CN116415734A - Ultra-short term load prediction method and system based on deep learning - Google Patents

Ultra-short term load prediction method and system based on deep learning Download PDF

Info

Publication number
CN116415734A
CN116415734A CN202310449828.4A CN202310449828A CN116415734A CN 116415734 A CN116415734 A CN 116415734A CN 202310449828 A CN202310449828 A CN 202310449828A CN 116415734 A CN116415734 A CN 116415734A
Authority
CN
China
Prior art keywords
load
trust
predicted
prediction
feature vector
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
CN202310449828.4A
Other languages
Chinese (zh)
Other versions
CN116415734B (en
Inventor
蒋雪冬
李晓波
汪超群
徐晓波
陈懿
王翔
迟长云
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.)
Zhejiang Zheda Energy Technology Co ltd
Original Assignee
Zhejiang Zheda Energy Technology 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 Zhejiang Zheda Energy Technology Co ltd filed Critical Zhejiang Zheda Energy Technology Co ltd
Priority to CN202310449828.4A priority Critical patent/CN116415734B/en
Publication of CN116415734A publication Critical patent/CN116415734A/en
Application granted granted Critical
Publication of CN116415734B publication Critical patent/CN116415734B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H02J3/003—Load forecast, e.g. methods or systems for forecasting future load demand
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • 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
    • G06N3/047—Probabilistic or stochastic 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
    • G06N3/048—Activation functions
    • 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
    • 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
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2123/00—Data types
    • G06F2123/02—Data types in the time domain, e.g. time-series data
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
    • H02J2203/10—Power transmission or distribution systems management focussing at grid-level, e.g. load flow analysis, node profile computation, meshed network optimisation, active network management or spinning reserve management
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
    • H02J2203/20—Simulating, e g planning, reliability check, modelling or computer assisted design [CAD]
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Molecular Biology (AREA)
  • Business, Economics & Management (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Water Supply & Treatment (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Public Health (AREA)
  • Probability & Statistics with Applications (AREA)
  • Primary Health Care (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Power Engineering (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention requests protection of a deep learning-based ultra-short-term load prediction method and a deep learning-based ultra-short-term load prediction system, wherein a plurality of historical predicted historical load of a current predicted load environment and work archive information of a target load terminal corresponding to the historical predictions are extracted by responding to a current predicted prediction operation in the current predicted load environment, and the target load terminal is a load terminal started in the historical predictions; generating a time-sharing characteristic vector corresponding to the current predicted load environment according to the work file information of the target load terminal and the historical predicted load; extracting a plurality of candidate prediction loads matched with each other; and confirming the trust characteristic vectors of the plurality of candidate predicted loads according to the time-sharing characteristic vectors and confirming the currently predicted trust predicted load from the plurality of candidate predicted loads. The method can fully utilize the historical operation data of the load terminal, and accurately predict the short-term load level in the corresponding range according to the specific requirements on the basis.

