
Learning is performed, and a forward GRU hidden vector is calculated in the horizontal direction +.>
And the inverted GRU concealment vector for each time step +.>
；
A2, combining
and
Obtaining pair->
Predicted outcome of->
，
、
Respectively indicate->
and
In calculating->
Weight of time, weight of time->
Representing the bias.
Preferably, the current hidden layer state of the BiGRU is input by the current
Output of hidden layer state forward at time t-1 +.>
And the output of the inverted hidden layer state +.>
Together, the hidden layer state of BiGRU at time t is determined by the forward hidden layer state +.>
And reverse hidden layer state->
And (5) obtaining weighted summation.
Preferably, the method for predicting each low-frequency subsequence by using the stacked LSTM networks specifically includes the steps of:
b1, definition of forget gate through the LSTM network is defined as
Information part to be filtered out in each of said low frequency subsequences +.>
；/>
B2, determining through the input gate of the LSTM network
Information part to be kept->
And updating the information part determined not to remain, and then +.>
Updated to->
；
B3, outputting the pair through the output gate of the LSTM network
Predicted outcome of->
。
Preferably, in step B1, the information part to be filtered is determined through the forgetting gate
，
According to the input of the current t moment +.>
And t-1 time status->
And by the activation function sigmoid it is determined that an output value between 0 and 1, a closer to 0 means that it should be discarded and a closer to 1 means that it should be preserved. The determination process is expressed as follows:
wherein ,
representing an activation function sigmoid;
Representing the weight;
Representing the bias.
Preferably, in step B2, the information of the hidden state of the previous layer and the information of the current input are firstly transferred into a sigmoid function, and the value is adjusted to be between 0 and 1 to determine
The information part to be kept +.>
0 represents unimportance, 1 represents importance, and its reserved expression is:
secondly, the information of the hidden state of the previous layer and the information input at present are transmitted to the tanh function to create a new candidate cell state, and the process is expressed as follows:
finally multiplying the output value of sigmoid with the output value of tanh, wherein the output value of sigmoid determines which information in the output value of tanh is important and needs to be preserved;
will be
Updated to->
The process of (2) is expressed as follows:
in the formula ,
representing the input signal;
Representing the weight;
Representing the bias;
Representing a candidate cell state;
Representing the weight;
Representing the bias;
Representing the current updated cell state;
Representing reservation information through a forget gate;
The cell state at time t-1 is shown.
Preferably, in step B3, the output gate outputs
The process of (1) comprises the steps of:
b31 determination by activating function sigmoid
Output part of->
The determination process is expressed as:
b32, multiplying the output part by the activation function tanh
Predicted value of +.>
The specific process is expressed as follows:
in the formula ,
representing the weight;
Representing the bias;
The state of the cell at the current time t is indicated.
The invention combines a CEEMDAN decomposition model with a WGAN and LSTM prediction model, and the CEEMDAN-WGAN-LSTM model provided uses a data decomposition technology and an advanced Machine Learning (ML) and Deep Learning (DL) model to identify the dependency relationship and network topology between data in solar irradiance time series. CEEMDAN decomposed the original univariate solar irradiance dataset. The single column of GHI data is converted into a plurality of sub-sequence signals and a residual signal. Next, the obtained sub-sequences are divided into high and low frequencies, and each sub-sequence is divided into a training set and a test set for a subsequent prediction model. The invention delivers high frequency classes through WGAN and low frequency classes through LSTM. Finally, the prediction results for each sub-sequence are accumulated to produce a final prediction result. Experimental results show that the prediction effect and the prediction stability of the method are obviously improved compared with the existing solar irradiance prediction method.
Drawings
In order to more clearly illustrate the technical solution of the embodiments of the present invention, the drawings that are required to be used in the embodiments of the present invention will be briefly described below. It is evident that the drawings described below are only some embodiments of the present invention and that other drawings may be obtained from these drawings without inventive effort for a person of ordinary skill in the art.
