CN112819224B - Unit output prediction and confidence evaluation method based on deep learning fusion model - Google Patents
Unit output prediction and confidence evaluation method based on deep learning fusion model Download PDFInfo
- Publication number
- CN112819224B CN112819224B CN202110134578.6A CN202110134578A CN112819224B CN 112819224 B CN112819224 B CN 112819224B CN 202110134578 A CN202110134578 A CN 202110134578A CN 112819224 B CN112819224 B CN 112819224B
- Authority
- CN
- China
- Prior art keywords
- model
- unit
- data
- training
- power grid
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000004927 fusion Effects 0.000 title claims abstract description 55
- 238000011156 evaluation Methods 0.000 title claims abstract description 11
- 238000013135 deep learning Methods 0.000 title claims abstract description 10
- 238000000034 method Methods 0.000 claims abstract description 35
- 238000007781 pre-processing Methods 0.000 claims abstract description 22
- 238000004364 calculation method Methods 0.000 claims abstract description 15
- 238000012937 correction Methods 0.000 claims abstract description 9
- 238000012549 training Methods 0.000 claims description 61
- 238000002790 cross-validation Methods 0.000 claims description 9
- 238000012360 testing method Methods 0.000 claims description 9
- 238000004804 winding Methods 0.000 claims description 9
- 238000010977 unit operation Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 5
- 210000002569 neuron Anatomy 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 4
- 238000004088 simulation Methods 0.000 claims description 4
- 150000001875 compounds Chemical class 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 2
- 238000005457 optimization Methods 0.000 claims description 2
- 238000013528 artificial neural network Methods 0.000 claims 1
- 125000004122 cyclic group Chemical group 0.000 claims 1
- 238000010248 power generation Methods 0.000 claims 1
- 238000000513 principal component analysis Methods 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 7
- 239000010410 layer Substances 0.000 description 10
- 238000010586 diagram Methods 0.000 description 6
- 230000007774 longterm Effects 0.000 description 6
- 230000009467 reduction Effects 0.000 description 4
- 238000005096 rolling process Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 230000008034 disappearance Effects 0.000 description 2
- 238000004880 explosion Methods 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000001276 controlling effect Effects 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000011229 interlayer Substances 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- 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)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Resources & Organizations (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Marketing (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Operations Research (AREA)
- Entrepreneurship & Innovation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Game Theory and Decision Science (AREA)
- Quality & Reliability (AREA)
- Development Economics (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Primary Health Care (AREA)
- Supply And Distribution Of Alternating Current (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses a unit output prediction and confidence evaluation method based on a deep learning fusion model, which comprises the following steps: s1, acquiring historical data of the power grid target unit; s2, fusing model input data for preprocessing; s3, constructing a DIndRNN and RVM based depth fusion model; s4, predicting the unit output by adopting a fusion model and calculating a confidence coefficient; s5, correcting the AGC instruction by calling the fusion model; and S6, dynamically updating the fusion model. According to the method, the fusion model is constructed and called to finish the correction of the AGC calculation instruction, and the regulation and control unit can accurately execute the AGC calculation instruction value, so that the AGC control scheme realizes the expected control effect.
Description
Technical Field
The invention relates to a power system scheduling control method, in particular to a unit output prediction and confidence evaluation method based on a deep learning fusion model.
Background
Automatic Generation Control (AGC) has matured and gained widespread use over the years of research. With the continuous increase of new energy power stations in recent years, the influence of the uncertainty of the new energy power station output on the power grid dispatching is not negligible. In actual engineering, the output of a thermal power generating unit is adjusted to ensure the active balance of a power grid. However, in some areas, due to the influence of thermal power generating unit equipment and environment, the deviation of the execution result of the unit to the AGC instruction is often large, which not only causes that the AGC control scheme cannot achieve the expected control effect, but also causes the out-of-limit situation of a boundary connecting line due to poor instruction execution capability. Therefore, the development of the AGC instruction execution effect prediction and confidence evaluation method has important significance for optimizing the control effect of the AGC control technology in a high-proportion new energy power grid.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a unit output prediction and confidence evaluation method based on a deep learning fusion model.
In order to achieve the purpose, the invention adopts the following technical scheme:
the unit output prediction and confidence evaluation method based on the deep learning fusion model comprises the following steps:
s1, acquiring historical data of the power grid target unit;
s2, fusing model input data for preprocessing;
s3, constructing a DIndRNN and RVM based depth fusion model;
s4, predicting the unit output by adopting a fusion model and calculating a confidence coefficient;
s5, correcting the AGC instruction by calling the fusion model;
and S6, dynamically updating the fusion model.
In order to optimize the technical scheme, the specific measures adopted further comprise:
further, the history data acquired at step S1 includes: the method comprises the steps of AGC instruction value, AGC instruction issuing time, unit output, unit historical output, total load of a power grid where the unit is located, frequency of the power grid where the unit is located, air temperature of the location of the unit, unit winding current and unit bus voltage.
Furthermore, the time span for acquiring the historical data is 6 months, the sampling frequency is 1 minute, and 262080 data sets are obtained; the output of the data set is a training label, and the other characteristics are training data.
Further, the preprocessing process of the model input data in step S2 is as follows:
s21, retrieving a data missing value, and replacing the data missing value by using the average value of the moment before and after the missing value;
s22, using a normalization formula to divide all data into the range of 0-1,
the formula is as follows:
in the formula (I), the compound is shown in the specification,
is normalized value, X is the original value of the feature, XminFor the minimum of this feature in the training set, XmaxThe maximum value of the feature in the training set;
s23, determining the parameter optimization range of the KPCA model: the kernel function adopts a Gaussian kernel function;
s24, determining parameter values by using a random cross validation grid search method;
s25, establishing a model by using the parameters determined in the step S24.
Further, the random cross-validation grid search method in step S24 includes the following steps:
s241, selecting a group of parameter combinations, constructing a KPCA model by using the parameter combinations and processing data;
s242, the processed training data is processed according to the following steps of 9: 1, randomly dividing into a training set and a testing set in proportion;
s243, training the multilayer perceptron model by using a training set;
s244, testing the multilayer perceptron model by using the test set to obtain a mean square error;
s245, repeating the steps S242-S244 for 10 times to obtain 10 mean square errors;
s246, calculating the average value of 10 mean square errors as a model error under the parameter combination;
s247, repeating the steps S241-S246 by using a group of parameter combinations until all the parameter combinations are calculated to obtain model errors under each parameter combination;
and S248, comparing the model errors, and combining the parameters when the model error is minimum into the optimal parameters.
Further, step S3 specifically includes the following steps:
s31, determining the type of the hyperparameter of the DIndRNN model: the method comprises the steps of counting neurons in each layer, counting the number of network layers and time step length;
s32, determining the value range of the hyperparameter of the DIndRNN model; the neuron number of each layer is 1-100, the network layer number is 1-100, and the time step length is 1-144;
s33, determining a model hyper-parameter value by using a K-fold cross validation grid search method;
s34, establishing a DIndRNN model according to the determined hyper-parameters;
s35, adding a full connection layer at the end of the DIndRNN model for mapping the training characteristics back to the training labels;
s36, processing the data by using the preprocessing model in the step S2;
s37, training the model by using the preprocessed data set: training a DIndRNN model by using a random gradient descent method; using early-stop techniques to prevent overfitting; configuring learning rate exponential decay to improve the convergence speed of the model;
s38, inputting the preprocessed data into the DIndRNN model trained in the step S37 to obtain data characteristics;
s39, training the RVM model by using the data characteristics;
s310, connecting the DIndRNN model and the RVM model in sequence to form a fusion model.
Further, step S4 specifically includes the following steps:
s41, acquiring current data of the power grid unit, wherein the current data comprise an AGC instruction value, AGC instruction issuing time, historical output of the unit, total load of a power grid where the unit is located, frequency of the power grid where the unit is located, air temperature of the location where the unit is located, winding current of the unit and bus voltage of the unit, and form current power grid state data;
s42, inputting the current power grid state data into a preprocessing model;
s43, inputting the preprocessed current power grid state data into a fusion model, and giving a unit output predicted value and confidence coefficient of the next period by the fusion model;
s44, adding the unit output predicted value given in the step S43 into the historical output characteristics of the unit in the current power grid state data, and keeping the other characteristics unchanged to form new power grid state data;
s45, repeating the steps S42-S44, inputting new power grid state data into the preprocessing model and the fusion model, and obtaining output of the next period;
and S46, obtaining the predicted value trend of the unit output in a future period of time, and giving a confidence expression.
Further, the confidence of the next cycle obtained in step S43 is added to the confidence of the previous cycle by the square to obtain a new confidence.
Further, the step S5 specifically includes the following steps:
s51 execution deviation delta P of instruction of more than one period of AGCbeforeAs a reference, simulating incremental output of the unit in the deviation direction, wherein the incremental step length is lambdaΔ＝ΔPbefore/NΔ，NΔTaking 100, and generating a simulation instruction set;
s52, calculating a unit output predicted value corresponding to all the instructions in the instruction set by using the models in the steps S2 and S3;
and S53, taking the analog command corresponding to the predicted value closest to the AGC calculation command as an AGC command correction value, and performing issuing control.
Further, the dynamic update method of the fusion model in step S6 is as follows:
s61, setting a training server on line and configuring relevant software;
s62, the training server reads the unit operation data in real time by taking 1 minute as a unit from an API (application programming interface) provided by the power grid, wherein the unit operation data comprises an AGC (automatic gain control) instruction value, AGC instruction issuing time, unit historical output, total load of the power grid where the unit is located, frequency of the power grid where the unit is located, air temperature of the location of the unit, winding current of the unit and bus voltage of the unit;
s63, when the data are accumulated for 24 hours, training the preprocessing model and the fusion model again according to the steps S2 and S3 by using the acquired data;
and S64, after the training is finished, updating the preprocessing model and the fusion model into an AGC system and covering the original model.
The invention has the beneficial effects that: according to the method, the fusion model is constructed and called, the AGC calculation instruction is corrected, and the AGC calculation instruction value is accurately executed by the control unit, so that the AGC control scheme achieves the expected control effect.
Drawings
FIG. 1 is a flow chart of the method for predicting the output of the unit and evaluating the confidence based on the deep learning fusion model.
FIG. 2 is a flow chart of the present invention for building a data pre-processing model.
FIG. 3 is a schematic diagram of the fusion model training principle of the present invention.
Fig. 4 is a schematic diagram of the medium-and-long-term multi-step rolling prediction principle of the invention.
Fig. 5 is a flow chart of the random cross-validation method of the present invention.
FIG. 6 is a schematic diagram of the model dynamic update principle of the present invention.
FIG. 7 is a comparison of different model predictive analyses in an embodiment of the invention.
FIG. 8 is a diagram illustrating a long-term prediction result of the fusion model in the embodiment of the present invention.
FIG. 9 is a schematic diagram of probabilities of different intervals in the embodiment of the present invention.
Detailed Description
The present invention will now be described in further detail with reference to the accompanying drawings.
It should be noted that the terms "upper", "lower", "left", "right", "front", "back", etc. used in the present invention are for clarity of description only, and are not intended to limit the scope of the present invention, and the relative relationship between the terms and the terms is not limited by the technical contents of the essential changes.
The invention provides a unit output prediction and confidence evaluation method based on a deep learning fusion model, which comprises the following steps:
and S1, establishing an offline training server, and configuring related software tools in the server.
S2, the training server acquires historical data of a certain set through an API (application programming interface) provided by a power grid platform, wherein the historical data comprises an AGC (automatic gain control) instruction value, AGC instruction issuing time, current output of the set, historical output of the set, total load of a power grid where the set is located, frequency of the power grid where the set is located, air temperature of the location of the set, winding current of the set and bus voltage of the set, and nine characteristics are summed; the data time span is half a year, the sampling frequency is 1 minute, and the data are 262080 pieces in total; and taking the current output of the unit as a model training label, and taking the rest data as model training input.
S3, training the fusion model by using a training server, and specifically comprising the following steps:
s31, dividing all data into a range of 0-1 by using a normalization formula;
the formula is as follows:
in the formula (I), the compound is shown in the specification,
is normalized value, X is the original value of the feature, XminFor the minimum of this feature in the training set, XmaxThe feature maximum is in the training set.
S32, determining KPCA dimensionality reduction model hyperparameters by using a random cross validation grid search method, wherein the Gaussian kernel parameter sigma value is 0-1, and the dimensionality reduction dimension is 1-original data dimension, and establishing a dimensionality reduction model;
as shown in fig. 5, determining the parameter values by using the random cross-validation grid search method includes the following steps:
1) taking a group of parameter combinations, constructing a KPCA model by using the parameter combinations and processing data;
2) randomly dividing the processed training data into a training set and a test set according to the ratio of 9: 1;
3) training a multilayer perceptron model by using a training set;
4) testing the multilayer perceptron model by using a test set to obtain a mean square error;
5) repeating the steps S242-S244 for 10 times to obtain 10 mean square errors;
6) calculating the average value of 10 mean square errors as the model error under the parameter combination;
7) repeating the steps by taking a group of parameter combinations until all the parameter combinations are calculated to obtain model errors under each parameter combination;
8) and comparing the model errors, and combining the parameters when the model errors are minimum into optimal parameters.
S33, sequentially connecting the normalization formula and the KPCA model with the determined parameters to form a preprocessing model;
s34, preprocessing the data by using a preprocessing model;
s35, establishing a DIndRNN model, and finally accessing the model to a full-connection regression layer for mapping the features extracted by the DIndRNN model back to the training labels;
s36, training a DIndRNN model by using the preprocessed data, and training the DIndRNN model (including a full connection layer) by using a Stochastic Gradient Descent (SGD) method; using early-stop techniques to prevent overfitting; configuring learning rate exponential decay to improve the convergence speed of the model;
s37, removing a full connection layer from the trained DIndRNN model; extracting training data characteristics by using a trained DIndRNN model (without a full connection layer);
s38, training the RVM model by using the extracted features;
s39, fusing the DIndRNN model and the RVM model.
And S4, sending a model training completion signal to the AGC system.
And S5, downloading the model from the training server after the AGC system receives the signal, inputting the current real-time data into the model, and obtaining the unit output predicted value and the prediction confidence coefficient in a future period of time in a rolling prediction mode. As shown in fig. 4, the steps are as follows:
s51, acquiring current data of the power grid unit, wherein the current data comprise 8 characteristics including AGC instruction values, AGC instruction issuing time, historical unit output, total load of a power grid where the unit is located, frequency of the power grid where the unit is located, air temperature of the location where the unit is located, winding current of the unit and bus voltage of the unit, and the current data form the state data of the power grid;
s52, inputting the current power grid state data into a preprocessing model;
s53, inputting the preprocessed current power grid state data into a fusion model, and giving a unit output predicted value and confidence coefficient in the next period (after 1 minute) by the fusion model;
s54, adding the output predicted value of the unit given in the step S53 into the historical output characteristic of the unit in the current power grid state data, and keeping the other characteristics unchanged to form new power grid state data;
s55, inputting the new power grid state data into the preprocessing model and the fusion model to obtain the output of the next period (after 2 minutes);
s56, circulating in the way, obtaining the predicted value trend of the unit output in a period of time in the future, and giving confidence expression;
wherein, the fusion model gives a random quantity which accords with normal distribution and is expressed by mean value and variance; the unit output predicted value is a mean value, and the confidence coefficient is expressed as a variance; forming new grid state data in step S52 only needs to consider the prediction mean, and the new confidence in step S53 needs to be squared and superimposed with the confidence in the previous cycle, as follows:
let the random variable X, Y obey a normal distribution, X-N (μ)1，σ1 2)；Y～N(μ2，σ2 2) (ii) a X is the output distribution of the unit in the previous period, Y is the correction value calculated by the fusion model, X + Y is the output of the unit in the next period, and according to the normal distribution addition formula, X + Y follows the following distribution X + Y-N (mu)3，σ3 2) In which μ3＝μ1+μ2，σ3 2＝σ1 2+σ2 2(ii) a The variance of the raw data collected from the grid is 0.
S6, the training server continuously reads the power grid unit data, when the data are accumulated for 24 hours, the newly acquired data are used as training data, the model is retrained on the original model weight, and after the model training is finished, the signals are sent to the AGC system again; and circulating in this way, and realizing the dynamic update of the model.
S7, implementing instruction correction by an AGC application model, and specifically comprising the following steps:
s71, before the AGC system issues the calculation instruction, the execution deviation (delta P) of the instruction of more than one periodbefore) As a reference, simulating incremental output of the unit in the deviation direction, wherein the incremental step length is lambdaΔ＝ΔPbefore/NΔ(NΔTaking 100) and generating a new simulation instruction set according to the 100) of the simulation instruction;
and S72, calling the fusion model, calculating the output predicted values of the units corresponding to all the instructions in the instruction set, taking the analog instruction corresponding to the predicted value closest to the AGC calculation instruction as an AGC instruction correction value, and issuing and controlling.
The schematic diagram of the model dynamic update principle of the present invention is shown in fig. 6.
1) Setting up a training server on line and configuring relevant software;
2) the method comprises the steps that a training server reads unit operation data in real time by taking 1 minute as a unit from an API (application programming interface) provided by a power grid, wherein the unit operation data comprises an AGC (automatic gain control) instruction value, AGC instruction issuing time, unit historical output, total load of the power grid where a unit is located, frequency of the power grid where the unit is located, air temperature of the location of the unit, unit winding current and unit bus voltage;
3) when the data are accumulated for 24 hours, training the preprocessing model and the fusion model again according to the steps S2 and S3 by using the acquired data;
4) and after the training is finished, updating the model into an AGC system and covering the original model.
The specific embodiment of the unit output prediction and confidence evaluation method based on the deep learning fusion model is as follows:
in the example, the real data of a certain power plant is used as an experimental object, and a comparison graph of the multi-step prediction results of the fusion model, the RNN and the LSTM is shown in FIG. 7.
As can be seen from fig. 7, the triangular mark curve deviates from the true value seriously at the initial moment, because the RNN model is limited by the model structure, and neurons share parameters in time, so that the problems of gradient disappearance and explosion occur, and thus a large deviation occurs at the initial stage of multi-step rolling prediction.
The star-shaped labeled curve and the square-shaped labeled curve represent the prediction results of the LSTM and the fusion model respectively, and as can be seen from FIG. 7, the two curves can better fit the real curve in the initial 10 minutes, but the star-shaped labeled curve has a larger prediction error after 10 minutes, and although a forgetting mechanism is introduced into the LSTM network, gradient interlayer attenuation still occurs, so that long sequence information cannot be captured by the star-shaped labeled curve and the square-shaped labeled curve. While the square signature curves generally fit the true curve well within the first 30 minutes despite slight deviations from the true curve (but no more than 5% deviation) within 10 to 30 minutes.
Therefore, it can be seen from the above analysis that the fitting effect of the fusion model of the present invention is slightly better than that of the other two depth models in the short-term prediction, and particularly the middle-term and long-term prediction capabilities are significantly better than that of the other two depth models, because the Relu unsaturated activation function is introduced into the DIndRNN layer in the fusion model, the problems of gradient disappearance and explosion are well solved, and therefore, the fusion model introduced by the present invention can further improve the middle-term and long-term prediction accuracy.
The results of the multi-step confidence evaluation are shown in FIG. 8.
In fig. 8, the light-to-dark colored regions correspond to the random value intervals of 99.73%, 95.45%, and 68.27%, respectively; it can be seen from the figure that with the increase of the prediction duration, the probability distribution curve gradually becomes gentle, the variance of the model prediction result follows the square superposition principle and gradually increases, the value intervals of different probabilities are enlarged, and the model has the characteristics of high short-term prediction accuracy and low long-term prediction accuracy.
In fig. 8, the longitudinal curves are predicted value probability distribution curves at different positions, and the probability distribution curves tend to be flat with the increase of the prediction duration. As can be seen from the figure, the fusion model has better long-term prediction capability and can better approximate the change of the real value in the first 30 minutes. With the advance of the time step, the probability distribution curve of the predicted value tends to be smooth, and the variance of the predicted value is enlarged.
The curve in fig. 9 is the probability that the true value falls within the range of the predicted value's fixed deviation. It can be seen from the figure that as the rolling prediction progresses, the probability that the actual value falls within each value range decreases. The probability is rapidly reduced greatly in the first ten minutes, and then the reduction speed tends to be flat, which means that the short-term prediction reliability is high and the long-term prediction reliability is low.
An example of the AGC system call model for instruction correction is as follows:
taking the actual operation section of a certain provincial power grid as an example, the AGC calculation instruction is as follows: the unit a is regulated to power 304.34MW (i.e., command a). Table 1 gives the actual output variation after the unit has executed command a (second column). As can be seen from the table, the unit output gradually becomes stable after 6 minutes, the average output in the stable period is 309.04MW, the output in the 10 th minute is 309.29MW, the difference value with the calculated issued command is 4.95MW, and the control deviation is large. Step S7 is adopted to correct the AGC calculation instruction, the instruction correction value is 302.67MW (namely instruction B), the output of the unit is 304.69MW after 10 minutes, the difference with the calculation instruction value issued by the AGC is only 0.35MW, the calculation instruction is basically consistent with the AGC calculation instruction, and the third column of the table 1 gives the prediction execution result.
TABLE 1 actual and predicted output of unit
From the analysis, the AGC system finishes the correction of the AGC calculation instruction by calling the fusion model, and further realizes the effect of accurately executing the AGC calculation instruction value by the control unit.
The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above-mentioned embodiments, and all technical solutions belonging to the idea of the present invention belong to the protection scope of the present invention. It should be noted that modifications and embellishments within the scope of the invention may be made by those skilled in the art without departing from the principle of the invention.
Claims (10)
1. The unit output prediction and confidence evaluation method based on the deep learning fusion model is characterized by comprising the following steps of:
s1, acquiring historical data of the power grid target unit;
s2, fusing model input data for preprocessing;
s3, constructing a depth fusion model based on DIndRNN and RVM, wherein the DIndRNN is a depth independent cyclic neural network, and the RVM is a relevance vector machine;
s4, predicting the unit output by adopting a fusion model and calculating a confidence coefficient;
s5, correcting an AGC instruction by calling a fusion model, wherein AGC is automatic power generation control;
and S6, dynamically updating the fusion model.
2. The method for crew contribution prediction and confidence assessment according to claim 1, wherein the historical data obtained in step S1 comprises: the method comprises the steps of AGC instruction value, AGC instruction issuing time, unit output, unit historical output, total load of a power grid where the unit is located, frequency of the power grid where the unit is located, air temperature of the location of the unit, unit winding current and unit bus voltage.
3. The unit capacity prediction and confidence assessment method according to claim 2, wherein the time span for obtaining the historical data is 6 months, the sampling frequency is 1 minute, and 262080 data sets are obtained; the output of the data set is a training label, and the other characteristics are training data.
4. The method for machine output prediction and confidence assessment according to claim 1, wherein the preprocessing of the model input data in step S2 is as follows:
s21, retrieving a data missing value, and replacing the data missing value by using the average value of the moment before and after the missing value;
s22, using a normalization formula to divide all data into the range of 0-1,
the formula is as follows:
in the formula (I), the compound is shown in the specification,
is normalized value, X is the original value of the feature, XminFor the minimum of this feature in the training set, XmaxThe maximum value of the feature in the training set;
s23, determining the parameter optimization range of the KPCA model, wherein the KPCA is a kernel principal component analysis: the kernel function adopts a Gaussian kernel function;
s24, determining parameter values by using a random cross validation grid search method;
s25, establishing a model by using the parameters determined in the step S24.
5. The method for unit capacity prediction and confidence assessment according to claim 4, wherein the random cross validation grid search method in step S24 comprises the following steps:
s241, selecting a group of parameter combinations, constructing a KPCA model by using the parameter combinations and processing data;
s242, the processed training data is processed according to the following steps of 9: 1, randomly dividing into a training set and a testing set in proportion;
s243, training the multilayer perceptron model by using a training set;
s244, testing the multilayer perceptron model by using the test set to obtain a mean square error;
s245, repeating the steps S242-S244 for 10 times to obtain 10 mean square errors;
s246, calculating the average value of 10 mean square errors as a model error under the parameter combination;
s247, repeating the steps S241-S246 by using a group of parameter combinations until all the parameter combinations are calculated to obtain model errors under each parameter combination;
and S248, comparing the model errors, and combining the parameters when the model error is minimum into the optimal parameters.
6. The method for crew contribution prediction and confidence assessment according to claim 1, wherein step S3 comprises the following steps:
s31, determining the type of the hyperparameter of the DIndRNN model: the method comprises the steps of counting neurons in each layer, counting the number of network layers and time step length;
s32, determining the value range of the hyperparameter of the DIndRNN model; the neuron number of each layer is 1-100, the network layer number is 1-100, and the time step length is 1-144;
s33, determining a model hyper-parameter value by using a K-fold cross validation grid search method;
s34, establishing a DIndRNN model according to the determined hyper-parameters;
s35, adding a full connection layer at the end of the DIndRNN model for mapping the training characteristics back to the training labels;
s36, processing the data by using the preprocessing model in the step S2;
s37, training the model by using the preprocessed data set: training a DIndRNN model by using a random gradient descent method; using early-stop techniques to prevent overfitting; configuring learning rate exponential decay to improve the convergence speed of the model;
s38, inputting the preprocessed data into the DIndRNN model trained in the step S37 to obtain data characteristics;
s39, training the RVM model by using the data characteristics;
s310, connecting the DIndRNN model and the RVM model in sequence to form a fusion model.
7. The method for crew contribution prediction and confidence assessment according to claim 1, wherein step S4 comprises the following steps:
s41, acquiring current data of the power grid unit, wherein the current data comprise an AGC instruction value, AGC instruction issuing time, historical output of the unit, total load of a power grid where the unit is located, frequency of the power grid where the unit is located, air temperature of the location where the unit is located, winding current of the unit and bus voltage of the unit, and form current power grid state data;
s42, inputting the current power grid state data into a preprocessing model;
s43, inputting the preprocessed current power grid state data into a fusion model, and giving a unit output predicted value and confidence coefficient of the next period by the fusion model;
s44, adding the unit output predicted value given in the step S43 into the historical output characteristics of the unit in the current power grid state data, and keeping the other characteristics unchanged to form new power grid state data;
s45, repeating the steps S42-S44, inputting new power grid state data into the preprocessing model and the fusion model, and obtaining output of the next period;
and S46, obtaining the predicted value trend of the unit output in a future period of time, and giving a confidence expression.
8. The method for machine output prediction and confidence assessment according to claim 7, wherein the confidence level of the next cycle obtained in step S43 is added to the confidence level of the previous cycle by a square method to obtain a new confidence level.
9. The method for unit capacity prediction and confidence assessment according to claim 1, wherein step S5 comprises the following steps:
s51 execution deviation delta P of instruction of more than one period of AGCbeforeAs a reference, simulating incremental output of the unit in the deviation direction, wherein the incremental step length is lambdaΔ＝ΔPbefore/NΔ，NΔTaking 100, and generating a simulation instruction set;
s52, calculating a unit output predicted value corresponding to all the instructions in the instruction set by using the models in the steps S2 and S3;
and S53, taking the analog command corresponding to the predicted value closest to the AGC calculation command as an AGC command correction value, and performing issuing control.
10. The method for machine output prediction and confidence assessment according to claim 1, wherein the dynamic updating method of the fusion model in step S6 is as follows:
s61, setting a training server on line and configuring relevant software;
s62, the training server reads the unit operation data in real time by taking 1 minute as a unit from an API (application programming interface) provided by the power grid, wherein the unit operation data comprises an AGC (automatic gain control) instruction value, AGC instruction issuing time, unit historical output, total load of the power grid where the unit is located, frequency of the power grid where the unit is located, air temperature of the location of the unit, winding current of the unit and bus voltage of the unit;
s63, when the data are accumulated for 24 hours, training the preprocessing model and the fusion model again according to the steps S2 and S3 by using the acquired data;
and S64, after the training is finished, updating the preprocessing model and the fusion model into an AGC system and covering the original model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110134578.6A CN112819224B (en) | 2021-01-29 | 2021-01-29 | Unit output prediction and confidence evaluation method based on deep learning fusion model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110134578.6A CN112819224B (en) | 2021-01-29 | 2021-01-29 | Unit output prediction and confidence evaluation method based on deep learning fusion model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112819224A CN112819224A (en) | 2021-05-18 |
| CN112819224B true CN112819224B (en) | 2021-10-08 |
Family
ID=75860956
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110134578.6A Active CN112819224B (en) | 2021-01-29 | 2021-01-29 | Unit output prediction and confidence evaluation method based on deep learning fusion model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112819224B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116436160B (en) * | 2023-03-29 | 2023-12-22 | 华能威海发电有限责任公司 | AGC performance index on-line monitoring system and method |
| CN116760026B (en) * | 2023-06-29 | 2024-05-17 | 南京工程学院 | Accurate prediction method for unit output considering parameter unrecognizable |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104933516A (en) * | 2015-05-27 | 2015-09-23 | 华南理工大学 | Multi-time-scale power system robustness scheduling system design method |
| CN105207253B (en) * | 2015-09-15 | 2017-11-21 | 重庆大学 | Consider wind-powered electricity generation and the probabilistic AGC stochastic and dynamics Optimization Scheduling of frequency |
| CN111738512A (en) * | 2020-06-22 | 2020-10-02 | 昆明理工大学 | Short-term power load prediction method based on CNN-IPSO-GRU hybrid model |
| CN112116162A (en) * | 2020-09-26 | 2020-12-22 | 国家电网公司华中分部 | Power transmission line icing thickness prediction method based on CEEMDAN-QFAOA-LSTM |
-
2021
- 2021-01-29 CN CN202110134578.6A patent/CN112819224B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104933516A (en) * | 2015-05-27 | 2015-09-23 | 华南理工大学 | Multi-time-scale power system robustness scheduling system design method |
| CN105207253B (en) * | 2015-09-15 | 2017-11-21 | 重庆大学 | Consider wind-powered electricity generation and the probabilistic AGC stochastic and dynamics Optimization Scheduling of frequency |
| CN111738512A (en) * | 2020-06-22 | 2020-10-02 | 昆明理工大学 | Short-term power load prediction method based on CNN-IPSO-GRU hybrid model |
| CN112116162A (en) * | 2020-09-26 | 2020-12-22 | 国家电网公司华中分部 | Power transmission line icing thickness prediction method based on CEEMDAN-QFAOA-LSTM |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112819224A (en) | 2021-05-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108764568B (en) | Data prediction model tuning method and device based on LSTM network | |
| CN112819224B (en) | Unit output prediction and confidence evaluation method based on deep learning fusion model | |
| CN110443447B (en) | Method and system for adjusting power system load flow based on deep reinforcement learning | |
| CN111461463A (en) | Short-term load prediction method, system and equipment based on TCN-BP | |
| CN113988481B (en) | Wind power prediction method based on dynamic matrix prediction control | |
| CN114022021A (en) | Online unit load prediction method based on ensemble learning | |
| CN112330012B (en) | Building energy consumption prediction method and equipment based on transfer learning | |
| CN104036328A (en) | Self-adaptive wind power prediction system and prediction method | |
| CN113919545A (en) | Photovoltaic power generation power prediction method and system with integration of multiple data models | |
| CN111290282B (en) | Predictive control method for thermal power generating unit coordination system | |
| CN114648176A (en) | Wind-solar power consumption optimization method based on data driving | |
| CN112651519A (en) | Secondary equipment fault positioning method and system based on deep learning theory | |
| Zaman et al. | Wind speed forecasting using ARMA and neural network models | |
| CN114154753A (en) | Load prediction method and system | |
| CN116880191A (en) | Intelligent control method of process industrial production system based on time sequence prediction | |
| Wang et al. | How deep is deep enough for deep belief network for approximating model predictive control law | |
| CN116562454A (en) | Manufacturing cost prediction method applied to BIM long-short-time attention mechanism network | |
| CN113836819B (en) | Bed temperature prediction method based on time sequence attention | |
| CN114759579A (en) | Power grid active power optimization control system, method and medium based on data driving | |
| CN115441475A (en) | Power emergency control method for power transmission line of power grid | |
| CN111950811B (en) | Regional photovoltaic power prediction method and system based on double-layer artificial neural network | |
| CN113935600A (en) | Deep learning-based adaptive economic dispatching system and method | |
| CN112183814A (en) | Short-term wind speed prediction method | |
| CN112615364A (en) | Novel wide-area intelligent cooperative control method for power grid stability control device | |
| CN116722548B (en) | Photovoltaic power generation prediction method based on time sequence model and related equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |