CN114971058B - Photovoltaic forecasting method based on depth attention network and clear sky radiation priori fusion - Google Patents
Photovoltaic forecasting method based on depth attention network and clear sky radiation priori fusion Download PDFInfo
- Publication number
- CN114971058B CN114971058B CN202210651311.9A CN202210651311A CN114971058B CN 114971058 B CN114971058 B CN 114971058B CN 202210651311 A CN202210651311 A CN 202210651311A CN 114971058 B CN114971058 B CN 114971058B
- Authority
- CN
- China
- Prior art keywords
- photovoltaic power
- forecast
- clear sky
- value
- historical
- 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.)
- Active
Links
- 230000005855 radiation Effects 0.000 title claims abstract description 117
- 238000013277 forecasting method Methods 0.000 title claims abstract description 15
- 230000004927 fusion Effects 0.000 title claims abstract description 14
- 238000012549 training Methods 0.000 claims abstract description 49
- 238000000034 method Methods 0.000 claims abstract description 38
- 238000012805 post-processing Methods 0.000 claims abstract description 5
- 230000008569 process Effects 0.000 claims description 24
- 230000007246 mechanism Effects 0.000 claims description 23
- 239000004973 liquid crystal related substance Substances 0.000 claims description 18
- 238000005315 distribution function Methods 0.000 claims description 16
- 239000011159 matrix material Substances 0.000 claims description 16
- 238000012360 testing method Methods 0.000 claims description 16
- 238000005070 sampling Methods 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 description 41
- 238000010248 power generation Methods 0.000 description 16
- 238000010586 diagram Methods 0.000 description 7
- 108090000672 Annexin A5 Proteins 0.000 description 4
- 238000000605 extraction Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000001419 dependent effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000007787 long-term memory Effects 0.000 description 2
- 230000009466 transformation 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
- 241000070928 Calligonum comosum Species 0.000 description 1
- OKTJSMMVPCPJKN-UHFFFAOYSA-N Carbon Chemical compound [C] OKTJSMMVPCPJKN-UHFFFAOYSA-N 0.000 description 1
- 238000010521 absorption reaction Methods 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 229910052799 carbon Inorganic materials 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000029087 digestion Effects 0.000 description 1
- 238000006386 neutralization reaction Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
- G06N3/045—Combinations of networks
-
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- 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)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Business, Economics & Management (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Strategic Management (AREA)
- Evolutionary Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Mathematical Analysis (AREA)
- Pure & Applied Mathematics (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Computational Mathematics (AREA)
- Mathematical Optimization (AREA)
- Water Supply & Treatment (AREA)
- Primary Health Care (AREA)
- Public Health (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
Abstract
The invention discloses a photovoltaic forecasting method based on depth attention network and clear sky radiation priori fusion, which aims to solve the problems of low forecasting precision and insufficient forecasting reliability of the existing photovoltaic power forecasting method and acquire historical photovoltaic power data and future clear sky radiation intensity; establishing a model, inputting the acquired historical photovoltaic power data and future clear sky radiation intensity into the model for training, outputting a photovoltaic power forecast value, obtaining a photovoltaic power forecast value sequence, and obtaining a trained model; performing post-processing on the photovoltaic power forecast value sequence by utilizing the nuclear density estimation to obtain a forecast probability interval of the photovoltaic power true value; the method comprises the steps of obtaining historical photovoltaic power data before photovoltaic power moment to be forecasted and clear sky radiation intensity after the photovoltaic power moment to be forecasted, inputting the historical photovoltaic power data and the clear sky radiation intensity into a model, and outputting a forecasting probability interval of a photovoltaic power forecasting value and a photovoltaic power true value of the photovoltaic power moment to be forecasted.
Description
Technical Field
The invention relates to a forecasting method, in particular to a method for realizing accurate point forecasting and probability interval forecasting of future photovoltaic power by utilizing historical photovoltaic power data and future clear sky radiation priori information, and belongs to the technical field of photovoltaic power generation forecasting.
Background
With the aim of carbon neutralization, the world enters an energy transformation and transformation period, and the strong development of renewable energy sources such as wind, light and the like has become a wide consensus among countries of the world. Photovoltaic power generation is an efficient way of utilizing solar energy. Due to the randomness, fluctuation and intermittence of the photovoltaic power, the problem of effective absorption of the photovoltaic power generation is increasingly serious along with the annual increase of the installed capacity of the photovoltaic power generation in China. Developing accurate photovoltaic power generation power forecast is an effective means for promoting the digestion of photovoltaic power generation and is widely paid attention to by researchers.
In the prior art, future photovoltaic power is forecasted by using a long short-term memory (LSTM) neural network and historical photovoltaic power data, the deep LSTM neural network can well model and describe dynamic characteristics in the historical data, and application in actual data shows that the deep LSTM neural network has a good forecasting effect in photovoltaic forecasting, but the method only considers the historical photovoltaic power data, does not fully consider fusion of inherent priori information of a photovoltaic power generation system and the historical photovoltaic power data, does not conduct adaptive fusion of the priori information and the historical photovoltaic power data, and can lead to insufficient forecasting accuracy and insufficient forecasting reliability of the photovoltaic power.
Disclosure of Invention
The invention aims to solve the problems of low forecasting precision and insufficient forecasting reliability of the existing photovoltaic power forecasting method, and further provides a photovoltaic forecasting method based on the prior fusion of a deep attention network and clear sky radiation.
The technical scheme adopted by the invention is as follows:
it comprises the following steps:
s1, acquiring historical photovoltaic power data and future clear sky radiation intensity;
s2, establishing a depth attention ConvLSTM model, wherein the depth attention ConvLSTM model sequentially comprises a ConvLSTM1 layer, an attention mechanism layer, a ConvLSTM2 layer, a flame layer and a full connection layer, inputting acquired historical photovoltaic power data and future clear sky radiation intensity into the depth attention ConvLSTM model for training, outputting a photovoltaic power forecast value, and obtaining a photovoltaic power forecast value sequence until loss converges, so as to obtain a trained depth attention ConvLSTM model;
s3, performing post-processing on the photovoltaic power forecast value sequence obtained in the S2 by utilizing nuclear density estimation to obtain a forecast probability interval of the photovoltaic power true value;
s4, acquiring historical photovoltaic power data before the photovoltaic power moment to be forecasted and clear sky radiation intensity after the photovoltaic power moment to be forecasted, inputting the acquired historical photovoltaic power data and the clear sky radiation intensity into a depth attention ConvLSTM model trained in S2, outputting a photovoltaic power forecast value of the photovoltaic power moment to be forecasted, and executing S3 to obtain a forecast probability interval of a photovoltaic power true value of the photovoltaic power moment to be forecasted.
