CN115577754A - Power load prediction method based on affair map - Google Patents

Power load prediction method based on affair map Download PDF

Info

Publication number
CN115577754A
CN115577754A CN202211345202.0A CN202211345202A CN115577754A CN 115577754 A CN115577754 A CN 115577754A CN 202211345202 A CN202211345202 A CN 202211345202A CN 115577754 A CN115577754 A CN 115577754A
Authority
CN
China
Prior art keywords
power
event
short
term
term time
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202211345202.0A
Other languages
Chinese (zh)
Inventor
刘真
雷智辉
栗鑫炜
卢思博
董宁华
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.)
Beijing Jiaotong University
Original Assignee
Beijing Jiaotong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Jiaotong University filed Critical Beijing Jiaotong University
Priority to CN202211345202.0A priority Critical patent/CN115577754A/en
Publication of CN115577754A publication Critical patent/CN115577754A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • 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
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Business, Economics & Management (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Economics (AREA)
  • Water Supply & Treatment (AREA)
  • Public Health (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Power Engineering (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a power load prediction method based on a physics map, and belongs to the field of power load prediction. According to the method, a power affair map, a rule base, a knowledge base and an MAT are constructed according to power related events and causal relationships among the events; after the power affair map is activated, selecting a power utilization scene, and reasoning on the power affair map to obtain a cause event set which may cause the power utilization scene to occur; traversing nodes in the reason event set according to MAT, selecting corresponding exogenous variables, and splicing with historical loads to obtain input variables of the prediction model; and then constructing a multivariate short-term power load prediction model MNLF, extracting long and short-term time sequence characteristics, namely embedding expression, from different time sequences by adopting different encoders, calculating a correlation coefficient, calculating the weight of the long-term time sequence, splicing the weight with the short-term time sequence embedding expression, and predicting a power load future value. The invention improves the accuracy of the prediction result.

Description

Power load prediction method based on affair map
Technical Field
The invention belongs to the field of power load prediction, and particularly relates to a power load prediction method based on a physics map.
Background
The power system load prediction is to predict a future power load based on historical data of the power load and auxiliary information. The auxiliary information is historical data of relevant variables having influence on the power load prediction, such as historical air temperature data, air pressure data, precipitation, social status and the like. The historical data is used as basic data, a theoretical model is built between the load and other relevant factors, the internal relation of the theoretical model is searched, and the future change situation is accurately estimated. The electric load is divided into commercial load, industrial load and residential electric load, and the fluctuation of the electric load has periodicity and continuity. The power load prediction is generally classified according to prediction time periods, and is classified into different prediction methods according to different periods, wherein the medium-short term prediction comprises daily prediction, monthly prediction and the like.
In the prior art, the medium and short term power load prediction method comprises a classical statistical model and an artificial intelligence method. The classical statistical model comprises a regression analysis method, a time sequence method, a gray model method and the like; the artificial intelligence method adopts an artificial neural network, a deep learning method and the like, and predicts the power load after carrying out big data intelligent analysis on historical data. However, the cause of the power load fluctuation is complicated, and it is necessary to recognize the power scenario in order to accurately predict the power load. The existing power load prediction method is lack of identification of power utilization scenes, multi-dimensional exogenous variables cannot be selected for the power utilization scenes, multi-dimensional factors influencing power loads cannot be comprehensively considered, and the prediction result is large in deviation and inaccurate.
Disclosure of Invention
In view of the above-mentioned defects or shortcomings in the prior art, the present invention aims to provide a power load prediction method based on a case map, which introduces the case map to identify a power scene, selects exogenous variables by means of a case map inference method and a mapping relation table between events and variables, uses long and short time sequence division to distinguish data with different distances in a time dimension, and adopts different encoders to extract respective features for the long and short time sequences, thereby improving the accuracy of prediction.
In order to achieve the above purpose, the embodiment of the present invention adopts the following technical solutions:
the invention provides a power load prediction method based on a matter map, which comprises the following steps:
step S1, constructing a power event map, a rule base P, a knowledge base K and an event and variable mapping relation table MAT according to power related events in power historical daily data and monthly data and causal relations among the events;
s2, activating the power event graph, traversing graph node events, and selecting a power utilization scene by combining the rule base P and the knowledge base K;
s3, reasoning on the power affair map based on the selected preset power utilization scene to obtain a cause event set C which may cause the power utilization scene to occur;
s4, traversing the nodes in the reason event set according to an event and variable mapping relation table MAT, selecting corresponding exogenous variables, and splicing the selected exogenous variables with historical loads to obtain input variables of the prediction model;
s5, constructing a multivariate short-term power load prediction model MNLF, wherein the prediction model divides the long and short-term time sequence of input data, and meanwhile, different encoders are adopted for different time sequences, and the characteristics of the long and short-term time sequence, namely the embedded expression, are extracted; and calculating a correlation coefficient between the short-term time sequence embedded representation and the long-term time sequence embedded representation, performing weighted representation as the weight of the long-term time sequence, splicing the result of the weighted representation and the short-term time sequence embedded representation, inputting the result into a model full-connection layer, and predicting a future value of the power load.
As one of the present inventionIn a preferred embodiment, the power event graph includes nodes and directed edges, denoted as G = (E, R), where E = { E = { (E, R) } 1 ,e 2 ,...,e n Is a set of nodes in the case graph, e i Represents a power event; r = { R = 1 ,r 2 ,...,r m Denotes a set of events and relationships between events, r i Representing events and causal relationships between events;
the rule base P is used for storing rules for determining whether the relevant nodes in the power event graph are activated or not, and the events bound by the nodes are likely to occur when the nodes are activated;
the knowledge base K is used for storing a predicted value of the exogenous variable required by rule judgment in a future period of time;
and the event and variable mapping relation table MAT is used for storing the mapping relation between the events and the variables.
As a preferred embodiment of the present invention, the selecting of the power utilization scenario specifically includes: the knowledge base K stores values of variables required for regular judgment in a future period of time; traversing the corresponding variable value in the knowledge base K within a period of time in the future by using the rule in the rule base P, judging whether the corresponding node event is possible to occur, and taking the corresponding event as a preset power utilization scene when the judgment result of the rule is yes.
As a preferred embodiment of the present invention, the inference is performed on the power event map, and the inference algorithm includes the following steps:
step S31, inputting a graph structure corresponding to the graph and an activated node A;
step S32, starting access from A and initializing a set C;
step S33, if the adjacent node of the vertex accessed currently has no accessed node, selecting one node to be accessed, if the node has no accessible node taking the node as the starting point, adding the node into the set C, and returning to the vertex accessed recently;
and step S34, until all the vertexes communicated with the initial vertex are visited, and returning to the set C.
As a preferred embodiment of the present invention, in step S5, for the short-term time series, using the LSTM as an encoder, the trend change information is extracted; for the long-term time sequence, the CNN is used as an encoder, the long-term time sequence is used as a picture object to be processed, and the mutual influence among different variables is extracted.
As a preferred embodiment of the present invention, the short term time series encoder includes an input layer, a concealment layer, and an output layer; wherein the content of the first and second substances,
the core of the LSTM is the cellular state, represented by the horizontal lines through the cell; the LSTM changes the state of the cell through an input gate, a forgetting gate and an output gate; the input gate is shown as a formula (1) and used for controlling the information input into the cell unit, the forgetting gate is shown as a formula (2) and used for controlling the information forgotten at the previous moment, and the output gate is shown as a formula (3) and used for controlling the information transmitted to the next moment by the cell unit; memory state s t As shown in equation (4), history information advantageous for predicting future data is memorized, where [ h ] t-1 ；x t ]Representing the hidden layer state h at the previous moment t-1 And the current input x t The current hidden layer state h is obtained through the formula (5) t ；
i t ＝δ(W i [h t-1 ；x t ]+b i ) (1)
f t ＝δ(W f [h t-1 ；x t ]+b f ) (2)
o t ＝δ(W o [h t-1 ；x t ]+b o ) (3)
Figure BDA0003918150810000041
Figure BDA0003918150810000042
Formulas (1) - (5)) In (i) t ,f t ,o t Respectively representing an input gate, a forgetting gate, an output gate and a memory state; w is a group of i ,W f ,W o ,W s Weight coefficients representing input gate, forgetting gate, output gate and memory state, respectively, b i ,b f ,b o ,b s Respectively representing the bias parameters of an input gate, a forgetting gate, an output gate and a memory state, and delta and tanh represent an activation function sigmod; s t ,s t-1 Memory states at time t and t-1, h t ,h t-1 Hidden states at time t and time t-1 respectively,
Figure BDA0003918150810000043
representing multiplication by element.
In a preferred embodiment of the present invention, the long-term time-series encoder uses convolution layers composed of a plurality of convolution kernels having variable dimensions, performs a convolution operation on time-series data in a two-dimensional matrix form, and extracts a temporal dependency of the data and an influence relationship between the variables.
As a preferred embodiment of the present invention, the convolution operation is described using equation (6):
h k ＝φ(W S *X s +b S ) (6)
in the formula (6), b S Denotes a bias parameter, h k Representing the result of the coding of the long-term time series, X S Is a time series, W S Phi is the activation function and phi is the convolution operation sign.
As a preferred embodiment of the invention, the multivariate short-term power load prediction model is constructed based on a memory network, and the model is a long-term time sequence { X } i }＝X 1 ,...,X n The embedded representation vector m encoded by the encoder i }＝m 1 ,...,m n Stored in the memory array, as shown in equation (7):
m i ＝Encoder(X i ) (7)
taking the short-term time sequence as a problem, obtaining an embedded vector representation u by an encoder, as shown in formula (8):
u＝Encoder(Q) (8)
in equations (7) and (8), the Encoder is denoted by Encoder, the long-term time-series Encoder is CNN, and the short-term time-series Encoder is LSTM.
As a preferred embodiment of the present invention, the MNLF calculates { m ] by using a short-term time series as a question and a long-term time series as an answer, based on a memory network model i }＝m 1 ,...,m n The correlation coefficient p of each vector in the vector set with the vector u i }＝p 1 ,...,p n As shown in formula (9):
p i ＝Softmax(u T m i ) (9)
in the formula (9), p i Representing a memory vector m i Attention weight coefficient, softmax denotes the Softmax function; p is to be i As a memory vector m i Attention to the weight coefficient, the vector m is memorized i And p i Multiplying to obtain a weighted output vector o i As shown in formula (10):
o i ＝p i ×m i (10)
finally, the embedded representation u of the short-term time sequence and the weighted output vector set { o } i Splicing is carried out to be used as the input of a full connection layer, so that the value of the future time t is predicted, as shown in a formula (11):
y t ＝W[u；o 1 ；o 2 ,…,o T ]+b (11)。
in formula (11), y t The predicted value is the power load at time t.
The technical scheme provided by the embodiment of the invention has the following beneficial effects:
according to the power load prediction method based on the physics map, the power physics map is combined with expert knowledge, a rule base, a knowledge base, an event and variable mapping relation table and the like are set, and the purposes of identifying a power scene, obtaining a cause event causing the power scene, selecting a proper exogenous variable and splicing a historical load to predict a future value are achieved; in the power load prediction stage, a multivariable power load prediction model based on a memory network is provided, not only are far and near time sequences distinguished in time dimension, but also respective important characteristics are effectively extracted, meanwhile, information which has important influence on the future is recorded by taking the memory network structure as a reference, and the prediction accuracy is improved.
Of course, not all of the advantages described above need to be achieved at the same time in the practice of any one product or method of the invention.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a flow chart of a method for predicting a power load based on a physics map according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of an embodiment of the invention based on a power event map for reasoning;
FIG. 3 is an example of a power event map in an embodiment of the present invention;
FIG. 4 is a block diagram of a short term time series encoder according to an embodiment of the present invention;
FIG. 5 is a block diagram of a long term time series encoder according to an embodiment of the present invention;
FIG. 6 is a schematic diagram of a multivariate short-term electrical load prediction model according to an embodiment of the present invention.
Detailed Description
The present inventors have found the above-described problems and have made extensive studies on a conventional power load prediction method. Researches find that the cause of power fluctuation is complex, the power load change is influenced by a plurality of factors, and in order to accurately predict the future power load, the power utilization scene needs to be identified; when the model input is a multi-dimensional variable, the power load prediction in different power utilization scenes depends on different dimensional variables, even some variables are noise in some situations, so that the prediction accuracy is influenced. Meanwhile, the conventional common time series prediction method for user power load prediction comprises a statistical model, a machine learning method and a deep learning-based method, but the conventional common time series prediction method is lack of organic combination with expert experience, and the prediction accuracy is poor. In addition, historical power loads in different time periods have different meanings for predicting future values, and the existing power load prediction method lacks of distinguishing the historical loads in different time periods. Therefore, there is still a problem in the prediction model.
Meanwhile, with the popularization and application of artificial intelligence and intelligent electric meters, electric power companies acquire massive historical data. The mass historical data are always restrained, but the objectivity and the accuracy of the data are not questionable and are accepted by the public. If the electricity utilization information can be effectively utilized, the electricity utilization condition of each family can be completely mastered, and the method has important significance for power enterprises and power consumers. For the power enterprises, the resources can be properly configured to balance supply and demand according to the prediction results, or the load can be shaped by adjusting demand response strategies such as dynamic pricing and the like, so that the shortage of infrastructure capacity is avoided.
It should be noted that the above prior art solutions have defects which are the results of practical and careful study by the inventors, and therefore, the discovery process of the above problems and the solutions proposed by the following embodiments of the present invention to the above problems should be the contribution of the inventors to the present invention in the course of the present invention.
The technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. It is to be understood that the described embodiments are merely a few embodiments of the invention, and not all embodiments. The components of embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations. It should be noted that the embodiments and features of the embodiments of the present invention may be combined with each other without conflict.
It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined or explained in subsequent figures. In the description of the present invention, the terms "first", "second", "third", "fourth", and the like are used only for distinguishing the description, and are not intended to indicate or imply relative importance.
After the deep analysis, the application provides a power load prediction method based on a physical map, long and short time sequence division is adopted, long and short time sequences are distinguished, and different encoders are adopted to obtain the characteristics of the sequences; for predicting future values, by taking the memory network structure as reference, setting a memory network component to store a long-term time sequence embedded representation, carrying out correlation calculation on the long-term time sequence embedded representation and a short-term time sequence embedded representation, carrying out weighted summation on the long-term time sequence embedded representation by calculating the correlation of different time sequence embedded representations as weight coefficients, and then splicing to predict future values, thereby achieving the purpose of capturing the dependency on the time dimension; meanwhile, a mapping relation table between the cause events and the variables is established by combining expert knowledge, and after the cause event set is obtained based on the event map, the corresponding exogenous variable set is selected facing different scenes through the mapping table.
The invention provides a power load prediction method framework based on a matter graph for the first time, the matter graph is activated by utilizing a knowledge base and a rule base, then a reasoning stage of the matter graph is entered, a cause event set capable of using an electric field scene is obtained, and then related variables are selected based on a mapping relation table between the events and the variables and are spliced with historical load data to be used as model input. And in the model prediction stage, the embedded expression of the time sequence is obtained by dividing the long time and the short time of the time sequence and using different encoders, the correlation between the embedded expression of the short time sequence and the embedded expression of the long time sequence is calculated and used as a weight coefficient to carry out weighted summation on the embedded expression of the long time sequence, the dependency on the time dimension is captured and spliced with the short time sequence, and a prediction model is input to predict future load.
Referring to fig. 1, the method for predicting a power load based on a physics map according to the embodiment of the present invention includes the following steps:
step S1, constructing a power event map, a rule base P, a knowledge base K and an event and variable mapping relation table MAT according to power related events in power historical daily data and monthly data and causal relations among the events.
In this step, the power event graph includes nodes and directed edges, and is represented by G = (E, R), where E = { E = { (E) } 1 ,e 2 ,...,e n Is a set of nodes in the case graph, e i Represents an event; r = { R = 1 ,r 2 ,...,r m Denotes a set of events and relationships between events, r i Representing events and causal relationships between events.
Wherein, the node is an event, and the directed edge is a causal relation between the events.
The rule base P is used for storing rules for determining whether the relevant nodes in the power event graph are activated, and the node activation indicates that the event of node binding is likely to occur. For example, in the node binding event, "the body-sensory temperature exceeds 28 degrees" and "the binding rule" the air temperature exceeds 28 degrees ", when the air temperature reaches 28 degrees, it is considered that" the body-sensory temperature exceeds 28 degrees "is likely to occur.
And the knowledge base K is used for storing the predicted values of the exogenous variables required by rule judgment in a future period of time. For example, the variable required for the judgment that "the air temperature exceeds 28 degrees" is the air temperature.
And the event and variable Mapping relation Table MAT (Mapping Table) is used for storing the Mapping relation between the event and the variable. For example, the event "rain comes" may affect humidity, temperature, pressure, etc., which are all attribute fields in the data set.
As shown in fig. 3, which is a typical power event graph, nodes include power load increases; the main cause categories of the increase of the electric load, such as the increase of the heating load; specific reasons for each category, such as a low sensible temperature that causes an increase in heating load, and a possible reason for the low sensible temperature.
Wherein the nodes represent events and the directed edges between the nodes represent causal relationships. For example, in fig. 3, the reasons for the increase in the electrical load are mainly classified into four major categories: the electrical load of entertainment and office equipment is increased, the electrical load of illumination is increased, the heating load is increased, and the cold electrical loads of air-conditioning fans and the like are increased. The reason why the lighting power load is increased is that the lighting intensity is insufficient, and the reason why the lighting intensity is insufficient is as follows: the rainy days in the cloudy days and seasonal factors cause long days and short nights. The event nodes bound by the rule base are mainly dark gray nodes, for example, the 'light intensity is insufficient' in the graph, the inference result is light gray nodes, for example, 'rainy days in cloudy days' and 'seasonal factors, long day and night' step S2, the power affair graph is activated, the graph node events are traversed, and the power utilization scene is selected by combining the rule base P and the knowledge base K.
In this step, the knowledge base K stores predicted values of specific variables in a future period of time, the rule base P stores determination rules for determining whether or not events of each node in the power event graph may occur, the rules are established by expert knowledge, and the specific variables are exogenous variables required for rule determination. And when traversing the rules in the rule base P, inspecting the corresponding variable values in the knowledge base K required by the corresponding rules within a period of time in the future, judging whether the corresponding node events are possible to occur, and when the judgment result of the rules is yes, taking the corresponding events as the preset power utilization scene. .
And S3, reasoning on the power affair map based on the selected power utilization scene to obtain a cause event set C which may cause the power utilization scene to occur.
In the step, the power affair map is a causal relationship among various power related events formed through expert knowledge guidance, after the affair map activation link, a preset power utilization scene is selected, the map inference stage is started, other possible reason events which can theoretically cause the event bound by the node are searched according to expert knowledge contained in the map, and a complete reason event set required for observing the event is obtained as far as possible.
Specifically, the inference algorithm comprises the following steps:
step S31, inputting a graph structure corresponding to the graph and an activated node A;
step S32, starting access from A and initializing a set C;
step S33, if the adjacent node of the vertex accessed currently has no accessed node, selecting one node to be accessed, if the node has no accessible node taking the node as the starting point, adding the node into the set C, and returning to the vertex accessed recently;
and step S34, until all the vertexes communicated with the initial vertex are visited, and returning to the set C.
The pseudo code corresponding to the algorithm is described as follows:
[ input: a case graph G = (E, R), and an activated node set A in the graph;
and (3) outputting: a reason node set C which can enable the nodes in the set A to be activated in the graph;
1.FOR point e i IN set A
node Access flag initialization in DO graph
DO Slave node e i Starting the depth map traversal, whenever a node with only in-degree and no out-degree is encountered and
4. when not in the set C, the information is put into the set C;
5.END FOR
6. return set C).
And S4, traversing the nodes in the reason event set according to an event and variable mapping relation table MAT, selecting corresponding exogenous variables, and splicing the selected exogenous variables with the historical load to obtain input variables of the prediction model. The length of the historical load is a set time window value, wherein the length of the exogenous variable is consistent with the length of the historical load, and therefore the historical load and the exogenous variable are spliced into a multi-dimensional variable.
In this step, as shown in table 1, for an example of a partial cause event and variable mapping table MAT:
TABLE 1 event and variable mapping Table (MAT) example
Figure BDA0003918150810000101
And selecting corresponding exogenous variables according to the MAT.
Step S5, constructing a multivariable Short-Term power Load prediction Model (MNLF), wherein the prediction model is used for carrying out long-Term and Short-Term time sequence division on input data, and meanwhile, different encoders are adopted for different time sequences, and long-Term and Short-Term time sequence features are extracted, namely embedding expression; and calculating a correlation coefficient between the short-term time sequence embedded representation and the long-term time sequence embedded representation, performing weighted representation as the weight of the long-term time sequence, splicing the result of the weighted representation and the short-term time sequence embedded representation, inputting the result into a model full-connection layer, and predicting a future value of the power load.
In the step, the long-term and short-term time sequences are divided, and different encoders are adopted for the long-term and short-term time sequences to obtain corresponding characteristics of the long-term and short-term time sequences, so that the historical data in different periods can be distinguished. In the time dimension, the meaning of the recent history data and the long-term history data to the prediction future value is different, so that the division of the long-term and short-term time sequence is performed in the step.
For a Short-Term time series, because the trend change is more meaningful for predicting a future value because the Short-Term time series is closer to the future value, a Long Short Term Memory (LSTM) is used as an encoder to extract trend change information. As shown in fig. 4, the short term time series encoder includes an input layer, a hidden layer, and an output layer. The core of the LSTM is the cellular state, represented by the horizontal lines running through the cell, which is like a conveyor belt, which runs through the entire cell but only very muchAnd the number of branches is small, so that the information can flow through the whole network under the condition of no change. LSTM changes the state of a cell through an input gate (equation (1)), a forgetting gate (equation (2)), and an output gate (equation (3)). The input gate is used for controlling and inputting information in the cell unit, the forgetting gate is used for controlling and forgetting information at the previous moment, and the output gate is used for controlling and transmitting information at the next moment to the cell unit. s is t For the memory state (equation (4)) historical information is memorized that is advantageous for predicting future data, where h t-1 ；x t ]Representing the hidden layer state h at the previous moment t-1 And the current input x t The current hidden layer state h is obtained through the formula (5) t 。
i t ＝δ(W i [h t-1 ；x t ]+b i ) (1)
f t ＝δ(W f [h t-1 ；x t ]+b f ) (2)
o t ＝δ(W o [h t-1 ；x t ]+b o ) (3)
Figure BDA0003918150810000111
Figure BDA0003918150810000112
In formulae (1) to (5), i t ,f t ,o t Respectively representing an input gate, a forgetting gate, an output gate and a memory state; w i ,W f ,W o ,W s Weight coefficients representing input gate, forgetting gate, output gate and memory status, respectively, b i ,b f ,b o ,b s Respectively representing the bias parameters of an input gate, a forgetting gate, an output gate and a memory state, and delta and tanh represent an activation function sigmod; s t ,s t-1 Memory states at time t and t-1, h t ,h t-1 Hidden states at time t and time t-1 respectively,
Figure BDA0003918150810000113
representing multiplication by element.
For the long-term time sequence, because the time span is large, the mutual influence among various richer variables is preserved, so the CNN is used as an encoder, and is used as a picture object to be processed, and the mutual influence among different variables is extracted. As shown in fig. 5, the data to be processed by the long-term time-series encoder is a multivariate time-series, is presented in a two-dimensional matrix form, and has a structure similar to picture data due to the mutual influence between its variables, so the multivariate time-series is subjected to a convolution operation using CNN. The convolution operation effect is influenced by the size of a convolution kernel, the selection of a pooling function, the setting of the number of convolution layers and the number of convolution channels of each layer. The long-term time sequence encoder uses a convolution layer consisting of a plurality of convolution kernels with variable dimensions, performs convolution operation on time sequence data in a two-dimensional matrix form, and extracts the dependency of data on time and the mutual influence relationship of the variables. Setting the time sequence as X S The parameter in the convolution kernel is W S The activation function is phi (the activation function is generally the ReLU function), the convolution operation symbol is x, and the above convolution operation is described by the formula, which can be seen in formula (6):
h k ＝φ(W S *X s +b S ) (6)
in the formula (6), b S Denotes a bias parameter, h k Representing the coding result of the long-term time series.
The calculating of the correlation coefficient between the short-term time-series embedded representation and the long-term time-series embedded representation, the weighting representation as the weight of the long-term time series, is a process for capturing the dependency on the data time dimension. The embedded representations of the long-term time series are weighted and summed for better prediction by the input model. Specifically, a memory component is firstly arranged to store the embedded representation of the long-term time sequence, then the correlation between the embedded representation of the short-term time sequence and the embedded representation of the long-term time sequence is calculated to be used as a weight coefficient, and the embedded representation of the long-term time sequence is subjected to weighted summation, so that the capture of the time dimension dependency is realized. And finally, the long-term time sequence embedded representation is spliced with the short-term time sequence embedded representation, and a full connection layer is input to predict a future value.
In the prediction, as shown in fig. 6, the multivariate short-term power load prediction model is constructed based on a memory network, and the model is a long-term time series { X } i }＝X 1 ,...,X n The embedded representation vector m resulting from the encoding by the encoder i }＝m 1 ,...,m n And storing the short-term time sequence in a memory array as formula (7), and taking the short-term time sequence as a problem to obtain an embedded vector expression u of the short-term time sequence through an encoder as formula (8).
m i ＝Encoder(X i ) (7)
u＝Encoder(Q) (8)
In equations (7) and (8), the Encoder is denoted by Encoder, the long-term time series Encoder is CNN, and the short-term time series Encoder is LSTM.
In predicting a future power load, the importance of a sequence of historical loads at different intervals on the time axis to the future prediction is different. The MNLF is based on popularization of a memory network model, takes a short-term time sequence as a question and a long-term time sequence as an answer, and then calculates { m } i }＝m 1 ,…,m n The correlation coefficient p of each vector in the vector set with the vector u i }＝p 1 ,…,p n As shown in formula (9):
p i ＝Softmax(u T m i ) (9)
in the formula (9), p i Representing a memory vector m i Attention weight coefficient, softmax denotes the Softmax function; p is to be i As a memory vector m i Attention to the weight coefficient, the vector m is memorized i And p i Multiplying to obtain a weighted output vector o i As shown in formula (10)：
o i ＝p i ×m i (10)
Finally, the embedded representation u of the short-term time sequence and the weighted output vector set { o } i Splicing is carried out to be used as the input of a full connection layer, so that the value of the t-th time in the future is predicted, as shown in a formula (11):
y t ＝W[u；o 1 ；o 2 ,…,o T ]+b (11)。
in equation (11), T represents the length of the set elapsed time window.
According to the technical scheme, the power load prediction method based on the event graph, provided by the embodiment of the invention, is characterized in that the power event graph is constructed based on historical data so as to integrate expert knowledge of causal relations among power events, and the knowledge base and the rule base component are arranged to activate the power event graph, so that the automation of distinguishing power utilization scenes according to the expert knowledge is realized; reasoning on the power event graph by using a power event graph depth traversal algorithm to obtain a cause event set which can cause a preset power utilization scene, so that the automation of power utilization scene cause analysis according to expert knowledge is realized; by using a mapping relation table between events and variables, exogenous variables related to a cause event set obtained by map inference are searched, an automatic process from distinguishing power utilization scenes to selecting exogenous variables is realized, and the capability of combining a case map with expert knowledge is utilized without being limited by the capability of the case map; meanwhile, a multivariable power load prediction model based on a memory network is provided, firstly, long-term and short-term time sequences are divided, then, different encoders are adopted for extracting characteristics of the time sequences with different lengths, the correlation of the long-term and short-term time sequences is calculated, the correlation is used as a weight coefficient to carry out weighted summation on long-term time sequence embedded representation, the long-term time sequence embedded representation and the short-term time sequence embedded representation are spliced, and a prediction future value of a full connection layer is input, so that the accuracy of a prediction result is improved.
The above description is only a preferred embodiment of the invention and an illustration of the applied technical principle and is not intended to limit the scope of the claimed invention but only to represent a preferred embodiment of the invention. It will be appreciated by those skilled in the art that the scope of the invention herein disclosed is not limited to the particular combination of features described above, but also encompasses other arrangements formed by any combination of the above features or their equivalents without departing from the spirit of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without making any creative effort, shall fall within the protection scope of the present invention.

Claims (10)

1. A power load prediction method based on a physics map is characterized by comprising the following steps:
step S1, constructing a power event map, a rule base P, a knowledge base K and an event and variable mapping relation table MAT according to power related events in power historical daily data and monthly data and causal relations among the events;
s2, activating the power event graph, traversing graph node events, and selecting a power utilization scene by combining the rule base P and the knowledge base K;
s3, reasoning on the power affair map based on the selected preset power utilization scene to obtain a cause event set C which may cause the power utilization scene to occur;
s4, traversing the nodes in the reason event set according to an event and variable mapping relation table MAT, selecting corresponding exogenous variables, and splicing the selected exogenous variables with historical loads to obtain input variables of the prediction model;
s5, constructing a multivariate short-term power load prediction model MNLF, wherein the prediction model divides the long and short-term time sequence of input data, and meanwhile, different encoders are adopted for different time sequences, and the characteristics of the long and short-term time sequence, namely the embedded expression, are extracted; and calculating a correlation coefficient between the short-term time sequence embedded representation and the long-term time sequence embedded representation, performing weighted representation as the weight of the long-term time sequence, splicing the result of the weighted representation and the short-term time sequence embedded representation, inputting the result into a model full-connection layer, and predicting a future value of the power load.
2. The event map-based power load prediction method according to claim 1,
the power physics graph includes nodes and directed edges, denoted as G = (E, R), where E = { E = { (E, R) } 1 ,e 2 ,...,e n Is a set of nodes in the case graph, e i Represents a power event; r = { R = 1 ,r 2 ,...,r m Denotes a set of events and relationships between events, r i Representing events and causal relationships between events;
the rule base P is used for storing rules for determining whether the relevant nodes in the power event graph are activated or not, and the events bound by the nodes are likely to occur when the nodes are activated;
the knowledge base K is used for storing a predicted value of the exogenous variable required by rule judgment in a future period of time;
and the event and variable mapping relation table MAT is used for storing the mapping relation between the events and the variables.
3. The event graph-based power load prediction method according to claim 1, wherein the selecting of the power usage scenario specifically comprises: the knowledge base K stores values of variables required for regular judgment in a future period of time; traversing the corresponding variable value in the knowledge base K within a period of time in the future by using the rule in the rule base P, judging whether the corresponding node event is possible to occur, and taking the corresponding event as a preset power utilization scene when the judgment result of the rule is yes.
4. The event graph-based power load forecasting method according to claim 1, characterized in that reasoning is performed on the power event graph, and the steps of a reasoning algorithm are as follows:
step S31, inputting a graph structure corresponding to the graph and an activated node A;
step S32, starting access from A and initializing a set C;
step S33, if the adjacent node of the vertex accessed currently has no accessed node, selecting one node to be accessed, if the node has no accessible node taking the node as the starting point, adding the node into the set C, and returning to the vertex accessed recently;
and step S34, until all the vertexes communicated with the initial vertex are visited, and returning to the set C.
5. The event graph-based power load prediction method according to claim 1, wherein in step S5, for a short-term time series, trend change information is extracted using LSTM as an encoder; for the long-term time sequence, the CNN is used as an encoder, the long-term time sequence is used as a picture object to be processed, and the mutual influence among different variables is extracted.
6. A situational map-based power load prediction method according to claim 5, characterized in that said short term time series encoder comprises an input layer, a hidden layer and an output layer; wherein, the first and the second end of the pipe are connected with each other,
the core of the LSTM is the cellular state, represented by the horizontal lines through the cell; the LSTM changes the state of the cell through an input gate, a forgetting gate and an output gate; the input gate is shown as formula (1) and is used for controlling the information input into the cell unit, the forgetting gate is shown as formula (2) and is used for controlling the information forgotten at the previous moment, and the output gate is shown as formula (3) and is used for controlling the information transmitted to the next moment by the cell unit; memory state s t As shown in equation (4), history information advantageous for predicting future data is memorized, where [ h ] t-1 ；x t ]Representing the hidden layer state h at the previous moment t-1 And current input x t The current hidden layer state h is obtained through a formula (5) t ；
i t ＝δ(W i [h t-1 ；x t ]+b i ) (1)
f t ＝δ(W f [h t-1 ；x t ]+b f ) (2)
o t ＝δ(W o [h t-1 ；x t ]+b o ) (3)
Figure FDA0003918150800000031
Figure FDA0003918150800000032
In formulae (1) to (5), i t ,f t ,o t Respectively showing an input gate, a forgetting gate and an output gate; w i ,W f ,W o ,W s Weight coefficients representing input gate, forgetting gate, output gate and memory status, respectively, b i ,b f ,b o ,b s Respectively representing the bias parameters of an input gate, a forgetting gate, an output gate and a memory state, and delta and tanh represent an activation function sigmod; s t ,s t-1 Memory states at time t and t-1, h t ,h t-1 Hidden states at time t and time t-1 respectively,
Figure FDA0003918150800000033
representing multiplication by element.
7. A method for event-map-based power load prediction as defined in claim 5 wherein the long-term time-series encoder uses a plurality of convolution kernels having variable dimensions in width to form a convolution layer, and performs a convolution operation on the time-series data in the form of a two-dimensional matrix while extracting the temporal dependency of the data and the relationship of the variables with respect to each other.
8. A situational map-based power load prediction method according to claim 7, characterized in that the convolution operation is described using equation (6):
h k ＝φ(W S *X s +b S ) (6)
in the formula (6), b S Denotes a bias parameter, h k Representing the result of the coding of the long-term time series, X S Is a time series, W S The parameters in the convolution kernel, phi, are the activation functions, and phi, the convolution operation signs.
9. The method of claim 8, wherein the multivariate short-term power load prediction model MNLF is constructed based on a memory network, and the model is a long-term time series { X } i }＝X 1 ,...,X n The embedded representation vector m resulting from the encoding by the encoder i }＝m 1 ,...,m n Stored in the memory array, as shown in equation (7):
m i ＝Encoder(X i ) (7)
using the short-term time sequence as a problem, obtaining an embedded vector representation u of the short-term time sequence by an encoder, as shown in equation (8):
u＝Encoder(Q) (8)
in equations (7) and (8), the Encoder is denoted by Encoder.
10. The event map-based power load prediction method according to claim 9, wherein the MNLF is based on a memory network model, takes a short-term time series as a question and a long-term time series as an answer, and then calculates { m } m i }＝m 1 ,...,m n The correlation coefficient p of each vector in the vector set with the vector u i }＝p 1 ,…,p n As shown in formula (9):
p i ＝Softmax(u T m i ) (9)
in the formula (9), p i Represents a memory vector m i Attention weight coefficient, softmax denotes the Softmax function; p is to be i As a memory vector m i Attention to the weight coefficient, the vector m is memorized i And p i Multiplying to obtain a weighted output vector o i As shown in formula (10):
o i ＝p i ×m i (10)
finally, the embedded expression u of the short-term time sequence and the weighted output vector set { o } i Splicing is carried out to be used as the input of a full connection layer, so that the value of the future time t is predicted, as shown in a formula (11):
y t ＝W[u；o 1 ；o 2 ,…,o T ]+b (11)。
in formula (11), y t The predicted value is the power load at time t.
CN202211345202.0A 2022-10-31 2022-10-31 Power load prediction method based on affair map Pending CN115577754A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211345202.0A CN115577754A (en) 2022-10-31 2022-10-31 Power load prediction method based on affair map

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211345202.0A CN115577754A (en) 2022-10-31 2022-10-31 Power load prediction method based on affair map

Publications (1)

Publication Number Publication Date
CN115577754A true CN115577754A (en) 2023-01-06

Family

ID=84589034

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211345202.0A Pending CN115577754A (en) 2022-10-31 2022-10-31 Power load prediction method based on affair map

Country Status (1)

Country Link
CN (1) CN115577754A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116204792A (en) * 2023-04-28 2023-06-02 北京航空航天大学 Training method for generating causal interpretation model

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116204792A (en) * 2023-04-28 2023-06-02 北京航空航天大学 Training method for generating causal interpretation model
CN116204792B (en) * 2023-04-28 2023-07-14 北京航空航天大学 Training method for generating causal interpretation model

Similar Documents

Publication Publication Date Title
Alhussein et al. Hybrid CNN-LSTM model for short-term individual household load forecasting
CN113962364B (en) Multi-factor power load prediction method based on deep learning
Szoplik Forecasting of natural gas consumption with artificial neural networks
Ahmad et al. Nonlinear autoregressive and random forest approaches to forecasting electricity load for utility energy management systems
Khafaga et al. Forecasting Energy Consumption Using a Novel Hybrid Dipper Throated Optimization and Stochastic Fractal Search Algorithm
Khotanzad et al. A neuro-fuzzy approach to short-term load forecasting in a price-sensitive environment
Zhou et al. A trigonometric grey prediction approach to forecasting electricity demand
Chen et al. A novel data-driven approach for residential electricity consumption prediction based on ensemble learning
Sun et al. Short‐term building load forecast based on a data‐mining feature selection and LSTM‐RNN method
CN109492830A (en) A kind of mobile pollution source concentration of emission prediction technique based on space-time deep learning
Peng et al. Dual-stage attention-based long-short-term memory neural networks for energy demand prediction
CN107133695A (en) A kind of wind power forecasting method and system
Verma et al. A comprehensive review on the nilm algorithms for energy disaggregation
CN111539841A (en) Electric quantity prediction method, device, equipment and readable storage medium
CN112149890A (en) Comprehensive energy load prediction method and system based on user energy label
CN113554466A (en) Short-term power consumption prediction model construction method, prediction method and device
Zhang et al. Predicting rooftop solar adoption using agent-based modeling
CN111882157A (en) Demand prediction method and system based on deep space-time neural network and computer readable storage medium
CN115577754A (en) Power load prediction method based on affair map
CN115983448B (en) Multi-energy load prediction method based on space-time diagram neural network
CN115759371A (en) GCN-LSTM-based short-term load prediction method for power system
CN115013859B (en) User portrait-based household regulation and control method for heat supply secondary network
Dagnely et al. Predicting hourly energy consumption. Can regression modeling improve on an autoregressive baseline?
Kim et al. Time-series clustering and forecasting household electricity demand using smart meter data
CN109104301A (en) A kind of method and system carrying out the prediction of network temperature for variety show based on deep learning model

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