CN115829126A - Photovoltaic power generation power prediction method based on multi-view self-adaptive feature fusion - Google Patents
Photovoltaic power generation power prediction method based on multi-view self-adaptive feature fusion Download PDFInfo
- Publication number
- CN115829126A CN115829126A CN202211560534.0A CN202211560534A CN115829126A CN 115829126 A CN115829126 A CN 115829126A CN 202211560534 A CN202211560534 A CN 202211560534A CN 115829126 A CN115829126 A CN 115829126A
- Authority
- CN
- China
- Prior art keywords
- representing
- matrix
- feature
- formula
- features
- 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.)
- Pending
Links
Images
Classifications
-
- 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
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a photovoltaic power generation power prediction method based on multi-view self-adaptive feature fusion, which can be used for mining potential correlation among information from four views of the whole, local, time sequence and space and predicting photovoltaic power generation power; dividing the collected data into a global view part and a local view part, and extracting future total features and historical total features from a time sequence view and a space view; in order to better fuse the feature information, a parameter sharing feature extraction module is constructed to extract public features, and an attention mechanism is introduced to distribute proper attention weight for the local features of each time window, so that the whole model can self-adaptively learn the most relevant information from the global and local features and fully fuse the information; and constructing a loss function by using the consistency constraint and the independence constraint to strengthen the model, so that the photovoltaic power generation power can be accurately predicted.
Description
Technical Field
The invention relates to the field of lighting equipment, in particular to a photovoltaic power generation power prediction method based on multi-view self-adaptive feature fusion.
Background
In recent years, under the drive of a double-carbon target, china obtains huge achievements in the field of new energy development, and as the result of the statistical data of the national energy bureau, the installed capacity of electricity generation in China is about 24.4 hundred million kilowatts and the installed capacity of electricity generation in China is increased by 8.1 percent on year by year 6, wherein the accumulated installed capacity of wind electricity is about 3.4 hundred million kilowatts, the accumulated installed capacity of photovoltaic electricity generation is about 3.1 hundred million kilowatts, the accumulated installed capacity of electricity generation is increased by 8.1 percent and 25.8 percent on year by year, the renewable energy sources such as photovoltaic, wind electricity, nuclear electricity and the like are in the top of the world, and the vigorous development of the renewable energy sources such as photovoltaic, wind electricity, nuclear electricity and the like becomes a necessary trend of low-carbon transformation development in China. Wherein, solar energy is a clean, safe, sustainable electricity generation resource, and the user not only can accomplish spontaneous self-service after installing photovoltaic power generation system equipment near the power consumption place, can also be with unnecessary electric quantity transmission for the country, balanced distribution system, just so can make full use of local solar energy resource, replace and reduce fossil energy consumption. However, photovoltaic power generation has several disadvantages: first, photovoltaic power generation needs to rely on sunlight, and only daytime generates electricity, and the power generation mode has intermittence. Second, photovoltaic power generation often fluctuates with changes in meteorological factors such as solar irradiance, temperature, etc. According to the two points, the photovoltaic power generation power has obvious intermittence and fluctuation, the power generation mode is not stable, the controllability of the output power is not ideal, and the safe and stable operation of a power grid can be impacted to a certain extent by the large-scale photovoltaic power generation access. Therefore, accurate photovoltaic power generation power prediction is one of key technologies for solving the problem and is also an important guarantee for safe and reliable operation of a power grid.
At present, many scholars have conducted a lot of research on the prediction of the generated power of a photovoltaic system, and the methods thereof are mainly divided into two categories: direct prediction methods and indirect prediction methods. The direct prediction method uses historical photovoltaic power station data and meteorological data as input, a mathematical model is established to predict photovoltaic power generation power, and finally the predicted photovoltaic power generation power is obtained through model calculation. The indirect prediction method is that solar radiation in historical meteorological data is predicted, and then a mathematical model is further established according to an electric power model of the photovoltaic power station to predict the generated power of the photovoltaic power station. A large amount of academic research is carried out by many researchers at home and abroad aiming at photovoltaic power generation power prediction based on machine learning, deep learning and other artificial intelligence technologies. A common machine learning prediction method includes a differential integration moving average autoregressive model, which is expressed in english as auto-regressive integrated moving average, referred to as ARIMA for short; the Support Vector Machine is expressed as Support Vector Machine in English, and is called SVM for short; markov Chain, its English expression is Markov Chain, abbreviated as MC, etc. The models can predict the output of the next time step by using a window of past information, and learn the mapping relation between input and output from a large number of historical samples, but the models neglect the mutual influence among characteristics, namely the spatial correlation among the characteristics, so the prediction accuracy of the photovoltaic power generation power still needs to be improved.
Disclosure of Invention
Aiming at the problems, the invention provides a photovoltaic power generation power prediction method based on multi-view self-adaptive feature fusion, which extracts feature representation of acquired data from a global view and a local view according to a time sequence and a space view respectively, and inputs the acquired data and the local information into a parameter sharing feature extraction module to extract common features by taking potential common features in the global information and the local information into consideration; and adaptively distributing proper attention weight to the local features of each time window by using an attention fusion module, and finally obtaining a photovoltaic power generation power prediction result by the obtained fusion features through a multilayer perceptron.
In order to achieve the above object, the present invention provides a photovoltaic power generation power prediction method based on multi-view adaptive feature fusion, which includes:
s1: defining the number of features of the local information as n 1 The time elapsed for each time window is T 1 Data granularity of T 2 Then a time window size is win = T 1 /T 2 W is the number of time windows, and the local information in each time window is
The time sequence feature extraction module adopts a GRU gate cycle unit to extract the time sequence feature, and the calculation formula is as follows:
in the formula, x t Local information indicating time t
h t-1 Indicating the hidden state at the previous moment, r t Denotes a reset gate, z t Representing the update gate, σ (-) representing the activation function, W r 、W z Representing a weight matrix, b r 、b z Representing a bias vector;
s2: according to the r t 、z t Calculating the hidden state h t The calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
represents a candidate hidden state, h t Representing the hidden state at the time t obtained by the GRU module, tanh (-) representing an activation function, representing the operation by elements, W h Representing a weight matrix, b h Representing a bias vector; the hidden state h t For the time-series characteristics of the local information,
extracting the time sequence characteristics of the local information of the module for the time sequence characteristics of the w time windows;
s3: the time-sequence characteristics of local information
The method comprises the following steps of inputting the data to a spatial feature extraction module and obtaining historical total features of local information, wherein the spatial feature extraction module adopts a graph convolution network to extract spatial features, and the calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
a contiguous matrix is represented that is,
the unit matrix is represented by a matrix of units,
is represented by an adjacency matrix A l And a unit matrix I l The sum of the reachable matrices, k representing the number of layers of the convolutional network of the graph, W l (k) A k-th layer weight matrix representing a graph convolution network, σ (-) representing an activation function,
represent
A diagonal matrix of (a);
the initial input of the graph convolution network representing the ith time window is also the timing feature module output
A feature matrix representing the convolution output of the k-th layer map of the ith time window, which is defined as a historical feature
Then
Extracting historical total features of local information of the module for the spatial features of the w time windows;
s4: specifying the number of features of global information as n 2 The time length of collecting the global information is T 3 Data granularity of T 4 Global information is formed
Inputting the data into a graph convolution network of the spatial feature extraction module, and obtaining future total features Z of global information after k-layer learning G The calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
a contiguous matrix is represented that is,
the unit matrix is represented by a matrix of units,
is represented by an adjacency matrix A g And identity matrix I g The sum of the resulting reachable matrices, k representing the number of layers of the graph convolution network,
a k-th layer weight matrix representing a graph convolution network, σ (-) representing an activation function,
to represent
The diagonal matrix of (a) is,
representing the initial input of the graph convolutional network, also the global information X g ，
Defining the feature matrix of the graph convolution network output of the k layer as the future total feature Z G ；
S5: the local information X is processed l And inputting the global information gas into a parameter sharing space extraction module to obtain sharing historical characteristics and sharing future characteristics, wherein the parameter sharing space extraction module adopts a sharing graph convolution network to extract the sharing characteristics, and the calculation formula is as follows:
in the formula, W c (k) A weight matrix representing a k-th layer shared parameter module,
meaning the same as in step S3,
and
meaning the same as in step S4, σ (-) denotes an activation function,
representing the history feature matrix of the k-th layer shared graph convolution network output of the ith time window, and defining the history feature matrix as the shared history feature
A future feature matrix representing k-th layer shared graph convolution network output, defined as shared future features
S6: the historical total characteristics of the local information are measured
Future gross characteristics of global information
Sharing future features
And shared history features
Inputting the information to an attention fusion module; by means of future general features
For each time window historical overall characteristics
Attention weights are assigned, the formula is as follows:
in the formula, att (. Alpha.) represents an attention function, alpha 1 ，α 2 ，...，α w ∈R h Respectively representing historical gross characteristics
The att function is specifically calculated as follows:
in the formula, ω i ∈R 1*h Representing the historical total characteristics of the ith time window
Values of interest in each dimension, ω ∈ R w*h Represents w timesOmega of the window i A matrix of values of interest is composed,
representing a weight matrix, b A ∈R h*h Represents the bias term, tanh (-) represents the activation function, g ∈ R h*1 Representing a vector of interest; then, the attention value matrix omega is normalized by columns by adopting a softmax activation function to obtain alpha 1 ，α 2 ，...，α w And then the final attention matrix A is obtained α The calculation formula is as follows:
in the formula, alpha i，j E R represents the attention coefficient of the jth dimension history total characteristic of the ith window, alpha i ∈R h The attention vector, representing the historical total features of the ith window, is the output of the att (-) function, A α ∈R w*h Denotes alpha 1 ，α 2 ，...，α w A composed attention matrix;
s7: will be provided with
Respectively by column averaging to obtain
And combine them into a fused historical feature Z' L ∈R w*h Attention matrix A α And fuse historical feature Z' L Obtaining the historical feature vector P after the attention fusion by calculating the inner product according to the columns L ∈R h ；
S8: according to the steps S6 and S7, calculating to obtain shared history characteristics
Corresponding attention matrix A α Further, the shared history feature vector P with the attention weight fused is obtained CL ∈R h ；
S9: will Z G 、Z CG Respectively according to the columnMean value gives vector Z' G 、Z′ CG ∈R h Introduction of said P into L 、P CL 、Z′ G 、Z′ CG Splicing to the final fusion feature Z = [ P = L ，P CL ，Z′ G ，Z′ CG ]∈R 4h ；
S10: inputting the fusion characteristic Z into a multilayer perceptron to obtain a final output value of the photovoltaic power generation power prediction model, wherein a calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
the method comprises the steps of representing a photovoltaic power generation power predicted value, wherein MLP represents a multilayer perceptron which is composed of a plurality of full-connection layers;
s11: constructing an overall objective function L of the photovoltaic power generation power prediction model:
L＝L M +γL c +βL d
in the formula, L M In order to take the mean square error MSE as a loss function for the prediction task, L c For consistency constraint, gamma is the multiplier corresponding to consistency constraint, L d Is independence constraint, beta is multiplier corresponding to independence constraint;
s12: forecasting and comparing experimental results, selecting different time granularities, forecasting the photovoltaic power generation power for two continuous days by adopting the method provided by the invention, comparing the power with actual power to draw a curve graph, and selecting evaluation indexes to compare and evaluate the quality of a model;
the evaluation indexes are MAE, MAPE and RMSE, and the formula is as follows:
in the formula (I), the compound is shown in the specification,
and Y i And respectively representing the predicted value and the true value of the photovoltaic power generation power, wherein n is the number of samples.
Preferably, the loss function L M The calculation formula of (2) is as follows:
in the formula (I), the compound is shown in the specification,
and Y i And respectively representing the predicted value and the true value of the photovoltaic power generation power, wherein n is the number of samples.
Preferably, the consistency constraint is:
wherein w represents the number of time windows, D KL Showing the difference of the two characteristic distributions as measured by KL divergence.
Preferably, the independence constraint is:
in the formula, HSIC represents that the independence between two characteristics is measured by using a Hilbert-Schmidt independence index.
Preferably, the local information includes at least one of: solar irradiance, temperature, humidity, soil humidity, CO in historical meteorological data 2 Air pressure, wind direction, instantaneous wind speed, 2 minute average wind speed, 10 minute average wind speed, evaporation, PM2.5, cumulative 1 day radiant quantity, and power station generated power, direct voltage, direct current, electricity usage voltage, battery power in the power station generated data.
Preferably, the global information includes at least one of the following: future solar irradiance, temperature, humidity, air pressure, wind speed, wind direction, precipitation, cloud cover in weather forecast data.
The invention has the beneficial effects that: extracting time sequence and space visual angle characteristic information from global and local visual angles respectively by adopting a series of deep learning technologies; a design parameter sharing feature extraction module extracts public features, adaptively fuses feature representation and learned weight by using an attention mechanism, and finally extracts the most relevant information from the network model, which is difficult to realize by the traditional graph network; consistency constraint and independence constraint are introduced to construct a loss function, effectiveness of the two constraints is verified through experiments, and the problem that the photovoltaic power generation power prediction accuracy is low can be solved through the model.
Drawings
FIG. 1 is a flow chart of a photovoltaic power generation power prediction method of the present invention;
FIG. 2 is a block diagram of a gate cycle unit GRU model of the present invention;
FIG. 3 is a graph of attention weight assignment according to the present invention;
fig. 4 is a comparison graph of predicted power and actual power using the method of the present invention.
Detailed Description
The specific embodiment is as follows:
and respectively extracting features of the original data from a global view and a local view. The collected data is divided into two categories: one type is dynamic time series data which changes along with time and can be regarded as local information, and the other type is non-time series data which does not change along with time and can be regarded as global information. The invention adopts historical meteorological data which comprises characteristics of solar irradiance, temperature, humidity, soil humidity and CO 2 Air pressure, wind direction, instantaneous wind speed, 2-minute average wind speed, 10-minute average wind speed, evaporation, PM2.5 and accumulated 1-day radiant quantity; the power station power generation data comprises power station power generation power, direct current voltage, direct current, power utilization voltage, battery voltage and battery power; the two data are used as local information; weather forecast data including characteristics of future solar irradiance, temperature, humidity, air pressure, wind speed, wind direction, precipitation and cloud cover are used as global informationAnd (4) information. And subsequently, different feature extraction methods are respectively adopted for the data under the two visual angles so as to obtain the feature representations of the data after the data are respectively subjected to information fusion.
As shown in fig. 1, for historical meteorological data and power station power generation data of local information, a series of deep learning techniques are adopted to extract features according to a visual angle sequence of a time sequence and a space, and the specific method comprises the following steps:
s1: defining the number of features of the local information as n 1 The time elapsed for each time window is T 1 Data granularity of T 2 Then a time window size is win = T 1 /T 2 W is the number of time windows, and the local information in each time window is calculated
The time sequence feature extraction module adopts a GRU gate cycle unit to extract the time sequence feature, as shown in FIG. 2, the calculation formula is as follows:
in the formula, x t Local information indicating time t
h t-1 Indicating the hidden state at the previous moment, r t Denotes a reset gate, z t Representing the update gate, σ (-) representing the activation function, W r 、W z Representing a weight matrix, b r 、b z Representing a bias vector;
s2: according to the r t 、z t Calculating the hidden state h t The calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
represents a candidate hidden state, h t Representing a hidden state at the time t obtained by the GRU module, tan h (phi) representing an activation function, W representing an operation by element h Representing a weight matrix, b h Representing a bias vector; the hidden state h t For the time-series characteristics of the local information,
extracting the time sequence characteristics of the local information of the module for the time sequence characteristics of the w time windows;
s3: the time-sequence characteristics of local information
The method comprises the following steps of inputting the information into a spatial feature extraction module, further aggregating feature spatial information and obtaining historical total features of local information, wherein the module usually adopts a GNN, GAT or GCN graph convolution network model, the spatial feature extraction module adopts a GCN graph convolution network to extract spatial features, and the calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
a contiguous matrix is represented that is,
the unit matrix is represented by a matrix of units,
is represented by an adjacency matrix A l And identity matrix I l The sum of the reachable matrices, k representing the number of layers of the convolutional network of the graph, W l (k) A k-th layer weight matrix representing a graph convolution network, σ (-) representing an activation function,
to represent
A diagonal matrix of (a);
the initial input of the graph convolution network representing the ith time window is also the timing feature module output
A feature matrix representing the convolution output of the k-th layer map of the ith time window, which is defined as a historical feature
Then
Extracting historical total features of local information of the module for the spatial features of the w time windows;
the method comprises the following steps of extracting future total features aiming at weather forecast data of global information and space structure features of the weather forecast data, wherein the future total features are specifically as follows:
s4: specifying the number of features of global information as n 2 The time length of collecting the global information is T 3 Data granularity of T 4 Global information is formed
Inputting the data into a graph convolution network of the spatial feature extraction module, and obtaining future total features Z of global information after k-layer learning G The calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
a contiguous matrix is represented that is,
the unit matrix is represented by a matrix of units,
is represented by an adjacency matrix A g And identity matrix I g The sum of the reachable matrices, k representing the number of layers of the convolutional network of the graph, W g (k) A k-th layer weight matrix representing a graph convolution network, σ (-) representing an activation function,
to represent
The diagonal matrix of (a) is,
representing the initial input of the graph convolutional network, also the global information X g ，
Defining the feature matrix of the graph convolution network output of the k layer as the future total feature Z G ；
Designing a parameter sharing spatial feature extraction module to extract public feature representation under two visual angles; there is also a potential association between features of the global and local views, so it is also necessary to extract common features shared by them, so that the model can be more adaptively applied to subsequent prediction tasks, which are specifically:
s5: the local information X is processed l And global information X g The method comprises the following steps of inputting the data to a parameter sharing space extraction module to obtain sharing historical characteristics and sharing future characteristics, wherein the parameter sharing space extraction module adopts a sharing graph convolution network to extract the sharing characteristics, and the calculation formula is as follows:
in the formula, W c (k) A weight matrix representing a k-th layer shared parameter module,
the physical meaning and the instituteThe same applies to the step S3 described above,
and
the physical meaning is the same as in the step S4, σ (-) denotes the activation function,
is the history characteristic matrix of the k-th layer shared graph convolution network output of the ith time window, and is defined as the shared history characteristic
A future feature matrix representing k-th layer shared graph convolution network output, defined as shared future features
As shown in fig. 3, an attention mechanism is further introduced, different attention weights are distributed according to the historical total features and the shared historical features obtained in the steps S3 and S5 and according to a time window, and fusion features are obtained after weighted fusion and used for final photovoltaic power generation power prediction; the attention mechanism can enable the model to focus more critical information on the current task in a plurality of input information, reduce the attention degree on other information, and filter irrelevant information, which specifically comprises the following steps:
s6: the historical total characteristics of the local information are measured
Future gross characteristics of global information
Sharing future features
And shared history features
Inputting the information to an attention fusion module; by means of future general features
For each time window historical overall characteristics
Attention weights are assigned, the formula is as follows:
in the formula, att (. Alpha.) represents an attention function, alpha 1 ，α 2 ，...，α w ∈R h Respectively representing historical gross characteristics
The specific calculation formula of att function is as follows:
in the formula, omega i ∈R 1*h Representing the historical total characteristics of the ith time window
Values of interest in each dimension, ω ∈ R w*h ω representing w time windows i A matrix of values of interest is composed,
representing a weight matrix, b A ∈R h*h Represents the bias term, tanh (-) represents the activation function, g ∈ R h*1 Representing a vector of interest; then, the attention value matrix omega is normalized by columns by adopting a softmax activation function to obtain alpha 1 ，α 2 ，...，α w And then the final attention matrix A is obtained α The calculation formula is as follows:
in the formula, alpha i，j E R represents the attention coefficient of the jth dimension history total feature of the ith window, and the larger the value is, the more important the feature of the jth dimension of the corresponding ith window is; alpha is alpha i ∈R h The attention vector representing the historical total features of the ith window, which is the output of the att (-) function, A α ∈R w*h Denotes alpha 1 ，α 2 ，...，α w A composed attention matrix;
s7: will be provided with
Respectively by column averaging to obtain
And combine them into a fused historical feature Z' L ∈R w*h Attention matrix A α And fuse historical feature Z' L Obtaining the historical feature vector P after the attention fusion by calculating the inner product according to the columns L ∈R h ；
S8: according to the steps S6 and S7, calculating to obtain shared history characteristics
Corresponding attention matrix A α Further, the shared history feature vector P with the attention weight fused is obtained CL ∈R h ；
S9: will Z G 、Z CG Respectively obtaining vector Z 'by column averaging' G 、Z′ CG ∈R h Introduction of said P into L 、P CL 、Z′ G 、Z′ CG Splicing to final fusion signature Z = [ P = [ ] L ，P CL ，Z′ G ，Z′ CG ]∈R 4h ；
S10: inputting the fusion characteristic Z into a multilayer perceptron to obtain a final output value of the photovoltaic power generation power prediction model, wherein a calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
the method comprises the steps of representing a photovoltaic power generation power predicted value, wherein MLP represents a multilayer perceptron which is composed of a plurality of full-connection layers;
s11: constructing an overall objective function L of the photovoltaic power generation power prediction model:
L＝L M +γL c +βL d
in the formula, L M For the loss function with mean square error MSE as the prediction task, L c For consistency constraint, gamma is the multiplier corresponding to consistency constraint, L d And beta is a multiplier corresponding to the independence constraint.
The loss function L M The calculation formula of (2) is as follows:
in the formula (I), the compound is shown in the specification,
and Y i And respectively representing the predicted value and the true value of the photovoltaic power generation power, wherein n is the number of samples.
For sharing the feature representation Z generated by the feature extraction module with the parameters CG And
further decoupling, using KL divergence to measure this constraint, the consistency constraint formula is as follows:
wherein w represents the number of time windows, D KL Showing the difference of the two characteristic distributions as measured by KL divergence.
To ensure Z G ，Z CG 、
Can capture different information, and measures and enhances the difference between two characteristics by using Hilbert-Schmidt Independence Criterion, HSIC for short, and the Independence constraint formula is as follows:
in the formula, HSIC represents that the independence between two characteristics is measured by using a Hilbert-Schmidt independence index. The objective function with the consistency constraint and the independence constraint can adaptively fuse the characteristic information in the model training, and further more accurate photovoltaic power generation power prediction is realized.
S12: prediction was performed and experimental results were compared. According to the method, historical meteorological data, power station power generation data and weather forecast data of 30 meshes from 2022, 4 and 1 to 2022, 6 and 6 provided by Dalian Chinese academy of sciences are selected as data sets, three time granularities of 10/30/60min are set, the photovoltaic power generation power of 30 meshes on 29 days from 2022 and 6 and months is predicted by the method provided by the invention, and the power is compared with actual power, as shown in figure 4, the prediction method is good in fitting effect under different time granularities. In order to compare and evaluate the model, MAE, MAPE and RMSE are selected as evaluation indexes, and the formula is as follows:
in the formula (I), the compound is shown in the specification,
and Y i And respectively representing the predicted value and the true value of the photovoltaic power generation power, wherein n is the number of samples. The model provided by the invention, the three variant models thereof, namely the model without consistency constraint, independence constraint or both constraint and the other two existing models, namely SVM and ARIMA models are compared, and the evaluation index calculation result is shown in table 1, so that the method can be seenCompared with SVM and ARIMA, the prediction result of the proposed model has smaller numerical value; the model with the consistency constraint and the independence constraint has the best prediction performance, the effectiveness of the method provided by the invention is verified, and the excellent prediction performance is shown for the photovoltaic power generation power prediction problem.
TABLE 1 calculation results of three evaluation indexes
The key point of the invention is that (1) the global and local feature representation of the existing data is constructed from the four visual angles of global, local, time sequence and space; (2) Constructing a parameter sharing feature extraction module to extract public features, and introducing an attention fusion module to adaptively distribute attention weights for local features of different time windows, so as to generate fusion features for predicting the photovoltaic power generation power; (3) In order to strengthen the performance of the model, a loss function with consistency constraint and independence constraint is designed to be used for more accurate photovoltaic power generation power prediction.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein. The scheme in the embodiment of the application can be implemented by adopting various computer languages, such as object-oriented programming language Java and transliterated scripting language JavaScript.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
While the preferred embodiments of the present application have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the present application.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present application without departing from the spirit and scope of the application. Thus, if such modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is intended to include such modifications and variations as well.
Claims (6)
1. A photovoltaic power generation power prediction method based on multi-view adaptive feature fusion is characterized by comprising the following steps:
s1: defining the number of features of the local information as n 1 The time elapsed for each time window is T 1 Data granularity of T 2 Then a time window size is win = T 1 /T 2 W is the number of time windows, and the local information in each time window is
The time sequence feature extraction module adopts a GRU gate cycle unit to extract the time sequence feature, and the calculation formula is as follows:
in the formula, x t Local information indicating time t
h t-1 Indicating the hidden state at the previous moment, r t Denotes a reset gate, z t Represents the update gate, σ (-) represents the activation function, W r 、W z Representing a weight matrix, b r 、b z Representing a bias vector;
s2: according to the r t 、z t Calculating the hidden state h t The calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
represents a candidate hidden state, h t Representing a hidden state at the time t obtained by the GRU module, tan h (phi) representing an activation function, W representing an operation by element h Representing a weight matrix, b h Representing a bias vector; the hidden state h t Is a time-series characteristic of the local information,
extracting the time sequence characteristics of the local information of the module for the time sequence characteristics of the w time windows;
s3: the time-sequence characteristics of local information
The method comprises the following steps of inputting the data to a spatial feature extraction module and obtaining historical total features of local information, wherein the spatial feature extraction module adopts a graph convolution network to extract spatial features, and the calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
a contiguous matrix is represented that is,
the matrix of the unit is expressed by,
is represented by an adjacency matrix A l And identity matrix I l The sum of the reachable matrices, k representing the number of layers of the convolutional network of the graph, W l (k) A k-th layer weight matrix representing a graph convolution network, σ (-) representing an activation function,
to represent
A diagonal matrix of (a);
the initial input of the graph convolution network representing the ith time window is also the timing feature module output
A feature matrix representing the convolution output of the k-th layer map of the ith time window, which is defined as a historical feature
Then
Extracting historical total features of local information of the module for the spatial features of the w time windows;
s4: specifying the number of features of global information as n 2 The time length of collecting the global information is T 3 Data granularity of T 4 Global information is provided
Inputting the data into a graph convolution network of the spatial feature extraction module, and obtaining future total features Z of global information after k-layer learning G The calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
a matrix of adjacency is represented by a matrix of adjacency,
the unit matrix is represented by a matrix of units,
is represented by an adjacency matrix A g And identity matrix I g The sum of the resulting reachable matrices, k representing the number of layers of the graph convolution network,
a k-th layer weight matrix representing a graph convolution network, sigma (-) represents an activation function,
to represent
The diagonal matrix of (a) is,
representing the initial input of the graph convolutional network, also the global information X g ，
Defining the feature matrix of the graph convolution network output of the k layer as the future total feature Z G ；
S5: the local information X is processed l And global information X g The method comprises the following steps of inputting the data to a parameter sharing space extraction module to obtain sharing historical characteristics and sharing future characteristics, wherein the parameter sharing space extraction module adopts a sharing graph convolution network to extract the sharing characteristics, and the calculation formula is as follows:
in the formula, W c (k) A weight matrix representing a k-th layer shared parameter module,
meaning the same as in step S3,
and
meaning the same as in step S4, σ (-) denotes an activation function,
represents the ith timeThe history feature matrix output by the k-th layer shared graph convolution network of the window is defined as the shared history feature
A future feature matrix representing the k-th layer shared graph convolution network output, defined as the shared future feature
S6: the historical total characteristics of the local information are measured
Future gross characteristics of global information
Sharing future features
And shared history features
Inputting the information to an attention fusion module; by means of future general features
For each time window historical overall characteristics
An attention weight is assigned, the formula is as follows:
in the formula, att (. Alpha.) represents an attention function, alpha 1 ,α 2 ,...,α w ∈R h Respectively representing historical gross characteristics
The att function is specifically calculated as follows:
in the formula, ω i ∈R 1*h Representing the historical total characteristics of the ith time window
Values of interest in each dimension, ω ∈ R w*h ω representing w time windows i A matrix of values of interest is composed,
representing a weight matrix, b A ∈R h'*h Represents the bias term, tanh (-) represents the activation function, q ∈ R h'*1 Representing a vector of interest; then, the attention value matrix omega is normalized by columns by adopting a softmax activation function to obtain alpha 1 ,α 2 ,...,α w And then the final attention matrix A is obtained α The calculation formula is as follows:
in the formula, alpha i,j E R represents the attention coefficient of the jth dimension history total characteristic of the ith window, alpha i ∈R h The attention vector, representing the historical total features of the ith window, is the output of the att (-) function, A α ∈R w*h Denotes alpha 1 ,α 2 ,...,α w A composed attention matrix;
s7: will be provided with
Respectively by column averaging to obtain
And combine them into a fused historical feature Z' L ∈R w*h Attention matrix A α And fuse historical feature Z' L Obtaining the historical feature vector P after the attention fusion by calculating the inner product according to the columns L ∈R h ；
S8: according to the steps S6 and S7, calculating to obtain shared history characteristics
Corresponding attention matrix A α Further, the shared history feature vector P after the fusion attention weight is obtained CL ∈R h ；
S9: will Z G 、Z CG Respectively averaging according to rows to obtain vector Z' G 、Z' CG ∈R h Introducing said P L 、P CL 、Z' G 、Z' CG Splicing to final fusion signature Z = [ P = [ ] L ,P CL ,Z' G ,Z' CG ]∈R 4h ；
S10: inputting the fusion characteristic Z into a multilayer perceptron to obtain a final output value of the photovoltaic power generation power prediction model, wherein a calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
the method comprises the steps of representing a photovoltaic power generation power predicted value, wherein MLP represents a multilayer perceptron which is composed of a plurality of full-connection layers;
s11: constructing an overall objective function L of the photovoltaic power generation power prediction model:
L＝L M +γL c +βL d
in the formula, L M For the loss function with mean square error MSE as the prediction task, L c For consistency constraint, gamma is the multiplier corresponding to consistency constraint, L d Is independence constraint, beta is multiplier corresponding to independence constraint;
s12: predicting and comparing experimental results, selecting different time granularities, predicting the photovoltaic power generation power of two continuous days by adopting the method provided by the invention, comparing the power with the actual power to draw a curve graph, and selecting evaluation indexes to compare and evaluate the quality of the model;
the evaluation indexes are MAE, MAPE and RMSE, and the formula is as follows:
in the formula (I), the compound is shown in the specification,
and Y i And respectively representing the predicted value and the true value of the photovoltaic power generation power, wherein n is the number of samples.
2. The method according to claim 1, wherein the loss function L is a function of a maximum power loss of the photovoltaic power generation system M The calculation formula of (2) is as follows:
in the formula (I), the compound is shown in the specification,
and Y i And respectively representing the predicted value and the true value of the photovoltaic power generation power, wherein n is the number of samples.
3. The multi-view adaptive feature fusion based photovoltaic power generation power prediction method according to claim 1, wherein the consistency constraint is:
wherein w represents the number of time windows, D KL Is represented by KL divergence measures the difference in the distribution of the two features.
4. The method according to claim 1, wherein the independence constraint is:
in the formula, HSIC represents that the independence between two characteristics is measured by using a Hilbert-Schmidt independence index.
5. The method according to claim 1, wherein the local information comprises at least one of the following: solar irradiance, temperature, humidity, soil humidity, CO in historical meteorological data 2 Air pressure, wind direction, instantaneous wind speed, 2 minute average wind speed, 10 minute average wind speed, evaporation, PM2.5, cumulative 1 day radiant quantity, and power station generated power, direct voltage, direct current, electricity usage voltage, battery power in the power station generated data.
6. The method according to claim 1, wherein the global information comprises at least one of the following: future solar irradiance, temperature, humidity, air pressure, wind speed, wind direction, precipitation, cloud cover in weather forecast data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211560534.0A CN115829126A (en) | 2022-12-07 | 2022-12-07 | Photovoltaic power generation power prediction method based on multi-view self-adaptive feature fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211560534.0A CN115829126A (en) | 2022-12-07 | 2022-12-07 | Photovoltaic power generation power prediction method based on multi-view self-adaptive feature fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115829126A true CN115829126A (en) | 2023-03-21 |
Family
ID=85545388
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211560534.0A Pending CN115829126A (en) | 2022-12-07 | 2022-12-07 | Photovoltaic power generation power prediction method based on multi-view self-adaptive feature fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115829126A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116757369A (en) * | 2023-08-22 | 2023-09-15 | 国网山东省电力公司营销服务中心(计量中心) | Attention mechanism-based carbon emission analysis method and system |
| CN117150326A (en) * | 2023-10-31 | 2023-12-01 | 深圳市大数据研究院 | New energy node output power prediction method, device, equipment and storage medium |
| CN117556379A (en) * | 2024-01-12 | 2024-02-13 | 西南石油大学 | Photovoltaic power generation power prediction method based on depth feature fusion under domain knowledge constraint |
| CN117592004A (en) * | 2024-01-19 | 2024-02-23 | 中国科学院空天信息创新研究院 | PM2.5 concentration satellite monitoring method, device, equipment and medium |
-
2022
- 2022-12-07 CN CN202211560534.0A patent/CN115829126A/en active Pending
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116757369A (en) * | 2023-08-22 | 2023-09-15 | 国网山东省电力公司营销服务中心(计量中心) | Attention mechanism-based carbon emission analysis method and system |
| CN116757369B (en) * | 2023-08-22 | 2023-11-24 | 国网山东省电力公司营销服务中心(计量中心) | Attention mechanism-based carbon emission analysis method and system |
| CN117150326A (en) * | 2023-10-31 | 2023-12-01 | 深圳市大数据研究院 | New energy node output power prediction method, device, equipment and storage medium |
| CN117150326B (en) * | 2023-10-31 | 2024-02-23 | 深圳市大数据研究院 | New energy node output power prediction method, device, equipment and storage medium |
| CN117556379A (en) * | 2024-01-12 | 2024-02-13 | 西南石油大学 | Photovoltaic power generation power prediction method based on depth feature fusion under domain knowledge constraint |
| CN117556379B (en) * | 2024-01-12 | 2024-04-09 | 西南石油大学 | Photovoltaic power generation power prediction method based on depth feature fusion under domain knowledge constraint |
| CN117592004A (en) * | 2024-01-19 | 2024-02-23 | 中国科学院空天信息创新研究院 | PM2.5 concentration satellite monitoring method, device, equipment and medium |
| CN117592004B (en) * | 2024-01-19 | 2024-04-12 | 中国科学院空天信息创新研究院 | PM2.5 concentration satellite monitoring method, device, equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | Random forest solar power forecast based on classification optimization | |
| Wang et al. | A day-ahead PV power forecasting method based on LSTM-RNN model and time correlation modification under partial daily pattern prediction framework | |
| Yu et al. | An LSTM short-term solar irradiance forecasting under complicated weather conditions | |
| Li et al. | Photovoltaic power forecasting with a hybrid deep learning approach | |
| US20220373984A1 (en) | Hybrid photovoltaic power prediction method and system based on multi-source data fusion | |
| CN115829126A (en) | Photovoltaic power generation power prediction method based on multi-view self-adaptive feature fusion | |
| CN112529282A (en) | Wind power plant cluster short-term power prediction method based on space-time graph convolutional neural network | |
| CN110909919A (en) | Photovoltaic power prediction method of depth neural network model with attention mechanism fused | |
| CN105868853B (en) | Method for predicting short-term wind power combination probability | |
| Wang et al. | Comparative study of machine learning approaches for predicting short-term photovoltaic power output based on weather type classification | |
| CN103218674A (en) | Method for predicating output power of photovoltaic power generation system based on BP (Back Propagation) neural network model | |
| CN105069521A (en) | Photovoltaic power plant output power prediction method based on weighted FCM clustering algorithm | |
| Wang et al. | Deep learning based irradiance mapping model for solar PV power forecasting using sky image | |
| CN114792156A (en) | Photovoltaic output power prediction method and system based on curve characteristic index clustering | |
| CN107679687A (en) | A kind of photovoltaic output modeling method and Generation System Reliability appraisal procedure | |
| Liu et al. | A short-term probabilistic photovoltaic power prediction method based on feature selection and improved LSTM neural network | |
| Li et al. | Short-term Forecasting Approach Based on bidirectional long short-term memory and convolutional neural network for Regional Photovoltaic Power Plants | |
| CN116629416A (en) | Photovoltaic power station power prediction method and device | |
| Chen et al. | MultiCycleNet: multiple cycles self-boosted neural network for short-term electric household load forecasting | |
| CN116345555A (en) | CNN-ISCA-LSTM model-based short-term photovoltaic power generation power prediction method | |
| Li et al. | Photovoltaic power prediction method for zero energy consumption buildings based on multi-feature fuzzy clustering and MAOA-ESN | |
| 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 | |
| CN108038518A (en) | A kind of photovoltaic generation power based on meteorological data determines method and system | |
| Li et al. | Deep learning model for short-term photovoltaic power forecasting based on variational mode decomposition and similar day clustering |
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 |