Preferably, the step S1 of acquiring historical photovoltaic power data and future clear sky radiation intensity includes the following specific steps:
the method comprises the steps of acquiring three years of historical photovoltaic power data and future clear air radiation intensities corresponding to the historical photovoltaic power data before a certain photovoltaic power moment, wherein the sampling interval time of two adjacent sets of historical photovoltaic power data is 15 minutes, taking the historical photovoltaic power data of the first year and the second year and the corresponding future clear air radiation intensities as training sets, and taking the historical photovoltaic power data of the third year and the corresponding future clear air radiation intensities as test sets.
Preferably, in the step S2, a deep attention ConvLSTM model is built, the deep attention ConvLSTM model sequentially includes a ConvLSTM1 layer, an attention mechanism layer, a ConvLSTM2 layer, a flame layer and a full connection layer, the obtained historical photovoltaic power data and the future clear sky radiation intensity are input into the deep attention ConvLSTM model for training, a photovoltaic power forecast value is output, a photovoltaic power forecast value sequence is obtained until loss converges, and a trained deep attention ConvLSTM model is obtained, and the specific process is as follows:
s21, inputting the historical photovoltaic power data in the training set into a ConvLSTM1 layer of a depth attention ConvLSTM model, and outputting correlation information among a plurality of variables at the same moment and time correlation information of adjacent moments of the same variable in the historical photovoltaic power data; the plurality of variables in the historical photovoltaic power data comprise photovoltaic power, actual solar radiation intensity, clear sky factor and temperature;
S22, inputting correlation information among a plurality of variables at the same time and time correlation information at the same variable adjacent time in the historical photovoltaic power data output in the S21 and future clear air radiation intensity in the training set into an attention mechanism layer of a deep attention ConvLSTM model, and outputting information obtained by combining the correlation information among the plurality of variables at the same time and the time correlation information at the same variable adjacent time in the future clear air radiation intensity and the historical photovoltaic power data;
s23, inputting the combined information output in the S22 into a ConvLSTM2 layer of a depth attention ConvLSTM model, and outputting a feature matrix of the combined information;
s24, inputting the feature matrix of the combined information into a flat layer of a depth attention ConvLSTM model, and outputting a unidimensional feature matrix;
s25, inputting the unidimensional feature matrix into a full-connection layer of the depth attention ConvLSTM model, and outputting a photovoltaic power forecast value to obtain a photovoltaic power forecast value sequence.
Preferably, in the step S22, the correlation information between the future clear sky radiation intensity in the training set and the multiple variables at the same time and the time correlation information between adjacent times of the same variables in the historical photovoltaic power data output in the step S21 are input into an attention mechanism layer of a deep attention ConvLSTM model, and the correlation information between the multiple variables at the same time and the time correlation information between adjacent times of the same variables in the future clear sky radiation intensity and the historical photovoltaic power data are output, where the specific process is as follows:
And acquiring correlation information among a plurality of variables at the same time in the historical photovoltaic power data in the training set and time correlation information among adjacent time of the same variable and weight proportion among clear sky radiation intensity in the historical photovoltaic power data by using an attention mechanism, and combining the correlation information among the plurality of variables at the same time in the historical photovoltaic power data and time correlation information among the adjacent time of the same variable and future clear sky radiation intensity information in the training set according to the weight proportion to obtain combined information.
Preferably, in the step S3, the sequence of photovoltaic power forecast values obtained in the step S2 is post-processed by using the kernel density estimation to obtain a forecast probability interval of the true photovoltaic power, and the specific process is as follows:
s31, calculating and estimating a joint probability density function by utilizing two-dimensional kernel density estimation according to the photovoltaic power true value sequence and the photovoltaic power forecast value sequence;
s32, calculating a conditional probability density function of the photovoltaic power true value by using the estimated joint probability density function;
s33, calculating a conditional probability distribution function of the photovoltaic power true value by using the conditional probability density function of the photovoltaic power true value;
s34, calculating a probability confidence interval of the photovoltaic power true value by using a conditional probability distribution function of the photovoltaic power true value to obtain a forecast probability interval of the photovoltaic power true value;
S35, inputting the test set into the trained depth attention ConvLSTM model in S2, outputting a photovoltaic power forecast value, judging whether the photovoltaic power forecast value is accurate or not by using a standardized root mean square error, if the photovoltaic power forecast value is smaller than 0.1, the photovoltaic power forecast value is accurate, otherwise, executing S2;
normalized root mean square error:
wherein Cap represents the rated power of the photovoltaic panel, y i Representing the true value of the photovoltaic power,
represents a photovoltaic power forecast value, n represents a photovoltaic power sample number, i=1, 2,..n;
when the photovoltaic power forecast value is accurate, executing S34 to obtain a forecast probability interval of the photovoltaic power true value, evaluating the forecast probability interval of the photovoltaic power true value by using the coverage probability of the forecast interval, if the coverage probability of the forecast interval is 85% -90%, the forecast probability interval of the photovoltaic power true value is accurate, otherwise, repeatedly executing S34;
prediction interval coverage probability:
wherein I is i An indicator variable representing the ith photovoltaic power sample,
preferably, in the step S31, a two-dimensional kernel density estimation is used to calculate and estimate a joint probability density function according to the real value sequence of the photovoltaic power and the forecast value sequence of the photovoltaic power, and the specific process is as follows:
Wherein, the liquid crystal display device comprises a liquid crystal display device,
representing an estimated joint probability density function;
x 1 representing a sequence of photovoltaic power true values;
x 2 representing a sequence of photovoltaic power forecast values;
n represents the number of samples and, n=1, 2,. -%, N;
b 1 and b 2 Representing window width;
ker (. Cndot.) represents a Gaussian kernel function,
x λ representing a sequence of true or predicted photovoltaic power values, x when λ=1 λ Is x 1 At this time, a sequence of true photovoltaic power values is represented; when λ=2, x λ I.e. x 2 At this time, a photovoltaic power forecast value sequence is represented;
x λn representing an nth sample in the sequence of photovoltaic power real values or the sequence of photovoltaic power forecast values, when λ=1, representing the nth sample in the sequence of photovoltaic power real values at this time; when λ=2, this represents the nth sample in the sequence of photovoltaic power forecast values;
b λ represents window width, b when λ=1 λ B is 1 When λ=2, b λ B is 2 ；
The optimal window width value is determined by the principle that the average integral square error between the true joint probability density function and the estimated joint probability density function is minimal.
Preferably, in the step S32, a conditional probability density function of the photovoltaic power real value is calculated by using the estimated joint probability density function, and the specific process is as follows:
wherein, the liquid crystal display device comprises a liquid crystal display device,
a conditional probability density function representing a true value of photovoltaic power;
Preferably, in S33, the conditional probability distribution function of the photovoltaic power real value is calculated by using the conditional probability density function of the photovoltaic power real value, and the specific process is as follows:
wherein, the liquid crystal display device comprises a liquid crystal display device,
a conditional probability distribution function representing the true value of the photovoltaic power.
Preferably, in the step S34, a probability confidence interval of the photovoltaic power real value is calculated by using a conditional probability distribution function of the photovoltaic power real value, so as to obtain a forecast probability interval of the photovoltaic power real value, which specifically includes the following steps:
wherein x is u Representing the upper probability interval limit at 90% confidence level;
x l representing the lower probability interval limit at 90% confidence level;
p { · } represents the size of the probability value;
X 1 representing a random variable corresponding to the true value of the photovoltaic power;
X 2 and representing the random variable corresponding to the photovoltaic power predicted value.
Preferably, in the step S4, the historical photovoltaic power data before the photovoltaic power moment to be forecasted and the clear sky radiation intensity after the photovoltaic power moment to be forecasted are obtained, the obtained historical photovoltaic power data and the clear sky radiation intensity are input into a training deep attention ConvLSTM model in the step S2, a photovoltaic power forecast value of the photovoltaic power moment to be forecasted is output, the step S3 is executed, and a forecast probability interval of a photovoltaic power true value of the photovoltaic power moment to be forecasted is obtained, and the specific process is as follows:
And acquiring historical photovoltaic power data of 96 sampling points before the photovoltaic power moment to be forecasted and clear sky radiation intensity of one hour after the photovoltaic power moment to be forecasted, inputting the acquired historical photovoltaic power data and the clear sky radiation intensity into a training depth attention ConvLSTM model in S2, outputting a photovoltaic power forecast value of the photovoltaic power moment to be forecasted, and executing S34 to obtain a forecast probability interval of a photovoltaic power true value of the photovoltaic power moment to be forecasted.
The beneficial effects are that:
the invention establishes a short-term forecasting model (depth attention ConvLSTM model) of photovoltaic power, wherein the input of the model is historical photovoltaic power data and future clear sky radiation priori information, and the output is a photovoltaic power forecasting value, so as to obtain a photovoltaic power forecasting value sequence. The depth attention ConvLSTM model fuses an attention mechanism with a ConvLSTM network, historical photovoltaic power data and future clear sky radiation prior information are combined in a self-adaptive mode through the attention mechanism, the future clear sky radiation prior information is represented through clear sky radiation intensity, and feature extraction is carried out until the forecasted photovoltaic power is output. The accuracy of photovoltaic power forecasting is improved, the forecasting probability interval of the actual value of the photovoltaic power is obtained by utilizing two-dimensional nuclear density estimation of the photovoltaic power, so that accurate point forecasting and probability interval forecasting of the photovoltaic power in the future are realized, and the reliability of forecasting can be ensured.
Drawings
FIG. 1 is a block diagram of a depth attention ConvLSTM model;
FIG. 2 is a block diagram of a ConvLSTM network;
FIG. 3 is a schematic diagram of an attention machine;
FIG. 4 is a graph of photovoltaic power data from example 2;
FIG. 5 is a schematic diagram of the radiation amount in example 2;
FIG. 6 is a schematic diagram of clear sky radiation amount in example 2;
FIG. 7 is a schematic diagram of clear sky factor in example 2;
FIG. 8 is a schematic diagram of the temperature in example 2;
FIG. 9 is a graph of the 15 minute photovoltaic power forecast results of example 2;
Detailed Description
The first embodiment is as follows: 1-3, the photovoltaic forecasting method based on the prior fusion of the depth attention network and the clear sky radiation according to the embodiment comprises the following steps:
s1, acquiring historical photovoltaic power data and future clear sky radiation intensity;
acquiring three years of historical photovoltaic power data and future clear air radiation intensities corresponding to the historical photovoltaic power data before a certain photovoltaic power moment, wherein the sampling interval time of two adjacent groups of historical photovoltaic power data is 15 minutes, dividing the historical photovoltaic power data and the clear air radiation intensities corresponding to the historical photovoltaic power data into a training set and a test set according to time sequence, taking the historical photovoltaic power data of the first year and the historical photovoltaic power data of the second year and the corresponding future clear air radiation intensities as the training set, and taking the historical photovoltaic power data of the third year and the corresponding future clear air radiation intensities as the test set. The training set is used for training the depth attention ConvLSTM model, and the test set is used for evaluating or checking the forecasting performance of the depth attention ConvLSTM model.
Currently, there are many models for estimating the radiation intensity of the clear sky in the future, the McClear model is one of the most popular models, and the McClear model can output the radiation intensity estimation result of the clear sky at any time in the year of a certain place only given the longitude and latitude of the place. The invention adopts the McClear model to predict the future clear sky radiation intensity. According to the invention, the physical priori information of the clear sky is represented by the radiation intensity of the clear sky in the future, so that the forecast is more accurate. Because the photovoltaic power data has obvious daily cycle characteristics, the invention predicts the photovoltaic power of one hour in the future by using the historical photovoltaic power data of 96 sampling points before the photovoltaic power moment to be forecasted and the clear sky radiation intensity of one hour after the photovoltaic power moment to be forecasted as the input of the deep attention ConvLSTM model in the training and application.
S2, establishing a depth attention ConvLSTM model, wherein the depth attention ConvLSTM model sequentially comprises a ConvLSTM1 layer, an attention mechanism layer, a ConvLSTM2 layer, a flame layer and a full connection layer, inputting acquired historical photovoltaic power data and future clear sky radiation intensity into the depth attention ConvLSTM model for training, outputting a photovoltaic power forecast value, and obtaining a photovoltaic power forecast value sequence until loss converges, so as to obtain a trained depth attention ConvLSTM model, wherein the specific process is as follows:
S21, inputting the historical photovoltaic power data in the training set into a ConvLSTM1 layer of a depth attention ConvLSTM model, and outputting correlation information among a plurality of variables at the same moment and time correlation information of adjacent moments of the same variable in the historical photovoltaic power data.
The historical photovoltaic power data comprise a plurality of variables such as photovoltaic power, actual solar radiation intensity, clear sky factor, temperature and the like, and the photovoltaic power, the actual solar radiation intensity, the clear sky factor and the temperature at the moment are all data of historical moments. For the extraction of information among a plurality of variables in historical photovoltaic power data, the patent adopts a convolution long-term memory (ConvLSTM) network to extract characteristics. The ConvLSTM network can adaptively extract effective information from the multi-variable time sequence dynamic data, so that the effective information for forecasting future photovoltaic power generation power can be extracted from a plurality of variables such as photovoltaic power, temperature, actual solar radiation intensity, clear sky factor, clear sky radiation intensity and the like at the historical moment in the historical photovoltaic power data, and the correlation information among the plurality of variables at the same moment and the time correlation information of the adjacent moment of the same variable can be extracted.
S22, inputting correlation information among a plurality of variables at the same time and time correlation information at the same variable adjacent time in the historical photovoltaic power data output in S21 in the training set into an attention mechanism layer of a deep attention ConvLSTM model, and outputting information after combining the correlation information among the plurality of variables at the same time and the time correlation information at the same variable adjacent time in the future clear air radiation intensity and the historical photovoltaic power data, wherein the specific process is as follows:
firstly, obtaining correlation information among a plurality of variables at the same time in historical photovoltaic power data in a training set and time correlation information of adjacent time of the same variable and weight proportion between clear sky radiation intensity in the historical photovoltaic power data by using an attention mechanism at an attention mechanism layer, adaptively combining the correlation information among the plurality of variables at the same time in the historical photovoltaic power data and time correlation information of adjacent time of the same variable and information of future clear sky radiation intensity (namely clear sky physical priori information) in the training set according to the weight proportion, and obtaining combined information.
S23, inputting the combined information output in S22 into a ConvLSTM2 layer of a depth attention ConvLSTM model, outputting a feature matrix of the combined information, and further extracting features of the combined information.
S24, inputting the feature matrix of the combined information into a flat layer of a depth attention ConvLSTM model, and outputting a unidimensional feature matrix.
The function of the flat layer is to convert the multidimensional feature matrix into a unidimensional feature matrix, and the transition from the convolution layer to the full connection layer is realized.
S25, inputting the unidimensional feature matrix into a full-connection layer of the depth attention ConvLSTM model, and outputting a photovoltaic power forecast value to obtain a photovoltaic power forecast value sequence.
And training the depth attention ConvLSTM model by using a training set, and updating convolution kernels and deviation items in the depth attention ConvLSTM model by error back propagation in the training process until loss converges, outputting a photovoltaic power forecast value, and obtaining a photovoltaic power forecast value sequence to obtain the trained depth attention ConvLSTM model.
S3, performing post-processing on the photovoltaic power forecast value sequence obtained in the S2 by utilizing nuclear density estimation to obtain a forecast probability interval of a photovoltaic power true value, wherein the specific process is as follows:
s31, calculating and estimating a joint probability density function by utilizing two-dimensional kernel density estimation according to the photovoltaic power true value sequence and the photovoltaic power forecast value sequence:
wherein, the liquid crystal display device comprises a liquid crystal display device, 
Representing an estimated joint probability density function;
x 1 representing a sequence of photovoltaic power true values;
x 2 representing a sequence of photovoltaic power forecast values;
n represents the number of samples and, n=1, 2,. -%, N;
b 1 and b 2 Representing window width;
ker (. Cndot.) represents a Gaussian kernel function,
x λ representing a sequence of true or predicted photovoltaic power values, x when λ=1 λ Is x 1 At this time, a sequence of true photovoltaic power values is represented; when λ=2, x λ I.e. x 2 At this time, a photovoltaic power forecast value sequence is represented;
x λn represents the nth sample in the sequence of true or predicted photovoltaic power values, when λ=1, representing light Fu GongAn nth sample in the sequence of rate realism values; when λ=2, this represents the nth sample in the sequence of photovoltaic power forecast values;
b λ represents window width, b when λ=1 λ B is 1 When λ=2, b λ B is 2 ；
Since this step is in the training phase, after obtaining the photovoltaic power forecast value in the training set, the photovoltaic power real value sequence x is obtained according to the known 1 And the photovoltaic power forecast value sequence x obtained in S25 2 And calculating and estimating a joint probability density function by using the two-dimensional kernel density estimation. When N is large enough, estimating joint probability density function
Can be converged to a known true joint probability density function f (x 1 ,x 2 ) The accuracy of the two-dimensional kernel density estimation at this time is not dependent on the type of kernel function Ker (). In two-dimensional kernel density estimation, the optimal window width value is determined by the principle that the average integral square error between the true joint probability density function and the estimated joint probability density function is minimum.
S32, calculating a conditional probability density function of the photovoltaic power true value by using the estimated joint probability density function:
wherein, the liquid crystal display device comprises a liquid crystal display device,
a conditional probability density function representing a true value of photovoltaic power;
S33, calculating a conditional probability distribution function of the photovoltaic power true value by using the conditional probability density function of the photovoltaic power true value:
wherein, the liquid crystal display device comprises a liquid crystal display device,
a conditional probability distribution function representing the true value of the photovoltaic power.
S34, calculating a probability confidence interval of the photovoltaic power true value by using a conditional probability distribution function of the photovoltaic power true value to obtain a forecast probability interval of the photovoltaic power true value, wherein the concrete process is as follows:
wherein x is u Representing the upper probability interval limit at 90% confidence level;
x l representing the lower probability interval limit at 90% confidence level;
p { · } represents the size of the probability value;
X 1 representing a random variable corresponding to the true value of the photovoltaic power;
X 2 Representing a random variable corresponding to the photovoltaic power predicted value;
the estimated conditional probability distribution can be obtained through training, and when the method is applied subsequently, the prediction probability confidence interval of the photovoltaic power true value can be obtained only through the photovoltaic power predicted value by utilizing the formula (4) and the formula (5).
S35, inputting the test set into the trained depth attention ConvLSTM model in S2, outputting a photovoltaic power forecast value, judging whether the photovoltaic power forecast value is accurate or not by using a standardized root mean square error, if the photovoltaic power forecast value is smaller than 0.1, the photovoltaic power forecast value is accurate, otherwise, executing S2;
normalized root mean square error (normalized Root Mean Square Error, nRMSE):
wherein Cap represents the rated power of the photovoltaic panel, y i Representing the true value of the photovoltaic power,
represents the photovoltaic power forecast value, n represents the number of photovoltaic power samples, i=1, 2. The smaller nRMSE, the higher the forecast accuracy.
And when the photovoltaic power forecast value is accurate, executing S34 to obtain a forecast probability interval of the photovoltaic power true value, evaluating the forecast probability interval of the photovoltaic power true value by using the coverage probability of the forecast interval, if the coverage probability of the forecast interval is 85% -90%, the forecast probability interval of the photovoltaic power true value is accurate, otherwise, repeatedly executing S34. The closer the PICP value is to 90%, the better the probability interval forecast result. Prediction interval coverage probability (prediction interval coverage probability, PICP):
Wherein I is i An indicator variable representing the ith photovoltaic power sample,
the resulting conditional probability distribution function of the true value of the photovoltaic power is applied to the test set, and the process is implemented on-line. Firstly, inputting a test set into a trained deep attention ConvLSTM model in S2 to obtain a photovoltaic power forecast value of the test set, judging whether the photovoltaic power forecast value is accurate or not by using a standardized root mean square error, thereby realizing the forecast performance of the deep attention ConvLSTM model, and then executing S34 on the photovoltaic power forecast value to obtain a probability confidence interval of the photovoltaic power true value, and evaluating the forecast probability interval of the photovoltaic power true value by using the coverage probability of the forecast interval to accurately obtain the probability interval forecast of the photovoltaic power true value.
