CN117035154A - Multi-element load prediction method and device for comprehensive energy system - Google Patents
Multi-element load prediction method and device for comprehensive energy system Download PDFInfo
- Publication number
- CN117035154A CN117035154A CN202310756350.XA CN202310756350A CN117035154A CN 117035154 A CN117035154 A CN 117035154A CN 202310756350 A CN202310756350 A CN 202310756350A CN 117035154 A CN117035154 A CN 117035154A
- Authority
- CN
- China
- Prior art keywords
- data
- load
- dimensional
- value
- model
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 81
- 230000002159 abnormal effect Effects 0.000 claims abstract description 55
- 239000011159 matrix material Substances 0.000 claims abstract description 53
- 238000004458 analytical method Methods 0.000 claims abstract description 38
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 27
- 238000000926 separation method Methods 0.000 claims abstract description 16
- 238000010219 correlation analysis Methods 0.000 claims abstract description 14
- 238000012545 processing Methods 0.000 claims abstract description 11
- 238000012549 training Methods 0.000 claims abstract description 11
- 238000005516 engineering process Methods 0.000 claims abstract description 8
- 230000003542 behavioural effect Effects 0.000 claims abstract description 6
- 230000000694 effects Effects 0.000 claims description 15
- 238000007781 pre-processing Methods 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 11
- 238000005070 sampling Methods 0.000 claims description 11
- 238000005311 autocorrelation function Methods 0.000 claims description 8
- 238000010276 construction Methods 0.000 claims description 8
- 238000004590 computer program Methods 0.000 claims description 7
- 230000007246 mechanism Effects 0.000 claims description 7
- 230000015654 memory Effects 0.000 claims description 7
- 238000001816 cooling Methods 0.000 claims description 6
- 239000000284 extract Substances 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 6
- 238000005065 mining Methods 0.000 claims description 6
- 238000010438 heat treatment Methods 0.000 claims description 5
- 230000005611 electricity Effects 0.000 claims description 4
- 238000004519 manufacturing process Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 238000007637 random forest analysis Methods 0.000 claims description 4
- 230000006403 short-term memory Effects 0.000 claims description 4
- 230000006399 behavior Effects 0.000 claims description 3
- 230000006386 memory function Effects 0.000 claims description 3
- 230000008439 repair process Effects 0.000 claims description 3
- 230000001960 triggered effect Effects 0.000 claims description 3
- 230000005856 abnormality Effects 0.000 claims description 2
- 230000019771 cognition Effects 0.000 claims description 2
- 239000002131 composite material Substances 0.000 claims description 2
- 238000000605 extraction Methods 0.000 claims description 2
- 238000002955 isolation Methods 0.000 claims description 2
- 238000012417 linear regression Methods 0.000 claims description 2
- 230000007787 long-term memory Effects 0.000 claims description 2
- 238000007670 refining Methods 0.000 claims description 2
- 238000003860 storage Methods 0.000 claims description 2
- 230000008878 coupling Effects 0.000 description 10
- 238000010168 coupling process Methods 0.000 description 10
- 238000005859 coupling reaction Methods 0.000 description 10
- 238000010586 diagram Methods 0.000 description 8
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 6
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 6
- 238000009826 distribution Methods 0.000 description 6
- 238000005457 optimization Methods 0.000 description 6
- 238000013461 design Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000004140 cleaning Methods 0.000 description 2
- 238000005520 cutting process Methods 0.000 description 2
- 238000012843 least square support vector machine Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000005192 partition Methods 0.000 description 2
- 238000011897 real-time detection Methods 0.000 description 2
- 238000010200 validation analysis Methods 0.000 description 2
- 241000726094 Aristolochia Species 0.000 description 1
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 230000008485 antagonism Effects 0.000 description 1
- BBFQZRXNYIEMAW-UHFFFAOYSA-N aristolochic acid I Chemical compound C1=C([N+]([O-])=O)C2=C(C(O)=O)C=C3OCOC3=C2C2=C1C(OC)=CC=C2 BBFQZRXNYIEMAW-UHFFFAOYSA-N 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000001627 detrimental effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000003912 environmental pollution Methods 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000000611 regression analysis Methods 0.000 description 1
- 230000001373 regressive effect Effects 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
- G06F18/15—Statistical pre-processing, e.g. techniques for normalisation or restoring missing data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Business, Economics & Management (AREA)
- Economics (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Tourism & Hospitality (AREA)
- Software Systems (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Mathematics (AREA)
- Computing Systems (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Pure & Applied Mathematics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Water Supply & Treatment (AREA)
- Quality & Reliability (AREA)
- Algebra (AREA)
- Probability & Statistics with Applications (AREA)
- Databases & Information Systems (AREA)
- Game Theory and Decision Science (AREA)
- Operations Research (AREA)
- Primary Health Care (AREA)
- Entrepreneurship & Innovation (AREA)
- Public Health (AREA)
- Development Economics (AREA)
Abstract
The application discloses a multi-element load prediction method of a comprehensive energy system, which comprises the following steps: firstly, constructing a high-dimensional data matrix for the acquired data through a separation window technology; secondly, performing exception processing on outliers in the high-dimensional data matrix based on an orphan forest algorithm; thirdly, analyzing and repairing the abnormal data hidden in the time sequence through a dynamic track method, triggering an alarm by a system after finding the abnormal value data, entering a message surface behavioural analysis link, visualizing a separation window, and judging the abnormal value type; fourthly, performing correlation analysis on the multi-element load and weather data thereof by using an MIC method, and constructing a high-dimensional feature matrix; fifthly, outputting predicted values of multiple loads based on the TLMMoE multi-task training network, wherein the method has the advantages that: the application provides a power load prediction method and device, which can improve the accuracy of a prediction result of power load prediction.
Description
Technical Field
The application relates to the technical field of comprehensive energy systems, in particular to the technical field of multi-element load prediction of a short-term comprehensive energy system.
Background
The problems of environmental pollution and energy shortage always restrict the sustainable development of human society. How to efficiently and economically realize interconnection and intercommunication in different energy links, strengthen energy cooperation and improve energy utilization rate, and become a hot problem of general attention in the industry. To meet the urgent need for complementary coordinated operation between various energy sources, integrated energy systems (Integrated Energy System, IES) have been developed. The short-term load prediction is a precondition for the comprehensive energy system to realize daily energy management and optimal scheduling, and the prediction accuracy directly influences the stability and economy of the comprehensive energy system in the running period.