Fig. 1 is a schematic diagram of a network structure of a WGAN model according to an embodiment of the present invention;
FIG. 2 is a flowchart of decomposing solar irradiance data and predicting solar irradiance using CEEMDAN-WGAN-LSTM model using CEEMDAN algorithm provided in the embodiment of the present invention, and an overall structure schematic of the CEEMDAN-WGAN-LSTM model used;
FIG. 3 is a graph of the evaluation index (MAE, MAPE, RMSE) quantification of four decomposition-integration models with good performance;
fig. 4 is a diagram of implementation steps of a solar irradiance prediction method based on multivariate time series ensemble learning according to an embodiment of the present invention.
Detailed Description
The technical scheme of the invention is further described below by the specific embodiments with reference to the accompanying drawings.
Wherein the drawings are for illustrative purposes only and are shown in schematic, non-physical, and not intended to be limiting of the present patent; for the purpose of better illustrating embodiments of the invention, certain elements of the drawings may be omitted, enlarged or reduced and do not represent the size of the actual product; it will be appreciated by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
The same or similar reference numbers in the drawings of embodiments of the invention correspond to the same or similar components; in the description of the present invention, it should be understood that, if the terms "upper", "lower", "left", "right", "inner", "outer", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, only for convenience in describing the present invention and simplifying the description, rather than indicating or implying that the apparatus or elements being referred to must have a specific orientation, be constructed and operated in a specific orientation, so that the terms describing the positional relationships in the drawings are merely for exemplary illustration and should not be construed as limiting the present patent, and that the specific meaning of the terms described above may be understood by those of ordinary skill in the art according to specific circumstances.
In the description of the present invention, unless explicitly stated and limited otherwise, the term "coupled" or the like should be interpreted broadly, as it may be fixedly coupled, detachably coupled, or integrally formed, as indicating the relationship of components; can be mechanically or electrically connected; can be directly connected or indirectly connected through an intermediate medium, and can be communication between the two parts or interaction relationship between the two parts. The specific meaning of the above terms in the present invention will be understood in specific cases by those of ordinary skill in the art.
The solar irradiance prediction method based on multivariate time series ensemble learning provided by the embodiment of the invention, as shown in (b) in fig. 4 and fig. 2, specifically comprises the following steps:
1) Collecting solar irradiance data of a certain region for one year, sequentially recording irradiance values according to the collection time sequence to obtain time sequence data, and recording as
，
Indicate the%>
Data of->
Representing the amount of data collected;
here, the time series is
The data in the method is historical solar irradiance data in the same area, and the time interval for data acquisition is preferably 5 minutes; />
2) And carrying out missing value and normalization processing on the acquired data, and eliminating the dimensional influence of the data. Specifically, for each time-series data
Data missing values and normalization processing are performed. Time series data->
First, a missing value processing is performed, for example, no sunlight is used at night, irradiance is 0, and time series data is added>
And deleting irradiance data acquired at night. Then, the time series data is added>
Normalization processing is performed. Time series data->
There are many existing methods for performing the missing values and normalization processing, and therefore, detailed description will not be given. Finally, the preprocessed time series data +.>
According to the seasonal characteristics, the data are divided into corresponding data sets according to four seasons of spring, summer, autumn and winter. Finally, dividing the data in each data set into a training set, a verification set and a test set according to a division mode of 3:1:1;
3) The divided training set is input into a CEEMDAN-WGAN-LSTM model shown in (c) of fig. 2 for training, and the specific training method is as follows:
first, solar irradiance time series data is obtained by CEEMDAN algorithm
Dividing the sequence into K eigenmode functions (Intrinsic Mode Functions, IMFs), wherein the K eigenmode functions have frequency differences, the first K/2 IMFs obtained by decomposition are defined as high-frequency subsequences, and the last K/2 IMFs are defined as low-frequency subsequences;
then, predicting a low frequency sub-sequence using the WGAN model shown in fig. 1, and predicting a low frequency sub-sequence using the LSTM network shown in (c) of fig. 2;
then, adding the prediction results of the subsequences to obtain a solar irradiance prediction result;
and then evaluating the prediction result obtained by adding by using the evaluation index, and when the prediction accuracy rate is judged to be inconsistent with the expectation, adjusting the model training parameters and then carrying out iterative training on the model until the expected prediction accuracy rate is reached, thus obtaining the final CEEMDAN-WGAN-LSTM model for predicting solar irradiance.
In this embodiment, as shown in FIG. 2 (a), the CEEMDAN model decomposes solar irradiance time series data
The method of (1) specifically comprises the steps of:
s11, time sequence data
White noise +.>
（
To follow the white noise of the normal distribution N (0, 1), it is expressed as extracting a value from one normal distribution at each time instant to form a white noise time series. Also, the parameters of this normal distribution are fixed and do not change over time. This is the case, therefore, of repeatedly extracting values from a fixed probability distribution to form a time series. ) Obtain->
, wherein ,
Representing the time series data after adding white noise +.>
， 
Representing a noise figure;
s12, an EMD modal decomposition algorithm is used (EMD is to decompose signals according to the time scale characteristics of the data, and no basis function is required to be preset, which is oneA time-frequency domain signal processing mode. The EMD decomposes the input signal into several eigenmode functions and a residual component. ) Decompose each
And averaging the components obtained by the decomposition to obtain a final eigenmode function and residual error. EMD algorithm decomposes each +.>
The process of (1) specifically comprises the following steps:
s121, defining the EMD modal decomposition algorithm to decompose the kth component as an operator
，
Is that
First order modal component sequence obtained via EMD, < >>
，
Representing time series data +.>
Adding the total number of Gaussian white noise with the mean value of 0, and then decomposing each +.>
And extracting a first IMF (first order eigenmode function obtained for the first decomposition +.>
），
Expressed as:
s122, calculating a first residual error r 1 (t)：
S123, decomposing residual error
Obtaining the second IMF as
S124, for the rest
Repeating the above steps, and calculating to obtain the final residual +.>
，
Expressed as:
4) The WGAN model shown in FIG. 1 was used to determine the frequency subsequence (hereinafter referred to as
) Training and prediction are performed. As shown in fig. 1, the WGAN model is composed of a generator G and a discriminator D, which are composed of stacked bi-directional gating cyclic units (bigrus) and multi-layer perceptrons (MLPs),the generator G and the discriminator D with the structure can solve the gradient problem in the neural network, and are helpful for stabilizing the model structure of the WGAN and improving the prediction performance of the model.
The WGAN adopts Wasserstein distance to judge the difference between the real sample and the generated sample distribution, when the difference between the two distributions is larger, the generator can still be ensured to update, and the problems that the original GAN adopts KL or JS divergence (the KL divergence is also called relative entropy and KL distance, the difference between two probability distributions P and Q can be simply understood as similarity, the more similar the two are, the smaller the KL divergence is a variation of the KL divergence, the more similar the JS divergence is as the KL divergence, the smaller the JS divergence is) as a loss function of the model exist, and the gradient vanishes and the model collapses, so that the generated data of the generator is not ideal are solved.
Under ideal conditions, wasserstein distance
Is continuously differentiable, and the loss function formula is as follows:
in formula (5): sup (-) represents the upper bound of the function value;
is Lipschitz constant;
For real data +.>
Is the generated data;
Representation function->
Satisfy K-Lipschitz continuous, function +.>
Fitting can be performed using a neural network;
Probability distribution representing real data x +.>
Representing production data +.>
Probability distribution of (2);
An expected function representing real data +.>
Representing a desired function of the generated data;
Representing the distribution of real data +.>
The representation generator generates a distribution of data.
The continuous generation of the generator in the WGAN network is used for continuous identification of the identifier, so that the generated data closest to the original solar irradiance data is obtained. Since the generator G is composed of BiGRU, high frequency subsequences
The final output is obtained by the generator biglu of WGAN as input signal>
. Since BiGRU learns the input data by the forward and reverse GRU layers, the forward GRU hidden vector is calculated in the horizontal direction>
And the inverted GRU concealment vector for each time step +.>
. By constructing multiple layers of BiGRUs, the input sequence is fully learned. The final output can be represented by the following formulas (6) - (8), wherein +.>
、
Representing weights +.>
Representing the bias:
the following pair of WGAN models predicts high frequency subsequences
The process of (2) is further described:
the WGAN is composed of a generator G for generating sample data conforming to the distribution of real data, and a discriminator D for judging and classifying input data, and outputting "1" if the input data is judged to be real data, and "0" if the input data is judged to be false data. The training of WGAN is divided into two phases, first training discriminator D and then training generator G. During the training process, the two models can continuously update the parameters of the models, so that the respective loss function and output error are minimized. The WGAN structure is brand new and customized, and the generator G and the discriminator D are respectively composed of the stacked BiGRU and the stacked MLP, so that the gradient problem existing in the neural network can be effectively solved, and the prediction performance of the model is improved. The final output of the high frequency sub-sequence by WGAN can be expressed as:
Prediction using stacked LSTM networks is defined as either simultaneously with or after prediction of the high frequency sub-sequences is completed
Is a low frequency subsequence of (a). Each cell unit in the LSTM network adopted by the invention comprises 3 parts of a forgetting gate, an input gate and an output gate, and filtering, storing and generating information are respectively determined. The following describes the door structure in detail:
a) Forget to leave the door. The invention determines each decomposed low-frequency subsequence through forgetting gate
The information part of the component that needs to be filtered out. Input of the current t moment +.>
And t-1 time status->
By activating a function sigmoid (expressed as +.>
) It is determined whether to filter. The closer the output value is between 0 and 1, the more it should be discarded, and the closer the output value is to 1, the more it should be retained. The formula is as follows:
in the formula (10), the amino acid sequence of the compound,
a discard value representing a forget gate;
Representing an activation function sigmoid;
Representing the weight;
The hidden layer state at the time t-1 is represented;
Representing the bias.
b) An input gate. Determining input information
Information part to be kept->
And updating the information part determined not to remain, and then +.>
Updated to->
. Firstly, the information of the hidden state of the previous layer and the information input currently are transferred into a sigmoid function, and the value is adjusted to be between 0 and 1 to determine +.>
The information part to be kept +.>
. 0. Not important, 1 is important. The reserved expression is:
and secondly, transmitting the information of the hidden state of the previous layer and the information input currently into the tanh function to create a new candidate cell state. The process is expressed as follows:
finally, the output value of sigmoid is multiplied by the output value of tanh, which determines which information in the output value of tanh is important and needs to be preserved. Will be
Updated to->
The process of (2) is expressed as follows:
in the formulae (11) to (13),
representing the weight;
Representing the bias;
Representing a candidate cell state;
Representing the weight;
Representing the bias;
Indicating the current renewing cellA state;
Representing forget gate discard information;
The cell state at time t-1 is shown.
c) And outputting a door. First by an activation function
And determining a unit output part, and multiplying the unit states through an activation function tanh output part to obtain a predicted value point of the model. The formula is as follows:
in the formulae (14) to (15),
representing the weight;
Representing the bias;
Representing the current updated cell state;
Representation pair
Is a predicted result of (a).
Eventually, it will
And->
Adding to obtain a t time pairTime series data->
Solar irradiance prediction of +.>
The method specifically comprises the following steps:
in order to evaluate the prediction performance of the CEE-WGAN-LSTM model on solar irradiance, the embodiment of the invention adopts any one or more of average absolute value error (MAE), average absolute percentage error (MAPE) and root mean square error (Root Mean Square Error, RMSE) evaluation methods to evaluate the prediction precision of the WGAN model, the LSTM model and the integral solar irradiance prediction model CEE-WGAN-LSTM. The evaluation process of each error evaluation method is expressed by the following formula:
in the formulas (17) - (19),
and
Representing the real value and the predicted value of the object model, respectively,/->
and
Respectively represent the firstaSum of true values and thaThe predicted values of the individual object models are,brepresenting the length of the test set.
Specifically, in step 7) of the CEE-WGAN-LSTM-based solar irradiance prediction method provided in this embodiment, according to the error evaluation of the prediction result of the initial target model, the model internal parameters are adjusted to minimize the minimum prediction errors, that is, the mean absolute value error (MAE), the Mean Absolute Percentage Error (MAPE), and the root mean square error (Root Mean Square Error, RMSE). In order to verify the performance of the CEEMDAN-WGAN-LSTM model provided by the invention, the invention selects a machine learning model which is popular in the field of time sequence prediction at present and a decomposition-integration mixed model for comparison. These models include GRU, RNN, LSTM, WGAN, transformer, CEE-LSTM, CEE-WGAN, CEEMDAN-LSTM-WGAN (hereinafter referred to as CEE-L-W in the tables). Tables 1,2,3, and 4 below show solar irradiance data for four seasons, spring, summer, autumn, and winter, respectively, for the evaluation comparison between the CEEMDAN-WGAN-LSTM (hereinafter referred to as CEE-W-L in the chart) model provided by the present invention and each of the models set forth above, with the evaluation index being MAE, MAPE, RMSE, respectively. Furthermore, to more intuitively demonstrate the predictive performance of our proposed model, we convert the quantized evaluation results of four better performing decomposition-integration models (CEE-LSTM, CEE-WGAN, CEE-L-W and CEE-W-L) into a histogram, as shown in FIG. 3. It can be seen very intuitively from the figure that the MAE, MAPE, RMSE model of CEEMDAN-WGAN-LSTM (shown as CEE-W-L) we propose is the lowest. All experiments are carried out on the same experimental platform by using the data set so as to ensure fairness of the experiments, and experimental results show that the CEEMDAN-WGAN-LSTM model prediction performance provided by the invention is obviously superior to that of a comparison model.
TABLE 1
TABLE 2
TABLE 3 Table 3
TABLE 4 Table 4
In summary, the invention combines CEEMDAN decomposition model with WGAN and LSTM prediction model, and provides CEEMDAN-WGAN-LSTM model that uses data decomposition technique and advanced Machine Learning (ML) and Deep Learning (DL) model to identify the dependency relationship and network topology between data in solar irradiance time series. CEEMDAN decomposed the original univariate solar irradiance dataset. The single column of GHI data is converted into a plurality of sub-sequence signals and a residual signal. Next, the obtained sub-sequences are divided into high and low frequencies, and each sub-sequence is divided into a training set and a test set for a subsequent prediction model. The invention delivers high frequency classes through WGAN and low frequency classes through LSTM. Finally, the prediction results for each sub-sequence are accumulated to produce a final prediction result. Experimental results show that the prediction effect and the prediction stability of the method are obviously improved compared with the existing solar irradiance prediction method.
It should be understood that the above description is only illustrative of the preferred embodiments of the present invention and the technical principles employed. It will be apparent to those skilled in the art that various modifications, equivalents, variations, and the like can be made to the present invention. However, such modifications are intended to fall within the scope of the present invention without departing from the spirit of the present invention. In addition, some terms used in the specification and claims of the present application are not limiting, but are merely for convenience of description.
Claims (10)
1. A solar irradiance prediction method based on multivariate time series ensemble learning is characterized by comprising the following steps:
s1, decomposing solar irradiance time series data by using CEEMDAN algorithm
Obtaining a high-frequency subsequence and a low-frequency subsequence;
s2, predicting each high-frequency subsequence by using an improved WGAN model, predicting each low-frequency subsequence by using a stacked LSTM network, and adding the prediction results of each subsequence to obtain a final solar irradiance prediction result.
2. The solar irradiance prediction method based on multivariate time series ensemble learning of claim 1, wherein in step S1, the time series data is decomposed
The method of (1) comprises the steps of:
s11, for the time sequence data
White noise +.>
Obtain->
, wherein ,
representing the time series data after adding white noise +.>
， 
Representing a noise figure;
s12, decomposing each by using an EMD modal decomposition algorithm
And averaging the components obtained by the decomposition to obtain the final eigenmode function +.>
And residual->
。
3. The solar irradiance prediction method based on multivariate time series ensemble learning of claim 2, wherein each is decomposed
The method of (1) comprises the steps of:
s121, defining the EMD modal decomposition algorithm to decompose the kth component as an operator
，
Is->
First order modal component sequence obtained via EMD, < >>
，
Representing time series data +.>
Adding the total number of Gaussian white noise with the mean value of 0, and then decomposing each +.>
And extract the firstFirst order eigenmode function obtained by secondary decomposition +.>
；
S122, calculating a first residual error r 1 (t)
；
S123, decomposing residual error
Obtaining
；
S124, for the rest
，
Decomposing the residual error by using the method of step S122-S123, and finally calculating to obtain the final residual error +.>
。
4. A solar irradiance prediction method based on multivariate time series ensemble learning according to claim 3, wherein the eigenmode functions and residuals of the first K/2 are the high frequency subsequences and the eigenmode functions and residuals of the remaining K/2 are the low frequency subsequences, ordered from high frequency to low frequency.
5. The solar irradiance prediction method based on multivariate time series ensemble learning of claim 1, wherein the method of predicting each of said high frequency sub-sequences using said WGAN model with modifications specifically comprises the steps of:
a1, will be defined as
Is input into a BiGRU layer of a generator G of the WGAN model, the BiGRU layer being populated with data from pairs of forward and reverse GRU layers>
Learning is performed, and a forward GRU hidden vector is calculated in the horizontal direction +.>
And the inverted GRU concealment vector for each time step +.>
；
A2, combining
and
Obtaining pair->
Predicted outcome of->
，
、
Respectively indicate->
and
In calculating->
Weight of time, weight of time->
Representing the bias. />
6. The solar irradiance prediction method of claim 5, wherein the current hidden layer state of biglu is input from the current
Output of hidden layer state forward at time t-1
And the output of the inverted hidden layer state +.>
Together, the hidden layer state of BiGRU at time t is determined by the forward hidden layer state +.>
And reverse hidden layer state->
And (5) obtaining weighted summation.
7. The solar irradiance prediction method based on multivariate time series ensemble learning of claim 1, wherein the method of predicting each of the low frequency sub-sequences using the stacked LSTM networks specifically comprises the steps of:
b1, definition of forget gate through the LSTM network is defined as
Information part to be filtered out in each of said low frequency subsequences +.>
；
B2, through the LSTM networkInput door determination
Information part to be kept->
And updating the information part determined not to remain, and then +.>
Updated to->
；
B3, outputting the pair through the output gate of the LSTM network
Predicted outcome of->
。
8. The solar irradiance prediction method based on multivariate time series ensemble learning of claim 7, wherein in step B1, the information portion to be filtered out is determined by the forgetting gate
，
According to the input at the current time t
And t-1 time status->
And by the activation function sigmoid it is determined that an output value between 0 and 1, a closer to 0 means that it should be discarded and a closer to 1 means that it should be preserved. The determination process is expressed as follows:
wherein ,
representing an activation function sigmoid;
Representing the weight;
Representing the bias.
9. The method for solar irradiance prediction based on multivariate time series ensemble learning of claim 7, wherein in step B2, first, information of a hidden state of a previous layer and information of a current input are transferred to a sigmoid function, and a value is adjusted to be between 0 and 1 to determine
The information part to be kept +.>
0 represents unimportance, 1 represents importance, and its reserved expression is:
secondly, the information of the hidden state of the previous layer and the information input at present are transmitted to the tanh function to create a new candidate cell state, and the process is expressed as follows:
finally multiplying the output value of sigmoid with the output value of tanh, wherein the output value of sigmoid determines which information in the output value of tanh is important and needs to be preserved;
will be
Updated to->
The process of (2) is expressed as follows:
in the formula ,
representing the input signal;
Representing the weight;
Representing the bias;
Representing a candidate cell state;
Representing the weight;
representing the bias;
Representing the current updated cell state;
Representing reservation information through a forget gate;
Cells at time t-1Status of the device.
10. The solar irradiance prediction method based on multivariate time series ensemble learning of claim 9, wherein in step B3, the output gate outputs
The process of (1) comprises the steps of:
b31 determination by activating function sigmoid
Output part of->
The determination process is expressed as: />
B32, multiplying the output part by the activation function tanh
Predicted value of +.>
The specific process is expressed as follows:
in the formula ,
representing the weight;
Representing the bias;
Representing the current time tCell status. />
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310441767.7A CN116167465A (en) | 2023-04-23 | 2023-04-23 | Solar irradiance prediction method based on multivariate time series ensemble learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310441767.7A CN116167465A (en) | 2023-04-23 | 2023-04-23 | Solar irradiance prediction method based on multivariate time series ensemble learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116167465A true CN116167465A (en) | 2023-05-26 |
Family
ID=86416690
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310441767.7A Withdrawn CN116167465A (en) | 2023-04-23 | 2023-04-23 | Solar irradiance prediction method based on multivariate time series ensemble learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116167465A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116451598A (en) * | 2023-06-20 | 2023-07-18 | 杭州经纬信息技术股份有限公司 | Solar irradiance prediction method based on denoising diffusion probability model |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107392368A (en) * | 2017-07-17 | 2017-11-24 | 天津大学 | Meteorological forecast-based office building dynamic heat load combined prediction method |
| CN110909926A (en) * | 2019-11-18 | 2020-03-24 | 中国计量大学 | TCN-LSTM-based solar photovoltaic power generation prediction method |
| CN112434891A (en) * | 2020-12-18 | 2021-03-02 | 中国计量大学上虞高等研究院有限公司 | Method for predicting solar irradiance time sequence based on WCNN-ALSTM |
| CN112668611A (en) * | 2020-12-08 | 2021-04-16 | 湖南工业大学 | Short-term photovoltaic power generation power prediction method based on Kmeans and CEEMD-PE-LSTM |
| CN113962441A (en) * | 2021-09-26 | 2022-01-21 | 华能澜沧江水电股份有限公司 | Short-term irradiance prediction method and prediction system based on historical data analysis |
| CN114201924A (en) * | 2022-02-16 | 2022-03-18 | 杭州经纬信息技术股份有限公司 | Solar irradiance prediction method and system based on transfer learning |
| CN114912718A (en) * | 2022-07-13 | 2022-08-16 | 汉谷云智(武汉)科技有限公司 | Ultra-short-term solar irradiance prediction method based on discrete wavelet transform feature extraction |
| CN115409258A (en) * | 2022-08-25 | 2022-11-29 | 东南大学 | Hybrid deep learning short-term irradiance prediction method |
| CN115442191A (en) * | 2022-11-08 | 2022-12-06 | 中国人民解放军战略支援部队航天工程大学 | Communication signal noise reduction method and system based on relative average generation countermeasure network |
| CN115511657A (en) * | 2022-08-23 | 2022-12-23 | 国网甘肃省电力公司电力科学研究院 | Wind power output and photovoltaic output evaluation method based on combined prediction model |
| CN115526376A (en) * | 2022-08-11 | 2022-12-27 | 山西大学 | Multi-feature fusion generation countermeasure network ultra-short-term wind power prediction method |
| CN115545265A (en) * | 2022-06-10 | 2022-12-30 | 四川大学 | Electricity sales prediction method based on data restoration and decomposition sequence prediction |
| CN115689055A (en) * | 2022-11-22 | 2023-02-03 | 河海大学 | Short-term solar irradiance prediction method and device |
-
2023
- 2023-04-23 CN CN202310441767.7A patent/CN116167465A/en not_active Withdrawn
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107392368A (en) * | 2017-07-17 | 2017-11-24 | 天津大学 | Meteorological forecast-based office building dynamic heat load combined prediction method |
| CN110909926A (en) * | 2019-11-18 | 2020-03-24 | 中国计量大学 | TCN-LSTM-based solar photovoltaic power generation prediction method |
| CN112668611A (en) * | 2020-12-08 | 2021-04-16 | 湖南工业大学 | Short-term photovoltaic power generation power prediction method based on Kmeans and CEEMD-PE-LSTM |
| CN112434891A (en) * | 2020-12-18 | 2021-03-02 | 中国计量大学上虞高等研究院有限公司 | Method for predicting solar irradiance time sequence based on WCNN-ALSTM |
| CN113962441A (en) * | 2021-09-26 | 2022-01-21 | 华能澜沧江水电股份有限公司 | Short-term irradiance prediction method and prediction system based on historical data analysis |
| CN114201924A (en) * | 2022-02-16 | 2022-03-18 | 杭州经纬信息技术股份有限公司 | Solar irradiance prediction method and system based on transfer learning |
| CN115545265A (en) * | 2022-06-10 | 2022-12-30 | 四川大学 | Electricity sales prediction method based on data restoration and decomposition sequence prediction |
| CN114912718A (en) * | 2022-07-13 | 2022-08-16 | 汉谷云智(武汉)科技有限公司 | Ultra-short-term solar irradiance prediction method based on discrete wavelet transform feature extraction |
| CN115526376A (en) * | 2022-08-11 | 2022-12-27 | 山西大学 | Multi-feature fusion generation countermeasure network ultra-short-term wind power prediction method |
| CN115511657A (en) * | 2022-08-23 | 2022-12-23 | 国网甘肃省电力公司电力科学研究院 | Wind power output and photovoltaic output evaluation method based on combined prediction model |
| CN115409258A (en) * | 2022-08-25 | 2022-11-29 | 东南大学 | Hybrid deep learning short-term irradiance prediction method |
| CN115442191A (en) * | 2022-11-08 | 2022-12-06 | 中国人民解放军战略支援部队航天工程大学 | Communication signal noise reduction method and system based on relative average generation countermeasure network |
| CN115689055A (en) * | 2022-11-22 | 2023-02-03 | 河海大学 | Short-term solar irradiance prediction method and device |
Non-Patent Citations (1)
| Title |
|---|
| QIANQIAN LI等: "A Solar Irradiance Forecasting Framework Based on the CEE-WGAN-LSTM Model", 《SENSORS》, vol. 23, no. 5, pages 1 - 16 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116451598A (en) * | 2023-06-20 | 2023-07-18 | 杭州经纬信息技术股份有限公司 | Solar irradiance prediction method based on denoising diffusion probability model |
| CN116451598B (en) * | 2023-06-20 | 2023-09-05 | 杭州经纬信息技术股份有限公司 | Solar Irradiance Prediction Method Based on Denoising Diffusion Probability Model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112348271A (en) | Short-term photovoltaic power prediction method based on VMD-IPSO-GRU | |
| CN109214575A (en) | A kind of super short-period wind power prediction technique based on small wavelength short-term memory network | |
| CN109635928A (en) | A kind of voltage sag reason recognition methods based on deep learning Model Fusion | |
| CN107730044A (en) | A kind of hybrid forecasting method of renewable energy power generation and load | |
| Dou et al. | Hybrid model for renewable energy and loads prediction based on data mining and variational mode decomposition | |
| CN108665112A (en) | Photovoltaic fault detection method based on Modified particle swarm optimization Elman networks | |
| CN108717579B (en) | Short-term wind power interval prediction method | |
| CN113554466A (en) | Short-term power consumption prediction model construction method, prediction method and device | |
| CN110837915B (en) | Low-voltage load point prediction and probability prediction method for power system based on hybrid integrated deep learning | |
| CN109583588B (en) | Short-term wind speed prediction method and system | |
| CN111506868B (en) | Ultra-short-term wind speed prediction method based on HHT weight optimization | |
| CN115859792A (en) | Medium-term power load prediction method and system based on attention mechanism | |
| CN117114160A (en) | Short-term photovoltaic power prediction method | |
| CN116167465A (en) | Solar irradiance prediction method based on multivariate time series ensemble learning | |
| CN116050605A (en) | Power load prediction method based on neural network and random forest method | |
| CN116629417A (en) | Rail transit traction load prediction method combined with convolution attention module | |
| CN118095891A (en) | Active power distribution network payload prediction method and system considering source load meteorological characteristic decoupling | |
| CN108694475B (en) | Short-time-scale photovoltaic cell power generation capacity prediction method based on hybrid model | |
| CN114611757A (en) | Electric power system short-term load prediction method based on genetic algorithm and improved depth residual error network | |
| CN112418504B (en) | Wind speed prediction method based on mixed variable selection optimization deep belief network | |
| CN117689082A (en) | Short-term wind power probability prediction method, system and storage medium | |
| CN117034762A (en) | Composite model lithium battery life prediction method based on multi-algorithm weighted sum | |
| CN115764855A (en) | Real-time adjustable capacity and available electric quantity prediction method for electric vehicle quick charging station | |
| Amreen et al. | Solar power forecasting using hybrid model | |
| Mellit et al. | Application of neural networks and genetic algorithms for predicting the optimal sizing coefficient of photovoltaic supply (PVS) systems |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20230526 |
|
| WW01 | Invention patent application withdrawn after publication |