+

CN113988437A - An Adaptive Interval Forecasting Method for Short-Term Residential Load - Google Patents

An Adaptive Interval Forecasting Method for Short-Term Residential Load Download PDF

Info

Publication number
CN113988437A
CN113988437A CN202111283707.4A CN202111283707A CN113988437A CN 113988437 A CN113988437 A CN 113988437A CN 202111283707 A CN202111283707 A CN 202111283707A CN 113988437 A CN113988437 A CN 113988437A
Authority
CN
China
Prior art keywords
prediction
interval
loss function
adaptive
hyperparameters
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111283707.4A
Other languages
Chinese (zh)
Other versions
CN113988437B (en
Inventor
龙寰
耿润昊
顾伟
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
Original Assignee
Southeast University
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 filed Critical Southeast University
Priority to CN202111283707.4A priority Critical patent/CN113988437B/en
Publication of CN113988437A publication Critical patent/CN113988437A/en
Application granted granted Critical
Publication of CN113988437B publication Critical patent/CN113988437B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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 OR CALCULATING; 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Economics (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Software Systems (AREA)
  • Tourism & Hospitality (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • General Business, Economics & Management (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • Marketing (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Primary Health Care (AREA)
  • Water Supply & Treatment (AREA)
  • Public Health (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a short-term resident load adaptive interval prediction method, which comprises the following steps: collecting processing data, and establishing an upper limit and a lower limit prediction model of a load interval; reducing the number of hyperparameters with a convex loss function through parameter transformation, taking the hyperparameters as loss functions for optimizing parameters of an upper limit prediction model and a lower limit prediction model, realizing automatic optimization of hyperparameters of the loss functions under iterative adjustment of a proposed adaptive updating strategy, and solving a minimized loss function; constructing an optimal prediction interval which achieves the expected coverage rate and has the narrowest average prediction interval width according to the minimum loss function; the prediction performance and stability of the model are further improved by integrating the single-edge coverage rate index into the proposed self-adaptive updating strategy. According to the prediction method, the upper and lower limit prediction models are constructed, the hyperparameters of the loss function are adjusted based on the unilateral coverage rate, the number of the hyperparameters is simplified through parameter transformation, and the hyperparameter adjustment difficulty is remarkably reduced.

Description

Self-adaptive interval prediction method for short-term resident load
Technical Field
The invention relates to the technical field of power load prediction, in particular to a short-term resident load adaptive interval prediction method.
Background
Short-term resident load prediction is an important component of power load prediction, and has high requirements on calculation efficiency, iteration frequency and prediction accuracy. Accurate short-term resident load prediction has important significance for safe and stable dispatching of the power grid. Most of the existing researches mainly focus on load point prediction, which can only provide a determined point prediction result, and lacks the evaluation on load data uncertainty, so that decision-making work faces risks. Therefore, a prediction interval based on a certain degree of confidence reflecting the fluctuation range and uncertainty of the load can reasonably describe the future change trend of the load.
At present, with the rapid increase of data volume, a deep learning method is widely applied to the field of load prediction, and beneficial information can be better extracted from massive data. And a high-quality prediction interval is constructed, and the calculation efficiency is limited by a large number of over-parameters.
In view of the above existing problems, an adaptive interval prediction method for short-term residential load is proposed.
Disclosure of Invention
The invention aims to provide a short-term resident load adaptive interval prediction method, which solves the problems that the instability evaluation of load data is lack, the evaluation is unreasonable, a large amount of hyperparameters influence the calculation efficiency and the like in the prior art.
The purpose of the invention can be realized by the following technical scheme:
an adaptive interval prediction method of short-term resident load, the prediction method comprising the steps of:
step one, collecting processing data, and establishing an upper limit and a lower limit prediction model of a load interval.
And step two, reducing the number of the hyperparameters of the partial convex loss function through parameter transformation, taking the hyperparameters as loss functions for optimizing parameters of an upper limit prediction model and a lower limit prediction model, realizing automatic optimization of the hyperparameters of the loss functions under iterative adjustment of the proposed adaptive updating strategy, and solving the minimized loss function.
And step three, constructing an optimal prediction interval which achieves the expected coverage rate and has the narrowest average prediction interval width according to the minimum loss function.
And step four, the prediction performance and the stability of the model are further improved by integrating the single-edge coverage rate index into the proposed self-adaptive updating strategy.
Further, the upper limit prediction model in the first step is
Figure BDA0003332207090000021
The lower limit prediction model is
Figure BDA0003332207090000022
E (-) is a modified biased loss function, PuAnd PlNetwork parameters, omega, of an upper-bound prediction model and a lower-bound prediction model, respectivelyuAnd ΩlHyperparameters, Ω, of loss functions of the upper and lower prediction models, respectivelyu＝{Wu,cuAnd Ωl＝{Wl,cl}，Wu、cuRespectively improving the weight and the translation coefficient of a penalty term with a partial convex loss function for an upper limit prediction model, Wl、clRespectively improving the weight and the translation coefficient of a penalty term with a convex loss function for the lower limit prediction model,
Figure BDA0003332207090000023
Figure BDA0003332207090000024
for the predicted deviation of the upper limit of the interval,
Figure BDA0003332207090000025
as a predicted deviation of the lower bound of the interval, ytFor the true value, u, corresponding to the t-th predicted pointt、ltAnd (4) predicting the upper limit and the lower limit of the interval for the t-th predicted point.
Further, the second step includes predicting an interval coverage ratio PICP as
Figure BDA0003332207090000026
N is the number of samples in the data set, function
Figure BDA0003332207090000027
Measuring the actual value y of the t predicted pointstWhether or not in the prediction interval [ l ]t,ut]In the above-mentioned manner,
Figure BDA0003332207090000028
average prediction interval width PINAW of
Figure BDA0003332207090000029
R is the span of the true value,
Figure BDA00033322070900000210
the overall prediction interval evaluation index CWC is
Figure BDA00033322070900000211
Gamma is used to determine whether to introduce an exponential term,
Figure BDA0003332207090000031
normalized root mean square width PINRW of prediction interval as
Figure BDA0003332207090000032
η1Linear expansion of the value of PINRW, eta2For the penalty coefficient of the PICP, the PINC is the expected coverage rate of a prediction interval preset when the model is tested.
Further, the method for reducing the number of the hyperparameters of the loss function by using parameter transformation in the second step comprises the following steps: and performing equivalent transformation on the biased convex loss function, eliminating a scaling coefficient under the scene of minimizing the loss function, removing a regularization coefficient, and reducing the number of the hyper-parameters from eight to four.
Further, the step of obtaining the optimal prediction interval in the third step is as follows: s1, when the coverage rate of the prediction interval reaches the start coverage rate, adjusting the hyper-parameters by using a self-adaptive updating strategy to construct and improve a biased loss function, otherwise, keeping the hyper-parameters unchanged; s2, solving a minimization loss function problem according to an Adam algorithm to obtain update parameters of an upper and lower limit prediction model; and S3, outputting by using an upper limit and a lower limit prediction model to form a prediction interval, and sequentially iterating and circulating the process until convergence to obtain an optimal prediction interval.
Furthermore, the single-sided coverage rate index in the fourth step equally divides the prediction interval into the upper and lower sub-prediction intervals, and the coverage rates of the upper and lower sub-prediction intervals all reach the expected coverage rate through a self-adaptive updating strategy, so that the true values are more uniformly distributed in the prediction interval, the control of the coverage rate of the prediction interval is finer, and the average width is further reduced.
Further, the adaptive updating strategy proposed in the fourth step directly utilizes the relation between the improved over-parameter of the biased-convex loss function and the coverage rate of the prediction interval to construct the over-parameter adaptive updating strategy, and in the process of solving the optimal parameters of the upper and lower limit prediction models by minimizing the improved biased loss function, the over-parameter of the loss function is optimized, so that the optimal adjustment of the over-parameter and the construction of the optimal prediction interval are realized simultaneously.
Further, the iterative update formula of the adaptive update strategy is as follows:
Figure BDA0003332207090000041
Figure BDA0003332207090000042
Figure BDA0003332207090000043
Figure BDA0003332207090000044
in the above formula: k is a radical of1And k2Controlling the iterative update rate, PICP*Preset for training modelsI denotes the current iteration number.
Further, the step one of collecting and processing data includes sequentially preprocessing input data including temperature forecast data and load history data of the previous week, extracting features, normalizing, and inputting the input data into the interval prediction model.
Furthermore, the upper limit prediction model and the lower limit prediction model are realized through a gate cycle network and a fully connected neural network, initialization parameters are obtained by an Xavier and He initialization method respectively, and the network structure is determined by performing cross validation on the model point prediction result.
The invention has the beneficial effects that:
1. the prediction method converts the short-term resident load interval prediction problem into two independent prediction subproblems of an upper limit and a lower limit, respectively constructs the upper limit and the lower limit of a prediction interval by using two independent networks, supervises the parameter learning of an upper limit prediction model and a lower limit prediction model by using two independent loss functions, adaptively adjusts the hyper-parameters of the two loss functions based on two independent unilateral coverage rates, and obviously improves the independence and the degree of freedom of model training;
2. the prediction method reduces the number of the hyper-parameters to four through parameter transformation and reasonable simplification, remarkably reduces the difficulty of hyper-parameter adjustment, simplifies calculation, constructs a self-adaptive updating method according to the internal relation between the hyper-parameters and the coverage rate of a prediction interval, automatically optimizes the hyper-parameters of a loss function in the training process, solves the defect of difficulty in preselecting the optimal hyper-parameters of biased loss functions, and improves the stability of the training effect;
3. according to the prediction method, the prediction interval is equally divided into two subintervals, and the prediction interval with more uniform true value distribution is favorably obtained by matching with a self-adaptive updating strategy, so that more detailed construction of the prediction interval is realized, the quality of the prediction interval is further improved, and the evaluation result is reasonable.
Drawings
The invention will be further described with reference to the accompanying drawings.
FIG. 1 is a flowchart of a short term resident load section prediction according to the present invention;
FIG. 2 is a graph comparing the predicted interval to the actual value for the present application at a single step lead of 95% predicted interval coverage;
FIG. 3 is an enlarged view of a portion of FIG. 2 of the present invention;
FIG. 4 is a graph comparing the predicted interval and actual values of the present application at four steps leading the coverage of the 95% predicted interval in accordance with the present invention;
FIG. 5 is an enlarged view of a portion of the invention shown in FIG. 4;
FIG. 6 is a graph illustrating the convergence curve of the loss function over-parameter under the single-step look-ahead 85% prediction interval coverage;
FIG. 7 is a graph showing the convergence curve of the loss function over-parameter under the single-step advance of 85% prediction interval coverage according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention discloses a short-term resident load adaptive interval prediction method, which is a framework of adaptive iterative optimization, and as shown in figure 1, the prediction method comprises the following steps:
s1, preprocessing data, extracting characteristics and generating tensor needed by training
Integrating input data including temperature forecast data and load historical data of the previous week into a data sample form containing input and output vectors, preprocessing and extracting features of the data, normalizing data variables to be in a [0,1] range, and generating a tensor required by training and suitable for a deep learning model based on the size of a time window obtained by correlation analysis.
S2, providing load prediction interval evaluation indexes and initializing overall parameters of the model
The evaluation indexes of the load prediction interval comprise average prediction interval width PINAW, prediction interval coverage rate PICP and interval overall evaluation index CWC, and the given training data set is { (x)1,y1),…,(xN,yN) And the prediction interval evaluation indexes are respectively shown as formulas (1) to (7):
Figure BDA0003332207090000061
Figure BDA0003332207090000062
Figure BDA0003332207090000063
Figure BDA0003332207090000064
Figure BDA0003332207090000065
Figure BDA0003332207090000066
Figure BDA0003332207090000067
in the above equations (1) to (7), N is the number of statistical data samples in the training data set, y1For the 1 st predicted point x1Actual value of (a), yNFor the Nth predicted point xNActual value, function of
Figure BDA0003332207090000068
Measuring the actual value y of the t predicted pointstWhether or not toIn the prediction interval lt,ut]Inner, ut、ltUpper and lower limits of prediction interval for the tth prediction point, R is span of true value, and gamma is used for judging whether to introduce exponential term, eta1Linear expansion of the value of PINRW, eta2For the penalty factor of the PICP, the PINC is a preset expected coverage.
Respectively obtaining initialization network parameters of a GRU network and a fully-connected neural network by utilizing an Xavier and He initialization method, carrying out cross validation on a model point prediction result to determine a network structure, and obtaining an update rate k by utilizing a cross validation method1、k2And training the set desired interval coverage PICP*Randomly generating a loss function hyperparameter, wherein the weight W of the penalty term of the loss function is [10,15 ]]Is randomly generated, and the translation coefficient c is in [2,4 ]]Is randomly generated.
S3, constructing an interval prediction model taking the improved partial convex loss function as an optimization target
The prediction model adopts a GRU network and a single-output fully-connected neural network, the GRU is used for extracting features from time sequence data, the extracted information is sent to the fully-connected neural network to generate a prediction interval boundary value, two independent networks are used as an interval upper limit prediction model and a lower limit prediction model respectively, and optimal parameters of the prediction model are obtained by solving and improving the minimization problem of a function with a convex loss.
The steps for constructing the improved bumpy loss function are as follows:
(a) for the training data set { (x)t,yt) Target y of 1, … N | t |tThe upper and lower limits of the prediction interval are utAnd ltAnd (4) showing. The prediction deviation of the upper limit and the lower limit of the interval is defined as formula (8):
θu,t＝yt-ut,θl,t＝lt-yt (8)
the upper limit deviation and the lower limit deviation corresponding to the true value falling in the prediction interval need to satisfy the formula (9):
θu,t≤0,θl,t≤0 (9)
with qtIn place of qu,tAnd q isl,tThen the initial partial convex loss function is of the form:
Figure BDA0003332207090000071
Figure BDA0003332207090000072
w is the weight of the penalty term S (-) with the offset loss function, and r and c are the scaling coefficient and the translation coefficient of the function S (-) with the offset loss function.
(b) Reducing the number of hyperparameters of loss functions through parameter transformation: and performing equivalent transformation on the biased convex loss function, eliminating a scaling coefficient under the scene of minimizing the loss function, removing a regularization coefficient, and reducing the number of the hyper-parameters from eight to four.
Parameter conversion of the above equation (10)
Figure BDA0003332207090000073
The hyperparameter r in S (-) can be eliminated and the transformed loss function is of the form:
Figure BDA0003332207090000081
Figure BDA0003332207090000082
in the context of minimizing the loss function,
Figure BDA0003332207090000083
and
Figure BDA0003332207090000084
act in unison and therefore use
Figure BDA0003332207090000085
Instead of the former
Figure BDA0003332207090000086
Figure BDA0003332207090000087
An improved bumpy loss function is obtained, as shown in equations (14) and (13):
Figure BDA0003332207090000088
therefore, the improved loss function only has two superparameters, W and c.
(c) An optimization objective function of upper and lower limit prediction model parameters is constructed based on the improved partial convex loss function, and the optimal parameters of the prediction model are obtained by solving the minimization problem, as shown in formulas (15) and (16):
Figure BDA0003332207090000089
Figure BDA00033322070900000810
wherein P isuAnd PlNetwork parameters, omega, of an upper-bound prediction model and a lower-bound prediction model, respectivelyuAnd ΩlHyperparameters, Ω, of loss functions of the upper and lower prediction models, respectivelyu＝{Wu,cuAnd Ωl＝{Wl, cl}，NbRepresenting the number of samples of the training batch, Wu、cuRespectively improving the weight and the translation coefficient of a penalty term with a partial convex loss function for an upper limit prediction model, Wl、clRespectively improving the weight and the translation coefficient of a penalty term with a convex loss function for the lower limit prediction model,
Figure BDA00033322070900000811
for the predicted deviation of the upper limit of the interval,
Figure BDA00033322070900000812
as a predicted deviation of the lower bound of the interval, ytFor the true value, u, corresponding to the t-th predicted pointt、ltThe calculation formula is the same as the formula (8) for the upper and lower limits of the prediction interval of the t-th prediction point.
S4 optimizing and adjusting loss function hyperparameters by utilizing unilateral self-adaptive updating strategy
According to the internal relation between the improved hyper-parameter with the convex loss function and the coverage rate of the prediction interval, a self-adaptive updating strategy is designed, and the optimal hyper-parameter is automatically searched in the iterative training process. The proportion of the penalty term in the loss function is controlled by the hyper-parameter W, the value of the penalty term is increased so as to apply larger penalty to points outside the prediction boundary, the prediction boundary is rapidly adjusted so as to include the point, the corresponding coverage rate is improved, otherwise, the coverage rate of the prediction interval can be reduced by reducing the value of W, and therefore, the self-adaptive updating formula of W is shown as the formula (17):
W←W-k1·(PICP-PICP*) (17)
PICP*representing the expected coverage set when training the model.
Because the coverage rate of the prediction interval does not need to reach 100%, the super-parameter c has the function of mitigating the effect of the penalty term on the target true value, the coverage rate of the prediction interval is improved by reducing the value of c, the coverage rate is reduced by increasing the value of c, and the adaptive updating formula of c is shown as the formula (18):
c←c+k2·(PICP-PICP*) (18)
in order to improve the fineness of the adaptive updating strategy for controlling the predictive performance and further improve the prediction quality, a prediction interval evaluation index of the unilateral coverage rate is provided, the prediction interval is equally divided into two subintervals to respectively calculate the coverage rate, and the calculation formula is shown in formulas (19) to (21):
Figure BDA0003332207090000091
Figure BDA0003332207090000092
Figure BDA0003332207090000093
in ideal case, the interval coverage satisfies PICPu＝PICPl＝PICP＝PICP*At this time, the true values are more uniformly distributed in the prediction interval, so that the average width of the prediction interval is optimized, and based on the unilateral coverage rate index, a self-adaptive updating formula of the loss function hyperparameter is obtained, as shown in formula (22):
Figure BDA0003332207090000094
where i represents the number of iterations.
The above formula is called a unilateral self-adaptive updating strategy, when the coverage rate of the prediction interval reaches the start coverage rate, the numerical value of the hyper-parameter is iteratively adjusted by using the above formula, then the numerical value is transmitted into a loss function, the parameter of the prediction model is obtained by solving, and the iteration is continuously carried out until convergence, so that the optimal prediction interval is obtained.
The average value data of the load power of two hundred or more residents in a certain area in the first quarter of a year is selected as a data source of an implementation case, and the time resolution of the data is 15 minutes/point. The raw data is preprocessed into data samples, and the data at the front 2/3 is used as training data, and the data at the back 1/3 is used as testing data. Prediction models with prediction steps of 15 minutes and 1 hour, i.e., single-step and four-step lead prediction models, were trained for nominal coverage of 85%, 90%, and 95%, respectively.
And (3) based on the prediction result of the test set, evaluating the reliability and stability of the prediction model by using an average coverage rate deviation index, wherein a calculation formula is shown as a formula (23), and performing overall evaluation on the prediction region by using an evaluation index CWC.
ACD＝PICP-PINC (23)
The evaluation results of the prediction intervals obtained by the method disclosed in the present application are shown in table 1.
TABLE 1 evaluation results of prediction intervals of the present application for each number of look-ahead steps and expected coverage
Figure BDA0003332207090000101
According to the evaluation result, the coverage rate can be carefully controlled to be close to the preset expected coverage rate PINC, the coverage rate requirement is met, and meanwhile, the excellent average interval width and the interval overall quality are guaranteed. Fig. 2 and 3 show partial prediction intervals with a desired coverage of 95% and advanced by one step, and fig. 4 and 5 show partial prediction intervals with a desired coverage of 95% and advanced by four steps. As can be seen from fig. 2 to 5, the method of the present invention can achieve better interval sharpness while ensuring good reliability.
Fig. 6 and 7 show the convergence performance curves of the method disclosed in the present application, and convergence can be achieved under 40 iterations, so as to meet the requirement of short-term load prediction on computation time.
Table 2 demonstrates the effect of the single-sided coverage index in the methods described herein through specific experiments. And (3) measuring the uniformity degree of the distribution of the real value in the prediction interval by adopting a unilateral coverage rate deviation index, wherein the calculation formula is shown as a formula (24), and the smaller the value is, the more uniform the distribution of the real value in the prediction interval is.
ΔPICP＝|PICPu-PINC|+|PICPl-PINC| (24)
The adaptive updating method adopting the whole coverage rate of the prediction interval is called a bilateral adaptive updating strategy, and the method adopting the unilateral coverage rate is called a unilateral adaptive updating strategy.
TABLE 2 comparison of two-sided and one-sided adaptive update strategies at single-step 90% expected coverage
Figure BDA0003332207090000111
According to the comparison result, the unilateral coverage rate index enables the self-adaptive updating strategy to have finer control on the coverage rate, the distribution of the true value in the prediction interval is more uniform, and the average width of the prediction interval is effectively reduced.
In the description herein, references to the description of "one embodiment," "an example," "a specific example," etc., mean that a particular feature, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
The foregoing shows and describes the general principles, essential features, and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed.

Claims (10)

1.一种短期居民负荷的自适应区间预测方法，其特征在于，所述预测方法包括以下步骤：1. An adaptive interval prediction method for short-term resident load, characterized in that the prediction method comprises the following steps: 步骤一、收集处理数据，建立负荷区间的上限以及下限预测模型；Step 1: Collect and process data, and establish a prediction model for the upper limit and lower limit of the load interval; 步骤二、通过参数变换削减有偏凸损失函数的超参数个数，并将其作为优化上限预测模型和下限预测模型参数的损失函数，在提出的自适应更新策略的迭代调整下，实现损失函数超参数的自动寻优，求解最小化损失函数；Step 2: Reduce the number of hyperparameters of the partial convex loss function through parameter transformation, and use it as the loss function for optimizing the parameters of the upper limit prediction model and the lower limit prediction model. Under the iterative adjustment of the proposed adaptive update strategy, the loss function is realized Automatic optimization of hyperparameters to minimize the loss function; 步骤三、根据最小化损失函数构建达到期望覆盖率且平均预测区间宽度最窄的最优预测区间；Step 3: Construct an optimal prediction interval that achieves the desired coverage and has the narrowest average prediction interval width according to the minimized loss function; 步骤四、通过将单边覆盖率指标融入提出的自适应更新策略，进一步提升模型的预测性能和稳定性。Step 4: By incorporating the unilateral coverage index into the proposed adaptive update strategy, the prediction performance and stability of the model are further improved. 2.根据权利要求1所述的一种短期居民负荷的自适应区间预测方法，其特征在于，所述步骤一中上限预测模型为
Figure FDA0003332207080000011
下限预测模型为
Figure FDA0003332207080000012
E(·)为改进的有偏凸损失函数，Pu和Pl分别为上限预测模型和下限预测模型的网络参数，Ωu和Ωl分别为上限预测模型和下限预测模型损失函数的超参数，Ωu＝{Wu,cu}和Ωl＝{Wl,cl}，Wu、cu分别为上限预测模型改进有偏凸损失函数惩罚项的权重、平移系数，Wl、cl分别为下限预测模型改进有偏凸损失函数惩罚项的权重、平移系数，
Figure FDA0003332207080000013
为区间上限的预测偏差，
Figure FDA0003332207080000014
为区间下限的预测偏差，yt为第t个预测点对应的真实值，ut、lt为第t个预测点预测区间的上下限。2. The self-adaptive interval prediction method for short-term resident load according to claim 1, wherein in the step 1, the upper limit prediction model is
Figure FDA0003332207080000011
The lower bound prediction model is
Figure FDA0003332207080000012
E( ) is the improved partial convex loss function, P u and P l are the network parameters of the upper and lower prediction models, respectively, Ω u and Ω l are the hyperparameters of the loss functions of the upper and lower prediction models, respectively , Ω u ={W u ,c u } and Ω l ={W l ,c l }, Wu and c u are the weights and translation coefficients of the penalty term of the improved partial convex loss function of the upper bound prediction model, W l , c l are the weight and translation coefficient of the penalty term for the improved partial convex loss function of the lower bound prediction model, respectively,
Figure FDA0003332207080000013
is the forecast deviation of the upper bound of the interval,
Figure FDA0003332207080000014
is the prediction deviation of the lower limit of the interval, y t is the true value corresponding to the t-th prediction point, and u t and l t are the upper and lower limits of the prediction interval of the t-th prediction point. 3.根据权利要求2所述的一种短期居民负荷的自适应区间预测方法，其特征在于，所述步骤二包括预测区间覆盖率PICP为
Figure FDA0003332207080000015
N为数据集中样本的个数，函数
Figure FDA0003332207080000016
衡量第t个预测点的实际值yt是否在预测区间[lt,ut]之内，
Figure FDA0003332207080000021
3. The self-adaptive interval prediction method of a short-term resident load according to claim 2, wherein the step 2 comprises that the prediction interval coverage rate PICP is:
Figure FDA0003332207080000015
N is the number of samples in the data set, the function
Figure FDA0003332207080000016
Measure whether the actual value y t of the t-th forecast point is within the forecast interval [l t , u t ],
Figure FDA0003332207080000021
 平均预测区间宽度PINAW为
Figure FDA0003332207080000022
R为真实值的跨度，
Figure FDA0003332207080000023
The mean prediction interval width PINAW is
Figure FDA0003332207080000022
R is the span of the true value,
Figure FDA0003332207080000023
 预测区间整体评价指标CWC为
Figure FDA0003332207080000024
γ用于判断是否引入指数项，
Figure FDA0003332207080000025
The overall evaluation index CWC of the prediction interval is
Figure FDA0003332207080000024
γ is used to judge whether to introduce an exponential term,
Figure FDA0003332207080000025
 预测区间归一化均方根宽度PINRW为
Figure FDA0003332207080000026
η1线性扩大PINRW的数值，η2为PICP的惩罚系数，PINC为测试模型时预设的预测区间期望覆盖率。The normalized root mean square width of the prediction interval, PINRW, is
Figure FDA0003332207080000026
η 1 linearly expands the value of PINRW, η 2 is the penalty coefficient of PICP, and PINC is the expected coverage rate of the prediction interval preset when testing the model. 4.根据权利要求3所述的一种短期居民负荷的自适应区间预测方法，其特征在于，所述步骤二中利用参数变换削减损失函数超参数个数的方法为：对有偏凸损失函数作等效变换，在最小化损失函数的场景下消去缩放系数，并去除正则化系数，将超参数个数从八个降至四个。4. The self-adaptive interval prediction method of a short-term resident load according to claim 3, wherein the method for reducing the number of hyperparameters of the loss function by using parameter transformation in the step 2 is: for a partial convex loss function As an equivalent transformation, the scaling factor is eliminated in the scenario of minimizing the loss function, and the regularization factor is removed, reducing the number of hyperparameters from eight to four. 5.根据权利要求3所述的一种短期居民负荷的自适应区间预测方法，其特征在于，所述步骤三中求取最优预测区间的步骤为：S1、当预测区间覆盖率达到启动覆盖率时，利用自适应更新策略调整超参数构建改进有偏损失函数，否则超参数不变；S2、根据Adam算法求解最小化损失函数问题得到上下限预测模型的更新参数；S3、利用上下限预测模型输出构成预测区间，此过程依次迭代循环直至收敛求得最优预测区间。5. The self-adaptive interval prediction method for short-term resident load according to claim 3, wherein the step of obtaining the optimal prediction interval in the step 3 is: S1. When the coverage rate of the prediction interval reaches the startup coverage When the rate is higher, use the adaptive update strategy to adjust the hyperparameters to construct an improved biased loss function, otherwise the hyperparameters remain unchanged; S2. Solve the problem of minimizing the loss function according to the Adam algorithm to obtain the update parameters of the upper and lower limit prediction models; S3, use the upper and lower bounds to predict The output of the model constitutes the prediction interval, and this process iterates and loops successively until convergence to obtain the optimal prediction interval. 6.根据权利要求3所述的一种短期居民负荷的自适应区间预测方法，其特征在于，所述步骤四中单边覆盖率指标将预测区间等分为上下子预测区间，通过自适应更新策略使得上下子预测区间覆盖率均达到期望覆盖率，以保证真实值更均匀地分布在预测区间之中，使得预测区间覆盖率的控制更为精细且平均宽度进一步减小。6. The self-adaptive interval prediction method for short-term resident load according to claim 3, characterized in that in the step 4, the unilateral coverage ratio index divides the prediction interval into upper and lower sub-prediction intervals equally, and is updated by adaptively updating the prediction interval. The strategy makes the coverage ratio of the upper and lower sub-prediction intervals reach the expected coverage ratio, so as to ensure that the real values are more evenly distributed in the prediction interval, so that the control of the prediction interval coverage ratio is more precise and the average width is further reduced. 7.根据权利要求6所述的一种短期居民负荷的自适应区间预测方法，其特征在于，所述步骤四中提出的自适应更新策略直接利用改进有偏凸损失函数超参数与预测区间覆盖率的关系，构建超参数自适应更新策略，在最小化改进有偏损失函数求取上下限预测模型最优参数的过程中，对损失函数超参数进行寻优，以同时实现超参数最优调节和最优预测区间构建。7. The adaptive interval prediction method of a short-term resident load according to claim 6, wherein the adaptive update strategy proposed in the step 4 directly utilizes the improved partial convex loss function hyperparameter and prediction interval coverage In the process of minimizing and improving the biased loss function to obtain the optimal parameters of the upper and lower bound prediction model, the hyperparameters of the loss function are optimized to realize the optimal adjustment of the hyperparameters at the same time. and the optimal prediction interval is constructed. 8.根据权利要求7所述的一种短期居民负荷的自适应区间预测方法，其特征在于，所述自适应更新策略的迭代更新公式为：8. The self-adaptive interval prediction method of a short-term resident load according to claim 7, wherein the iterative update formula of the self-adaptive update strategy is:
Figure FDA0003332207080000031
Figure FDA0003332207080000031
Figure FDA0003332207080000032
Figure FDA0003332207080000032
Figure FDA0003332207080000033
Figure FDA0003332207080000033
Figure FDA0003332207080000034
Figure FDA0003332207080000034
 上式中：k1和k2控制迭代更新速率，PICP*为训练模型所预设的预测区间期望覆盖率，i表示当前迭代次数。In the above formula: k 1 and k 2 control the iterative update rate, PICP * is the expected coverage rate of the prediction interval preset by the training model, and i represents the current number of iterations. 9.根据权利要求1所述的一种短期居民负荷的自适应区间预测方法，其特征在于，所述步骤一中收集处理数据包括对包含温度预报数据、前一周该时刻负荷历史数据的输入数据依次进行预处理、特征提取和归一化处理，再输入区间预测模型。9 . The method for self-adaptive interval prediction of short-term residential load according to claim 1 , wherein collecting and processing data in the step 1 comprises inputting data including temperature forecast data and historical load data at this moment in the previous week. 10 . Preprocessing, feature extraction and normalization are performed in sequence, and then the interval prediction model is input. 10.根据权利要求2-9任意一项所述的一种短期居民负荷的自适应区间预测方法，其特征在于，所述上限预测模型和下限预测模型通过门循环网络以及全连接神经网络实现，初始化参数分别由Xavier和He初始化方法得到，网络结构通过对模型点预测结果进行交叉验证确定。10. The self-adaptive interval prediction method for short-term resident load according to any one of claims 2-9, wherein the upper-limit prediction model and the lower-limit prediction model are implemented by a gated recurrent network and a fully connected neural network, The initialization parameters are obtained by the Xavier and He initialization methods, respectively, and the network structure is determined by cross-validation of the model point prediction results.
CN202111283707.4A 2021-11-01 2021-11-01 An adaptive interval forecasting method for short-term residential load Active CN113988437B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111283707.4A CN113988437B (en) 2021-11-01 2021-11-01 An adaptive interval forecasting method for short-term residential load

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111283707.4A CN113988437B (en) 2021-11-01 2021-11-01 An adaptive interval forecasting method for short-term residential load

Publications (2)

Publication Number Publication Date
CN113988437A true CN113988437A (en) 2022-01-28
CN113988437B CN113988437B (en) 2025-01-14

Family

ID=79745431

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111283707.4A Active CN113988437B (en) 2021-11-01 2021-11-01 An adaptive interval forecasting method for short-term residential load

Country Status (1)

Country Link
CN (1) CN113988437B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110598929A (en) * 2019-09-10 2019-12-20 河海大学 Wind power nonparametric probability interval ultrashort term prediction method
CN111461404A (en) * 2020-03-09 2020-07-28 华能四川水电有限公司 Short-term load and hydropower prediction method based on neural network prediction interval
CN111626468A (en) * 2020-04-09 2020-09-04 东南大学 Photovoltaic interval prediction method based on partial convex loss function
WO2021007812A1 (en) * 2019-07-17 2021-01-21 深圳大学 Deep neural network hyperparameter optimization method, electronic device and storage medium
AU2020104000A4 (en) * 2020-12-10 2021-02-18 Guangxi University Short-term Load Forecasting Method Based on TCN and IPSO-LSSVM Combined Model

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021007812A1 (en) * 2019-07-17 2021-01-21 深圳大学 Deep neural network hyperparameter optimization method, electronic device and storage medium
CN110598929A (en) * 2019-09-10 2019-12-20 河海大学 Wind power nonparametric probability interval ultrashort term prediction method
CN111461404A (en) * 2020-03-09 2020-07-28 华能四川水电有限公司 Short-term load and hydropower prediction method based on neural network prediction interval
CN111626468A (en) * 2020-04-09 2020-09-04 东南大学 Photovoltaic interval prediction method based on partial convex loss function
AU2020104000A4 (en) * 2020-12-10 2021-02-18 Guangxi University Short-term Load Forecasting Method Based on TCN and IPSO-LSSVM Combined Model

Also Published As

Publication number Publication date
CN113988437B (en) 2025-01-14

Similar Documents

Publication Publication Date Title
Wang et al. Nonparametric probabilistic forecasting for wind power generation using quadratic spline quantile function and autoregressive recurrent neural network
CN108900346B (en) Wireless network traffic prediction method based on LSTM network
CN108876039B (en) A Power Quality Prediction Method for Distribution Networks with Distributed Power Sources Based on Support Vector Machines
CN110707763B (en) AC/DC power distribution network load prediction method based on ensemble learning
Ramos et al. Residential energy consumption forecasting using deep learning models
CN109615124B (en) SCADA master station load prediction method based on deep learning
CN114139305A (en) Optimization method of flow characteristics of single valve based on pressure prediction model of steam turbine regulating stage
CN117592593A (en) Short-term power load prediction method based on improved quadratic modal decomposition and WOA optimization BILSTM-intent
CN112085254A (en) Prediction method and model based on multi-fractal cooperative measurement gating cycle unit
CN116663404A (en) A Flood Forecasting Method and System Coupling Artificial Intelligence and Bayesian Theory
CN114970343B (en) A multi-objective optimization method for cement grinding process based on search space variation
CN112434848A (en) Nonlinear weighted combination wind power prediction method based on deep belief network
CN109816167A (en) Runoff Forecast method and Runoff Forecast device
CN119598410A (en) Data processing method, device and equipment based on artificial intelligence technology and data model
CN117952242A (en) Self-adaptive prediction method and system for residual chlorine in water outlet of water works
CN107729988B (en) Blue algae bloom prediction method based on dynamic deep belief network
CN110119846A (en) A kind of multiple target deepness belief network for mid-term electric load forecasting
CN119472448B (en) Control method and system for semiconductor manufacturing machine
CN116520702A (en) A self-adaptive constraint multi-objective operation optimization control method for sewage treatment process
CN111832818A (en) A multi-step prediction method for LSTM network power generation based on correlation analysis
CN114219122B (en) Ultra-short-term prediction method of small hydropower station power generation based on grid meteorological data
CN119026988B (en) Multi-factor ship health dynamic prediction method
CN113988437A (en) An Adaptive Interval Forecasting Method for Short-Term Residential Load
CN119442856A (en) Ultra-short-term power forecasting optimization method for wind farms based on lidar
CN109523386A (en) A kind of investment portfolio risk prediction technique of GMM in conjunction with LSTM

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
点击 这是indexloc提供的php浏览器服务，不要输入任何密码和下载