The comprehensive energy system load prediction work comprises four links of data characteristic preprocessing, model selection, optimization strategy and algorithm selection. For a long time, extensive and intensive studies have been conducted in the industry on short-term load prediction methods of integrated energy systems, with significant results. The existing prediction method mainly comprises traditional machine learning, integrated learning, deep learning, reinforcement learning, transfer learning and the like. Aiming at the problem of low prediction accuracy caused by the defects of low convergence rate and the like of a traditional Wavelet Neural Network (WNN) comprehensive energy system load prediction model, the short-term load prediction method of the WNN comprehensive energy system based on Improved Particle Swarm (IPSO) is provided. By considering the coupling relation among heat, electricity and gas loads in industry and analyzing the multi-element data such as weather information, history information and the like, the coupling relation among the data is mined, and a prediction method based on deep learning and multi-task learning is provided, so that the prediction capability of a model is improved, and a prediction result with higher precision is obtained. In consideration of high-dimensional time dynamic characteristics, an encoder-decoder model based on a long-short-term memory network (LSTMED) is provided, and the deep learning method is verified to be well applicable to comprehensive energy load prediction. By analyzing the coupling relation among different subsystems and utilizing the concept of a multi-task learning weight sharing mechanism and a least square support vector machine, an electric, thermal, cold and gas load combined prediction method based on the multi-task learning and the least square support vector machine is provided. The comprehensive load prediction model based on the Bi-directional generation antagonism network (Bi-GAN) data enhancement and migration learning technology is provided, and the model considers the condition of data deficiency in an information system, solves the problem when a new user joins a comprehensive energy system while maintaining higher prediction precision, improves the robustness of the comprehensive energy system, and expands the generalized nature of the comprehensive energy system.
In recent years, with the massive access of various intelligent systems and distributed energy sources, the energy interactive structure is diversified, and the energy utilization characteristics of users under the multi-structure are real-time and complicated. In the actual operation process, the comprehensive energy system can operate under extreme conditions such as abrupt change of external environment, equipment failure and the like, and load data generated at the moment belongs to unconventional data. Because the data volume is less in the unconventional mode, the parameter optimization difficulty is high, and the neural network is sensitive to abnormal data. When the load is affected by some complex factors and exhibits strong randomness and non-stationary characteristics, the prediction accuracy of the neural network will be significantly reduced. So far, most of students carry out conventional design on the comprehensive energy system in the links of model selection, optimization strategy and algorithm selection, but cannot meet the operation requirements of the comprehensive energy system with various modes and abrupt change of working conditions. On the basis of the method, on one hand, the design is carried out in the data characteristic preprocessing link, abnormal data are removed, meanwhile, in order to meet the requirements of actual working conditions, the influence of event-based driving factors is required to be considered, and the historical energy utilization characteristics of a user are analyzed so as to improve the overall stability of the system. In another aspect, the appropriate parameters are selected in the model selection, optimization strategy and algorithm selection links. And through the scheme design of a plurality of links, the load prediction precision in the conventional mode is improved.
At present, abnormal value identification methods for time series data are also continuously developed, and common cleaning methods comprise Gaussian distribution based on statistics, a box graph, a clustering method based on machine learning and the like. On the basis of judging whether the error of the measured data meets the Gaussian distribution of zero mean value or not, the abnormal value is detected by using the 3 sigma criterion, and the method is simple and convenient, but is influenced by various factors in the operation process of the comprehensive energy system, the output data information has the characteristics of high dimensionality, randomness, intermittence and uncertainty, and the identification precision of the abnormal data by the method is greatly reduced and only simple outliers can be identified. The clustering algorithm is used as a machine learning algorithm, and can identify most of outliers, but is difficult to identify for abnormal points hidden in a sequence, and cannot meet the requirement of data analysis quality.
Disclosure of Invention
In view of the above, the invention provides a method and a device for predicting multiple loads of a comprehensive energy system, which can remove abnormal data in a time sequence in a conventional data mode and are suitable for predicting multiple loads of the comprehensive energy system.
The technical solution adopted by the invention for solving the technical problems is as follows:
A multi-element load prediction method of a comprehensive energy system comprises the following steps:
firstly, constructing a high-dimensional data matrix for the acquired data through a separation window technology;
secondly, performing exception processing on outliers in the high-dimensional data matrix based on an orphan forest algorithm;
thirdly, analyzing and repairing the abnormal data hidden in the time sequence through a dynamic track method, triggering an alarm by a system after finding the abnormal value data, entering a message surface behavioural analysis link, visualizing a separation window, and judging the abnormal value type;
fourthly, performing correlation analysis on the multi-element load and weather data thereof by using an MIC method, and constructing a high-dimensional feature matrix;
and fifthly, outputting predicted values of multiple loads based on the TLMMoE multi-task training network.
Preferably, the first step is to build a high-dimensional data matrix on the collected data by a separation window technology, and build the high-dimensional data matrix based on historical cold load, heat load, electric load and meteorological data:
(1) Building a high-dimensional data matrix: in the first step of data feature preprocessing, a reasonable high-dimensional data matrix X needs to be constructed T Firstly, constructing and selecting N-dimensional nodes of cold, heat, electric load and weather data to sample the data, wherein the N-dimensional sampling nodes can form a column vector at a certain sampling time t:
x t ＝[ x 1,t x 2,t … x N,t ] T (1)
As sampling time increases, column vector data increases, so that the following time sequence X can be formed:
X＝[ x 1 x 2 … x t …] (2)
in order to meet the requirement of real-time calculation, the acquired data are sequentially intercepted by a sliding time window for analysis, the width of the sliding time window is set to be tau, namely, the data at the t-th moment and the historical data of tau-t before the t-th moment are acquired, and a high-dimensional data matrix X is constructed by all the data in the time window T The following are provided:
X T ＝[ x t-τ+1 … x t ] (3)
where τ is the time window,
the time window is selected to satisfy the following conditions: a time window with proper width is set, so that the real-time detection requirement can be met, and all measurement data of all the partitions can be contained;
(2) Data normalization: construction of a high-dimensional data matrix X T All data are normalized, i.e. dimensionalized and normalized.
Preferably, the second step performs exception processing on outliers in the high-dimensional data matrix based on an orphan forest algorithm:
the isolation degree of the data points is estimated by constructing a random forest, so as to judge whether the data points are abnormal points, the orphan forest algorithm is sensitive to abnormal value points in a high-dimensional data matrix consisting of cold load, heat load, electric load and meteorological data in the comprehensive energy system, can be isolated, separated and segmented out efficiently and quickly,
Typically, the basic steps of the orphan forest algorithm are as follows:
firstly, randomly selecting 1 characteristic and randomly selecting a cutting point in a range of 1 characteristic value;
secondly, taking the selected characteristics and the cutting points as a segmentation rule, and segmenting the data points into two subsets;
third, recursively repeating the first and second steps until there are only 1 data points in each subset, or a predefined maximum depth of the tree is reached;
fourth, constructing a plurality of random trees and forming 1 random forest;
fifth, for each data point, calculating the path length in the random forest, namely the average value of the number of edges passing from the root node to the data point;
sixth, using the path length to measure the degree of abnormality of the data point, wherein a shorter path length represents a normal point that is easier to divide, and a longer path length represents an abnormal point that is easier to isolate;
seventh, by setting a threshold, the path length can be compared to the probability of outliers to determine which data points are classified as outliers.
Preferably, the step three is to analyze and repair the abnormal data hidden in the time sequence through a dynamic orbit method, if the abnormal value data is found, the system triggers an alarm, enters a message surface behavioural analysis link, visualizes a separation window, and judges the abnormal value type, and the specific steps are as follows:
(1) The invention provides a dynamic orbit method (Dynamic Railway Line, DRL) suitable for a comprehensive energy system for carrying out abnormal repair on data, wherein the dynamic orbit consists of a base orbit (Middle Railway Line, MRL), an upper orbit (Upper Railway Line, URL) and a lower orbit (Lower Railway Line, LRL), the shape of the dynamic orbit looks like a channel, and a data value is in the range of the channel and is regarded as a normal value; when the data value passes through the channel and appears outside the channel, the data value is judged to be an abnormal value;
(2) The calculation of the dynamic orbit method mainly comprises two steps:
1) Firstly, establishing a base rail of a dynamic rail by a moving average method;
2) Then, the upper rail and the lower rail are calculated, and on the basis of the determination of the base rail, the upper rail and the lower rail are calculated according to the following specific calculation formula:
URL＝(1+M1)*MRL (4)
LRL＝(1-M2)*MRL (5)
wherein M1 is the margin value of the upper track, and M2 is the margin value of the lower track;
preferably, the weight weighted moving average line index is selected in the present invention. Starting from the data at the first time point in the time series, the weight assigned to the time series data increases linearly with the increase of the time series.
Removing and repairing the abnormal value in the time sequence through a dynamic track method, and when the abnormal value exceeds the upper track value of the dynamic track, assigning the base track value as a corresponding value of a high-dimensional data matrix; when the abnormal value exceeds the lower rail value of the dynamic rail, the base rail value is assigned to the corresponding value of the high-dimensional data matrix;
(3) When abnormal alarm occurs, a message surface behavior analysis function of the system is triggered, the system can analyze by combining with activity conditions in the system to judge whether an emergency occurs, such as large-scale activity holding or new building and equipment are put into use, experience values of extreme events are provided for future production and life, and help is provided for developing an integrated prediction algorithm.
Preferably, in the fourth step, correlation analysis is performed on the multi-element load and weather data thereof by a MIC method, and a high-dimensional feature matrix is constructed:
(1) Correlation analysis of multiple loads:
the comprehensive energy system is a composite system which is composed of a cold system, a hot system and an electric energy system and is combined with a plurality of information systems of a meteorological system and an economic system, and is used as a dynamic time-varying system with complex and huge information, a plurality of energy sources in the system are mutually coupled and cooperatively complemented, and simultaneously, along with the continuous expansion of the scale and the increase of the complexity of the comprehensive energy system, the time dimension of the operation data of the comprehensive energy system is increased, mass data with time sequence state characteristics and time-space dimension characteristics are formed, the collected high-dimensional characteristic data are analyzed, the general cognition of the whole comprehensive energy system is facilitated, and the direction and the range are qualitatively indicated for the construction of a prediction model;
(2) Autoregressive system analysis:
in the invention, an autoregressive model (Auto Regressive Model, AR model) is adopted to analyze the time sequence state characteristics of a short-term time sequence, a basis is provided for the moving average day number l of a dynamic orbit method base rail, the autoregressive model uses itself as a regression variable, namely, the linear regression relation of random variables at a later moment is described by using the linear combination of information data of the autoregressive model, and a time sequence { Xt } is set so as to satisfy the following conditions:
X t ＝a 0 +aX t-1 +…+a p X t-p +ε t (6)
in which a is 0 ,a 1 ,…,a p Is tied in a way thatNumber, ε t Is a white noise value. The time-series status characteristics of the load data are analyzed according to an autocorrelation function (Autocorrelation Coefficient Function, ACF) and a partial autocorrelation function (Partial Auto correlation Coefficient Function, PACF) of the AR model.
(3) Maximum information coefficient analysis
The MIC method can measure the strength of the linear and nonlinear association degree between data, and the association relation between different types of load data is evaluated, so that the found wide relation type is obtained. Therefore, the invention adopts the maximum information coefficient (The Maximal Information Coefficient, MIC) to carry out space-time dimension characteristic analysis on a plurality of systems so as to construct a high-dimensional characteristic matrix input by a model. The degree of correlation between the large features is in the range of [0,1], and when the value is smaller than 0.3, the degree of correlation between the large features is weak; when the value is larger than 0.7, the relationship between the two is strong.
(4) Construction of input feature sets
And carrying out multidimensional correlation analysis on the high-dimensional data based on MIC, wherein the following data characteristics are taken as input data, and the granularity is 15min:
input1: predicting the cold load one year before the reference day;
input2: predicting the heat load one year before the reference day;
input3: predicting an electrical load one year before a baseline day;
input4: predicting temperature data one year before a reference day;
input5: predicting dew point data one year ago on a baseline day;
input6: predicting cold building data one year before the reference day;
input7: predicting heat utilization building data one year before the reference day;
input8: electricity usage building data one year before the baseline day is predicted.
Preferably, in the fifth step, the TLMMoE-based multitasking training network outputs a predicted value of the multiple load:
the invention adopts a TCN-LSTM-MMoE multi-task learning framework, which is called TLMMoE model for short, to build a multi-element load prediction model,
(1) After the network model receives the input high-dimensional feature matrix, the time sequence convolutional neural network (Temporal Convolutional Neural Network, TCN) is utilized to excavate and extract the time sequence characteristics and the space-time characteristics of the high-dimensional feature;
(2) The TCN network inputs the learned characteristic information into a Long Short-term memory network (Long Short-TermMemory, LSTM), further learns the characteristic matrix, and extracts time sequence links from the characteristic matrix by utilizing the unique memory function of the LSTM unit;
(3) And (3) transmitting the processed data characteristics into an MMoE (Multi-gate media-of-expertise) Multi-task learning model, and predicting and outputting power values of the cooling load, the heating load and the electric load by utilizing an expert network layer and a unique gating mechanism of the MMoE model.
The invention provides a multi-element load prediction device of a comprehensive energy system, which comprises the following components: the system comprises a high-dimensional data receiving module, a high-dimensional data preprocessing module, a multi-dimensional characteristic analysis module, a data characteristic mining module, a multi-dimensional characteristic refining and extracting module and a data information prediction output module,
the high-dimensional data receiving module collects a data set from the measuring device, slides at intervals of separation windows based on a separation window technology, and builds a high-dimensional data matrix;
the high-dimensional data preprocessing module adopts an orphan forest algorithm and a dynamic orbit method to reject outliers and abnormal values in time sequences;
the multidimensional feature analysis module performs correlation analysis on the cold load, the heat load, the electric load and the meteorological data of the data set by using an MIC analysis method, and selects input features;
the data feature mining module is used for mining and extracting input data by utilizing a time convolution unit in the TCN network layer, and extracting time and space features in time sequence data.
The multidimensional feature refinement extraction module refines and learns the input data of the front layer by using the memory unit in the LSTM network layer, and excavates long-term and short-term memory relation.
And the data information prediction output module predicts and outputs power values of the cold load, the hot load and the electric load by utilizing an expert network layer and a unique gating mechanism of the MMoE model.
The application also provides a computer device of the multi-element load prediction device of the comprehensive energy system, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, wherein the processor realizes the steps of the power load prediction method when executing the computer program.
The present application also provides a computer readable storage medium storing a computer program which when executed by a processor implements the steps of the power load prediction method as described above.
The beneficial effects achieved by adopting the technical proposal of the application are as follows:
the application provides a power load prediction method and device, which can improve the accuracy of a prediction result of power load prediction. Aiming at the problem of multi-element load prediction of the comprehensive energy system, the design is carried out in a data characteristic preprocessing link and a model preferential link, and the comprehensive energy system multi-element load prediction method based on an orphan forest algorithm and a dynamic orbit method is provided.
Through analysis of examples, the following conclusion is obtained:
1) The orphan forest algorithm can process the problem of outliers in high-dimensional big data, and has the characteristics of high processing precision and high calculation speed;
2) The dynamic orbit method can effectively remove hidden abnormal values in the time sequence, has good cleaning effect and provides a good basis for a neural network prediction model;
3) The coupling relation between load data in the comprehensive energy system is complex, the maximum information coefficient analysis method can better analyze the time sequence characteristics and the space characteristics between loads, and a high-dimensional feature matrix with strong correlation is constructed through analysis, so that a foundation is laid for improving the precision of a prediction model;
4) Through the reasonable design of the TLMMoE network structure, from three links of data feature grabbing, learning and multi-task distribution, the coupling feature of historical data is better learned, the prediction precision is improved, and the effectiveness of the algorithm in a time sequence feature sequence is proved.
The research result provides a new solution idea and method for the multi-element load prediction problem of the comprehensive energy system, and has important significance for promoting the intellectualization and high efficiency of the comprehensive energy system.
Drawings
FIG. 1 is a diagram of the operation of the electrical load of the present invention prior to processing;
FIG. 2 is a graph of the electrical load contour profile of the present invention;
FIG. 3 is a graph of the electrical load operation after processing in accordance with the present invention;
FIG. 4 is a graph of ACF and PACF analysis for the cold load of the present invention;
FIG. 5 is an ACF and PACF analysis chart of the electrical load of the present invention;
FIG. 6 is a MIC thermodynamic diagram of a feature sequence of the present invention;
FIG. 7 is a diagram of a model framework of the present invention;
FIG. 8 is a graph comparing the error of the IDTLMMoE model and the ITLMMoE model of the present invention;
FIG. 9 is a graph comparing the cooling loads of the IDCLSTM model and the IDTLMMoE model of the present invention;
FIG. 10 is a graph comparing thermal loads of the IDCLSTM model and the IDTLMMoE model of the present invention;
FIG. 11 is a graph comparing electrical loads of the IDCLSTM model and the IDTLMMoE model of the present invention;
FIG. 12 is a graph of the error contrast between the IDTLMMoE model and the IDCLSTM model of the present invention.
Detailed Description
The calculation load data are from IES cold, heat and electric load data and weather characteristic data of 2022, 2023, 5, 1 of the university of Aristolochia State university, and the sampling granularity is 15min, and the load data are from a project network database of the school, wherein the school is provided with 288 buildings, more than 5 ten thousand teachers and students and energy conversion equipment such as Combined Cooling and heating (CCHP, heating and Power), an electric boiler, a Gas boiler, an electric converter P2G (Power-to-Gas) and the like. The weather data is derived from published weather data of the weather website at corresponding times. On the division of the data set, 75% was taken as the training set, 15% as the validation set, and 10% as the test set.
The experiment platform is a hardware platform of Xeon4213R CPU and NVIDIA RTX 3090GPU, a tensorflow, keras deep learning framework is built, and python language is used for realizing the experiment.
1. Construction of high-dimensional data matrix based on separation window technique
A high-dimensional data matrix is constructed based on historical cold load, heat load, electrical load, and meteorological data.
Building a high-dimensional data matrix: in the first step of data feature preprocessing, a reasonable high-dimensional data matrix Xp needs to be constructed. In the invention, 21-dimensional nodes of cold, heat, electric load and weather data are firstly constructed and selected for data sampling, and the data characteristic information is shown in fig. 6.
At a certain sampling time t 1 The 21-dimensional sampling nodes may form a column vector:
x t ＝[ x 1,t x 2,t … x 21,t1 ] T (1)
as sampling time increases, column vector data increases, so that the following time sequence X can be formed:
X＝[ x 1 x 2 … x 21 ] (2)
in order to meet the requirement of real-time calculation, the acquired data are sequentially intercepted by a sliding time window for analysis, and the width of the sliding time window is set as tau, namely, at the t-th acquisition time 1 Time of day data and t 1 τ -t before the moment 1 Is used to construct a high-dimensional data matrix X from all data within this time window T The following are provided:
X T ＝[ x t1-τ+1 … x t1 ] (3)
where τ is the time window.
The time window is selected to satisfy the following conditions: and a time window with proper width is set, so that the real-time detection requirement can be met, and all measurement data of all the partitions can be contained.
(2) Data normalization: all data are normalized, i.e. dimensionalized and normalized.
The data in the examples all have different dimensions, for example, the unit of the cooling load is Ton/hrs, and the numerical range is between the interval of 1000 and 9000; the heat load unit is mmBTU/hr, and the numerical range is between 1 and 30; the electrical load is in kW and ranges in value between 7000 and 26000. It can be seen that the three different loads have different dimensional values, which is detrimental to the convergence of the neural network load prediction model. Therefore, all the data are standardized, and the processed data of each characteristic data after normalization conform to standard normal distribution, namely the mean value is 0, and the standard deviation is 1.
2. Outlier identification based on orphan forest algorithm
In the selected example data, the statistical feature quantity is 21, and the operation data is 46000. In the face of massive experimental data, the method uses the orphan forest algorithm to process global data, and the algorithm has the characteristics of high processing precision and high calculation speed, and is suitable for processing outlier problems in big data. Meanwhile, the processed data is more beneficial to the implementation of a dynamic orbit method.
Because of the limited space, the processing is only illustrated by taking electrical load data as an example, and the necessity analysis is performed. As shown in fig. 1, a historical data operation chart of the electric load is shown.
It is found by observation that there are extreme points in the data so that the dynamically changing power load data presents a straight line and there is a magnitude difference between the extreme points and the normal data.
In order to more clearly express the magnitude relation, a contour distribution diagram of the electric load is drawn, and as shown in fig. 2, white points in the diagram are clearly observed to be normal values, the white points are distributed in a contour distribution cluster of a blue region, the darker the color is, the more the color is deviated to the outer side of the contour, and the lighter the color is. In the figure, black dots are abnormal values and are dispersed in dark green areas.
And detecting and eliminating the abnormal points to obtain the actual operation data of the electric load after elimination, which is shown in fig. 3. By comparing the operation diagrams before and after the pretreatment of the power load, the abnormal value existing in the data is eliminated, and the power load operation data is trended and regularized, so that the method can efficiently process the abnormal value points existing in the data.
3. And analyzing and repairing the abnormal data hidden in the time sequence by a dynamic track method. If abnormal value data are found, the system triggers an alarm, enters a message surface behavioural analysis link, visualizes a separation window and judges the abnormal value type.
(1) First, the data is autoregressive through an AR algorithm, and a parameter l is determined, so that a base rail value of a dynamic track is determined.
For reasons of text space limitations, the electrical load data and the cold load data are taken as examples, and the cold load autocorrelation coefficient analysis chart and the partial autocorrelation coefficient analysis chart are shown in fig. 4, and the electrical load autocorrelation coefficient analysis chart and the partial autocorrelation coefficient analysis chart are shown in fig. 5.
As shown in fig. 4, the blue region is a confidence interval of the correlation coefficient map, where the confidence interval is taken to be 95% and the time granularity is 15 minutes. From the autocorrelation coefficient diagram of the cold load, it can be seen that, considering the whole series of influence factors, there is a strong coupling correlation between the loads in the first 13 time intervals; from the partial autocorrelation of the cold load, it can be seen that there is a strong coupling correlation between the loads during the first 3 time intervals. Comparing the cold load and the electrical load autocorrelation coefficient plots, the cold load was found to be more highly trending than the electrical load for a short period. Meanwhile, the cold load and the electric load are also periodically characterized. Therefore, the value of l for the cold load is set to 8 and the value of l for the electric load is set to 3.
(2) Next, the upper and lower rails are calculated. The specific calculation formula is as follows:
URL＝(1+M1)*MRL (4)
LRL＝(1-M2)*MRL (5)
Wherein M1 is 10% of the margin value of the upper track, and M2 is 10% of the margin value of the lower track.
Preferably, the weight weighted moving average line index is selected in the present invention. Starting from the data at the first time point in the time series, the weight assigned to the time series data increases linearly with the increase of the time series.
Removing and repairing the abnormal value in the time sequence through a dynamic track method, and when the abnormal value exceeds the upper track value of the dynamic track, assigning the base track value as a corresponding value of a high-dimensional data matrix; when the outlier exceeds the lower track value of the dynamic track, the base track value is assigned to the corresponding value of the high-dimensional data matrix.
(3) When abnormal alarm occurs, a message surface behavior analysis function of the system is triggered, the system can analyze by combining with activity conditions in the system to judge whether an emergency occurs, such as large-scale activity holding or new building and equipment are put into use, experience values of extreme events are provided for future production and life, and help is provided for developing an integrated prediction algorithm.
4. Carrying out correlation analysis on the multi-element load and weather data thereof by an MIC method, and constructing a high-dimensional feature matrix:
(1) Correlation analysis of multiple loads:
This example describes the analysis of the correlation of the individual variables using MIC analysis.
The linear correlation of the data can be effectively measured by using the pearson correlation coefficient or the spearman correlation coefficient and the like, and even the mathematical formulas of the linear relation and the simple nonlinear relation can be determined by regression analysis. However, in the integrated energy system, the coupling relationship between the various systems is complex, and not only is a linear relationship exist between the data, but also a large number of nonlinear relationships exist, and the nonlinear relationship cannot be expressed simply by using a mathematical formula.
The MIC method can measure the strength of the linear and nonlinear association degree between data, and the association relation between different types of load data is evaluated, so that the found wide relation type is obtained. Therefore, the invention adopts the maximum information coefficient (The Maximal Information Coefficient, MIC) to carry out space-time dimension characteristic analysis on a plurality of systems so as to construct a high-dimensional characteristic matrix input by a model. As shown in fig. 6, there is a MIC thermodynamic diagram of a plurality of systems such as an electric energy system, a cold system, a hot system, and a weather system.
The right hand side of FIG. 6 is illustrated as the degree of correlation between large features, with a range of values of [0,1], when the value is less than 0.3, the correlation between the two is weak; when the value is larger than 0.7, the relationship between the two is strong.
As shown in the figure, the cold system, the hot system and the electric system are used as three basic systems in the comprehensive energy system, wherein the correlation coefficient of the cold load and the hot load is 0.57, and the correlation is strong, which indicates that the cold system and the hot system have a relatively close coupling relationship. The correlation coefficient of the cold load and the electric load is 0.7, which indicates that the cold system and the electric system are particularly closely related in the production and life of the integrated energy system.
The correlation coefficient of the thermal system and the electrical system is 0.48, which indicates that the coupling relationship exists between the thermal system and the electrical system in the integrated energy system, and the relationship between the thermal system and the electrical system is relatively weaker. In addition, it can be observed that the correlation between the cold, heat and electric system and the temperature and dew point is also relatively tight. In addition, the building usage of each system and the respective system show a strong correlation.
(2) Construction of input feature sets
Carrying out multidimensional correlation analysis on the high-dimensional data based on MIC, taking the conditions of model optimization speed and the like into consideration, and taking the following data characteristics as input data (granularity is 15 min):
input1: predicting the cold load one year before the reference day;
input2: predicting the heat load one year before the reference day;
input3: predicting an electrical load one year before a baseline day;
Input4: predicting temperature data one year before a reference day;
input5: predicting dew point data one year ago on a baseline day;
input6: predicting cold building data one year before the reference day;
input7: predicting heat utilization building data one year before the reference day;
input8: electricity usage building data one year before the baseline day is predicted.
The specific list is shown in table 1:
5. the TLMMoE-based multitasking training network outputs predicted values of multiple loads.
The invention adopts a TCN-LSTM-MMoE multi-task learning framework, hereinafter referred to as CLMMOE model, to build a multi-element load prediction model, and the model framework is shown in figure 7. Firstly, after a network model receives an input high-dimensional feature matrix, a TCN network excavates and extracts time sequence characteristics and space-time characteristics of high-dimensional feature; and then, the TCN network inputs the learned characteristic information into the LSTM network, further learns the characteristic matrix, and extracts the time sequence link of the characteristic matrix by utilizing the unique memory function of the LSTM unit. And then, the processed data features are transmitted into an MMoE multi-task learning model, and the power values of the cooling load, the heating load and the electric load are predicted and output by utilizing an expert network layer and a unique gating mechanism of the MMoE model.
(1) Evaluation index
The invention adopts average absolute percentage error (Mean Absolute Percentage Error, MAPE) as main evaluation index of the predictive performance of each model, and selects average absolute error (Mean Absolute Error, MAE) as auxiliary evaluation index. The calculation formulas of MAPE and MAE are respectively as follows:
wherein: y is t Is the actual power load data;is predicted electrical load data; />Is the average value of the power load; n is the number of samples of the electrical load data.
(2) Super parameter setting
In the network model, super parameters exist in the TCN network layer, the LSTM network layer and the multi-task learning layer. And determining the hyper-parameters in the model by adopting a control variable method according to the learning effect of model training. The finally determined optimization parameters are shown in table 2 after multiple parameter adjustments.
6. Performance analysis of different predictive models
The method provided by the invention is different from the traditional multitasking method in both data preprocessing link and model structure, and in order to further comprehensively evaluate the correctness and effectiveness of the method based on the orphan forest algorithm and the dynamic orbit method in the multi-element load prediction of the comprehensive energy system, 3 groups of experimental models are objectively designed in this section, and the method is respectively as follows:
1) Based on an iForest-TCN-LSTM-MMoE multi-element load prediction model (ITLMMoE model);
2) Based on an iForest-DRL-CNN-LSTM multi-element load prediction model (IDCLSTM model);
3) Based on an iForest-DRL-TCN-LSTM-MMoE multi-element load prediction model (IDTLMMoE model).
To ensure fairness of the experiment, all models employ the same data set allocation ratio to divide the training set, validation set and test set. In addition, in order to ensure the training effect of the comparison model, the setting of the network layer adopts the optimal configuration.
(1) Comparing the IDTLMMoE model with the ITLMMoE model
In order to verify that the IDTLMMoE model provided by the invention can eliminate abnormal value points in a data preprocessing link and improve load prediction precision, the invention sets a comparison ITLMMoE model. And the comparison model is processed by using an orphan forest algorithm only in a data preprocessing link, and then load prediction is performed by using a TLMMoE network layer.
As shown in fig. 8, the error for both models in the test set is demonstrated. Therefore, after the DRL module is added, the prediction errors of the cold load and the electric load are obviously reduced, and the dynamic orbit method can effectively remove the abnormal value hidden in the time sequence, so that the interference of the abnormal value point on the neural network is reduced, and the prediction accuracy is improved.
As shown in Table 3, in comparison with the predicted error results of the IDTLMMoE model and the ITLMMoE model, the PAPE indexes of the cold load, the heat load and the electric load are respectively reduced by 2.57%,0.11% and 0.49%, and the MAE error values of the cold load, the heat load and the electric load are also reduced to a certain extent. Therefore, the model provided by the invention is better, and the dynamic orbit method is proved to be capable of removing hidden abnormal values in the time sequence, reducing the prediction error and improving the load prediction precision.
(2) Comparing the IDTLMMoE model with the ITLMMoE model
In order to verify the effectiveness of the IDTLMMoE model in the network layer structure TLMMoE module, the invention sets a comparison IDCLSTM model. The control model is a traditional multitasking load prediction model based on parameter hard sharing. The traditional multi-task learning mechanism takes an LSTM layer as a sharing layer, and then connects a full-connection layer as prediction output, so that multi-element load prediction is realized.
As shown in fig. 9, 10 and 11, we demonstrate a comparison of performance of the IDCLSTM model and the IDTLMMoE model for a certain period of time selected for cold, hot, and electrical loads on the test set. It can be seen that in the prediction effect of the electrical load, the provided IDTLMMoE model and the traditional multi-task IDCLSTM model can obtain better prediction effects, the fitting degree of the prediction curve to the real curve is higher, and compared with the IDCLSTM model, the IDTLMMoE model has better prediction effects. However, by comparing part of details, the IDTLMMoE model can still keep a better fitting effect in the part with severe curve fluctuation, the predicted value is more similar to the true value, the predicted error of the IDCLSTM model is improved, and the fitting effect is poor.
In addition, since fluctuations in the cold and hot loads are more severe than the electrical loads, training of the model is also more difficult and over-fitting or under-fitting easily occurs. As can be seen from fig. 12, the IDCLSTM model is more conservative in the prediction results of the cold and hot loads, and the prediction curve has a flatter volatility in the face of the cold and hot loads with larger fluctuation. The cold load MAPE percentage value in the IDCLSTM model is 6.98%, while the IDTLMMoE model still maintains better prediction capability, and the MAPE index percentage is 3.8%, compared with the IDCLSTM model, the MAPE index percentage value is reduced by 3.18 percentage points. The reason for this is that taking LSTM as a hard sharing layer cannot take into account the correlation difference between multiple loads, resulting in providing the same weight value for the subtask layer, thereby affecting the performance of the model. In contrast, the proposed IDTLMMoE model enables feature information sharing between multiple tasks by setting up multiple private subnets. Meanwhile, the gating layer is arranged to avoid interference of weak related information on subtasks. In addition, the IDTLMMoE model also sets different loss functions according to the fluctuation characteristics of subtasks and the correlation difference between multiple loads. Therefore, the IDTLMMoE model achieves better prediction effect.
As can be seen from FIG. 12, the IDTLMMoE model of the present invention has better performance in testing the heat, cold and electric loads than the IDCLSTM model, and the MAPE error percentage and MAE error value are reduced. The method also shows that compared with the CNN layer of the IDCLSTM model, the TCN layer in the IDTLMMoE model can acquire the characteristic information of a long-time sequence, and the data are mined, so that when the characteristic information is transmitted to the LSTM layer, the LSTM unit layer in the IDTLMMoE model can better refine and learn the time sequence characteristics and the space characteristics. Therefore, the prediction effect of the IDTLMMoE model is better.
Claims (9)
1. The method for predicting the multi-element load of the comprehensive energy system is characterized by comprising the following steps of:
firstly, constructing a high-dimensional data matrix for the acquired data through a separation window technology;
secondly, performing exception processing on outliers in the high-dimensional data matrix based on an orphan forest algorithm;
thirdly, analyzing and repairing the abnormal data hidden in the time sequence through a dynamic track method, triggering an alarm by a system after finding the abnormal value data, entering a message surface behavioural analysis link, visualizing a separation window, and judging the abnormal value type;
fourthly, performing correlation analysis on the multi-element load and weather data thereof by using an MIC method, and constructing a high-dimensional feature matrix;
And fifthly, outputting predicted values of multiple loads based on the TLMMoE multi-task training network.
2. The method for predicting multiple loads of integrated energy system according to claim 1, wherein in the first step, a high-dimensional data matrix is built on the collected data by a separation window technology, and the high-dimensional data matrix is built based on historical cold load, heat load, electric load and meteorological data:
(1) Building a high-dimensional data matrix: firstly, constructing and selecting N-dimensional nodes of cold, heat, electric load and weather data to sample the data, wherein the N-dimensional sampling nodes can form a column vector at a certain sampling time t:
x t ＝[x 1,t x 2,t … x N,t ] T (1)
as sampling time increases, column vector data increases, so that the following time sequence X can be formed:
X＝[x 1 x 2 … x t …] (2)
sequentially intercepting collected data by using a sliding time window for analysis, setting the width of the sliding time window as tau, namely collecting data at the t moment and historical data of tau-t before the t moment, and constructing a high-dimensional data matrix X by using all data in the sliding time window T The following are provided:
X T ＝[x t-τ+1 … x t ] (3)
wherein τ is a time window;
(2) Data normalization: construction of a high-dimensional data matrix X T All data are normalized, i.e. dimensionalized and normalized.
3. The method for predicting the multiple loads of the integrated energy system according to claim 1, wherein in the second step, outliers in the high-dimensional data matrix are subjected to exception processing based on an orphan forest algorithm:
the isolation degree of the data points is estimated by constructing a random forest, so that whether the data points are abnormal points or not is judged, and the orphan forest algorithm is sensitive to abnormal value points in a high-dimensional data matrix consisting of cold load, heat load, electric load and meteorological data in the comprehensive energy system, and can be isolated, separated and segmented out efficiently and rapidly.
4. The method for predicting the multiple loads of the comprehensive energy system according to claim 1, wherein in the third step, the abnormal data hidden in the time sequence is analyzed and repaired by a dynamic orbit method, if the abnormal value data is found, the system triggers an alarm, enters a message surface behavioural analysis link, visualizes a separation window, and judges the abnormal value type, and the method specifically comprises the following steps:
(1) The dynamic track method DRL is suitable for a comprehensive energy system to repair the abnormality of the data, wherein the dynamic track consists of a base track MRL, an upper track URL and a lower track LRL, and is shaped like a channel, and the data value is in the range of the channel and is regarded as a normal value; when the data value passes through the channel and appears outside the channel, the data value is judged to be an abnormal value;
(2) The calculation of the dynamic orbit method mainly comprises two steps:
1) Firstly, establishing a base rail of a dynamic rail by a moving average method;
2) Then, the upper rail and the lower rail are calculated, and on the basis of the determination of the base rail, the upper rail and the lower rail are calculated according to the following specific calculation formula:
URL＝(1+M1)*MRL (4)
LRL＝(1-M2)*MRL (5)
wherein M1 is the margin value of the upper track, and M2 is the margin value of the lower track;
removing and repairing the abnormal value in the time sequence through a dynamic track method, and when the abnormal value exceeds the upper track value of the dynamic track, assigning the base track value as a corresponding value of a high-dimensional data matrix; when the abnormal value exceeds the lower rail value of the dynamic rail, the base rail value is assigned to the corresponding value of the high-dimensional data matrix;
(3) When abnormal alarm occurs, a message surface behavior analysis function of the system is triggered, the system can analyze by combining with the activity conditions in the system, judge whether the emergency occurs, provide experience values of extreme events for future production and life, and provide help for developing an integrated prediction algorithm.
5. The method for predicting the multiple loads of the comprehensive energy system according to claim 1, wherein in the fourth step, correlation analysis is performed on the multiple loads and weather data thereof by a MIC method, and a high-dimensional feature matrix is constructed:
(1) Correlation analysis of multiple loads:
the comprehensive energy system is a composite system which is composed of a cold system, a hot system and an electric energy system and is combined with a plurality of information systems of a meteorological system and an economic system, and is used as a dynamic time-varying system with complex and huge information, a plurality of energy sources in the system are mutually coupled and cooperatively complemented, and simultaneously, along with the continuous expansion of the scale and the increase of the complexity of the comprehensive energy system, the time dimension of the operation data of the comprehensive energy system is increased, mass data with time sequence state characteristics and time-space dimension characteristics are formed, the collected high-dimensional characteristic data are analyzed, the general cognition of the whole comprehensive energy system is facilitated, and the direction and the range are qualitatively indicated for the construction of a prediction model;
(2) Autoregressive system analysis:
adopting an autoregressive model AR to analyze the time sequence state characteristics of the short-term time sequence, providing a basis for the moving average days l of the dynamic orbit method base rail, adopting the autoregressive model to use the autoregressive model as a regression variable, namely using the linear combination of the information data of the prior autoregressive model to describe the linear regression relation of random variables at a later moment, setting a time sequence { Xt }, and meeting the following conditions:
X t ＝a 0 +aX t-1 +…+a p X t-p +ε t (6)
In which a is 0 ,a 1 ,…,a p Is a coefficient, epsilon t For the white noise value, analyzing the time sequence state characteristics of the load data according to an autocorrelation function ACF and a partial autocorrelation function PACF of the AR model;
(3) Maximum information coefficient analysis
The MIC method can measure the strength of the linear and nonlinear association degree between data, and evaluate the association relation between different types of load data to find out the wide relation type, so that the maximum information coefficient MIC is adopted to carry out space-time dimension characteristic analysis on a plurality of systems so as to construct a high-dimensional characteristic matrix input by a model, the association degree between all large characteristics is in the range of 0,1, and the correlation between the two is shown to be weak when the value is smaller than 0.3; when the numerical value is larger than 0.7, the relation between the two is strong;
(4) Construction of input feature sets
And carrying out multidimensional correlation analysis on the high-dimensional data based on MIC, wherein the following data characteristics are taken as input data, and the granularity is 15min:
input1: predicting the cold load one year before the reference day;
input2: predicting the heat load one year before the reference day;
input3: predicting an electrical load one year before a baseline day;
input4: predicting temperature data one year before a reference day;
input5: predicting dew point data one year ago on a baseline day;
Input6: predicting cold building data one year before the reference day;
input7: predicting heat utilization building data one year before the reference day;
input8: electricity usage building data one year before the baseline day is predicted.
6. The method for predicting multiple loads of integrated energy system according to claim 1, wherein in the fifth step, the predicted value of the multiple loads is output based on the TLMMoE multitask training network:
a TCN-LSTM-MMoE multitask learning framework, hereinafter called TLMMoE model, is adopted to build a multi-element load prediction model,
(1) After the network model receives the input high-dimensional feature matrix, the TCN network is utilized to excavate and extract the time sequence characteristic and the space-time characteristic of the high-dimensional feature;
(2) The TCN network inputs the learned characteristic information into the LSTM network, further learns the characteristic matrix, and extracts the time sequence link of the characteristic matrix by utilizing the unique memory function of the LSTM unit;
(3) And transmitting the processed data characteristics into an MMoE multi-task learning model, and predicting and outputting power values of the cooling load, the heating load and the electric load by utilizing an expert network layer and a unique gating mechanism of the MMoE model.
7. A complex energy system multi-load prediction device, the device comprising: the system comprises a high-dimensional data receiving module, a high-dimensional data preprocessing module, a multi-dimensional characteristic analysis module, a data characteristic mining module, a multi-dimensional characteristic refining and extracting module and a data information prediction output module,
The high-dimensional data receiving module collects a data set from the measuring device, slides at intervals of separation windows based on a separation window technology, and builds a high-dimensional data matrix;
the high-dimensional data preprocessing module adopts an orphan forest algorithm and a dynamic orbit method to reject outliers and abnormal values in time sequences;
the multidimensional feature analysis module performs correlation analysis on the cold load, the heat load, the electric load and the meteorological data of the data set by using an MIC analysis method, and selects input features;
the data characteristic mining module is used for mining and extracting input data by utilizing a time convolution unit in the TCN network layer, and extracting time and space characteristics in time sequence data;
the multidimensional feature refinement extraction module refines and learns the input data of the front layer by using a memory unit in the LSTM network layer, and excavates a long-term and short-term memory relation;
and the data information prediction output module predicts and outputs power values of the cold load, the hot load and the electric load by utilizing an expert network layer and a unique gating mechanism of the MMoE model.
8. A computer device for a complex energy system multi-load prediction apparatus, comprising a memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the steps of the method according to any one of claims 1 to 6.
9. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310756350.XA CN117035154A (en) | 2023-06-25 | 2023-06-25 | Multi-element load prediction method and device for comprehensive energy system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310756350.XA CN117035154A (en) | 2023-06-25 | 2023-06-25 | Multi-element load prediction method and device for comprehensive energy system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117035154A true CN117035154A (en) | 2023-11-10 |
Family
ID=88637883
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310756350.XA Pending CN117035154A (en) | 2023-06-25 | 2023-06-25 | Multi-element load prediction method and device for comprehensive energy system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117035154A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117669839A (en) * | 2024-02-01 | 2024-03-08 | 山东赛马力发电设备有限公司 | Distributed load prediction method and system for comprehensive energy system |
| CN117786370A (en) * | 2024-02-26 | 2024-03-29 | 北京国旺盛源智能终端科技有限公司 | Information intelligent analysis system for gridding service terminal |
-
2023
- 2023-06-25 CN CN202310756350.XA patent/CN117035154A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117669839A (en) * | 2024-02-01 | 2024-03-08 | 山东赛马力发电设备有限公司 | Distributed load prediction method and system for comprehensive energy system |
| CN117669839B (en) * | 2024-02-01 | 2024-04-30 | 山东赛马力发电设备有限公司 | Distributed load prediction method and system for comprehensive energy system |
| CN117786370A (en) * | 2024-02-26 | 2024-03-29 | 北京国旺盛源智能终端科技有限公司 | Information intelligent analysis system for gridding service terminal |
| CN117786370B (en) * | 2024-02-26 | 2024-04-19 | 北京国旺盛源智能终端科技有限公司 | Information intelligent analysis system for gridding service terminal |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN117035154A (en) | Multi-element load prediction method and device for comprehensive energy system | |
| CN113962364A (en) | Multi-factor power load prediction method based on deep learning | |
| CN107918830B (en) | Power distribution network running state evaluation method based on big data technology | |
| CN111639823B (en) | Building cold and heat load prediction method constructed based on feature set | |
| Zheng et al. | Interpretable building energy consumption forecasting using spectral clustering algorithm and temporal fusion transformers architecture | |
| CN116167527B (en) | Pure data-driven power system static safety operation risk online assessment method | |
| CN113326966A (en) | CEEMD-LSTM-based multi-load prediction method for comprehensive energy system | |
| CN116187640B (en) | Power distribution network planning method and device based on grid multi-attribute image system | |
| CN111859283A (en) | Scene generation method considering multi-energy charge-time sequence and correlation | |
| CN116169670A (en) | Short-term non-resident load prediction method and system based on improved neural network | |
| CN103049669A (en) | Method for forecasting water bloom and analyzing factors on basis of multivariate cyclostationary time sequence analysis and grey theory | |
| Al-Ja’afreh et al. | An enhanced CNN-LSTM based multi-stage framework for PV and load short-term forecasting: DSO scenarios | |
| Leng et al. | Integrated energy system evaluation method based on dimensionality reduction and indexes updating with incomplete information | |
| CN117634678A (en) | Low-carbon park carbon emission prediction method based on actual operation scene | |
| CN112508278A (en) | Multi-connected system load prediction method based on evidence regression multi-model | |
| CN117033923A (en) | Method and system for predicting crime quantity based on interpretable machine learning | |
| Chen et al. | Air quality prediction based on Kohonen Clustering and ReliefF feature selection | |
| CN115496264A (en) | Method for predicting generated power of wind turbine generator | |
| Zhao et al. | Real-time health status evaluation for electric power equipment based on cloud model | |
| CN113887831A (en) | Novel power load prediction influence factor correlation analysis method | |
| Su et al. | Combined-LSTM based user electricity consumption prediction in a smart grid system | |
| Ping et al. | Forecasting the Electricity Demand in Malaysia using ARIMA Model | |
| Hao et al. | Demand forecasting for rush repair spare parts of power equipment using fuzzy C-means clustering and the fuzzy decision tree | |
| CN117039843A (en) | Power load prediction method | |
| CN117350594A (en) | Electric energy substitution technology evaluation method and device based on cloud model-entropy weight method |
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 |