Description

Ultra-short term load prediction method and system based on deep learning
Technical Field
The invention belongs to the field of power load prediction, and particularly relates to an ultra-short-term load prediction method and system based on deep learning.
Background
Along with the continuous improvement of the scale and complexity of the power system, the accuracy of the short-term load prediction of the power system plays a key role in effectively reducing the power generation cost and implementing the power system optimization control in each region. Short-term load prediction is mainly used for scheduling power generation, and has highest timeliness compared with long-term load prediction. The load variation speed is high, and the load variation is greatly influenced by abrupt factors such as temperature difference, humidity and the like, and belongs to a dynamic nonlinear time sequence. Because of such characteristics of short term loads, it is difficult to achieve accurate predictions. With the implementation of new electricity change, the competition of electricity selling market is deepened continuously, and new requirements are put on prediction precision. Therefore, it is necessary to provide a method for predicting short-term loads of an electric power system with high accuracy.
Accurate short term load prediction (STLF) based is one of the key challenges in developing power supply plans and power supply-demand balances. It considers predictions for several days in the future, which is an essential basis for the operation and planning of the electric market. Improving the accuracy of short-term load prediction is helpful to improve the utilization rate of the power equipment, reduce the energy consumption, and relieve the unbalance between the power supply end and the demand end.
Disclosure of Invention
According to a first aspect of the present invention, the present invention claims an ultra-short term load prediction method based on deep learning, which is characterized in that the method comprises:
responding to the current predicted operation in the current predicted load environment, extracting historical predicted loads of a plurality of historical predictions of the current predicted load environment and work archive information of a target load terminal corresponding to the historical predictions, wherein the target load terminal is a load terminal started in the historical predictions;
generating a time-sharing characteristic vector corresponding to the current predicted load environment according to the work file information of the target load terminal and the historical predicted load;
extracting a plurality of candidate predicted loads matched with the current predicted load;
confirming trust feature vectors of the candidate predictive loads according to the time-division feature vectors;
and confirming the currently predicted trust prediction load from the candidate prediction loads according to the trust feature vector.
Further, the generating the time-period feature vector corresponding to the current predicted load environment according to the working archive information of the target load terminal and the historical predicted load includes:
Confirming a first load vector corresponding to the work file information of a plurality of target load terminals corresponding to a prediction target and a second load vector corresponding to the predicted load of the prediction target, wherein the prediction target is any one of a plurality of history predictions;
confirming a prediction vector corresponding to the prediction target according to the first load vector and the second load vector;
confirming a time-sharing characteristic vector corresponding to the current predicted load environment according to the prediction vectors corresponding to the plurality of historical predictions;
the identifying a first load vector corresponding to the work profile information of the plurality of target load terminals corresponding to the predicted targets and a second load vector corresponding to the predicted loads of the predicted targets includes:
confirming first predicted loads contained in the work profile information of the plurality of target load terminals;
inputting the first predicted load corresponding to the target load terminals into a trained first LSTM neural network deep learning model for processing according to the starting sequence of the target load terminals to obtain a first load vector corresponding to the first predicted load;
and inputting the predicted load of the predicted target into a trained second LSTM neural network deep learning model for processing to obtain a second load vector corresponding to the predicted load of the predicted target.
Further, the determining a prediction vector corresponding to the prediction target according to the first load vector and the second load vector includes:
respectively confirming weights of the first load vector and the second load vector through a first attention mechanism;
carrying out weighted summation processing on the first load vector and the second load vector according to the corresponding weights to obtain a prediction vector corresponding to the prediction target;
the step of confirming the time-period feature vector corresponding to the current predicted load environment according to the prediction vectors corresponding to the plurality of historical predictions comprises the following steps:
respectively inputting the prediction vectors corresponding to the plurality of historical predictions into a trained third LSTM neural network deep learning model for processing to obtain a plurality of processed prediction vectors;
confirming weights of the plurality of processed prediction vectors through a second attention mechanism;
and carrying out weighted summation processing on the plurality of processed prediction vectors according to the corresponding weights to obtain the time-period feature vector corresponding to the current prediction load environment.
Further, the validating the trust feature vector of the plurality of candidate predicted loads according to the time-sharing feature vector includes:
Inputting the time-division feature vectors and the candidate predictive loads into a trained vector inverse encoder model for processing to obtain trust feature vectors of the candidate predictive loads;
the extracting a plurality of candidate predicted loads matched with the current predicted load includes:
extracting a plurality of target load areas, wherein in a predicted load environment database, a predicted sequence corresponding to the target load areas is adjacent to a predicted sequence corresponding to the predicted load;
and confirming candidate predicted loads matched with the current predicted load from the target load areas according to the predicted frequencies of the target load areas.
Further, the validating the currently predicted trust prediction load from the plurality of candidate prediction loads according to the trust feature vector includes:
dividing the trust feature vector into a plurality of trust sub-feature vectors to extract the trust sub-feature vector features of each trust sub-feature vector;
generating sequence information through a significance module according to the trust sub-feature vector features of each trust sub-feature vector so as to obtain the semantic relevance of the trust sub-feature vector;
Coding each trust sub-feature vector and the time-sharing information thereof according to the semantic relevance of the trust sub-feature vector so as to obtain the time-sharing information feature; and
extracting trust characteristic value information, and obtaining characteristic expression information according to the trust characteristic value information and the time-period information characteristics to obtain a trust prediction load identification result;
the differentiating the trust feature vector into a plurality of trust sub-feature vectors further comprises:
normalizing all trust feature vectors in the historical prediction set and the to-be-predicted set to the same size;
dividing the normalized trust feature vector into a plurality of trust sub-feature vectors through a sliding window;
extracting the trust sub-feature vector features of each trust sub-feature vector by using a convolutional neural network;
the calculation formula of the relevance is as follows:
f＝relu(W v V+W h h t-1 +W e e t-1 )；
Figure BDA0004197139490000031
Figure BDA0004197139490000032
wherein f is the combination of trust sub-feature vector features V, LSTM neural network hidden layer features h and feature vector semantic features e, W through a perceptron v 、W h 、W e Is the parameter of the corresponding sensor, wf i Parameters of the perceptron representing the ith trust sub-feature vector, alpha represents the probability of the trust sub-feature vector predicted by the next time period module, alpha i Representing the probability of predicting the ith trust sub-feature vector in the next time period, L is the number of all trust sub-feature vectors, z is regarded as the probability feature of the trust sub-feature vector in the next time period, T is the sequence number, and by setting T iterations, the module outputs the sequence feature { z } 1 ,z 2 ,..z t }，z t Respectively with z t-1 And z t+1 Strong relevance is achieved at the semantic level;
the encoding of each trust sub-feature vector and its time-period information according to the semantic relevance of the trust sub-feature vector further comprises:
extracting characteristics with processing time sequence information by using an LSTM neural network;
adding a gate module, and filtering out the characteristic without distinguishing performance in the trust sub-feature vector by utilizing the LSTM neural network;
encoding the trust sub-feature vector and the time-period information by means of the LSTM neural network storage;
and extracting features of the whole feature vector by utilizing the convolutional neural network according to the trust feature value information, and extracting the time-sharing information features based on the trust sub-feature vector.
According to a second aspect of the present invention, the present invention claims an ultra-short term load prediction system based on deep learning, characterized in that the system comprises:
The first extraction module is used for responding to the current predicted operation in the current predicted load environment, extracting historical predicted loads of a plurality of historical predictions of the current predicted load environment and working profile information of a target load terminal corresponding to the historical predictions, wherein the target load terminal is a load terminal started in the historical predictions;
the generation module is used for generating a time-sharing characteristic vector corresponding to the current predicted load environment according to the work file information of the target load terminal and the historical predicted load;
the second extraction module is used for extracting a plurality of candidate predicted loads matched with the current predicted load;
the first confirming module is used for confirming the trust feature vectors of the candidate prediction loads according to the time-sharing feature vectors;
and the second confirming module is used for confirming the currently predicted trust prediction load from the plurality of candidate prediction loads according to the trust feature vector.
Further, the generating module includes:
a first confirming sub-module, configured to confirm a first load vector corresponding to the work profile information of a plurality of target load terminals corresponding to a prediction target and a second load vector corresponding to a predicted load of the prediction target, where the prediction target is any one of the plurality of history predictions;
A second confirming sub-module for confirming a prediction vector corresponding to the prediction target according to the first load vector and the second load vector;
a third confirming sub-module, configured to confirm a time-period feature vector corresponding to the current predicted load environment according to the prediction vectors corresponding to the plurality of historical predictions;
the first acknowledgement submodule is further configured to:
confirming first predicted loads contained in the work profile information of the plurality of target load terminals;
inputting the first predicted load corresponding to the target load terminals into a trained first LSTM neural network deep learning model for processing according to the starting sequence of the target load terminals to obtain a first load vector corresponding to the first predicted load;
and inputting the predicted load of the predicted target into a trained second LSTM neural network deep learning model for processing to obtain a second load vector corresponding to the predicted load of the predicted target.
Further, the second confirmation sub-module is further configured to:
respectively confirming weights of the first load vector and the second load vector through a first attention mechanism;
carrying out weighted summation processing on the first load vector and the second load vector according to the corresponding weights to obtain a prediction vector corresponding to the prediction target;
The third confirmation sub-module is further configured to:
respectively inputting the prediction vectors corresponding to the plurality of historical predictions into a trained third LSTM neural network deep learning model for processing to obtain a plurality of processed prediction vectors;
confirming weights of the plurality of processed prediction vectors through a second attention mechanism;
and carrying out weighted summation processing on the plurality of processed prediction vectors according to the corresponding weights to obtain the time-period feature vector corresponding to the current prediction load environment.
Further, the first confirmation module includes:
the processing submodule is used for inputting the time-sharing characteristic vector and the plurality of candidate predictive loads into a trained vector inverse encoder model for processing to obtain trust characteristic vectors of the plurality of candidate predictive loads;
the second extraction module comprises:
an extraction sub-module, configured to extract a plurality of target load areas, where in a predicted load environment database, a predicted order corresponding to the target load areas is adjacent to a predicted order corresponding to the predicted load;
and a fourth confirming sub-module for confirming candidate predicted loads matched with the current predicted load from the plurality of target load areas according to the predicted frequencies of the plurality of target load areas.
Further, the second confirming module is configured to confirm the currently predicted trust prediction load from the plurality of candidate prediction loads according to the trust feature vector, and specifically includes:
dividing the trust feature vector into a plurality of trust sub-feature vectors to extract the trust sub-feature vector features of each trust sub-feature vector;
generating sequence information through a significance module according to the trust sub-feature vector features of each trust sub-feature vector so as to obtain the semantic relevance of the trust sub-feature vector;
coding each trust sub-feature vector and the time-sharing information thereof according to the semantic relevance of the trust sub-feature vector so as to obtain the time-sharing information feature; and
and extracting trust characteristic value information, and obtaining characteristic expression information according to the trust characteristic value information and the time-period information characteristics to obtain a trust prediction load identification result.
The differentiating the trust feature vector into a plurality of trust sub-feature vectors further comprises:
normalizing all trust feature vectors in the historical prediction set and the to-be-predicted set to the same size;
dividing the normalized trust feature vector into a plurality of trust sub-feature vectors through a sliding window;
Extracting the trust sub-feature vector features of each trust sub-feature vector by using a convolutional neural network;
the calculation formula of the relevance is as follows:
f＝relu(W v V+W h h t-1 +W e e t-1 )；
Figure BDA0004197139490000061
Figure BDA0004197139490000062
wherein f is the combination of trust sub-feature vector features V, LSTM neural network hidden layer features h and feature vector semantic features e, W through a perceptron v 、W h 、W e Is the parameter of the corresponding sensor, wf i Parameters of the perceptron representing the ith trust sub-feature vector, alpha represents the probability of the trust sub-feature vector predicted by the next time period module, alpha i Representing the probability of predicting the ith trust sub-feature vector in the next time period, L is the number of all trust sub-feature vectors, z is regarded as the probability feature of the trust sub-feature vector in the next time period, T is the sequence number, and by setting T iterations, the module outputs the sequence feature { z } 1 ,z 2 ,..z t }，z t Respectively with z t-1 And z t+1 Strong association at semantic levelSex;
the encoding of each trust sub-feature vector and its time-period information according to the semantic relevance of the trust sub-feature vector further comprises:
extracting characteristics with processing time sequence information by using an LSTM neural network;
adding a gate module, and filtering out the characteristic without distinguishing performance in the trust sub-feature vector by utilizing the LSTM neural network;
Encoding the trust sub-feature vector and the time-period information by means of the LSTM neural network storage;
and extracting features of the whole feature vector by utilizing the convolutional neural network according to the trust feature value information, and extracting the time-sharing information features based on the trust sub-feature vector.
The invention requests protection of a deep learning-based ultra-short-term load prediction method and a deep learning-based ultra-short-term load prediction system, wherein a plurality of historical predicted historical load of a current predicted load environment and work archive information of a target load terminal corresponding to the historical predictions are extracted by responding to a current predicted prediction operation in the current predicted load environment, and the target load terminal is a load terminal started in the historical predictions; generating a time-sharing characteristic vector corresponding to the current predicted load environment according to the work file information of the target load terminal and the historical predicted load; extracting a plurality of candidate prediction loads matched with each other; and confirming the trust characteristic vectors of the plurality of candidate predicted loads according to the time-sharing characteristic vectors and confirming the currently predicted trust predicted load from the plurality of candidate predicted loads. The method can fully utilize the historical operation data of the load terminal, and accurately predict the short-term load level in the corresponding range according to the specific requirements on the basis.
Drawings
FIG. 1 is a workflow diagram of a deep learning-based ultra-short term load prediction method in accordance with the present invention;
FIG. 2 is a second workflow diagram of a deep learning-based ultra-short term load prediction method in accordance with the present invention;
FIG. 3 is a third workflow diagram of a deep learning-based ultra-short term load prediction method in accordance with the present invention;
FIG. 4 is a block diagram of a deep learning-based ultra-short term load prediction system according to the present invention;
fig. 5 is a second structural block diagram of an ultra-short term load prediction system based on deep learning according to the present invention.
Detailed Description
Various exemplary embodiments, features and aspects of the disclosure will be described in detail below with reference to the drawings. In the drawings, like reference numbers indicate identical or functionally similar elements. Although various aspects of the embodiments are illustrated in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated.
The term "exemplary" as used herein means "serving as an example, embodiment, or illustration. Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.
In addition, numerous specific details are set forth in the following detailed description in order to provide a better understanding of the present disclosure. It will be understood by those skilled in the art that the present disclosure may be practiced without some of these specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art have not been described in detail in order not to obscure the present disclosure.
Fig. 1 shows a flowchart of an ultra-short term load prediction method based on deep learning according to an embodiment of the present disclosure. As shown in fig. 1, the method may include the steps of:
and step S11, responding to the current predicted operation in the current predicted load environment, extracting historical predicted loads of a plurality of historical predictions of the current predicted load environment and working profile information of a target load terminal corresponding to the historical predictions, wherein the target load terminal is a load terminal started in the historical predictions.
Embodiments of the present disclosure may be applied to applying trust prediction loads, such as predictions, for example, in power systems (e.g., WEB servers). The power system may extract a historical predicted load environment in response to a current predicted operation in the current predicted load environment. For example, the power system may capture log information of the current power terminal, where each adjacent prediction is guaranteed to be within a preset period. By analyzing the log information, a plurality of history predictions are extracted from the log information, and a plurality of history prediction loads and a plurality of history prediction target load terminal work files can be extracted from the plurality of history predictions.
The current predicted load environment may be a set of predictions of a predicted time interval within a preset time interval, for example: the preset time interval is half an hour. The power system can ensure that each adjacent prediction in the extracted log information is within half an hour.
For example, the power system may extract 3 historical predictions from the extracted log information in response to a currently predicted prediction operation (e.g., a click operation on a control that triggers the power system to perform the prediction operation) corresponding to the predicted load "neural network. The 1 st historical predicted load is the local hospital power load, and the work file information of the target load terminal is respectively: "annual hospital cost summary", "county local hospital power load" and "local hospital power load ledger"; the 2 nd historical predicted load is the power load of the current province school, and the work file information of the target load terminal is respectively: "the number of schools in this province" and "the scale of schools in this province".
And step S12, generating a time-sharing characteristic vector corresponding to the current predicted load environment according to the work file information of the target load terminal and the historical predicted load.
The power system can respectively generate the work file information of the target load terminal and the load vector corresponding to the historical predicted load, and further can generate the time-sharing characteristic vector corresponding to the current predicted load environment according to the load vector.
And S13, extracting a plurality of candidate predicted loads matched with the current predicted load.
For example, the power system may extract candidate predicted loads adjacent to the predicted load corresponding to the current prediction from a predicted load environment database. The predicted load environment database may be used to store historical predictions created by all power terminals during a preset period of time.
And step S14, confirming the trust characteristic vectors of the candidate prediction loads according to the time-sharing characteristic vector.
The power system may confirm the trust feature vectors of the plurality of candidate predicted loads based on the above-described time-period feature vectors confirmed from the plurality of historically predicted target load terminals and the historically predicted loads.
In one possible implementation manner, the identifying trust feature vectors of the candidate prediction loads according to the time-division feature vectors may include the following steps:
And inputting the time-division feature vectors and the candidate predictive loads into a trained vector inverse encoder model for processing to obtain trust feature vectors of the candidate predictive loads.
Wherein the vector inverse encoder model may refer to the following formula.
Figure BDA0004197139490000081
Wherein ωn may represent the n-th candidate predicted load, and n is a positive integer. S (Q) may represent a trust feature vector of the candidate predicted load, S may represent all the history predicted information (history predicted load, target load terminal), p (ω) n |ω 1：n-1 S) may represent likelihood probabilities of candidate predicted loads ωn. Wherein p (omega) n |ω 1：n-1 S) can be obtained by the following formula.
p(ω n |ω 1：n-1 ,S)＝softmax(ω n f(d m,n-1 ,ω n-1 ))
Wherein f (d) m,n-1 ,ω n-1 ) Can be obtained by the following formula.
f(d m,n-1 ,ω n-1 )＝Hd m,n-1 +Eω n-1 +b 0 ；
Wherein m is a positive integer, dm, n-1 may represent a load vector corresponding to the candidate predicted load ωn-1 in the mth prediction, H, E and b 0 Are all constant. The load vector corresponding to the candidate predicted load ωn can be obtained by the following formula.
d m,n ＝GRU dec (d m,n-1 ,ω n )；
Where dm, n may represent a load vector corresponding to the candidate predicted load ωn in the mth prediction.
When n is 1, dm,0 can be obtained by the following formula.
d m, ＝DS m +b 0 ；
Wherein Sm may represent a time-sharing feature vector corresponding to the current predicted load environment, and D is a constant.
In this way, the power system may be able to derive a trust score for each candidate predicted load, and the trust score for each candidate predicted load may be maximized during the training process.
And step S15, confirming the currently predicted trust prediction load from the plurality of candidate prediction loads according to the trust feature vector.
The power system may determine, from among the plurality of candidate predicted loads, a currently predicted trust predicted load based on the trust feature vectors of the plurality of candidate predicted loads, for example, determine that a candidate predicted load having a larger trust feature vector is a trust predicted load.
For example, the power system may confirm that the candidate predicted load with the top N-bit trust feature vector is the trust predicted load, where N is the number of preset trust predicted loads. For example: the number of preset candidate predicted loads is 9, the power system can extract the candidate predicted loads with the trust feature vector arranged in the first 9 bits from the candidate predicted loads, confirm that the 9 candidate predicted loads are the trust predicted loads, and display the 9 candidate predicted loads in a display interface when the display interface corresponding to the current predicted load environment is displayed.
Therefore, the power system can combine the work file information of the target load terminal started in the history prediction and the history prediction load to establish the time-sharing characteristic vector of the current prediction load environment. After extracting a plurality of candidate predicted loads matched according to the predicted load currently predicted, the power system may confirm trust feature vectors of the plurality of candidate predicted loads according to the time-period feature vectors of the current predicted load environment, and may confirm the trust predicted load from the candidate predicted loads according to the trust feature vectors. Compared with the problem that the reliability of the obtained trust prediction content is not high in the prior art by analyzing the trust prediction load through the history prediction load, the ultra-short-term load prediction method based on the deep learning provided by the embodiment of the invention can be used for confirming the trust prediction load according to the history prediction of the current prediction load environment and the working archive information of the load terminal started in the history prediction.
Fig. 2 shows a flowchart of a deep learning-based ultra-short term load prediction method according to an embodiment of the present disclosure.
In one possible implementation manner, referring to fig. 2, the step S12, generating a time-period feature vector corresponding to the current predicted load environment according to the work profile information of the target load terminal and the historical predicted load may include the following steps:
step S121, a first load vector corresponding to the work profile information of a plurality of target load terminals corresponding to a prediction target and a second load vector corresponding to the predicted load of the prediction target are confirmed, wherein the prediction target is any one of the plurality of history predictions.
The power system may confirm the first load vector corresponding to the work profile information of the plurality of target load terminals corresponding to each of the history predictions, and the second load vector corresponding to the history prediction load of each of the history predictions.
In one possible implementation manner, the step S121 may include:
and confirming first predicted loads contained in the work profile information of the target load terminals.
The power system can split the work file information of each target load terminal of the prediction target, and obtain first prediction loads corresponding to the target load terminals after removing invalid loads in the work file information. The dead load may include automatically filtering out the load as confirmed prior to processing the work profile information of the target load terminal in order to improve trust efficiency.
The step S121 may include: and inputting the first predicted load corresponding to the target load terminals into a trained first LSTM neural network deep learning model for processing according to the starting sequence of the target load terminals, so as to obtain a first load vector corresponding to the first predicted load.
The step S121 may include: and inputting the predicted load of the predicted target into a trained second LSTM neural network deep learning model for processing to obtain a second load vector corresponding to the predicted load of the predicted target.
It should be noted that, the last output of the first LSTM neural network deep learning model may be used as the first input of the second LSTM neural network deep learning model. Assume that the last output of the first LSTM neural network deep learning model is the first input in the second LSTM neural network deep learning model.
It should be noted that, the embodiments of the present disclosure use a simple feature method to represent the original input and directly learn to the optimization objective. In this disclosure, each load's most primitive input is represented by a load vector, the load vector length being equal to the size of the entire corpus (the prediction library is used to store the historical predicted load). All load vectors are mapped into a low-dimensional representation vector by a mapping matrix. The mapping matrix is used as a model learning parameter, is continuously adjusted in the training process, and finally learns a first LSTM neural network deep learning model, a second LSTM neural network deep learning model and a third LSTM neural network deep learning model. At the output end, the embodiment of the disclosure takes the m < th > prediction as a learning target of the previous m-1 historical predictions, so that the LSTM neural network deep learning model can maximize the output prediction target when seeing the previous historical predictions.
In this embodiment, a first LSTM neural network deep learning model may be used to generate a first advanced feature set from the preprocessed target load image. Likewise, a second LSTM neural network deep learning model may be used to generate a second advanced feature set from the preprocessed reference load image. Wherein the first LSTM neural network deep learning model and the second LSTM neural network deep learning model may enable an abstract description of the target load image and the reference load image by, for example, combining multiple layers of low-level features (pixel-level features). Here, the advanced features merely indicate the advanced nature of the features with respect to the primary features (e.g., pixel-level features) of the original image after the processing by the artificial neural network, but generally speaking, the higher the level and the more abstract the trend is as the neural network goes deeper through the neural network processing. In addition, a feature set is generally defined as including two or more features, and may be referred to as a "feature matrix" in the present invention. In addition, in some special cases, a feature set may have only 1 feature, such as an intermediate result, where a "feature set" may refer to only a single "feature".
In this embodiment, the third LSTM neural network deep learning model may be used to generate a determination result of the load change according to the result of feature fusion (feature combination set). The third LSTM neural network deep learning model may form a determination result for the input target load image based on the result obtained by the feature combination. That is, the third LSTM neural network deep learning model generates a determination result of the load change according to the feature combination set. In this embodiment, the output dimension of the third LSTM neural network deep learning model coincides with the category to be classified (e.g., load change type). That is, for example, when the category to be classified is two categories of overload and reasonable load, the output dimension of the third LSTM neural network deep learning model may be 2; the third LSTM neural network deep learning model may have an output dimension of 6 if the class to be classified is overloaded and rationalized (e.g., 5). In addition, the output dimension of the third LSTM neural network deep learning model can be adjusted according to actual conditions.
Step S122, confirming a prediction vector corresponding to the prediction target based on the first load vector and the second load vector.
After the power system confirms the first load vector and the second load vector, the first load vector and the second load vector can be weighted and summed to obtain a prediction vector corresponding to the prediction target. In one possible implementation manner, the step S122 may include: the weights of the first load vector and the second load vector are respectively confirmed through a first attention mechanism.
The step S122 may further include: and carrying out weighted summation processing on the first load vector and the second load vector according to the corresponding weights to obtain a prediction vector corresponding to the prediction target.
Step S123, confirming a time-sharing characteristic vector corresponding to the current predicted load environment according to the prediction vectors corresponding to the plurality of historical predictions.
After obtaining the prediction vectors corresponding to the plurality of historical predictions, the power system may confirm the time-sharing feature vector corresponding to the current predicted load environment according to the plurality of prediction vectors.
In one possible implementation manner, the step S123 may include: and respectively inputting the prediction vectors corresponding to the plurality of historical predictions into a trained third LSTM neural network deep learning model for processing to obtain a plurality of processed prediction vectors.
The attention mechanism is actually a weight self-learning method, and since the historical predicted load used in the embodiment of the disclosure contains many unimportant loads in the first predicted load in the click load terminal information, for one historical prediction, not all the predicted loads play the same role, so that it is necessary to perform weighted representation processing on the predicted loads. Embodiments of the present disclosure apply the attention mechanism to both load and prediction. The weighting representation strategy can be learned in training, and different weighting representations are given to the new predictions. After the weights are obtained, weighting all the load vectors, more accurate vector representation can be obtained.
In this way, in the embodiment of the disclosure, the importance weight can be effectively calculated from the predicted load level and the predicted load environment level by adopting the attention mechanism, the time-division feature vector of the current predicted load environment is better represented, the behavior information of the power terminal in the prediction operation process can be more carefully utilized, and the important part in the behavior information can be accurately extracted, so that the historical information for analyzing the trust predicted load is enriched, and the accuracy of the trust predicted load can be improved.
Fig. 3 shows a flowchart of a deep learning-based ultra-short term load prediction method according to an embodiment of the present disclosure.
In a possible implementation manner, referring to fig. 3, the step S13, extracting a plurality of candidate predicted loads that match the current predicted load may include the following steps.
And step S131, extracting a plurality of target load areas, wherein in a predicted load environment database, the predicted sequence corresponding to the target load areas is adjacent to the predicted sequence corresponding to the predicted load.
For example, the above-described predicted load environment database may be used to store historical predicted operation information of all power terminals, which may include relevant information of historical predictions. The power system may search a predicted load corresponding to the predicted load in a predicted load environment database according to the current predicted load, and confirm the predicted load as the first predicted load. And searching second predictions with the prediction orders adjacent to the first prediction orders in a predicted load environment database, and confirming that predicted loads corresponding to the adjacent second predictions are target load areas.
Step S132, according to the predicted frequencies of the target load areas, confirming candidate predicted loads matched with the current predicted load from the target load areas.
The power system may confirm the predicted frequency of each target load region described above, which may be used to represent the ratio of the number of target load regions in all target load regions, for example: there are 10 target load regions, of which 5 target load regions are identical, and are each "each region load use", so the predicted frequency of "each region load use" is 50%. The power system may confirm the candidate predicted load from the plurality of target load regions according to the predicted frequency of the target load region.
For example, if the preset number of candidate predicted loads is M, the power system may confirm that M target load areas, of which the magnitude of the predicted frequency is arranged in the first M bits, among the target load areas are candidate predicted loads.
In this way, the power system can confirm the candidate predicted load from the target load area, and calculate the trust feature vector of the candidate predicted load, so as to confirm the trust predicted load from the candidate predicted load according to the trust feature vector, thereby reducing the calculation amount of the power system, accelerating the calculation efficiency of the power system, and further improving the trust efficiency of the trust predicted load.
The embodiment of the disclosure utilizes an LSTM neural network in deep learning to construct a set of ultra-short-term load prediction model based on deep learning. The model is added with the work archive information of the log load terminal clicked by the power terminal on the basis of the predicted load which is only predicted previously by the power terminal. The information is used as auxiliary information for current prediction of the power terminal, and prediction content which meets the prediction requirement of the power terminal can be more effectively trusted for the power terminal in prediction trust. The embodiment of the disclosure can also effectively calculate the importance weight from the load level and the prediction level by using the attention mechanism, so that the time-sharing information vector can be better represented. Compared with the existing prediction trust method, the method and the device have the advantages that the power terminal behavior information in the search is utilized more carefully, important parts in the power terminal behavior information are extracted, and the prediction trust is realized more accurately.
Fig. 4 shows a schematic structural diagram of an ultra-short term load prediction system based on deep learning according to an embodiment of the present disclosure. As shown in fig. 4, the system may include:
the first extraction module 41 may be configured to extract, in response to a current predicted operation in a current predicted load environment, a plurality of historical predicted historic predicted loads of the current predicted load environment and work profile information of a target load terminal corresponding to the plurality of historic predictions, where the target load terminal is a load terminal started in the historic predictions;
the generating module 42 may be configured to generate a time-period feature vector corresponding to a current predicted load environment according to the work profile information of the target load terminal and the historical predicted load;
a second extraction module 43, configured to extract a plurality of candidate predicted loads that match the current predicted load;
a first confirmation module 44, configured to confirm the trust feature vectors of the plurality of candidate predicted loads according to the time-division feature vectors;
the second confirming module 45 may be configured to confirm the currently predicted trust prediction load from the plurality of candidate prediction loads according to the trust feature vector.
Therefore, the power system can combine the work file information of the target load terminal started in the history prediction and the history prediction load to establish the time-sharing characteristic vector of the current prediction load environment. After extracting a plurality of candidate predicted loads matched according to the predicted load currently predicted, the power system may confirm trust feature vectors of the plurality of candidate predicted loads according to the time-period feature vectors of the current predicted load environment, and may confirm the trust predicted load from the candidate predicted loads according to the trust feature vectors. Compared with the problem that the reliability of the obtained trust prediction content is not high in the prior art by analyzing the trust prediction load through the history prediction load, the deep learning-based ultra-short-term load prediction system provided by the embodiment of the invention can confirm the trust prediction load according to the history prediction of the current prediction load environment and the working archive information of the load terminal started in the history prediction.
Fig. 5 shows a schematic structural diagram of an ultra-short term load prediction system based on deep learning according to an embodiment of the present disclosure.
In one possible implementation, as shown in fig. 5, the generating module 42 may include:
the first confirmation sub-module 421 may be configured to confirm a first load vector corresponding to the work profile information of a plurality of target load terminals corresponding to a prediction target and a second load vector corresponding to a predicted load of the prediction target, where the prediction target is any one of the plurality of history predictions;
a second confirming sub-module 422, configured to confirm a prediction vector corresponding to the prediction target according to the first load vector and the second load vector;
the third confirming sub-module 423 may be configured to confirm the time-sharing feature vector corresponding to the current predicted load environment according to the prediction vectors corresponding to the plurality of historical predictions.
In one possible implementation manner, the first acknowledgement submodule 421 may be further configured to:
confirming first predicted loads contained in the work profile information of the plurality of target load terminals;
inputting the first predicted load corresponding to the target load terminals into a trained first LSTM neural network deep learning model for processing according to the starting sequence of the target load terminals to obtain a first load vector corresponding to the first predicted load;
And inputting the predicted load of the predicted target into a trained second LSTM neural network deep learning model for processing to obtain a second load vector corresponding to the predicted load of the predicted target.
In one possible implementation, the second acknowledgement submodule 422 may also be configured to:
respectively confirming weights of the first load vector and the second load vector through a first attention mechanism;
and carrying out weighted summation processing on the first load vector and the second load vector according to the corresponding weights to obtain a prediction vector corresponding to the prediction target.
In one possible implementation manner, the third acknowledgement submodule 423 may be further configured to:
respectively inputting the prediction vectors corresponding to the plurality of historical predictions into a trained third LSTM neural network deep learning model for processing to obtain a plurality of processed prediction vectors;
confirming weights of the plurality of processed prediction vectors through a second attention mechanism;
and carrying out weighted summation processing on the plurality of processed prediction vectors according to the corresponding weights to obtain the time-period feature vector corresponding to the current prediction load environment.
In one possible implementation, the first confirmation module 44 may include:
The processing submodule 441 may be configured to input the time-division feature vector and the plurality of candidate prediction loads into a trained vector inverse encoder model to process the time-division feature vector and the plurality of candidate prediction loads, so as to obtain trust feature vectors of the plurality of candidate prediction loads.
In one possible implementation, the second extraction module 43 may include:
an extraction sub-module 431, configured to extract a plurality of target load areas, where in the predicted load environment database, a predicted order corresponding to the target load areas is adjacent to a predicted order corresponding to the predicted load;
the fourth confirming sub-module 432 may be configured to confirm, from the plurality of target load areas, a candidate predicted load matching the currently predicted load according to the predicted frequencies of the plurality of target load areas.
The trust prediction load identification method based on the feature vector time-sharing information in one embodiment of the invention comprises the following steps:
in step S601, the trust feature vector is divided into a plurality of trust sub-feature vectors to extract the trust sub-feature vector features of each trust sub-feature vector.
Further, in an embodiment of the present invention, differentiating the trust feature vector into the plurality of trust sub-feature vectors may further include: normalizing all trust feature vectors in the historical prediction set and the to-be-predicted set to the same size; dividing the normalized trust feature vector into a plurality of trust sub-feature vectors through a sliding window; and extracting the trust sub-feature vector features of each trust sub-feature vector by using a convolutional neural network.
For example, all feature vectors in a given historical prediction set and a set to be predicted are normalized to the same size, then the feature vectors are divided into a plurality of trust sub-feature vectors through a sliding window, and the feature of each trust sub-feature vector is extracted by using a convolutional neural network.
In step S602, sequence information is generated by a saliency module according to the trust sub-feature vector feature of each trust sub-feature vector, so as to obtain the semantic relevance of the trust sub-feature vector.
Further, in one embodiment of the present invention, the calculation formula of the correlation is:
f＝relu(W v V+W h h t-1 +W e e t-1 )；
Figure BDA0004197139490000151
Figure BDA0004197139490000152
wherein f is the combination of trust sub-feature vector features V, LSTM neural network hidden layer features h and feature vector semantic features e, W through a perceptron v 、W h 、W e Is the parameter of the corresponding sensor, wf i Parameters of the perceptron representing the ith trust sub-feature vector, alpha represents the probability of the trust sub-feature vector predicted by the next time period module, alpha i Representing the probability of predicting the ith trust sub-feature vector in the next time period, L is the number of all trust sub-feature vectors, z is regarded as the probability feature of the trust sub-feature vector in the next time period, T is the sequence number, and by setting T iterations, the module outputs the sequence feature { z } 1 ,z 2 ,..z t }，z t Respectively with z t-1 And z t+1 Strong relevance is achieved on the semantic level.
In particular, by outputting a series of trust sub-feature vector features V, these are geometrically related, but semantically may not have a relevance. And designing a significance module to generate sequence information, wherein each feature in the sequence has strong semantic relevance with adjacent elements. The module finds the feature e most relevant to the current trust sub-feature vector feature, such as surrounding information or pair appearance in certain trust prediction loads, based on the multi-layer perceptron structure with the hidden layer feature ht-1 in the trust sub-feature vector features V, S2 and the time-division feature et-1 as inputs. The following is shown:
f＝relu(W v V+W h h t-1 +W e e t-1 )；
Figure BDA0004197139490000153
Figure BDA0004197139490000154
wherein f is the combination of trust sub-feature vector feature V, LSTM neural network hidden layer feature h and feature vector through a perceptronSemantic features e, W v 、W h 、W e Is the parameter of the corresponding sensor, wf i Parameters of the perceptron representing the ith trust sub-feature vector, alpha represents the probability of the trust sub-feature vector predicted by the next time period module, alpha i Representing the probability of predicting the ith trust sub-feature vector in the next time period, L is the number of all trust sub-feature vectors, z is regarded as the probability feature of the trust sub-feature vector in the next time period, T is the sequence number, and by setting T iterations, the module outputs the sequence feature { z } 1 ,z 2 ,..z t }，z t Respectively with z T-1 And z T+1 Strong relevance is achieved on the semantic level.
In step S603, each trust sub-feature vector and its time-sharing information are encoded according to the semantic relevance of the trust sub-feature vector to obtain the time-sharing information feature.
Further, in one embodiment of the present invention, the encoding of each trust sub-feature vector and its time-period information according to the semantic relevance of the trust sub-feature vector may further include: extracting characteristics with processing time sequence information by using an LSTM neural network; adding a gate module, filtering out the characteristic that the trust sub-feature vector does not have discrimination performance by using an LSTM neural network; the trust sub-feature vector and the time-period information are encoded by means of LSTM neural network storage.
It can be understood that, according to the step S1 in the step S602, the correlation of the trust sub-feature vector in terms of semantics is found, in order to combine the trust sub-feature vector with the time-division information, the characteristic of processing time sequence information is utilized by using the LSTM neural network, and the gate module is added, so that the improved LSTM neural network filters out the feature of the trust sub-feature vector, which does not have the distinguishing performance, on the one hand, and meanwhile, encodes the trust sub-feature vector and the time-division information thereof by means of the capability of the LSTM neural network to store the information.
In step S604, trust feature value information is extracted, and feature expression information is obtained according to the trust feature value information and the time-sharing information feature, so as to obtain a trust prediction load identification result.
Further, in one embodiment of the present invention, the trust eigenvalue information utilizes a convolutional neural network to perform feature extraction on the whole eigenvector, and the time-division information features are extracted based on trust sub-eigenvectors.
It can be understood that the segmentation period information features are extracted based on the trust sub-feature vector features according to the steps S101, S102 and S103, and the feature vector trust feature value information and the partial period information are fused to generate the final feature expression.
The following details are details of the advantages of the trust prediction load identification method based on feature vector time-sharing information according to the embodiments of the present invention, which are specifically as follows:
1. the embodiment of the invention combines the trust characteristic value information of the trust prediction load characteristic vector with the local time-sharing information, and can utilize various information in the trust prediction load to improve the characteristic representation of the network on the trust prediction load.
2. When capturing time-division information of the feature vector, the embodiment of the invention utilizes the saliency module to find out the feature closely related to the trust sub-feature vector, generates sequence information, and each element (trust sub-feature vector feature) in the sequence has strong relevance to the adjacent element, can reflect the semantic relevance of the trust sub-feature vector feature, and is convenient for enhancing the feature vector characterization capability.
3. According to the embodiment of the invention, the LSTM neural network is improved by adding the gate module, on one hand, the trust sub-feature vector features without distinguishing performance are removed, and meanwhile, the trust sub-feature vector features with distinguishing capability and the time-division information thereof are encoded, so that the distinguishing capability of the feature vector features can be improved, and the time-division information of the feature vector can be captured.
According to the trust prediction load identification method based on the feature vector time-sharing information, the feature vector trust feature value information and the local time-sharing information are fused, the feature vector time-sharing information is fully utilized, the object features and the time-sharing features of the object features in the trust prediction load are combined, the expression capacity of the network to the trust prediction load is effectively improved by utilizing the multiple aspects of features, and the performance is greatly improved.
Next, a trust prediction load recognition system based on feature vector time-sharing information according to an embodiment of the present invention will be described with reference to the accompanying drawings.
Unless defined otherwise, all terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
The foregoing is illustrative of the present disclosure and is not to be construed as limiting thereof. Although a few exemplary embodiments of this disclosure have been described, those skilled in the art will readily appreciate that many modifications are possible in the exemplary embodiments without materially departing from the novel teachings and advantages of this disclosure. Accordingly, all such modifications are intended to be included within the scope of this disclosure as defined in the claims. It is to be understood that the foregoing is illustrative of the present disclosure and is not to be construed as limited to the specific embodiments disclosed, and that modifications to the disclosed embodiments, as well as other embodiments, are intended to be included within the scope of the appended claims. The disclosure is defined by the claims and their equivalents.
In the description of the present specification, reference to the terms "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, 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, schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
While embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that: many changes, modifications, substitutions and variations may be made to the embodiments without departing from the spirit and principles of the invention, the scope of which is defined by the claims and their equivalents.

Claims (10)

1. An ultra-short term load prediction method based on deep learning, which is characterized by comprising the following steps:
responding to the current predicted operation in the current predicted load environment, extracting historical predicted loads of a plurality of historical predictions of the current predicted load environment and work archive information of a target load terminal corresponding to the historical predictions, wherein the target load terminal is a load terminal started in the historical predictions;
generating a time-sharing characteristic vector corresponding to the current predicted load environment according to the work file information of the target load terminal and the historical predicted load;
extracting a plurality of candidate predicted loads matched with the current predicted load;
confirming trust feature vectors of the candidate predictive loads according to the time-division feature vectors;
and confirming the currently predicted trust prediction load from the candidate prediction loads according to the trust feature vector.
2. The method according to claim 1, wherein the generating the time-sharing feature vector corresponding to the current predicted load environment according to the work profile information of the target load terminal and the historical predicted load includes:
confirming a first load vector corresponding to the work file information of a plurality of target load terminals corresponding to a prediction target and a second load vector corresponding to the predicted load of the prediction target, wherein the prediction target is any one of a plurality of history predictions;
confirming a prediction vector corresponding to the prediction target according to the first load vector and the second load vector;
confirming a time-sharing characteristic vector corresponding to the current predicted load environment according to the prediction vectors corresponding to the plurality of historical predictions;
the identifying a first load vector corresponding to the work profile information of the plurality of target load terminals corresponding to the predicted targets and a second load vector corresponding to the predicted loads of the predicted targets includes:
confirming first predicted loads contained in the work profile information of the plurality of target load terminals;
inputting the first predicted load corresponding to the target load terminals into a trained first LSTM neural network deep learning model for processing according to the starting sequence of the target load terminals to obtain a first load vector corresponding to the first predicted load;
And inputting the predicted load of the predicted target into a trained second LSTM neural network deep learning model for processing to obtain a second load vector corresponding to the predicted load of the predicted target.
3. The method of claim 2, wherein the identifying a prediction vector corresponding to the prediction target based on the first load vector and the second load vector comprises:
respectively confirming weights of the first load vector and the second load vector through a first attention mechanism;
carrying out weighted summation processing on the first load vector and the second load vector according to the corresponding weights to obtain a prediction vector corresponding to the prediction target;
the step of confirming the time-period feature vector corresponding to the current predicted load environment according to the prediction vectors corresponding to the plurality of historical predictions comprises the following steps:
respectively inputting the prediction vectors corresponding to the plurality of historical predictions into a trained third LSTM neural network deep learning model for processing to obtain a plurality of processed prediction vectors;
confirming weights of the plurality of processed prediction vectors through a second attention mechanism;
and carrying out weighted summation processing on the plurality of processed prediction vectors according to the corresponding weights to obtain the time-period feature vector corresponding to the current prediction load environment.
4. A method according to claim 3, wherein said validating the trust feature vector of the plurality of candidate predicted loads from the time-lapse feature vector comprises:
inputting the time-division feature vectors and the candidate predictive loads into a trained vector inverse encoder model for processing to obtain trust feature vectors of the candidate predictive loads;
the extracting a plurality of candidate predicted loads matched with the current predicted load includes:
extracting a plurality of target load areas, wherein in a predicted load environment database, a predicted sequence corresponding to the target load areas is adjacent to a predicted sequence corresponding to the predicted load;
and confirming candidate predicted loads matched with the current predicted load from the target load areas according to the predicted frequencies of the target load areas.
5. The method of claim 4, wherein said validating said currently predicted trust prediction load from said plurality of candidate prediction loads based on said trust feature vector comprises:
dividing the trust feature vector into a plurality of trust sub-feature vectors to extract the trust sub-feature vector features of each trust sub-feature vector;
Generating sequence information through a significance module according to the trust sub-feature vector features of each trust sub-feature vector so as to obtain the semantic relevance of the trust sub-feature vector;
coding each trust sub-feature vector and the time-sharing information thereof according to the semantic relevance of the trust sub-feature vector so as to obtain the time-sharing information feature; and
extracting trust characteristic value information, and obtaining characteristic expression information according to the trust characteristic value information and the time-period information characteristics to obtain a trust prediction load identification result;
the differentiating the trust feature vector into a plurality of trust sub-feature vectors further comprises:
normalizing all trust feature vectors in the historical prediction set and the to-be-predicted set to the same size;
dividing the normalized trust feature vector into a plurality of trust sub-feature vectors through a sliding window;
extracting the trust sub-feature vector features of each trust sub-feature vector by using a convolutional neural network;
the calculation formula of the relevance is as follows:
f＝relu(W v V+W h h t-1 +W e e t-1 )；
Figure FDA0004197139470000031
Figure FDA0004197139470000032
wherein f is the combination of trust sub-feature vector features V, LSTM neural network hidden layer features h and feature vector semantic features e, W through a perceptron v 、W h 、W e Is the parameter of the corresponding sensor, wf i Parameters of the perceptron representing the ith trust sub-feature vector, alpha represents the probability of the trust sub-feature vector predicted by the next time period module, alpha i Representing the probability of predicting the ith trust sub-feature vector in the next time period, L is the number of all trust sub-feature vectors, z is regarded as the probability feature of the trust sub-feature vector in the next time period, T is the sequence number, and by setting T iterations, the module outputs the sequence feature { z } 1 ,z 2 ,..z t }，z t Respectively with z t-1 And z t+1 Strong relevance is achieved at the semantic level;
the encoding of each trust sub-feature vector and its time-period information according to the semantic relevance of the trust sub-feature vector further comprises:
extracting characteristics with processing time sequence information by using an LSTM neural network;
adding a gate module, and filtering out the characteristic without distinguishing performance in the trust sub-feature vector by utilizing the LSTM neural network; encoding the trust sub-feature vector and the time-period information by means of the LSTM neural network storage;
and extracting features of the whole feature vector by utilizing the convolutional neural network according to the trust feature value information, and extracting the time-sharing information features based on the trust sub-feature vector.
6. An ultra-short term load prediction system based on deep learning, the system comprising:
the first extraction module is used for responding to the current predicted operation in the current predicted load environment, extracting historical predicted loads of a plurality of historical predictions of the current predicted load environment and working profile information of a target load terminal corresponding to the historical predictions, wherein the target load terminal is a load terminal started in the historical predictions;
the generation module is used for generating a time-sharing characteristic vector corresponding to the current predicted load environment according to the work file information of the target load terminal and the historical predicted load;
the second extraction module is used for extracting a plurality of candidate predicted loads matched with the current predicted load;
the first confirming module is used for confirming the trust feature vectors of the candidate prediction loads according to the time-sharing feature vectors;
and the second confirming module is used for confirming the currently predicted trust prediction load from the plurality of candidate prediction loads according to the trust feature vector.
7. The system of claim 6, wherein the generating module comprises:
A first confirming sub-module, configured to confirm a first load vector corresponding to the work profile information of a plurality of target load terminals corresponding to a prediction target and a second load vector corresponding to a predicted load of the prediction target, where the prediction target is any one of the plurality of history predictions;
a second confirming sub-module for confirming a prediction vector corresponding to the prediction target according to the first load vector and the second load vector;
a third confirming sub-module, configured to confirm a time-period feature vector corresponding to the current predicted load environment according to the prediction vectors corresponding to the plurality of historical predictions;
the first acknowledgement submodule is further configured to:
confirming first predicted loads contained in the work profile information of the plurality of target load terminals;
inputting the first predicted load corresponding to the target load terminals into a trained first LSTM neural network deep learning model for processing according to the starting sequence of the target load terminals to obtain a first load vector corresponding to the first predicted load;
and inputting the predicted load of the predicted target into a trained second LSTM neural network deep learning model for processing to obtain a second load vector corresponding to the predicted load of the predicted target.
8. The system of claim 7, wherein the second validation sub-module is further configured to:
respectively confirming weights of the first load vector and the second load vector through a first attention mechanism;
carrying out weighted summation processing on the first load vector and the second load vector according to the corresponding weights to obtain a prediction vector corresponding to the prediction target;
the third confirmation sub-module is further configured to:
respectively inputting the prediction vectors corresponding to the plurality of historical predictions into a trained third LSTM neural network deep learning model for processing to obtain a plurality of processed prediction vectors;
confirming weights of the plurality of processed prediction vectors through a second attention mechanism;
and carrying out weighted summation processing on the plurality of processed prediction vectors according to the corresponding weights to obtain the time-period feature vector corresponding to the current prediction load environment.
9. The system of claim 8, wherein the first confirmation module comprises:
the processing submodule is used for inputting the time-sharing characteristic vector and the plurality of candidate predictive loads into a trained vector inverse encoder model for processing to obtain trust characteristic vectors of the plurality of candidate predictive loads;
The second extraction module comprises:
an extraction sub-module, configured to extract a plurality of target load areas, where in a predicted load environment database, a predicted order corresponding to the target load areas is adjacent to a predicted order corresponding to the predicted load;
and a fourth confirming sub-module for confirming candidate predicted loads matched with the current predicted load from the plurality of target load areas according to the predicted frequencies of the plurality of target load areas.
10. The system of claim 9, wherein the second validating module is configured to validate the currently predicted trust prediction load from the plurality of candidate prediction loads based on the trust feature vector, and specifically comprises: dividing the trust feature vector into a plurality of trust sub-feature vectors to extract the trust sub-feature vector features of each trust sub-feature vector;
generating sequence information through a significance module according to the trust sub-feature vector features of each trust sub-feature vector so as to obtain the semantic relevance of the trust sub-feature vector;
coding each trust sub-feature vector and the time-sharing information thereof according to the semantic relevance of the trust sub-feature vector so as to obtain the time-sharing information feature; and
Extracting trust characteristic value information, and obtaining characteristic expression information according to the trust characteristic value information and the time-period information characteristics to obtain a trust prediction load identification result;
the differentiating the trust feature vector into a plurality of trust sub-feature vectors further comprises:
normalizing all trust feature vectors in the historical prediction set and the to-be-predicted set to the same size;
dividing the normalized trust feature vector into a plurality of trust sub-feature vectors through a sliding window;
extracting the trust sub-feature vector features of each trust sub-feature vector by using a convolutional neural network;
the calculation formula of the relevance is as follows:
f＝relu(W v V+W h h t-1 +W e e t-1 )；
Figure FDA0004197139470000051
Figure FDA0004197139470000052
wherein f is the combination of trust sub-feature vector features V, LSTM neural network hidden layer features h and feature vector semantic features e, W through a perceptron v 、W h 、W e Is the parameter of the corresponding sensor, wf i Parameters of the perceptron representing the ith trust sub-feature vector, alpha represents the probability of the trust sub-feature vector predicted by the next time period module, alpha i Predict ith on behalf of next periodThe probability of each trust sub-feature vector, L is the number of all trust sub-feature vectors, z is regarded as the probability feature of the trust sub-feature vector of the next time period, T is the sequence number, and by setting T iterations, the module outputs the sequence feature { z } 1 ,z 2 ,..z t }，z t Respectively with z t-1 And z t+1 Strong relevance is achieved at the semantic level;
the encoding of each trust sub-feature vector and its time-period information according to the semantic relevance of the trust sub-feature vector further comprises:
extracting characteristics with processing time sequence information by using an LSTM neural network;
adding a gate module, and filtering out the characteristic without distinguishing performance in the trust sub-feature vector by utilizing the LSTM neural network; encoding the trust sub-feature vector and the time-period information by means of the LSTM neural network storage;
and extracting features of the whole feature vector by utilizing the convolutional neural network according to the trust feature value information, and extracting the time-sharing information features based on the trust sub-feature vector.
CN202310449828.4A 2023-04-18 2023-04-18 Ultra-short term load prediction method and system based on deep learning Active CN116415734B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310449828.4A CN116415734B (en) 2023-04-18 2023-04-18 Ultra-short term load prediction method and system based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310449828.4A CN116415734B (en) 2023-04-18 2023-04-18 Ultra-short term load prediction method and system based on deep learning

Publications (2)

Publication Number Publication Date
CN116415734A true CN116415734A (en) 2023-07-11
CN116415734B CN116415734B (en) 2024-07-12

Family

ID=87054475

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310449828.4A Active CN116415734B (en) 2023-04-18 2023-04-18 Ultra-short term load prediction method and system based on deep learning

Country Status (1)

Country Link
CN (1) CN116415734B (en)

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040249775A1 (en) * 2003-05-13 2004-12-09 Dingguo Chen Very short term load prediction in an energy management system
CN110119826A (en) * 2018-02-06 2019-08-13 天津职业技术师范大学 A kind of power-system short-term load forecasting method based on deep learning
CN111931989A (en) * 2020-07-10 2020-11-13 国网浙江省电力有限公司绍兴供电公司 Power system short-term load prediction method based on deep learning neural network
CN113205207A (en) * 2021-04-19 2021-08-03 深圳供电局有限公司 XGboost algorithm-based short-term power consumption load fluctuation prediction method and system
CN113222279A (en) * 2021-05-28 2021-08-06 国网浙江桐乡市供电有限公司 Short-term load prediction method considering demand response
CN114091758A (en) * 2021-11-23 2022-02-25 四川信天启智能技术有限公司 Medium-short term load prediction method, system and storage medium
CN114169625A (en) * 2021-12-13 2022-03-11 国网江苏省电力有限公司经济技术研究院 Short-term load prediction method and device for power system
CN114444660A (en) * 2022-01-24 2022-05-06 南京电力设计研究院有限公司 Short-term power load prediction method based on attention mechanism and LSTM

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040249775A1 (en) * 2003-05-13 2004-12-09 Dingguo Chen Very short term load prediction in an energy management system
CN110119826A (en) * 2018-02-06 2019-08-13 天津职业技术师范大学 A kind of power-system short-term load forecasting method based on deep learning
CN111931989A (en) * 2020-07-10 2020-11-13 国网浙江省电力有限公司绍兴供电公司 Power system short-term load prediction method based on deep learning neural network
CN113205207A (en) * 2021-04-19 2021-08-03 深圳供电局有限公司 XGboost algorithm-based short-term power consumption load fluctuation prediction method and system
CN113222279A (en) * 2021-05-28 2021-08-06 国网浙江桐乡市供电有限公司 Short-term load prediction method considering demand response
CN114091758A (en) * 2021-11-23 2022-02-25 四川信天启智能技术有限公司 Medium-short term load prediction method, system and storage medium
CN114169625A (en) * 2021-12-13 2022-03-11 国网江苏省电力有限公司经济技术研究院 Short-term load prediction method and device for power system
CN114444660A (en) * 2022-01-24 2022-05-06 南京电力设计研究院有限公司 Short-term power load prediction method based on attention mechanism and LSTM

Also Published As

Publication number Publication date
CN116415734B (en) 2024-07-12

Similar Documents

Publication Publication Date Title
CN111753092B (en) Data processing method, model training method, device and electronic equipment
CN114743020B (en) Food identification method combining label semantic embedding and attention fusion
US11954590B2 (en) Artificial intelligence job recommendation neural network machine learning training based on embedding technologies and actual and synthetic job transition latent information
CN112989761B (en) Text classification method and device
CN113065358A (en) Text-to-semantic matching method based on multi-granularity alignment for bank consultation service
CN112990530B (en) Regional population quantity prediction method, regional population quantity prediction device, electronic equipment and storage medium
Vilakone et al. Movie recommendation system based on users’ personal information and movies rated using the method of k-clique and normalized discounted cumulative gain
CN117635238A (en) Commodity recommendation method, device, equipment and storage medium
CN113987161A (en) Text sorting method and device
CN111597401A (en) Data processing method, device, equipment and medium based on graph relation network
CN112417886B (en) Method, device, computer equipment and storage medium for extracting intention entity information
CN118245745A (en) Task processing, object recommendation, model training method and computing device
CN117194771B (en) Dynamic knowledge graph service recommendation method for graph model characterization learning
CN116415734B (en) Ultra-short term load prediction method and system based on deep learning
CN117094835A (en) Multi-target group classification method for social media content
CN117933226A (en) Context-aware dialogue information extraction system and method
CN116702976A (en) Enterprise resource prediction method and device based on modeling dynamic enterprise relationship
CN116681078A (en) Keyword generation method based on reinforcement learning
CN116028319A (en) Prediction method and device based on user interaction behavior and storage medium
Li et al. A hierarchical tracker for multi-domain dialogue state tracking
CN113626537A (en) Entity relationship extraction method and system for knowledge graph construction
CN113343666B (en) Method, device, equipment and storage medium for determining confidence of score
CN118484516B (en) Industry large model-oriented multi-level theme type search enhancement generation method and system
CN116777400B (en) Engineering consultation information whole-flow management system and method based on deep learning
CN118171110B (en) Business log text processing method and system applied to intelligent government affairs

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