S4, acquiring historical photovoltaic power data before the photovoltaic power moment to be forecasted and clear sky radiation intensity after the photovoltaic power moment to be forecasted, inputting the acquired historical photovoltaic power data and the clear sky radiation intensity into a training depth attention ConvLSTM model in S2, outputting a photovoltaic power forecast value of the photovoltaic power moment to be forecasted, and executing S3 to obtain a forecast probability interval of a photovoltaic power true value of the photovoltaic power moment to be forecasted, wherein the specific process is as follows:
And acquiring historical photovoltaic power data of 96 sampling points before the photovoltaic power moment to be forecasted and clear sky radiation intensity of one hour after the photovoltaic power moment to be forecasted, inputting the acquired historical photovoltaic power data and the clear sky radiation intensity into a training depth attention ConvLSTM model in S2, outputting a photovoltaic power forecast value of the photovoltaic power moment to be forecasted, and executing S34 to obtain a forecast probability interval of a photovoltaic power true value of the photovoltaic power moment to be forecasted.
Example 1
The basic principle of photovoltaic power generation is to convert solar energy into electric energy by using a photovoltaic panel assembly. Actual output P of photovoltaic panel at time t t The method comprises the following steps:
wherein P is stc Representing standard conditions (I) stc ＝1000W/m 2 ,T stc Output of photovoltaic panel at =25℃), for a given model of photovoltaic panel P stc Typically a constant; i r,t Representing the actual solar radiation intensity at time t; i stc Representing the intensity of solar radiation under standard conditions; alpha T Representing the power of a photovoltaic panelA temperature coefficient; t (T) t The temperature of the photovoltaic panel at the time T is represented by T stc The temperature of the photovoltaic panel under standard conditions is indicated.
Therefore, the photovoltaic power generation power is closely related to the temperature and the radiation intensity. Under the influence of factors such as cloud shielding and the like, the actual solar radiation intensity I r Less than solar radiation intensity I without cloud shielding a The intensity of solar radiation without cloud shielding is generally referred to as clear sky radiation intensity. Definition of the actual solar radiation intensity I r With clear sky radiation intensity I a The ratio of (2) is a clear sky factor, the clear sky factor reflects the degree of cloud cover shielding, and is generally smaller than or equal to 1, and when the cloud cover shielding is completely absent, the clear sky factor is 1. Because the rotation and revolution of the earth are regular, the future clear sky radiation intensity has strong certainty, the future clear sky radiation intensity of a place with given longitude and latitude can be accurately estimated by using a model in the prior art, at present, a plurality of mature models for estimating the clear sky radiation intensity exist, the McClear model is one of the most popular models, the longitude and latitude of a place with given McClear model can output the clear sky radiation intensity estimation result of the place at any moment in the future year, and the invention adopts the McClear model to predict the future clear sky radiation intensity.
The photovoltaic power generation power is closely related to the temperature, the actual solar radiation intensity, the clear sky factor and the clear sky radiation intensity. Let X ghi (t),X cghi (T), k (T), T (T) and p (T) respectively represent solar radiation intensity, clear sky factor, temperature and photovoltaic power generation power at the time of history T, and historical photovoltaic power data Z (T) = [ X ] is defined ghi (t),X cghi (t),k(t),T(t),p(t)]And forecasting future photovoltaic power generation power by using i photovoltaic power data samples between the current moment and a moment before the current moment. Due to the clear sky radiation intensity X at a future time cghi ′(t+1),X cghi ' t+2.) the estimation can be performed by the McClear model. The problem of photovoltaic power forecasting of the present invention is therefore defined as follows, taking into account the physical priors of clear sky radiation:given historical photovoltaic power data Z (t), Z (t-1), Z (t-i+1) and clear sky radiation intensity X at a future time instant cghi ′(t+1),X cghi ' t+2, & gt, forecasting future photovoltaic power generation p (t+1), p (t+2) & gt. The photovoltaic power forecast needs to fuse a plurality of variable information in historical photovoltaic power data and future clear sky radiation physical priori information. The invention uses the clear sky radiation intensity to represent the clear sky radiation physical priori information.
For extracting a plurality of variable information in historical photovoltaic power data, the characteristic extraction is carried out by adopting a convolution long-term memory (ConvLSTM) network, and the ConvLSTM network can adaptively extract effective information from multi-variable time sequence dynamic data, so that the method is suitable for extracting the effective information for forecasting future photovoltaic power generation power from a plurality of variables of the historical photovoltaic power data such as photovoltaic power, temperature, actual solar radiation intensity, clear sky factor, clear sky radiation intensity and the like at historical moments. The structure of ConvLSTM network is shown in figure 1, and its mathematical principle is shown in formula (2).
Wherein i is t Is an input door; w (W) xi ,W hi ,W ci ，W xf ,W hf ,W cf ，W xc ,W hc ,W xo ，W co And W is ho Are weight items; f (f) t Is a forgetful door; c (C) t Is in an intermediate state; o (o) t Is an output door; h is a t The final output of the ConvLSTM network; b f ,b i ,b c And b o Is a bias term; sign symbol
Representing an element-by-element product; symbol represents a convolution operation; x is x t An input representing a ConvLSTM network at time t; h is a t-1 The output of the ConvLSTM network at the time t-1 is represented; sigma () is the activation function shown in equation (3); tanh () is the activation function shown in equation (4);
in order to introduce clear sky radiation physical prior information in the forecasting process, the invention introduces a attention mechanism in a ConvLSTM network. The traditional ConvLSTM network distributes the same weight to all feature variables, and the attention mechanism can adaptively distribute weights to different feature variables to characterize the influence of different variables on the forecasting task, so that the attention weight can be adaptively adjusted through the back propagation in the training process of the deep attention ConvLSTM network by introducing the attention mechanism into the ConvLSTM network to adaptively adjust the weight proportion of information extracted from historical photovoltaic data and physical prior information of future clear sky solar radiation in future photovoltaic power forecasting.
The principle of the attention mechanism is shown in fig. 2. When applying the attention mechanism, for each forecasting step, the context vector c i And attention weight a ij The calculation formula of (2) is as follows:
wherein k represents the input sequence length; h is a i Representing the hidden state of the decoder at time i; s is(s) j Representing the hidden state of the encoder at time j; f (·) represents a predefined similarity function; a, a ij And the attention weight for adjusting the proportion of the historical data information and the future clear sky radiation prior information is represented.
The invention combines ConvLSTM network and attention mechanism as above, and proposes a deep attention ConvLSTM network, which comprises ConvLSTM1 layer, attention mechanism layer, convLSTM2 layer, flatten layer and full connection layer in sequence as shown in figure 3.
Given three years of historical photovoltaic power data and corresponding clear sky radiation intensities, the historical photovoltaic power data and the corresponding clear sky radiation intensities are divided into a training set and a test set according to time sequence, the first year of historical photovoltaic power data and the second year of historical photovoltaic power data and the corresponding clear sky radiation intensities are used as the training set, and the third year of historical photovoltaic power data and the corresponding clear sky radiation intensities are used as the test set. The training set is used for training the deep attention ConvLSTM network, and the test set is used for evaluating or checking the forecasting performance of the deep attention ConvLSTM network.
Training the deep attention ConvLSTM network by using a training set, wherein the single training process specifically comprises the following steps: firstly, inputting historical photovoltaic power data into a first layer ConvLSTM1 in a deep attention ConvLSTM network, and extracting correlation information and time correlation information among a plurality of variables in the historical photovoltaic power data at the same time in the layer; and then, the attention mechanism is utilized to self-adaptively acquire the weight proportion between the information extracted from the historical photovoltaic power data and the clear sky physical prior information from the historical photovoltaic power data, the information extracted from ConvLSTM1 is multiplied by the attention weight, and then the information extracted from the historical photovoltaic power data and the information of the clear sky radiation physical prior information are self-adaptively combined. And transmitting the combined information to a second ConvLSTM2 layer for further feature extraction, inputting the extracted features into a flat layer, outputting a unidimensional feature matrix, inputting the unidimensional feature matrix into a full-connection layer, outputting a photovoltaic power forecast value, and obtaining a photovoltaic power forecast value sequence. By using the depth attention ConvLSTM network, the correlation information, the time correlation information and the clear sky radiation physical priori information among a plurality of variables in the historical photovoltaic data can be optimally combined to obtain the predicted photovoltaic power and a predicted value sequence, so that more accurate photovoltaic power generation power prediction is realized.
Finally, carrying out post-processing on the forecast photovoltaic power output by the deep attention ConvLSTM network by adopting two-dimensional kernel density estimation to obtain a corresponding probability confidence interval, thereby realizing the forecast of the probability interval, and specifically comprising the following steps:
according to a known sequence of photovoltaic power true values x 1 ＝{x 11 ,x 12 ,...，x 1N Sum of calculated photovoltaic power forecast value sequence x 2 ＝{x 21 ,x 22 ,...，x 2N Using two-dimensional kernel density estimation to obtain estimated joint probability density function by equation (7)
Wherein Ker (·) represents the kernel function; n represents the number of samples and, n=1, 2,. -%, N; b 1 And b 2 The window width is the same; x is x λ Representing a sequence of true or predicted photovoltaic power values, x when λ=1 λ Is x 1 At this time, a sequence of true photovoltaic power values is represented; when λ=2, x λ I.e. x 2 At this time, a photovoltaic power forecast value sequence is represented; x is x λn Representing an nth sample in the sequence of photovoltaic power real values or the sequence of photovoltaic power forecast values, when λ=1, representing the nth sample in the sequence of photovoltaic power real values at this time; when λ=2, this represents the nth sample in the sequence of photovoltaic power forecast values; b λ Represents window width, b when λ=1 λ Namely b 1 When λ=2, b λ Namely b 2 ；
If N is sufficiently large, a joint probability density function is estimated 
Can be converged to the true joint probability density function f (x 1 ,x 2 ) The accuracy of the two-dimensional kernel density estimation at this time is not dependent on the type of kernel function Ker (). The invention adopts the method as shown in the prior artA gaussian kernel function represented by formula (8):
in equation (7), the optimal window width value is determined by the principle that the average integral square error between the true joint probability density function and the estimated joint probability density function is minimal.
After obtaining the estimated joint probability density function by using the formula (9) and the formula (10), the conditional probability density function of the actual photovoltaic power can be calculated
Then calculating the conditional probability distribution function of the actual photovoltaic power through the formula (11)>
Conditional probability distribution function of actual photovoltaic power to be obtained
The method is applied to a test set, a photovoltaic power forecast value of the test set is obtained through calculation, whether the photovoltaic power forecast value is accurate or not is judged by using a standardized root mean square error, if the photovoltaic power forecast value is smaller than 0.1, the photovoltaic power forecast value is accurate, otherwise, the photovoltaic power forecast value is calculated repeatedly; the root mean square error is normalized (normalized Root Mean Square Error,nRMSE)：
wherein Cap represents the rated power of the photovoltaic panel, y i Representing the true value of the photovoltaic power,
represents the photovoltaic power forecast value, n represents the number of photovoltaic power samples, i=1, 2. The smaller nRMSE, the higher the forecast accuracy.
When the photovoltaic power forecast value is accurate, obtaining the probability interval upper limit x of the photovoltaic power true value under the given 90% confidence level through the formula (13) and the formula (14) u And a probability interval lower limit x l And obtaining a probability confidence interval of the photovoltaic power true value, thereby realizing probability interval forecast of the photovoltaic power true value.
And (3) obtaining a prediction probability interval, evaluating the prediction probability interval of the photovoltaic power true value by using the prediction interval coverage probability, if the prediction interval coverage probability is 85% -90%, the prediction probability interval of the photovoltaic power true value is accurate, otherwise, repeating the step (S34). The closer the PICP value is to 90%, the better the probability interval forecast result. Prediction interval coverage probability (prediction interval coverage probability, PICP):
wherein I is i An indicator variable representing the ith photovoltaic power sample,
example 2
In order to detect the effectiveness of the method in photovoltaic power forecast, experimental verification is carried out on 3-year operation history data of a certain photovoltaic power station. The obtained photovoltaic power data are shown in fig. 4-8, the accuracy of the photovoltaic power forecast value is shown in table 1, and the forecast interval probability of the photovoltaic power true value is shown in table 2.
The experimental results of the present invention are shown in table 1, table 2 and fig. 9:
TABLE 1 accuracy of forecast values of the invention
TABLE 2 prediction interval probability of the invention
As can be seen from table 1, table 2 and fig. 9, the method provided by the invention can obtain better prediction accuracy than the traditional method, the prediction value and the true value are very close, the true value is almost in the given probability confidence interval, and a good prediction effect is obtained.
Claims (7)
1. The photovoltaic forecasting method based on the depth attention network and clear sky radiation prior fusion is characterized by comprising the following steps of: it comprises the following steps:
s1, acquiring historical photovoltaic power data and future clear sky radiation intensity, wherein the specific process is as follows:
acquiring three years of historical photovoltaic power data and corresponding future clear sky radiation intensity before a certain photovoltaic power moment, wherein the sampling interval time of two adjacent groups of historical photovoltaic power data is 15 minutes, the historical photovoltaic power data of the first year and the historical photovoltaic power data of the second year and the corresponding future clear sky radiation intensity are used as training sets, and the historical photovoltaic power data of the third year and the corresponding future clear sky radiation intensity are used as test sets;
s2, establishing a depth attention ConvLSTM model, wherein the depth attention ConvLSTM model sequentially comprises a ConvLSTM1 layer, an attention mechanism layer, a ConvLSTM2 layer, a flame layer and a full connection layer, the historical photovoltaic power data acquired in a training set and the future clear sky radiation intensity are input into the depth attention ConvLSTM model for training, a photovoltaic power forecast value is output, a photovoltaic power forecast value sequence is obtained until loss converges, and a trained depth attention ConvLSTM model is obtained, and the specific process is as follows:
S21, inputting the historical photovoltaic power data in the training set into a ConvLSTM1 layer of a depth attention ConvLSTM model, and outputting correlation information among a plurality of variables at the same moment and time correlation information of adjacent moments of the same variable in the historical photovoltaic power data; the plurality of variables in the historical photovoltaic power data comprise photovoltaic power, actual solar radiation intensity, clear sky factor and temperature;
s22, inputting correlation information among a plurality of variables at the same time and time correlation information at adjacent times of the same variable in the future clear sky radiation intensity and the historical photovoltaic power data output in S21 in the training set into an attention mechanism layer of a deep attention ConvLSTM model, and outputting information obtained by combining the correlation information among the plurality of variables at the same time and the time correlation information at adjacent times of the same variable in the future clear sky radiation intensity and the historical photovoltaic power data output in S21;
s23, inputting the combined information output in the S22 into a ConvLSTM2 layer of a depth attention ConvLSTM model, and outputting a feature matrix of the combined information;
s24, inputting the feature matrix of the combined information into a flat layer of a depth attention ConvLSTM model, and outputting a unidimensional feature matrix;
S25, inputting the unidimensional feature matrix into a full-connection layer of a depth attention ConvLSTM model, and outputting a photovoltaic power forecast value to obtain a photovoltaic power forecast value sequence;
s3, performing post-processing on the photovoltaic power forecast value sequence obtained in the S2 by utilizing nuclear density estimation to obtain a forecast probability interval of a photovoltaic power true value, wherein the specific process is as follows:
s31, calculating and estimating a joint probability density function by utilizing two-dimensional kernel density estimation according to the photovoltaic power true value sequence and the photovoltaic power forecast value sequence;
s32, calculating a conditional probability density function of the photovoltaic power true value by using the estimated joint probability density function;
s33, calculating a conditional probability distribution function of the photovoltaic power true value by using the conditional probability density function of the photovoltaic power true value;
s34, calculating a probability confidence interval of the photovoltaic power true value by using a conditional probability distribution function of the photovoltaic power true value to obtain a forecast probability interval of the photovoltaic power true value;
the estimated conditional probability distribution can be obtained through training, and when the method is applied subsequently, the prediction probability confidence interval of the true photovoltaic power value can be obtained only through the photovoltaic power predicted value;
s35, inputting the test set into the trained depth attention ConvLSTM model in S2, outputting a photovoltaic power forecast value, judging whether the photovoltaic power forecast value is accurate or not by using a standardized root mean square error, if the photovoltaic power forecast value is smaller than 0.1, the photovoltaic power forecast value is accurate, otherwise, executing S2;
Normalized root mean square error:
wherein Cap represents the rated power of the photovoltaic panel, y i Representing the true value of the photovoltaic power,
representing a photovoltaic power forecast value, n representing the number of photovoltaic power samples, i=1, 2, …, n;
when the photovoltaic power forecast value is accurate, executing S34 to obtain a forecast probability interval of the photovoltaic power true value, evaluating the forecast probability interval of the photovoltaic power true value by using the coverage probability of the forecast interval, if the coverage probability of the forecast interval is 85% -90%, the forecast probability interval of the photovoltaic power true value is accurate, otherwise, repeatedly executing S34;
prediction interval coverage probability:
wherein I is i An indicator variable representing the ith photovoltaic power sample,
x u representing the upper probability interval limit at 90% confidence level; x is x l Representing the lower probability interval limit at 90% confidence level;
s4, acquiring historical photovoltaic power data before the photovoltaic power moment to be forecasted and corresponding future clear sky radiation intensity, inputting the acquired historical photovoltaic power data before the photovoltaic power moment to be forecasted and the corresponding future clear sky radiation intensity into the depth attention ConvLSTM model trained in S2, outputting a photovoltaic power forecast value of the photovoltaic power moment to be forecasted, and obtaining a forecast probability interval of a photovoltaic power true value of the photovoltaic power moment to be forecasted.
2. The photovoltaic forecasting method based on depth attention network and clear sky radiation prior fusion as claimed in claim 1, wherein: in the step S22, the correlation information between the future clear sky radiation intensity in the training set and the multiple variables at the same time in the historical photovoltaic power data output in the step S21 and the time correlation information between adjacent times of the same variable are input into the attention mechanism layer of the deep attention ConvLSTM model, and the correlation information between the multiple variables at the same time in the future clear sky radiation intensity and the historical photovoltaic power data output in the step S21 and the time correlation information between adjacent times of the same variable are output, which comprises the following specific steps:
and acquiring correlation information among a plurality of variables at the same time in the historical photovoltaic power data in the training set and time correlation information among adjacent time of the same variable and weight proportion among the historical photovoltaic power data in the training set and corresponding future clear sky radiation intensity by using an attention mechanism, combining the correlation information among the plurality of variables at the same time in the historical photovoltaic power data in the training set and the time correlation information among the adjacent time of the same variable and the information of the future clear sky radiation intensity in the training set according to the weight proportion, and obtaining combined information.
3. The photovoltaic forecasting method based on depth attention network and clear sky radiation prior fusion as claimed in claim 2, wherein: in the step S31, a two-dimensional kernel density estimation calculation estimation joint probability density function is utilized according to a photovoltaic power true value sequence and a photovoltaic power forecast value sequence, and the specific process is as follows:
wherein, the liquid crystal display device comprises a liquid crystal display device,
representing an estimated joint probability density function;
x 1 representing a sequence of photovoltaic power true values;
x 2 representing a sequence of photovoltaic power forecast values;
n represents the number of samples, N' =1, 2, …, N;
b 1 and b 2 Representing window width;
ker (. Cndot.) represents a Gaussian kernel function;
x λ representing a sequence of true or predicted photovoltaic power values, x when λ=1 λ Is x 1 At this time, a sequence of true photovoltaic power values is represented; when λ=2, x λ I.e. x 2 At this time, a photovoltaic power forecast value sequence is represented;
x λn′ representing a sequence of true values of photovoltaic power or photovoltaic powerAn nth 'sample in the sequence of predicted values, when λ=1, representing an nth' sample in the sequence of real photovoltaic power values; when λ=2, this represents the nth sample in the sequence of photovoltaic power forecast values;
b λ represents window width, b when λ=1 λ B is 1 When λ=2, b λ B is 2 ；
The optimal window width value is determined by the principle that the average integral square error between the true joint probability density function and the estimated joint probability density function is minimal.
4. A photovoltaic forecasting method based on depth attention network and clear sky radiation prior fusion as claimed in claim 3, characterized in that: in the step S32, a conditional probability density function of the photovoltaic power true value is calculated by using the estimated joint probability density function, and the specific process is as follows:
wherein, the liquid crystal display device comprises a liquid crystal display device,
a conditional probability density function representing a true value of photovoltaic power;
5. The photovoltaic forecasting method based on depth attention network and clear sky radiation prior fusion as claimed in claim 4, wherein: in the step S33, a conditional probability distribution function of the photovoltaic power real value is calculated by using a conditional probability density function of the photovoltaic power real value, and the specific process is as follows:
wherein, the liquid crystal display device comprises a liquid crystal display device,
a conditional probability distribution function representing the true value of the photovoltaic power.
6. The photovoltaic forecasting method based on depth attention network and clear sky radiation prior fusion as claimed in claim 5, wherein: in the step S34, a probability confidence interval of the photovoltaic power true value is calculated by using a conditional probability distribution function of the photovoltaic power true value, and a prediction probability interval of the photovoltaic power true value is obtained, which comprises the following specific steps:
Wherein P { · } represents the size of the probability value;
X 1 representing a random variable corresponding to the true value of the photovoltaic power;
X 2 and representing the random variable corresponding to the photovoltaic power forecast value.
7. The photovoltaic forecasting method based on depth attention network and clear sky radiation prior fusion as claimed in claim 6, wherein: the step S4 is to acquire historical photovoltaic power data before the photovoltaic power moment to be forecasted and corresponding future clear sky radiation intensity, input the acquired historical photovoltaic power data before the photovoltaic power moment to be forecasted and the corresponding future clear sky radiation intensity into a training depth attention ConvLSTM model in the step S2, output a photovoltaic power forecast value of the photovoltaic power moment to be forecasted, and obtain a forecast probability interval of a photovoltaic power true value of the photovoltaic power moment to be forecasted, wherein the specific process is as follows:
the method comprises the steps of obtaining historical photovoltaic power data of 96 sampling points before a photovoltaic power moment to be forecasted and future clear sky radiation intensity of one hour after the photovoltaic power moment to be forecasted, inputting the obtained historical photovoltaic power data of 96 sampling points before the photovoltaic power moment to be forecasted and the obtained future clear sky radiation intensity of one hour after the photovoltaic power moment to be forecasted into a trained deep attention ConvLSTM model in S2, outputting a photovoltaic power forecast value of the photovoltaic power moment to be forecasted, and obtaining a forecast probability interval of a photovoltaic power true value of the photovoltaic power moment to be forecasted.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210651311.9A CN114971058B (en) | 2022-06-09 | 2022-06-09 | Photovoltaic forecasting method based on depth attention network and clear sky radiation priori fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210651311.9A CN114971058B (en) | 2022-06-09 | 2022-06-09 | Photovoltaic forecasting method based on depth attention network and clear sky radiation priori fusion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114971058A CN114971058A (en) | 2022-08-30 |
| CN114971058B true CN114971058B (en) | 2023-07-14 |
Family
ID=82960704
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210651311.9A Active CN114971058B (en) | 2022-06-09 | 2022-06-09 | Photovoltaic forecasting method based on depth attention network and clear sky radiation priori fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114971058B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112149879A (en) * | 2020-09-02 | 2020-12-29 | 上海电力大学 | New energy medium-and-long-term electric quantity prediction method considering macroscopic volatility classification |
| CN112507793A (en) * | 2020-11-05 | 2021-03-16 | 上海电力大学 | Ultra-short-term photovoltaic power prediction method |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018065045A1 (en) * | 2016-10-05 | 2018-04-12 | Telecom Italia S.P.A. | Method and system for estimating energy generation based on solar irradiance forecasting |
| CN109886492A (en) * | 2019-02-26 | 2019-06-14 | 浙江鑫升新能源科技有限公司 | Photovoltaic power generation power prediction model and its construction method based on Attention LSTM |
| CN113469427A (en) * | 2021-06-24 | 2021-10-01 | 国网山东省电力公司东营供电公司 | Convolution LSTM-based photovoltaic power station day-ahead power generation amount prediction method and system |
| CN114006369B (en) * | 2021-10-28 | 2022-08-30 | 华北电力大学 | Regional wind and light station power joint prediction method and device, electronic equipment and storage medium |
| CN114492945A (en) * | 2021-12-31 | 2022-05-13 | 安徽电力交易中心有限公司 | Short-term photovoltaic power prediction method, medium and equipment in electric power market background |
-
2022
- 2022-06-09 CN CN202210651311.9A patent/CN114971058B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112149879A (en) * | 2020-09-02 | 2020-12-29 | 上海电力大学 | New energy medium-and-long-term electric quantity prediction method considering macroscopic volatility classification |
| CN112507793A (en) * | 2020-11-05 | 2021-03-16 | 上海电力大学 | Ultra-short-term photovoltaic power prediction method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114971058A (en) | 2022-08-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106529814B (en) | Distributed photovoltaic ultra-short term prediction method based on Adaboost clustering and Markov chain | |
| CN103117546B (en) | A kind of Ultrashort-term slide prediction method for wind power | |
| CN110070226A (en) | Photovoltaic power prediction technique and system based on convolutional neural networks and meta learning | |
| CN104978613A (en) | Short-period forecasting method for photovoltaic output in consideration of assembly temperature | |
| CN112381282B (en) | Photovoltaic power generation power prediction method based on width learning system | |
| CN111242355A (en) | Photovoltaic probability prediction method and system based on Bayesian neural network | |
| CN115347571B (en) | Photovoltaic power generation short-term prediction method and device based on transfer learning | |
| CN115660887A (en) | Photovoltaic output prediction method and system based on limited weather forecast information | |
| CN113642767A (en) | Multi-dimensional feature combination prediction method based on MI-VMD-DA-EDLSTM-VEC | |
| Jobayer et al. | A systematic review on predicting PV system parameters using machine learning | |
| CN116739172B (en) | Method and device for ultra-short-term prediction of offshore wind power based on climbing identification | |
| CN114971058B (en) | Photovoltaic forecasting method based on depth attention network and clear sky radiation priori fusion | |
| Nijhuis et al. | Prediction of power fluctuation classes for photovoltaic installations and potential benefits of dynamic reserve allocation | |
| CN116167465A (en) | Solar irradiance prediction method based on multivariate time series ensemble learning | |
| Fen et al. | Short‐term photovoltaic power probability forecasting based on OLPP‐GPR and modified clearness index | |
| CN111612289B (en) | New energy multi-scene risk feature oriented power system risk assessment method | |
| Park et al. | Zero-Shot Photovoltaic Power Forecasting Scheme Based on a Deep Learning Model and Correlation Coefficient | |
| CN115545256A (en) | Small-scale photovoltaic power prediction method based on multi-dimensional data feature learning | |
| CN113837485A (en) | Light storage micro-grid system source-load characteristic prediction method, system, device and storage medium | |
| CN113537573A (en) | Wind power operation trend prediction method based on dual space-time feature extraction | |
| Xia et al. | Research on Solar Radiation Estimation based on Singular Spectrum Analysis-Deep Belief Network | |
| CN112734073A (en) | Photovoltaic power generation short-term prediction method based on long and short-term memory network | |
| Coroama et al. | A study on wind energy generation forecasting using connectionist models | |
| Ali et al. | Machine learning based solar power forecasting techniques: Analysis and comparison | |
| CN116776921B (en) | Solar radiation prediction method and device based on improved patch-informer |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |