CN117787078A - Multi-element load prediction method based on ESAM-MTL model - Google Patents
Multi-element load prediction method based on ESAM-MTL model Download PDFInfo
- Publication number
- CN117787078A CN117787078A CN202311333610.9A CN202311333610A CN117787078A CN 117787078 A CN117787078 A CN 117787078A CN 202311333610 A CN202311333610 A CN 202311333610A CN 117787078 A CN117787078 A CN 117787078A
- Authority
- CN
- China
- Prior art keywords
- model
- esam
- mtl
- load prediction
- prediction
- 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 42
- 230000007246 mechanism Effects 0.000 claims abstract description 46
- 238000012549 training Methods 0.000 claims abstract description 33
- 239000011159 matrix material Substances 0.000 claims abstract description 8
- 238000000605 extraction Methods 0.000 claims abstract description 6
- 238000011156 evaluation Methods 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 15
- 238000012795 verification Methods 0.000 claims description 15
- 230000008569 process Effects 0.000 claims description 10
- 238000012360 testing method Methods 0.000 claims description 7
- 238000002474 experimental method Methods 0.000 claims description 5
- 238000012216 screening Methods 0.000 claims description 5
- 102100038591 Endothelial cell-selective adhesion molecule Human genes 0.000 claims description 3
- 101000882622 Homo sapiens Endothelial cell-selective adhesion molecule Proteins 0.000 claims description 3
- 238000009826 distribution Methods 0.000 claims description 3
- 238000005259 measurement Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 238000009827 uniform distribution Methods 0.000 claims description 3
- 238000002790 cross-validation Methods 0.000 claims description 2
- 238000013450 outlier detection Methods 0.000 claims description 2
- 238000011160 research Methods 0.000 description 7
- 238000004364 calculation method Methods 0.000 description 6
- 238000010276 construction Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000011176 pooling Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000005611 electricity Effects 0.000 description 2
- 238000005265 energy consumption Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 239000002699 waste material Substances 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 239000003245 coal Substances 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000004146 energy storage Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000003208 petroleum Substances 0.000 description 1
- 238000010248 power generation Methods 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000013432 robust analysis Methods 0.000 description 1
- 239000004575 stone Substances 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a multi-load prediction method of a multi-load prediction method model based on a high-efficiency self-attention mechanism and a multi-task learning model, which comprises the steps of constructing the high-efficiency self-attention mechanism based on the self-attention mechanism and sharing information among a plurality of load prediction tasks by utilizing the multi-task learning framework, wherein the high-efficiency self-attention mechanism reduces the time complexity of the traditional self-attention mechanism by not calculating the attention generated by redundant data based on the sparse characteristic of a self-attention matrix; the multi-task learning architecture considers the association characteristics existing among multiple loads and the sharing of the feature extraction work, predicts different loads as a plurality of subtasks, accelerates the training and predicting speed of the model, reduces the risk of overfitting, and enhances the generalization capability of the model. The invention utilizes the high-efficiency self-attention mechanism, considers the relevance of the multi-element load prediction task, and realizes the high-efficiency and accurate prediction of the multi-element load based on the ESAM-MTL multi-element load prediction model.
Description
Technical Field
The invention relates to the field of load prediction, in particular to a multi-element load prediction method based on an ESAM-MTL model.
Background
In recent decades, energy supply structures are actively regulated in various countries in the world, the use ratio of various clean energy sources including renewable energy sources is improved, the consumption proportion of primary energy sources such as petroleum, coal and the like is reduced, and the strategic significance of completing the construction of a novel energy system is more important when China is used as the country for producing and consuming energy sources. In order to complete the construction of a novel energy system, renewable energy sources are developed and utilized on a large scale, and the improvement of the energy utilization efficiency is the core content and the necessary choice of the energy system revolution in China at present. The integrated energy system (Integrated Energy Systems, IES) not only can overcome the defects of low energy utilization efficiency, poor flexibility and the like in the traditional energy system, but also is an ideal way for realizing energy cascade utilization and renewable energy consumption, and has great development prospect. The regional comprehensive energy system takes an electric power system as a core, and uses distributed power generation, renewable energy, energy storage and demand response to promote supply and demand interaction of various energy sources such as electricity, cold, heat, gas and the like so as to improve the regional energy utilization efficiency and the renewable energy consumption level.
The research on the multi-element load prediction of the regional comprehensive energy system is a basic stone of a theoretical system of the whole regional comprehensive energy system. The multi-element load prediction of the regional comprehensive energy system is based on historical data of various loads including electricity, cold, heat, gas and the like in the regional comprehensive energy system, and is aided with historical data of relevant factors such as weather information, calendar information, social information and the like, so that a change rule of various load time sequences is mined and analyzed, a correct mapping relation is established, and a demand value of various loads in the future is obtained. The prediction result plays a vital role in the operation of the regional comprehensive energy system, provides data support for energy regulation and control work, and has the research significance as follows: (1) Through multi-element load prediction, the demand conditions of various energy sources at the future moment can be known, and more economic decisions can be made in the energy market, so that the operation risk of the energy source market is reduced; (2) The multi-element load prediction is beneficial to planning resources and equipment required in the future, and the accurate prediction result can ensure the energy supply and demand balance, so that the energy utilization rate is improved. (3) And the operation optimization and control of the regional comprehensive energy system are facilitated through multi-element load prediction. In summary, in order to complete the construction of the novel energy system in China, the development of the regional comprehensive energy system becomes a necessary way, and the research of the multi-element load prediction of the comprehensive energy system becomes a research hot spot in the current energy field because the research is the foundation and the center of the research field of the regional comprehensive energy system.
The existing multi-element load prediction method has the problems of long training time of a self-attention mechanism model and memory overflow. The problem of excessive space-time complexity exists when the traditional self-attention mechanism is applied to the multi-element load prediction problem; the relevance of multiple loads is not considered for single load prediction, and weather and calendar data are extracted by one-time feature extraction when each load is predicted, so that the waste of calculation resources is caused. How to solve the above problems is a main goal of the skilled person.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a multi-element load prediction method based on an ESAM-MTL model, so as to overcome the defects of excessively high space-time complexity, long model training time, memory overflow and no consideration of relevance of multi-element loads and waste of calculation resources in load prediction.
The technical scheme for achieving the purpose is as follows: the multi-element load prediction method based on the ESAM-MTL model is characterized by comprising an efficient self-attention mechanism ESAM constructed based on a self-attention mechanism and a multi-task learning MTL framework, wherein the efficient self-attention mechanism is used for reducing the time complexity of a traditional self-attention mechanism by not calculating the attention generated by redundant data based on the sparse characteristic of a self-attention matrix; the multi-task learning architecture considers the association characteristics existing among multiple loads and the sharing of the feature extraction work, considers the prediction work of different loads as a plurality of subtasks, accelerates the training and prediction speed of the model, reduces the risk of overfitting, and enhances the generalization capability of the model.
Further, the high-efficiency self-attention mechanism includes screening lnL M (q i K) lowest q i And traditional self-attention mechanisms; measuring k (q) using KL divergence i ,k j ) The proximity of the distribution to the uniform distribution, and thus the measurement q i Redundancy of (2):
further, the screening method is q i Redundancy evaluation index function M (q i ,K)：
Wherein M (q) i K) by selecting lnL K j Calculated, M (q i The greater K) the greater q i The higher the redundancy of (a), the more efficient self-attention mechanism chooses only ln L M (q i K) lowest q i Calculate efficient self-attention EA (Q, K, V):
furthermore, the multi-task learning realizes an ESAM-MTL model by utilizing a hard sharing mechanism MTL architecture, the multi-element load prediction model based on ESAM-MTL consists of an input layer, a sharing layer, an independent layer and an output layer, a plurality of load prediction tasks share an encoder and a decoder, each of the load prediction tasks independently share a full connection layer, the characteristics in the input layer are extracted through the shared encoder and decoder, and the output of each load prediction task is generated through the independently shared full connection layer.
Further, the ESAM-MTL model comprises three parts, namely a training process, a prediction process and an evaluation index; the ESAM-STL model parameter has complex structure and adopts an MTL architecture based on a hard sharing mechanism.
Further, the training process of the ESAM-STL model comprises the following steps of (1) using a two-stage outlier detection method, considering a missing value filling method of a multi-element load characteristic to process missing values and outliers in historical data, and performing dimensionless processing on a processed data set; (2) Selecting main influencing factors of the multi-element load by utilizing a MIC combined minimum redundancy maximum correlation algorithm; (3) Setting super parameters of the model, a minimum verification error, an upper limit of iteration steps and an early termination threshold P; (4) After the training of each step of model is finished, evaluating the training result of the current model by using a verification set, and if the error is greater than or equal to the minimum verification error, stopping counting in advance and adding 1; otherwise, updating the minimum verification error and storing the current model; and if the verification errors of the continuous P times of training are all greater than or equal to the minimum verification error or the iteration times reach the upper limit, terminating the model training.
Further, the prediction process of the ESAM-MTL model comprises the following steps: (1) Acquiring real-time data required by prediction according to main influencing factors selected during model training; (2) The same method and parameters as those used in model training are used for carrying out dimensionless treatment on real-time main influence factor data; (3) And predicting the multi-element load by using a model with the minimum verification error during model training to obtain a multi-element load prediction result.
Further, the evaluation indexes of the ESAM-MTL model comprise a prediction error evaluation index and a model stability evaluation index, wherein the prediction error evaluation index is divided into an average absolute percentage error function, a root mean square error function and an average absolute error function.
Furthermore, the model stability evaluation index of the ESAM-MTL model divides the data set D into K parts by a K-fold cross validation method, one part is used as a test set, and the other K-1 parts are used as a training set, and an experiment is performed to obtain a test error. Repeating the above steps K times to obtain a set { e } i Using the set { e } i Sample standard deviation S as an evaluation index of model stability.
Compared with the prior art, the invention has the following technical effects:
1. on the basis of the Self-attention mechanism, a high-efficiency Self-attention mechanism (ESAM) is provided, compared with the traditional Self-attention mechanism, the Self-attention mechanism has lower space-time complexity, the problems of long training time, memory overflow and the like of a Self-attention mechanism model are solved, and information in multi-element load historical data can be more fully mined.
2. Compared with the existing single load prediction method, the method considers the relevance among multiple load prediction tasks, utilizes a multi-task learning (Mult i-TaskLearning, MTL) architecture to share information among the multiple load prediction tasks, and provides the multiple load prediction method based on the ESAM-MTL model, so that efficient and accurate prediction on multiple loads can be realized.
Drawings
FIG. 1 is a diagram of data set information used in the present invention;
FIG. 2 is a computational flow diagram of an efficient self-attention mechanism in the present invention;
FIG. 3 is a graph comparing the results of multiple load prediction with the single load prediction and the time of use in the present invention;
FIG. 4 is a graph of a multivariate load prediction model based on ESAM-MTL in the present invention.
Detailed Description
In order to better understand the technical solution of the present invention, the following detailed description is given by way of specific examples:
as shown in fig. 1 and 2, which are a preferred embodiment of the present invention, the example process mainly includes four parts of hyper-parameter setting, prediction result, multi-load prediction versus single load prediction result, and efficient self-attention versus traditional self-attention prediction result.
The invention provides a multi-element load prediction method based on an ESAM-MTL model, which is added with a multi-element load prediction and a high-efficiency self-attention mechanism, solves the problems of long training time, memory overflow and the like of the self-attention mechanism model, and realizes the high-efficiency and accurate prediction of the multi-element load. Based on the technical scheme of the invention, the detailed steps of model prediction are as follows:
firstly, dividing a training set, a verification set and a test set according to the sequence of 8:1:1 in two data sets, and respectively carrying out experiments of ultra-short-term load prediction and short-term load prediction on each data set, wherein the ultra-short-term load prediction predicts multiple loads at the next moment, and the short-term load prediction predicts multiple loads at each moment of the next day. Wherein the weight calculation is performed according to the following formula:
secondly, the super parameters of the model determine the specific structure of the model, and comprehensively consider the training time costAnd prediction accuracy, after multiple experimental tests, setting an initial learning rate theta 0 ＝3×10 -3 Adam optimization algorithm is adopted to adjust the learning rate, the batch size is set to be 32, the maximum iteration number is set to be 100, the early stop strategy is used for avoiding overfitting, the early termination threshold is set to be 5, and the rest super parameters are shown in table 1:
TABLE 1ESAM-MTL model Supermarameter settings
And thirdly, selecting a Naive model, a differential integration moving average autoregressive (ARIMA) model, a BP-NN model, a CNN-LSTM model and a transducer model as comparison models. Wherein the method comprises the steps ofThe model uses the load value of the last moment as a predicted value during ultra-short-term load prediction, and uses the load value of the same moment of the last day as a predicted value during short-term load prediction, and the ultra-parameter p of the ARIMA model enumerates intervals [0,4 ]]The integer in the range, super parameter d, q enumerates intervals [0,2 ]]The whole number in the model is set up to the initial learning rate theta by the other deep learning models 0 ＝1×10 -3 The learning rate is adjusted by adopting an Adam optimization algorithm, the batch size is set to be 32, the maximum iteration number is set to be 100, the early stop strategy is used for avoiding overfitting, the early termination threshold is set to be 5, and the prediction results of all models can be obtained:
(1) In the short-term load prediction experiment, the weighted RMSE and the weighted MAE of the ARIMA model are larger than those of the Naive model, so that the short-term load prediction problem is complex, and an ideal prediction result is difficult to obtain by using a traditional model;
(2) The weighted RMSE and the weighted MAE of the BP-NN model are lower than the Naive model, but higher than the CNN-LSTM model and the transducer model, because the BP-NN model has weaker extraction capacity to time characteristics than the CNN-LSTM model and the transducer model;
(3) The weighted RMSE and the weighted MAE of the CNN-LSTM model in the ultra-short-term load prediction task are smaller than those of the transducer model, but when the short-term load prediction task with the length of the input sequence increased is dealt with, the prediction effect of the transducer model is better, so that the CNN-LSTM model is more suitable for the ultra-short-term load prediction problem, and the transducer model is more suitable for the short-term load prediction problem;
(4) Compared with the optimal results in other models, the ESAM-MTL model has the advantages that the weighting RMSE is respectively reduced by 6.15%, 6.09%, 12.5% and 3.94% in four learning tasks, and the weighting MAE is respectively reduced by 2.9%, 4.62%, 3.91% and 1.34%, so that the ESAM-MTL model provided by the invention overcomes the limitation of the CNN-LSTM model in short-term load prediction and the converter model in ultra-short-term load prediction, and is more suitable for multi-load prediction work.
Fourth, in order to verify the validity of the multi-load prediction and multi-task learning structure, the present invention compares the prediction effect and time of the single load prediction model (ESAM-STL) and the multi-load prediction model (ESAM-MTL), and the result is shown in fig. 3. Wherein the super parameter settings of the ESAM-STL model are exactly the same as ESAM-MTL.
The ESAM-STL model consists essentially of the following three parts:
(1) An encoder: the function of the encoder is to extract features for load prediction from the encoder input, consisting of an input layer, an efficient self-attention layer and a pooling layer.
The ELU activation function in the high-efficiency self-attention layer is expressed as:
(2) A decoder: the decoder mainly extracts weather and calendar characteristics of the moment to be predicted and characteristics available for load prediction before the moment to be predicted, and performs query matching with the output of the encoder. The decoder is composed of an input layer, an efficient self-attention layer, a pooling layer and an attention layer, wherein the efficient self-attention layer and the pooling layer are the same as the encoder, and the attention layer uses a standard attention mechanism.
(3) Full tie layer: the full junction layer is located at the tail of the model and functions to integrate all features and give the prediction result. The fully-connected layer comprises two layers L1 and L2.
The input/output of the L1 layer is expressed as:
h L1 ＝f L1 (w L1 y de +b L1 )
the input/output of the L2 layer is expressed as:
h L2 ＝f L2 (w L2 h L1 +b L2 )
the activation function ReLU function is expressed as:
as shown in FIG. 4, the ESAM-MTL multi-element load prediction model consists of the following four parts: the load prediction system comprises an input layer, a sharing layer, an independent layer and an output layer, wherein a plurality of load prediction tasks share an encoder and a decoder, each load prediction task independently shares a full connection layer, features in the input layer are extracted through the shared encoder and decoder, and output of each load prediction task is generated through the independently shared full connection layer. The structure of the encoder and decoder of the ESAM-MTL model is the same as that of the ESAM-STL model, and the input is the union of the inputs of the ESAM-STL model.
As can be seen from fig. 3: (1) All multi-element load prediction results of the ESAM-MTL model are superior to those of the ESAM-STL model, and MAE of the ultra-short-term load prediction task of the data set 1 is taken as an example, compared with the ESAM-STL model, the ESAM-MTL model has the advantages that the prediction errors of electric, cold, hot and gas loads are respectively reduced by 9.66%, 18.28%, 16.97% and 12.7%, and the combined prediction of the multi-element load has the capability of improving the prediction precision compared with the isolated prediction of each single load under the background that the multi-element load has a correlation;
(2) Compared with the ESAM-STL model, the ESAM-MTL model respectively reduces 72.55%, 73.39%, 61.48% and 63.19% of training time in four learning tasks, which shows that the multi-task learning framework effectively reduces the time required by model training by sharing the characteristic extraction layer of multiple loads and independently extracting the characteristics compared with each load.
Fifth, to verify the effectiveness of the high-efficiency self-attention mechanism in reducing the space-time complexity, this section uses data set 2 as an example, and compares the ESAM-MTL model based on the high-efficiency self-attention mechanism with the SAM-MTL model based on the conventional self-attention mechanism, and the prediction results and the time are shown in table 2. Wherein the SAM-MTL model is obtained by replacing the high-efficiency self-attention layer in the ESAM-MTL model with a conventional self-attention layer.
The traditional self-attention mechanism calculation method comprises the following steps:
the calculation method of the high-efficiency self-attention mechanism comprises the following steps:
(1) Measuring k (q) using KL divergence i ,k j ) The proximity of the distribution to the uniform distribution, and thus the measurement q i Redundancy of (2):
(2) Select q i Redundancy evaluation index function M (q i ,K)：
(3) Screening lnL M (q) i K) lowest q i Computing efficient self-attention EA (Q, K, V)
The time complexity of the efficient self-attention mechanism is reduced compared to the traditional self-attention mechanism, enabling longer input sequences to be received. In the traditional self-attention mechanism, after matrix multiplication, normalization and normalization exponential functions are completed by a query matrix Q and a keyword matrix K, a value matrix V is added to complete matrix multiplication, and when the traditional self-attention is calculated, all qi and all kj are required to be matched and calculated.
As can be seen from table 2: (1) Under the same prediction length and encoder input length, the ESAM-MTL model has a slightly higher prediction error than the SAM-MTL model, because part of the features may be ignored when the ESAM-MTL model calculates the high-efficiency attention;
(2) When the prediction length is given, compared with the SAM-MTL model, the ESAM-MTL model can accept longer encoder input to obtain lower prediction errors, the historical data is more fully mined, and the high-efficiency self-attention mechanism can improve the maximum prediction precision of the model by reducing the space complexity;
(3) In three experiments in which no memory overflow occurs in the SAM-MTL model, compared with the SAM-STL model, the ESAM-MTL model is respectively reduced by 88.56%, 94.29% and 93.49% during training, which shows that the high-efficiency self-attention mechanism has lower time complexity compared with the traditional self-attention mechanism and can be better applied to the production practice of IES.
TABLE 2 efficient self-attention vs. traditional self-attention prediction results and time-of-use comparisons
Note that: ' indicate that training cannot be completed due to memory overflow
The final step is robustness analysis of the model, electric, cold, hot and gas loads in the comprehensive energy system are influenced by meteorological data, and prediction data of the meteorological data are needed in the execution process of multi-load prediction, so that multi-load prediction errors caused by prediction errors of the meteorological data cannot be ignored. In the existing weather forecast research, the 15min ultra-short term weather forecast error is less than 10%; the ultra-short term weather forecast error is less than 20% in 1 h; the short-term weather forecast error is less than 40%. Therefore, in order to explore the robustness of the ESAM-MTL model when the weather data errors are dealt with, 10%, 20%, 40% and 100% of random errors are respectively added to the weather data of the test set, and the multi-element load is predicted again, according to the calculation result, the weighted RMSE value of the ESAM-MTL model is respectively increased by only 0.1%, 0.49%, 0.01% and 0.45% after the weather prediction errors corresponding to the prediction tasks are added, and the value of the weighted MAE is respectively increased by only 0.04%, 0.26%, 0.01% and 0.16%, so that the ESAM-MTL model has enough robustness to deal with the weather prediction errors.
It will be appreciated by persons skilled in the art that the above embodiments are provided for illustration only and not for limitation of the invention, and that variations and modifications of the above described embodiments are intended to fall within the scope of the claims of the invention as long as they fall within the true spirit of the invention.
Claims (9)
1. The multi-element load prediction method based on the ESAM-MTL model is characterized by comprising an efficient self-attention mechanism ESAM constructed based on a self-attention mechanism and a multi-task learning MTL framework, wherein the efficient self-attention mechanism is used for reducing the time complexity of a traditional self-attention mechanism by not calculating the attention generated by redundant data based on the sparse characteristic of a self-attention matrix; the multi-task learning architecture considers the association characteristics existing among multiple loads and the sharing of the feature extraction work, considers the prediction work of different loads as a plurality of subtasks, accelerates the training and prediction speed of the model, reduces the risk of overfitting, and enhances the generalization capability of the model.
2. The method for multiple load prediction based on ESAM-MTL model according to claim 1, wherein said efficient self-attention mechanism comprises screening lnL M (q i K) lowest q i And traditional self-attention mechanisms; measuring k (q) using KL divergence i ,k j ) The proximity of the distribution to the uniform distribution, and thus the measurement q i Redundancy of (2):
3. the method for multiple load prediction based on ESAM-MTL model of claim 2 wherein the screening method is a q i Redundancy evaluation index function M (q i ,K)：
Wherein M (q) i K) by selecting lnL K j Calculated, M (q i The greater K) the greater q i The higher the redundancy of (a), the more efficient self-attention mechanism chooses only lnL M (q i K) lowest q i Calculate efficient self-attention EA (Q, K, V):
4. the multi-load prediction method based on the ESAM-MTL model according to claim 1, wherein the multi-task learning realizes the ESAM-MTL model by utilizing a hard sharing mechanism MTL framework, the multi-load prediction model based on the ESAM-MTL consists of an input layer, a sharing layer, an independent layer and an output layer, a plurality of load prediction tasks share an encoder and a decoder, each of the plurality of load prediction tasks independently share a full connection layer, features in the input layer are extracted through the shared encoder and decoder, and output of each load prediction task is generated through the independent full connection layer.
5. The multi-component load prediction method based on the ESAM-MTL model according to claim 4, wherein the ESAM-MTL model comprises three parts of a training process, a prediction process and an evaluation index; the ESAM-STL model parameter has complex structure and adopts an MTL architecture based on a hard sharing mechanism.
6. The method for predicting multiple loads based on ESAM-MTL model according to claim 5, wherein the training process of ESAM-STL model comprises (1) using two-stage outlier detection method, processing missing values and outliers in historical data by considering missing value filling method of multiple load characteristics, and performing dimensionless processing on the processed data set; (2) Selecting main influencing factors of the multi-element load by utilizing a MIC combined minimum redundancy maximum correlation algorithm; (3) Setting super parameters of the model, a minimum verification error, an upper limit of iteration steps and an early termination threshold P; (4) After the training of each step of model is finished, evaluating the training result of the current model by using a verification set, and if the error is greater than or equal to the minimum verification error, stopping counting in advance and adding 1; otherwise, updating the minimum verification error and storing the current model; and if the verification errors of the continuous P times of training are all greater than or equal to the minimum verification error or the iteration times reach the upper limit, terminating the model training.
7. The multi-component load prediction method based on the ESAM-MTL model according to claim 5, wherein the prediction flow of the ESAM-MTL model comprises the following steps: (1) Acquiring real-time data required by prediction according to main influencing factors selected during model training; (2) The same method and parameters as those used in model training are used for carrying out dimensionless treatment on real-time main influence factor data; (3) And predicting the multi-element load by using a model with the minimum verification error during model training to obtain a multi-element load prediction result.
8. The method for multi-component load prediction based on ESAM-MTL model according to claim 5, wherein the evaluation index of ESAM-MTL model comprises a prediction error evaluation index and a model stability evaluation index, wherein the prediction error evaluation index is divided into an average absolute percentage error function, a root mean square error function and an average absolute error function.
9. The method for predicting multiple loads based on ESAM-MTL model as set forth in claim 8, wherein the model stability evaluation index of ESAM-MTL model is obtained by dividing data set D into K parts by K-fold cross validation method, and one of the K parts is used as one partAnd (3) taking the test set and K-1 parts as training sets, and performing an experiment to obtain a test error. Repeating the above steps K times to obtain a set { e } i Using the set { e } i Sample standard deviation S as an evaluation index of model stability.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311333610.9A CN117787078A (en) | 2023-10-16 | 2023-10-16 | Multi-element load prediction method based on ESAM-MTL model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311333610.9A CN117787078A (en) | 2023-10-16 | 2023-10-16 | Multi-element load prediction method based on ESAM-MTL model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117787078A true CN117787078A (en) | 2024-03-29 |
Family
ID=90378953
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311333610.9A Pending CN117787078A (en) | 2023-10-16 | 2023-10-16 | Multi-element load prediction method based on ESAM-MTL model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117787078A (en) |
-
2023
- 2023-10-16 CN CN202311333610.9A patent/CN117787078A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108596242B (en) | Power grid meteorological load prediction method based on wavelet neural network and support vector machine | |
| CN109858673A (en) | A kind of photovoltaic generating system power forecasting method | |
| CN113128113B (en) | Lean information building load prediction method based on deep learning and transfer learning | |
| CN106295899B (en) | Wind power probability density Forecasting Methodology based on genetic algorithm Yu supporting vector quantile estimate | |
| CN109558975A (en) | A kind of integrated approach of a variety of prediction results of electric load probability density | |
| Ye et al. | Combined approach for short-term wind power forecasting based on wave division and Seq2Seq model using deep learning | |
| CN109636054A (en) | Solar energy power generating amount prediction technique based on classification and error combination prediction | |
| CN109376950A (en) | A kind of polynary Load Forecasting based on BP neural network | |
| CN109492818A (en) | Based on energy development and the entitled electricity demand forecasting method of Shapley value | |
| CN115238993A (en) | Reservoir group multi-objective optimization scheduling scheme decision system based on NSGA-II algorithm | |
| CN112149890A (en) | Comprehensive energy load prediction method and system based on user energy label | |
| CN113591368A (en) | Comprehensive energy system multi-energy load prediction method and system | |
| CN115238854A (en) | Short-term load prediction method based on TCN-LSTM-AM | |
| CN108694475B (en) | Short-time-scale photovoltaic cell power generation capacity prediction method based on hybrid model | |
| CN114091776A (en) | K-means-based multi-branch AGCNN short-term power load prediction method | |
| CN113887809A (en) | Power distribution network supply and demand balance method, system, medium and computing equipment under double-carbon target | |
| Liao et al. | Multicore parallel dynamic programming algorithm for short-term hydro-unit load dispatching of huge hydropower stations serving multiple power grids | |
| CN101930566A (en) | Hydrological experimental simulation system and method based on parallel system | |
| Yang et al. | China's Energy Demand Forecasting Based on the Hybrid PSO-LSSVR Model. | |
| CN108615091B (en) | Electric power meteorological load data prediction method based on cluster screening and neural network | |
| Fu et al. | A spatial forecasting method for photovoltaic power generation combined of improved similar historical days and dynamic weights allocation | |
| Chen et al. | A prediction model based on unbiased grey Markov for airport energy consumption prediction | |
| Yong et al. | Short-term building load forecasting based on similar day selection and LSTM network | |
| CN117787078A (en) | Multi-element load prediction method based on ESAM-MTL model | |
| CN115375014A (en) | Source-load combination probability prediction method, device and storage medium |
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 |