CN111626468B - Photovoltaic interval prediction method based on biased convex loss function - Google Patents

Photovoltaic interval prediction method based on biased convex loss function Download PDF

Info

Publication number
CN111626468B
CN111626468B CN202010274717.0A CN202010274717A CN111626468B CN 111626468 B CN111626468 B CN 111626468B CN 202010274717 A CN202010274717 A CN 202010274717A CN 111626468 B CN111626468 B CN 111626468B
Authority
CN
China
Prior art keywords
interval
prediction
prediction model
loss function
photovoltaic
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
Application number
CN202010274717.0A
Other languages
Chinese (zh)
Other versions
CN111626468A (en
Inventor
龙寰
张琛
吴在军
胡伟
荆江平
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southeast University
State Grid Jiangsu Electric Power Co Ltd
Original Assignee
Southeast University
State Grid Jiangsu Electric Power Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Southeast University, State Grid Jiangsu Electric Power Co Ltd filed Critical Southeast University
Priority to CN202010274717.0A priority Critical patent/CN111626468B/en
Publication of CN111626468A publication Critical patent/CN111626468A/en
Application granted granted Critical
Publication of CN111626468B publication Critical patent/CN111626468B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • 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)
  • Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Economics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • General Physics & Mathematics (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Game Theory and Decision Science (AREA)
  • Development Economics (AREA)
  • Public Health (AREA)
  • Water Supply & Treatment (AREA)
  • Operations Research (AREA)
  • Primary Health Care (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Quality & Reliability (AREA)
  • Biomedical Technology (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a photovoltaic interval prediction method based on a biased convex loss function, and belongs to the technical field of calculation, calculation or counting. An extreme learning machine is used as a basic prediction model for predicting the upper and lower limits of the interval. The training optimization target of the interval prediction model is constructed based on the proposed biased convex loss function, and the training of the proposed interval prediction model can be regarded as a double-layer optimization problem. In the upper layer problem, determining the optimal combination of upper and lower limit prediction models by evaluating the interval prediction performance constructed under different candidate super parameters; in the lower layer optimization problem, super parameters of a biased loss function are given, and an upper limit prediction model and a lower limit prediction model of a prediction interval are trained through a convex optimization method. The method and the device realize quick prediction of the photovoltaic output interval, and overcome the defect that heuristic algorithm is directly optimized and easily falls into local optimum.

Description

Photovoltaic interval prediction method based on biased convex loss function
Technical Field
The invention discloses a photovoltaic interval prediction method based on a biased convex loss function, relates to a renewable energy output prediction technology, and belongs to the technical field of calculation, calculation or counting.
Background
In order to realize sustainable development of energy, adjust energy structure and protect ecological environment, the active development of renewable energy power generation technology has become the main melody of energy development in new period around the world. Photovoltaic power generation is an important expression form of solar power generation, is inexhaustible, is not limited by regions, and has wide application prospect. However, the photovoltaic power generation capacity is closely related to factors such as weather, seasons, geographical positions and the like, and the photovoltaic power generation has the output characteristics of randomness and intermittence. As the photovoltaic permeability of the power grid is continuously increased, the randomness and uncertainty of the photovoltaic output can have an increasingly serious negative effect on the safe, stable and economic operation of the power grid.
In the traditional renewable energy grid-connected research, the power generation prediction mode is mainly point prediction, but the uncertainty of the output of the point prediction is not consistent with the essential characteristics of the renewable energy because the amount of information contained in the point prediction is too low and the output is difficult to quantify. The predicted interval can describe the fluctuation range and uncertainty of the output of the renewable energy source, and the utilization rate of the predicted information is improved. Therefore, the prediction interval and the error based on a certain confidence level can better serve the economic operation and the safe and stable control of the power system.
Because the confidence interval of the prediction target is unknown, the traditional point prediction optimization method cannot be directly used for model optimization of interval prediction. The method aims at providing an objective function which can adapt to the learning of the section prediction model for the loss function of quantile regression, fully playing the learning performance of the prediction model and realizing the section prediction with higher precision.
Disclosure of Invention
The invention aims to overcome the defects of the background art, and provides a photovoltaic interval prediction method based on a biased convex loss function, which realizes the optimization and adjustment of model parameters on the basis of a certain confidence level, reduces the width of a prediction interval on the premise of reaching a desired interval confidence level, and solves the technical problem that the traditional point prediction optimization method cannot be directly used for model optimization of interval prediction.
The invention adopts the following technical scheme for realizing the purposes of the invention: the invention provides a photovoltaic combined interval prediction method based on a bias convex loss function, which is a double-layer optimization framework, wherein an interval prediction problem is decomposed into an upper limit prediction sub-problem and a lower limit prediction sub-problem, an upper limit prediction model and a lower limit prediction model of a photovoltaic interval are established, bias convex loss functions used for determining output weights under super parameter constraint are respectively established for the upper limit prediction model and the lower limit prediction model so as to achieve the expected coverage rate of a predicted interval and the minimum average prediction interval width is a target search optimal super parameter combination, the output weights when the bias convex loss function value of the upper limit prediction model is minimum or the bias convex loss function value of the lower limit prediction model is minimum are determined under the constraint of the optimal super parameter combination, the upper limit prediction model and the lower limit prediction model of the output weights are combined to obtain an interval prediction model, and input data are processed by the interval prediction model to obtain a photovoltaic output interval prediction result.
Further, the upper limit prediction model is as follows
Figure BDA0002444359500000021
The lower limit prediction model is +.>
Figure BDA0002444359500000022
Figure BDA0002444359500000023
βOutput weights of upper limit prediction model and lower limit prediction model respectively, < >>
Figure BDA0002444359500000024
β＝[β 1 ，β 2 ，…，β l ] T L is the total number of neurons of the hidden layer of the prediction model,
Figure BDA0002444359500000025
Δy i ＝L i -y i ，/>
Figure BDA0002444359500000026
as a predicted deviation of the upper limit of the interval, deltay i As a predicted deviation from the lower limit of the interval,
Figure BDA0002444359500000027
y i for the actual value of the ith predicted point, U i 、L i C is the upper limit and the lower limit of the actual value of the ith predicted point u And C l Regularization coefficients of the upper limit prediction model and the lower limit prediction model, respectively. V (V) u And V l Super-parameters of the upper limit prediction model and the lower limit prediction model respectively, V u ＝{W u ,r u ,c u ,C u Sum V l ＝{W l ,r l ,c l ,C l }，W u 、r u 、c u Weights, scaling coefficients, translation coefficients, W, respectively, of the partial convex loss function term of the upper limit prediction model l 、r l 、c l The lower limit prediction model is provided with a weight, a scaling coefficient and a translation coefficient of a bias convex loss function item respectively.
Still further, the prediction interval coverage PICR is
Figure BDA0002444359500000028
Average prediction interval width PINAW is +.>
Figure BDA0002444359500000029
N is the number of statistical data samples in the training data set, y 1 Is the 1 st predicted point x 1 Actual value of y N For the Nth predicted point x N Is an exponential function->
Figure BDA00024443595000000210
Measuring the actual value y of the ith predicted point i Whether or not in the prediction interval [ L ] i ,U i ]Inside (I)>
Figure BDA00024443595000000211
Figure BDA00024443595000000212
Representing taking the largest actual value in the training dataset,/->
Figure BDA00024443595000000213
Representing taking the smallest actual value in the training dataset.
Further, in the process of determining the output weight when the upper limit prediction model has the minimum partial convex loss function value or the lower limit prediction model has the minimum partial convex loss function value under the constraint of the optimal super-parameter combination, the method for minimizing the partial convex loss function is to solve the root of the gradient of the partial convex loss function.
Further, the method for searching the optimal super-parameter combination comprises the following steps: and respectively carrying out sensitivity analysis on the super parameters of the upper limit prediction model and the lower limit prediction model to obtain a sensitivity interval, generating candidate super parameter search networks of the upper limit prediction model and the lower limit prediction model in the sensitivity interval, calculating the coverage rate and the width of the prediction interval under different super parameter combinations in parallel, and selecting the optimal super parameter combination which reaches the expected coverage rate of the prediction interval and has the narrowest average prediction interval width.
Still further, the optimization problem of searching for the optimal super-parameter combination with the goal of achieving the expected coverage rate of the prediction interval and the narrowest average prediction interval width is:
Figure BDA0002444359500000031
PICR * and (5) the coverage rate is expected for a preset prediction interval.
Further, the input data comprising the photovoltaic conversion rate, the photovoltaic output from the previous moment to the photovoltaic output from the previous moment are sequentially preprocessed, feature extracted and normalized, and then are input into the section prediction model.
Further, the upper limit prediction model and the lower limit prediction model are realized through a neural network, the input weight of the neural network is randomly generated, and the total number of neurons of the hidden layer is determined through cross verification of the point prediction result. The invention adopts the technical scheme and has the following beneficial effects:
(1) According to the method, the photovoltaic interval prediction problem is converted into the combination of the upper limit prediction sub problem and the lower limit prediction sub problem, the supervised learning prediction model with the biased convex loss function as a target is used as the upper limit prediction model and the lower limit prediction model, the super parameter set meeting the precision requirement is firstly determined through a double-layer optimization framework, then the output weight of the upper limit prediction model and the output weight of the lower limit prediction model are determined through a convex optimization technology, the upper limit prediction model and the lower limit prediction model after the output weight is combined to obtain the interval prediction result, and the problems that the traditional heuristic algorithm directly optimizes the prediction model possibly causes the falling into local optimum, the optimization result is unstable and the like are solved through the optimization structure of the super parameter of the limited objective function.
(2) In the double-layer optimization process, the optimal super parameter set meeting the precision requirement is determined by searching the candidate super parameter grids through parallel calculation, the convex optimization is realized by independently solving the output weights of the upper and lower limit prediction models through parallel calculation, the calculation efficiency is greatly improved, and the width of the prediction interval can be obviously reduced on the premise of guaranteeing the reliability of coverage rate of the obtained interval.
Drawings
Fig. 1 is a flow chart of the present application predicting photovoltaic intervals.
Fig. 2 is a graph comparing the predicted interval of the present application with the actual value at 95% interval prediction coverage in summer.
Fig. 3 is a partial enlarged view of fig. 2.
Fig. 4 is a predicted interval of the method of the present invention at a winter 85% interval prediction coverage.
Fig. 5 is a partial enlarged view of fig. 4.
Detailed Description
The technical scheme of the invention is described in detail below with reference to the accompanying drawings.
The photovoltaic combined interval prediction method based on the biased convex loss function is a double-layer optimization framework, and as shown in fig. 1, the method comprises the following 5 steps:
(1) Data preprocessing and feature extraction
Integrating data comprising photovoltaic conversion rate, photovoltaic output from the previous moment to the previous three moments into a data sample form containing input and output vectors, preprocessing the data and extracting features, and normalizing data variables to be within the range of [0,1 ].
(2) Providing accuracy measurement standard of photovoltaic prediction interval and setting expected coverage rate
The precision measurement standard of the photovoltaic prediction interval comprises an average prediction interval width PINAW and a prediction interval coverage rate PICR, and the expected coverage rate is recorded as PICR * Given a training dataset of { (x) 1 ,y 1 ),…,(x N ,y N ) The prediction interval precision measurement indexes are respectively shown in the formulas (1) to (3):
Figure BDA0002444359500000041
Figure BDA0002444359500000042
Figure BDA0002444359500000043
in the formulas (1) to (3), N is the number of statistical data samples in the training data set, y 1 Is the 1 st predicted point x 1 Actual value of y N For the Nth predicted point x N An exponential function
Figure BDA0002444359500000044
Measuring the actual value y of the ith predicted point i Whether or not in the prediction interval [ L ] i ,U i ]Inside U i 、L i Upper and lower limit of the actual value of the ith predicted point,/->
Figure BDA0002444359500000045
Representing taking the largest actual value in the training dataset,/->
Figure BDA0002444359500000046
Representing taking the smallest actual value in the training dataset.
(3) Initializing interval upper and lower limit prediction model
The method comprises the steps of dividing interval prediction problematic into two sub-problems of interval upper limit prediction and interval lower limit prediction, respectively carrying out prediction combination on an upper limit prediction model and a lower limit prediction model by a neural network-based extreme learning machine to construct an interval prediction model, initializing connection weights from an input layer to a hidden layer by using a random generation mode aiming at the neural network-based extreme learning machine, namely inputting weight parameters, and carrying out cross-validation optimization on point prediction to determine the number L of neurons of the hidden layer of the neural network.
The specific method for randomly initializing the neural network input weight parameters comprises the following steps: given a training dataset of { (x) 1 ,y 1 ),…,(x N ,y N ) Generating an extreme learning machine prediction model with a hidden layer number L, wherein the input weight of the network is [ -1,1]Hidden layer bias at [0,1]Randomly generated within the range of (2). The d-dimensional input is mapped into the L-dimensional hidden layer mapping space through the input weight:
h＝g(ax+b) (4)，
in the formula (4), a= [ a ] 1 ,a 2 ,…,a L ]Input weight matrix in d×l dimension, b= [ b ] 1 ,…,b L ] T Is an L x 1-dimensional bias matrix.
(4) Constructing an interval prediction model with a biased convex loss function as an optimization target
The steps for constructing the biassed loss function are as follows:
(a) For training data set { (x) i ,y i ) Target y of i=1, … N } i The corresponding upper limit and lower limit of the prediction interval are respectively U i And L i And (3) representing. The prediction bias of the upper and lower limits of the interval is defined as formula (5):
Figure BDA0002444359500000051
the upper and lower limits of the acceptable interval deviation need to satisfy the formula (6):
Figure BDA0002444359500000052
(b) The objective function of the convex loss used by the upper limit and the lower limit extreme learning machine for predicting the neural network is defined as follows:
Figure BDA0002444359500000053
Figure BDA0002444359500000054
in the formulas (7) and (8),
Figure BDA0002444359500000055
β＝[β 1 ，β 2 ，…，β l ] T the output weights of the upper-limit prediction model and the lower-limit prediction model are respectively the decision variables of the objective function, C u And C l For regularization coefficient, V u ＝{W u ,r u ,c u ,C u Sum V l ＝{W l ,r l ,c l ,C l Super parameters of upper limit and lower limit prediction models respectively;
(c) The function E (θ|w, r, c) is defined as follows:
E(θ|W,r,c)＝θ 2 +W·S(θ|r,c) (9)，
Figure BDA0002444359500000061
in the formulas (9) and (10), θ is the predicted deviation, that is
Figure BDA0002444359500000062
Or deltay i W is the weight of a biased convex loss function term S (-) in a function E (theta|W, r, c), r and c are scaling factors and shifting factors of the function S (-), r and c are used for regulating the shape and trend of a curve, and W and c are positive values. It can be shown that S (·) is a continuous and second order derivative convex function.
(5) Simplifying the training of the interval prediction model into a double-layer optimization problem and then solving
The original training data is divided into training data and verification data to assist in model optimization training. In the upper layer problem, the optimal combination of the upper and lower limit prediction models is determined by evaluating the section prediction performance constructed under different candidate hyper-parameters. In the lower layer problem, super parameters of a biased loss function are given, and an interval upper and lower limit prediction model is trained through a convex optimization method. The specific steps for solving the double-layer optimization problem are as follows:
(a) Independent variables of the interval prediction model include upper and lower limit predictionsSuper parameter set Θ= { V of model u ,V l And corresponding output weight matrix
Figure BDA0002444359500000063
Wherein, the upper layer problem uses grid search and cross verification technology to obtain the optimal super parameter set theta, and the lower layer problem uses convex optimization technology to obtain +.>
Figure BDA0002444359500000064
Thus, to obtain the prediction interval at the desired coverage, the double-layer optimization problem is established as follows:
Figure BDA0002444359500000065
in formula (11), PICR * Is a preset desired coverage rate;
(b) For upper layer optimization, firstly, sensitivity analysis is used for searching a sensitive section of a super parameter set, namely, a section of which the performance of a prediction model can be sensitively changed by the change of super parameters, the super parameter setting is prevented from sinking into a saturated region, a candidate super parameter searching grid of an upper limit prediction model and a candidate super parameter of a lower limit prediction model is generated in the sensitive section, secondly, parallel computing technology is used for optimizing the upper limit prediction model and the lower limit prediction model under different super parameter setting, PINAW and PICR of the section prediction model combined by different upper limit prediction models and lower limit prediction models are evaluated based on different super parameter sets, and an optimal super parameter set Θ is selected according to an optimization target;
(c) In the lower-layer optimization problem, constructing a biased loss function F of an upper-lower-limit prediction model based on an upper-layer problem candidate super-parameter set u (. Cndot.) and F l (. Cndot.) minimizing function F using convex optimization method u (. Cndot.) and F l (. Cndot.) the problem is equivalent to solving the function F u (. Cndot.) and F l The root of the gradient of (-) is represented by formula (12) and formula (13):
Figure BDA0002444359500000071
Figure BDA0002444359500000072
in the lower layer problem, the output weight optimization of the upper limit prediction model and the lower limit prediction model is simultaneously performed by adopting parallel calculation; in the upper layer problem, the lower layer optimization corresponding to different super parameter sets is also performed simultaneously. The application of the parallel computing technology in the parameter optimization process effectively improves the computing efficiency, and overcomes the defects of time consumption, low efficiency and easy sinking into local optimum caused by multi-step iterative computation of the heuristic algorithm.
The published real photovoltaic power data of a photovoltaic power station in Australia in 2016-2018 12 is selected as a data source of an implementation case. The time resolution of the data is 30 minutes/point, and the data comprises photovoltaic historical power generation data and gas image historical actual measurement data. Preprocessing the original data into data samples, taking data from 2016 years 12 to 2017 years 12 as training data, and taking data from 2017 years 12 to 2018 years 12 as test data. The original data are divided according to seasons, and a prediction model is independently built for training and testing. Data from 10:00pm to 4:00am were excluded and predictions were made for 18 hours of data during the day. Predictive models with a predictive step size of 30 minutes were trained for nominal coverage of 85%, 90% and 95%, respectively.
And evaluating the reliability of the interval coverage rate of the prediction model based on the prediction result of the test set. The error between the actual coverage and the expected coverage as shown in the formula (14) is also one of the evaluation criteria, and the calculation formula is shown below. The comparison of the performance of the prediction interval obtained by the method disclosed in the present application is shown in table 1.
ACD＝PICR-PICR * (14)。
Table 1 evaluation results of prediction intervals of the present application in each season
Figure BDA0002444359500000073
Figure BDA0002444359500000081
As can be seen from the comparison result, the method and the device can meet the preset coverage rate PICR * The prediction of the narrowest interval width is achieved under the condition of (a). Fig. 2 and 3 show a partial prediction interval in summer with a nominal coverage of 95% and fig. 4 and 5 show a partial prediction interval in winter with a nominal coverage of 85%. As can be seen from fig. 2-5, the method of the present invention enables a better interval sharpness to be achieved with good reliability.

Claims (6)

1. A photovoltaic interval prediction method based on a bias loss function is characterized in that an upper limit prediction model and a lower limit prediction model of a photovoltaic interval are established, a bias loss function for determining output weights under the constraint of super parameters is respectively constructed for the upper limit prediction model and the lower limit prediction model, so that the expected coverage rate of the prediction interval is achieved, the width of an average prediction interval is narrowest, an optimal super parameter combination is searched for, the output weight when the bias loss function value of the upper limit prediction model is minimum or the bias loss function value of the lower limit prediction model is minimum is determined under the constraint of the optimal super parameter combination, the upper limit prediction model and the lower limit prediction model of the output weight are combined to obtain an interval prediction model, and the input data is processed by the interval prediction model to obtain a photovoltaic output interval prediction result;
the upper limit prediction model is as follows
Figure FDA0004058467450000011
The lower limit prediction model is
Figure FDA0004058467450000012
Figure FDA0004058467450000013
βOutput weights of upper limit prediction model and lower limit prediction model respectively, < >>
Figure FDA0004058467450000014
β＝[β 1 ,β 2 ,…，β l ] T L is the total number of neurons of the hidden layer of the predictive model, < >>
Figure FDA0004058467450000015
Δy i ＝L i -y i ，/>
Figure FDA0004058467450000016
As a predicted deviation of the upper limit of the interval, deltay i For the prediction bias of the lower interval limit, +.>
Figure FDA0004058467450000017
Δy i ≤0，y i For the actual value of the ith predicted point, U i 、L i C is the upper limit and the lower limit of the actual value of the ith predicted point u And C l Regularization coefficients, V, of the upper and lower prediction models, respectively u And V l Super-parameters of the upper limit prediction model and the lower limit prediction model respectively, V u ＝{W u ,r u ,c u ,C u Sum V l ＝{W l ,r l ,c l ,C l }，W u 、r u 、c u Weights, scaling coefficients, translation coefficients, W, respectively, of the partial convex loss function term of the upper limit prediction model l 、r l 、c l The weight, the scaling coefficient and the translation coefficient of the partial convex loss function item of the lower limit prediction model are respectively provided;
the coverage rate PICR of the prediction interval is
Figure FDA0004058467450000018
Average prediction interval width PINAW of
Figure FDA0004058467450000019
N is the number of statistical data samples in the training data set, exponential function +.>
Figure FDA00040584674500000110
Measuring the actual value y of the ith predicted point i Whether or not in the prediction interval [ L ] i ,U i ]Within the above-mentioned, the first and second,
Figure FDA00040584674500000111
Figure FDA00040584674500000112
representing taking the largest actual value in the training dataset,/->
Figure FDA00040584674500000113
Representing taking the smallest actual value in the training dataset.
2. The photovoltaic interval prediction method based on the biased convex loss function according to claim 1, wherein in the process of determining the output weight when the upper-limit prediction model has the minimum biased convex loss function value or the lower-limit prediction model has the minimum biased convex loss function value under the constraint of the optimal super-parameter combination, the method of minimizing the biased convex loss function is to solve the root of the gradient of the biased convex loss function.
3. The photovoltaic interval prediction method based on the biased convex loss function according to claim 1, wherein the method for searching the optimal super-parameter combination is as follows: and respectively carrying out sensitivity analysis on the super parameters of the upper limit prediction model and the lower limit prediction model to obtain a sensitivity interval, generating candidate super parameter search networks of the upper limit prediction model and the lower limit prediction model in the sensitivity interval, calculating the coverage rate and the width of the prediction interval under different super parameter combinations in parallel, and selecting the optimal super parameter combination which reaches the expected coverage rate of the prediction interval and has the narrowest average prediction interval width.
4. The method of claim 1, wherein the prediction interval period is reached by a photovoltaic interval prediction method based on a biased loss functionThe optimization problem of searching the optimal super-parameter combination by taking the narrowest average prediction interval width as the target of the coverage rate is as follows:
Figure FDA0004058467450000021
PICR * and (5) the coverage rate is expected for a preset prediction interval.
5. The photovoltaic interval prediction method based on the biased convex loss function according to claim 1, wherein the input data comprising the photovoltaic conversion rate, the photovoltaic output from the previous moment to the photovoltaic output from the previous moment are sequentially preprocessed, feature extracted and normalized to be input into the interval prediction model.
6. The photovoltaic interval prediction method based on the biased convex loss function according to claim 1, wherein the upper limit prediction model and the lower limit prediction model are implemented through a neural network, input weights of the neural network are randomly generated, and the total number of neurons of the hidden layer is determined through cross-validation of a point prediction result.
CN202010274717.0A 2020-04-09 2020-04-09 Photovoltaic interval prediction method based on biased convex loss function Active CN111626468B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010274717.0A CN111626468B (en) 2020-04-09 2020-04-09 Photovoltaic interval prediction method based on biased convex loss function

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010274717.0A CN111626468B (en) 2020-04-09 2020-04-09 Photovoltaic interval prediction method based on biased convex loss function

Publications (2)

Publication Number Publication Date
CN111626468A CN111626468A (en) 2020-09-04
CN111626468B true CN111626468B (en) 2023-07-14

Family

ID=72259712

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010274717.0A Active CN111626468B (en) 2020-04-09 2020-04-09 Photovoltaic interval prediction method based on biased convex loss function

Country Status (1)

Country Link
CN (1) CN111626468B (en)

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108921339B (en) * 2018-06-22 2021-08-10 南京工程学院 Quantile regression-based photovoltaic power interval prediction method for genetic support vector machine
CN108985965A (en) * 2018-06-22 2018-12-11 华北电力大学 A kind of photovoltaic power interval prediction method of combination neural network and parameter Estimation
CN109376951B (en) * 2018-11-21 2020-10-02 华中科技大学 Photovoltaic probability prediction method
CN110059862A (en) * 2019-03-25 2019-07-26 国网浙江省电力有限公司电力科学研究院 A kind of photovoltaic interval prediction method and system based on from coding and extreme learning machine
CN110110912B (en) * 2019-04-26 2021-09-28 华北电力大学 Photovoltaic power multi-model interval prediction method

Also Published As

Publication number Publication date
CN111626468A (en) 2020-09-04

Similar Documents

Publication Publication Date Title
CN107766990B (en) Method for predicting power generation power of photovoltaic power station
CN109002915B (en) Photovoltaic power station short-term power prediction method based on Kmeans-GRA-Elman model
CN107194495B (en) Photovoltaic power longitudinal prediction method based on historical data mining
CN110909919A (en) Photovoltaic power prediction method of depth neural network model with attention mechanism fused
CN109165774A (en) A kind of short-term photovoltaic power prediction technique
CN106251001A (en) A kind of based on the photovoltaic power Forecasting Methodology improving fuzzy clustering algorithm
CN116937579B (en) Wind power interval prediction considering space-time correlation and interpretable method thereof
CN114970952B (en) Photovoltaic output short-term prediction method and system considering environmental factors
CN113554466A (en) Short-term power consumption prediction model construction method, prediction method and device
Liu et al. Heating load forecasting for combined heat and power plants via strand-based LSTM
CN110866633B (en) Micro-grid ultra-short-term load prediction method based on SVR support vector regression
CN114462718A (en) CNN-GRU wind power prediction method based on time sliding window
Li et al. Deep spatio-temporal wind power forecasting
CN116341613A (en) Ultra-short-term photovoltaic power prediction method based on Informar encoder and LSTM
CN116842337A (en) Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model
CN116757057A (en) Air quality prediction method based on PSO-GA-LSTM model
CN112288157A (en) Wind power plant power prediction method based on fuzzy clustering and deep reinforcement learning
CN116341717A (en) Wind speed prediction method based on error compensation
CN116187508A (en) Wind turbine generator system fault prediction method integrating knowledge graph and deep learning
CN116345555A (en) CNN-ISCA-LSTM model-based short-term photovoltaic power generation power prediction method
Meng et al. A new PV generation power prediction model based on GA-BP neural network with artificial classification of history day
CN112836876B (en) Power distribution network line load prediction method based on deep learning
CN113962441A (en) Short-term irradiance prediction method and prediction system based on historical data analysis
CN117132132A (en) Photovoltaic power generation power prediction method based on meteorological data
CN111626468B (en) Photovoltaic interval prediction method based on biased convex loss function

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