CN117474151A - Method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion model - Google Patents

Method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion model Download PDF

Info

Publication number
CN117474151A
CN117474151A CN202311406549.6A CN202311406549A CN117474151A CN 117474151 A CN117474151 A CN 117474151A CN 202311406549 A CN202311406549 A CN 202311406549A CN 117474151 A CN117474151 A CN 117474151A
Authority
CN
China
Prior art keywords
time period
cnn
lstm
lightgbm
weight
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202311406549.6A
Other languages
Chinese (zh)
Inventor
夏勇军
陈莉娟
徐文
阮羚
吴颖波
施志勇
赵立华
卫婧怡
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hubei Central China Technology Development Of Electric Power Co ltd
State Grid Hubei Electric Power Co Ltd
Original Assignee
Hubei Central China Technology Development Of Electric Power Co ltd
State Grid Hubei Electric Power Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hubei Central China Technology Development Of Electric Power Co ltd, State Grid Hubei Electric Power Co Ltd filed Critical Hubei Central China Technology Development Of Electric Power Co ltd
Priority to CN202311406549.6A priority Critical patent/CN117474151A/en
Publication of CN117474151A publication Critical patent/CN117474151A/en
Pending legal-status Critical Current

Links

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
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

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

Abstract

A method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion models, comprising: calculating a first average absolute percentage error per day in a first time period by adopting a CNN-LSTM algorithm; calculating a second average absolute percentage error per day in the first time period using a lightGBM algorithm; calculating a CNN-LSTM weight and a lightGBM weight according to the first average absolute percentage error and the second average absolute percentage error, and further calculating a CNN-LSTM dynamic weight and a lightGBM dynamic weight; and obtaining a final predicted value of the second time period corresponding to each day according to the predicted value of the second time period predicted by the CNN-LSTM algorithm and the lightGBM algorithm and the CNN-LSTM dynamic weight and the lightGBM dynamic weight. The invention can improve the prediction stability of the model and reduce the average error.

Description

Method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion model
Technical Field
The invention relates to the field of artificial intelligence algorithms, in particular to a method for actual prediction of short-term load based on a CNN-LSTM and lightGBM dynamic weighted fusion model.
Background
The actual prediction of the short-term load (which means that the actual future value can be predicted) is the basis of the economic operation of the power system, and can support the dispatching of a power supply company and improve the economical efficiency and the safety of the operation of a power grid. The existing method is generally realized by adopting a single algorithm based on statistics, machine learning, artificial intelligence and the like. However, due to the influence of emergencies and the limitation of the operation mechanism of the single model, errors of the prediction days (such as 3 rd day, 6 th day, etc.) suddenly rise and suddenly fall in the actual prediction process, inaccurate prediction is caused from the first holiday to the holiday, and the like, so that a difficulty is brought to the actual application.
Disclosure of Invention
In view of the above, the invention provides a method for actual prediction of short-term load based on CNN-LSTM and lightGBM dynamic weighted fusion model, which can improve the prediction stability of the model and reduce average error.
A method for short-term load actual prediction based on a CNN-LSTM and lightGBM dynamic weighted fusion model comprises the following steps:
carrying out real prediction on the daily load in a first time period and a second time period by adopting a CNN-LSTM algorithm, and obtaining a first average absolute percentage error of each day in the first time period according to a predicted value of the first time period predicted by the CNN-LSTM algorithm and a real value of the first time period, wherein the first time period comprises at least 2 time segments, and each time segment is equal to the number of days in the second time period;
carrying out real prediction on the daily load in the first time period and the second time period by adopting a lightGBM algorithm, and obtaining a second average absolute percentage error of each day in the first time period according to the predicted value of the first time period predicted by the lightGBM algorithm and the real value of the first time period;
obtaining a CNN-LSTM weight and a lightGBM weight of each day in a first time period according to the first average absolute percentage error and the second average absolute percentage error;
obtaining a CNN-LSTM dynamic weight and a lightGBM dynamic weight of each day in the first time period according to the CNN-LSTM weight and the lightGBM weight of each day in the first time period;
and obtaining a final predicted value of each day corresponding to the second time period according to the predicted value of the second time period predicted by the CNN-LSTM algorithm and the lightGBM algorithm and the CNN-LSTM dynamic weight and the lightGBM dynamic weight of each day in the first time period.
Further, the CNN-LSTM weight and lightGBM weight of each day in the first period of time are obtained according to the first average absolute percentage error and the second average absolute percentage error, and the specific calculation formula is as follows:
wherein W is 1 For CNN-LSTM weight, W per day in the first time period 2 For the lightGBM weight, e, per day in the first time period 1 E is the first average absolute percentage error per day in the first period of time 2 Is the second average absolute percentage error per day over the first period of time.
Further, the obtaining the CNN-LSTM dynamic weight and the lightGBM dynamic weight of each day in the first period according to the CNN-LSTM weight and the lightGBM weight of each day in the first period, includes:
and obtaining the CNN-LSTM dynamic weight by averaging the CNN-LSTM weight of the corresponding date of each time segment in the first time period, and obtaining the lightGBM dynamic weight by averaging the lightGBM weight of the corresponding date of each time segment in the first time period.
Further, the predicted value of the second time period predicted according to the CNN-LSTM algorithm and the lightGBM algorithm, and the CNN-LSTM dynamic weight and the lightGBM dynamic weight of each day in the first time period, obtain a final predicted value of the second time period corresponding to each day, and the specific calculation formula is as follows:
F t ＝W 1 f1 t +W 2 f2 t t＝1，2，...，n
wherein F is t Representing the final predicted value of the second time period corresponding to each day, f1 t And f2 t Representing predicted values, W, of the second time period predicted by the CNN-LSTM algorithm and the LightGBM algorithm, respectively 1 And W is 2 The CNN-LSTM dynamic weight and the lightGBM dynamic weight of each day in the first time period are respectively represented, t represents the number of days sequence of each time segment in the first time period and the second time period, and n represents the number of days of each time segment in the second time period.
The invention has the core that 2 algorithms are dynamically weighted and fused, and the method can effectively relieve the 4-point defect of the existing algorithm model:
1. the stability of a single algorithm is poor. The sensitivity of each single algorithm to prediction errors of different holidays, emergencies and the like is different, so that the errors are suddenly high and suddenly low, the stability is poor, and the robustness of the model after dynamic weighting fusion is enhanced;
2. serial fusion has a high security risk. Serial mode refers to the running order among 2 or more algorithms, each algorithm performing only a portion of its tasks. Parallel mode refers to running results in parallel between 2 or more algorithms, each individual algorithm being able to run results. When any algorithm in the combined model fails due to the influence of various factors such as external and internal boundaries, the model in the serial mode can be stopped and swung directly, and the model in the parallel mode can still obtain a result although the error is possibly reduced (if 2 algorithms are used, the other algorithm can still output the result directly, and if a plurality of algorithms are used, the residual algorithms do not influence parallel fusion).
3. The weighting value is fixed. The existing weighting method has the weighting value of a fixed value mostly, and the weighting value is dynamic, namely different every day, and the dynamic weight is more beneficial to reducing errors due to timely change every moment.
4. And cannot be predicted practically. In the fusion process of the existing literature, the data at the unknown moment cannot be weighted, and only the existing historical data can be weighted. This results in insufficient model utility and an inability to truly predict future values. Not intended to be predicted, but not predicted. Because in the fusion process, the error is needed to be known for weighting, but in the actual prediction, the error cannot be calculated because the future value is not known. In order to solve the problem, the invention calculates the errors of 2 algorithms in historical data, calculates the weights of 2 algorithms at the future time according to the daily error expression on the historical data based on the characteristic that the errors of rolling cycle prediction and multi-model machine learning prediction are different in each day.
Drawings
FIG. 1 is a flow chart of a method of the present invention for short-term load actual prediction based on CNN-LSTM and LightGBM dynamic weighted fusion models;
FIG. 2 is a graph of predicted results using the CNN-LSTM algorithm in accordance with an embodiment of the present invention;
FIG. 3 is a graph of predicted results using the LightGBM algorithm according to an embodiment of the present invention;
FIG. 4 is a graph of the results of the CNN-LSTM weights and lightGBM weights calculated by an embodiment of the present invention for a first time period per day;
FIG. 5 is a calculation of dynamic weighting in accordance with an embodiment of the present invention;
FIG. 6 is a table of final fusion results calculated by an embodiment of the present invention;
fig. 7 is a representation of the final fusion result calculated by an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The whole flow of the embodiment of the invention comprises 2 major parts: CNN-LSTM actual prediction model and lightGBM actual prediction model. And then carrying out dynamic weighted fusion on the results of the 2 models, wherein the determination of the dynamic weight is the core of the invention. The CNN-LSTM actual prediction model and the lightGBM actual prediction model are all actually predicted 5 times (3.1-3.7, 4.1-4.7,5.1-5.7,6.1-6.7, and 6.8-6.14), the first 4 times are used for calculating dynamic weighting values (the prediction process is strictly performed according to the actual predicted scene, and the actual value comparison is added after the event in order to calculate the map), and the last 1 times are the final results.
Wherein the actual prediction results A1 and B1 are the results of No. 3.1-3.7, no. 4.1-4.7, no. 5.1-5.7 and No. 6.1-6.7; the actual predicted results A2 and B2 are the results of 6.8-6.14, and the final result generated by weighting and fusing the 2 results is the result of 6.8-6.14.
Referring to fig. 1, an embodiment of the present invention provides a method for short-term load actual prediction based on a CNN-LSTM and lightGBM dynamic weighted fusion model, which includes the following steps:
step 1, carrying out real prediction on daily loads in a first time period (No. 3.1-3.7, no. 4.1-4.7, no. 5.1-5.7, no. 6.1-6.7) and a second time period (No. 6.8-6.14) by adopting a CNN-LSTM algorithm, and obtaining a first average absolute percentage error of each day in the first time period according to a predicted value of the first time period predicted by the CNN-LSTM algorithm and a real value of the first time period, wherein the first time period comprises at least 2 time segments (4 time segments in the embodiment are respectively No. 3.1-3.7, no. 4.1-4.7, no. 5.1-5.7 and No. 6.1-6.7), and each time segment is equal to the number of days in the second time period (No. 6.8-6.14);
the process of predicting by adopting the CNN-LSTM algorithm in the step 1 is as follows:
1.1 acquiring data
The data set originates from a data center
The characteristic dimension of the data center table is as follows: date, electricity usage, of a certain area 20220101-20230630.
1.2 data cleaning
Processing the vacant data, and filling each row of vacant data by adopting a relatively convenient average filling method;
the data is normalized 0-1 such that the entire column of data is compressed to within the range of 0-1.
1.3 modeling
Data set partitioning:
the first 60% data set was used for training, 20% data set for testing, and the last 20% data set for validation;
setting parameters:
in the construction of the CNN convolution network, one-dimensional convolution Conv1D is adopted, pooling layer downsampling is used, the pooling kernel size is [1,2], namely downsampling is only carried out on the characteristic dimension, and the sequence window is kept unchanged. The channel functions need to be adjusted, the channel information fused, the number of channels reduced to 1, and then the channel dimensions squeezed out before the feature data is input from CNN to LSTM. Then, the prediction result is outputted through a full link layer with time series information (the sequence period is 14) of the data processed through the LSTM.
1.4 model results
The real predictions are respectively carried out on 3.1-3.7,4.1-4.7,5.1-5.7,6.1-6.7, and the parameters are adjusted for multiple times in the process.
Then, the real prediction is carried out on 6.8 to 6.14, as shown in figure 2,
1.5 obtaining a first average absolute percentage error e of each day in the first time period according to the predicted value of the first time period and the actual value of the first time period predicted by the CNN-LSTM algorithm 1 。
And 2, carrying out real prediction on the daily load in the first time period and the second time period by adopting a lightGBM algorithm, and obtaining a second average absolute percentage error of each day in the first time period according to the predicted value of the first time period predicted by the lightGBM algorithm and the real value of the first time period.
The specific implementation process of the step 2 is as follows:
2.1 obtaining data
The data set is derived from a data center and weather websites
The characteristic dimension of the data center table is as follows: date of a certain area 20220101-20230630, electricity consumption, power supply quantity, average load rate and photovoltaic power generation quantity.
The characteristic dimensions of the weather website are as follows: the date of a certain area 20220101-20230630, the highest temperature, the lowest temperature.
2.2 data cleaning
(1) Processing the vacant data, and filling each row of vacant data by adopting a relatively convenient average filling method;
(2) Calculating average load rate, wherein the load rate of the original data per day has 96 moments, and calculating average load rate per day;
(3) Calculating daily electric quantity data, wherein the original data is the daily electric quantity, power supply quantity, lost electric quantity and photovoltaic power generation quantity of each station area, and the daily electric quantity data are calculated by grouping and summing according to the date,
(4) The data is normalized 0-1 such that each column of data is compressed to within the range of 0-1.
2.3 engineering of features
A sliding window. Because the electricity utilization characteristic has certain periodicity, the power supply quantity, the load rate and the photovoltaic power generation quantity are respectively carried out on a sliding window before 7 days and a sliding window before 30 days, namely the current power consumption quantity, the power supply quantity before 7 days, the load rate and the photovoltaic power generation quantity.
Adding holidays, judging by adopting a Chinese_calendar module, wherein the holidays are 1, and the non-holidays/workdays are 0.
Adding air temperature, and converting the original highest temperature and lowest temperature, wherein the formula is as follows:
highest temperature 1=electricity consumption/(|highest temperature+50|)
Minimum temperature 1=electricity consumption/(|minimum temperature-50|)
2.4 modeling
(1) The data set is divided from 20220101 ～ 20230607 to about 500 pieces of data, the first 400 pieces are divided into training sets, and the last 100 pieces are divided into test sets.
(2) Setting the lightGBM algorithm parameters, wherein the tasks are regression type tasks and the past experience, the objective selects the regression, the boosting_type selects the gbdt, the max_depth selects 6, the learning_rate selects 0.05, and the rest parameters are mostly defaults.
(3) Since the fields use+1st to use+7st correspond to the dependent variables Y1 to Y7, training and prediction of use+1st to use+7st are sequentially required, and the obtained results are finally spliced.
2.5 results display
The real prediction and model verification are respectively carried out on 3.1-3.7,4.1-4.7,5.1-5.7,6.1-6.7, and the parameters are optimized for many times in the process.
The true predictions are then made for 6.8-6.14, as shown in fig. 3.
2.6 deriving a second average absolute percentage error e of each day in the first time period from the predicted value of the first time period predicted by the lightGBM algorithm and the actual value of the first time period 2 。
And step 3, obtaining the daily CNN-LSTM weight and the lightGBM weight in the first time period according to the first average absolute percentage error and the second average absolute percentage error.
After 3.1-3.7,4.1-4.7,5.1-5.7,6.1-6.7 are respectively and actually predicted before, the map error of each real prediction of 2 models is obtained, and a dynamic weighting value is required to be calculated, wherein a weighting formula is shown below, and the dynamic weighting is performed by adopting an error reciprocal method:
wherein W is 1 For CNN-LSTM weight per day in the first time period, i.e. weight 1, W 2 For the lightGBM weight per day in the first period, i.e. weight 2, e 1 E is the first average absolute percentage error per day in the first period of time 2 Is the second average absolute percentage error per day over the first period of time. Through e 1 And e 2 Can sequentially calculate the W of each day 1 And W is 2 As shown in fig. 4.
Step 4, according to the daily CNN-LSTM weight W in the first time period 1 And lightGBM weight W 2 The CNN-LSTM dynamic weight and the lightGBM dynamic weight are obtained daily in the first time period.
Step 3 calculates not the final weights, but also combines the fact that the accuracy of different models on different prediction days is different, for example, the rolling prediction model is more accurate on the first prediction day, and the prediction may be less accurate on several days after the prediction, so that the daily weights also need to be averaged:
for the first day of prediction, since CNN-LSTM is a rolling loop single step length prediction, lightGBM is a multi-model single step length prediction, and the first day of rolling loop prediction is the most accurate, it is possible to weight CNN-LSTM directly to 0.99, the weight of lightGBM to 0.01;
average weights W for t+2 days were calculated by averaging the next day of each prediction (i.e., averages e1 and e2 for four days 3.2,4.2,5.2,6.2) 1 And W is 2 ；
Average weights W for t+3 days were calculated by averaging the third day of each prediction (i.e., the averages e1 and e2 for four days 3.3,4.3,5.3,6.3) 1 And W is 2 ；
Similarly, average weights W for t+7 days were calculated by averaging the seventh day of each prediction (i.e., the averages e1 and e2 for the four days 3.7,4.7,5.7,6.7) 1 And W is 2 。
The dynamic weighting calculation is shown in fig. 5.
And 5, obtaining a final predicted value of each day corresponding to the second time period according to the predicted value of the second time period predicted by the CNN-LSTM algorithm and the lightGBM algorithm and the CNN-LSTM dynamic weight and the lightGBM dynamic weight of each day in the first time period.
The result of the real prediction of 6.8-6.14 is weighted according to the dynamic weight value to calculate the final predicted value, see the following formula:
F t ＝W 1 f1 t +W 2 f2 t t＝1,2,...,n
wherein f1 represents the true predicted value of CNN-LSTM in the range of 6.8-6.14, f2 represents the true predicted value of LightGBM in the range of 6.8-6.14, W 1 And W is 2 The final fusion result F can be finally calculated by the calculation in the above step, as shown in fig. 6 and 7.
From fig. 6, it can be calculated that the daily errors of 6.8-6.14 are 3.8%,7.2%,8.4%,4.1%,4.2%,3.7%,3.7%, respectively, and the average value of these 7 days is calculated to give a final map error of 5.0%, wherein the first day map error is 3.8%.
As can be seen from fig. 7, the lightGBM prediction models between 5.28 and 5.30 and between 5.31 and 6.5 perform less stably, with very large errors between 5.28 and 5.30 and very small errors between 5.31 and 6.5. The dashed line CNN-LSTM prediction model is contrary, the error is very small between 5.29 and 5.30, and the error is larger between 6.1 and 6.3. Because the result is fused with the predicted values of 2 models, the influence degree of a certain emergency and the like on different single models is different, and the combined model is more stable. The fusion predicted value is between 2 model predicted values (shown in 6.8-6.13) after weighting, so that the influence of the accidental occurrence of extreme values of a single model is avoided.
The invention has the following characteristics:
(1) Parallel mode fusion. When any algorithm in the combined model fails due to the influence of various factors such as the external and internal boundaries, the model in the serial mode can be stopped directly, and the model in the parallel mode can be reduced in error and still obtain a result.
(2) The weights are dynamic. The weight refinement is different every day, which is also the origin of the "dynamic" weighting, which is more advantageous to reduce errors due to the timely change of each moment than the fixed weight.
(3) Can be used for actual prediction. In the fusion process of the existing literature, the data at the unknown moment cannot be weighted, and only the existing historical data can be weighted. This results in insufficient model utility and an inability to truly predict future values. Not intended to be predicted, but not predicted. Because in the fusion process, the error is needed to be known for weighting, but in the actual prediction, the error cannot be calculated because the future value is not known. In order to solve the problem, the invention calculates the errors of 2 algorithms in historical data, calculates the weights of 2 algorithms at the future time according to the daily error expression on the historical data based on the characteristic that the errors of rolling cycle prediction and multi-model machine learning prediction are different in each day, so that the actual prediction can be carried out.
The foregoing is merely illustrative embodiments of the present invention, and the present invention is not limited thereto, and any changes or substitutions that may be easily contemplated by those skilled in the art within the scope of the present invention should be included in the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims (4)

1. A method for short-term load actual prediction based on a CNN-LSTM and lightGBM dynamic weighted fusion model, comprising the steps of:
carrying out real prediction on the daily load in a first time period and a second time period by adopting a CNN-LSTM algorithm, and obtaining a first average absolute percentage error of each day in the first time period according to a predicted value of the first time period predicted by the CNN-LSTM algorithm and a real value of the first time period, wherein the first time period comprises at least 2 time segments, and each time segment is equal to the number of days in the second time period;
carrying out real prediction on the daily load in the first time period and the second time period by adopting a lightGBM algorithm, and obtaining a second average absolute percentage error of each day in the first time period according to the predicted value of the first time period predicted by the lightGBM algorithm and the real value of the first time period;
obtaining a CNN-LSTM weight and a lightGBM weight of each day in a first time period according to the first average absolute percentage error and the second average absolute percentage error;
obtaining a CNN-LSTM dynamic weight and a lightGBM dynamic weight of each day in the first time period according to the CNN-LSTM weight and the lightGBM weight of each day in the first time period;
and obtaining a final predicted value of each day corresponding to the second time period according to the predicted value of the second time period predicted by the CNN-LSTM algorithm and the lightGBM algorithm and the CNN-LSTM dynamic weight and the lightGBM dynamic weight of each day in the first time period.
2. The method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion model according to claim 1, wherein: the CNN-LSTM weight and the lightGBM weight of each day in the first time period are obtained according to the first average absolute percentage error and the second average absolute percentage error, and the specific calculation formula is as follows:
wherein W is 1 For CNN-LSTM weight, W per day in the first time period 2 For the lightGBM weight, e, per day in the first time period 1 E is the first average absolute percentage error per day in the first period of time 2 Is the first day of the first time periodAverage absolute percentage error.
3. A method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion models according to claim 1 or 2, characterized in that: the method for obtaining the CNN-LSTM dynamic weight and the lightGBM dynamic weight of each day in the first time period according to the CNN-LSTM weight and the lightGBM weight of each day in the first time period comprises the following steps:
and obtaining the CNN-LSTM dynamic weight by averaging the CNN-LSTM weight of the corresponding date of each time segment in the first time period, and obtaining the lightGBM dynamic weight by averaging the lightGBM weight of the corresponding date of each time segment in the first time period.
4. A method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion models according to claim 1 or 3, characterized in that: the predicted value of the second time period predicted according to the CNN-LSTM algorithm and the lightGBM algorithm, and the CNN-LSTM dynamic weight and the lightGBM dynamic weight of each day in the first time period obtain the final predicted value of the second time period corresponding to each day, and the specific calculation formula is as follows:
F t ＝W 1 f1 t +W 2 f2 t t＝1，2，...，n
wherein F is t Representing the final predicted value of the second time period corresponding to each day, f1 t And f2 t Representing predicted values, W, of the second time period predicted by the CNN-LSTM algorithm and the LightGBM algorithm, respectively 1 And W is 2 The CNN-LSTM dynamic weight and the lightGBM dynamic weight of each day in the first time period are respectively represented, t represents the number of days sequence of each time segment in the first time period and the second time period, and n represents the number of days of each time segment in the second time period.
CN202311406549.6A 2023-10-27 2023-10-27 Method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion model Pending CN117474151A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311406549.6A CN117474151A (en) 2023-10-27 2023-10-27 Method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311406549.6A CN117474151A (en) 2023-10-27 2023-10-27 Method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion model

Publications (1)

Publication Number Publication Date
CN117474151A true CN117474151A (en) 2024-01-30

Family

ID=89637130

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311406549.6A Pending CN117474151A (en) 2023-10-27 2023-10-27 Method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion model

Country Status (1)

Country Link
CN (1) CN117474151A (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110084424A (en) * 2019-04-25 2019-08-02 国网浙江省电力有限公司 A kind of Methods of electric load forecasting based on LSTM and LGBM
CN110245801A (en) * 2019-06-19 2019-09-17 中国电力科学研究院有限公司 A kind of Methods of electric load forecasting and system based on combination mining model
CN113935244A (en) * 2021-10-22 2022-01-14 广东电网有限责任公司广州供电局 Method and system for predicting short-term power load of urban rural distribution transformer
CN115600500A (en) * 2022-10-31 2023-01-13 南昌工程学院（Cn） Ultrashort-term probability wind power prediction method based on space-time multiscale and K-SDW
US20230095676A1 (en) * 2021-09-27 2023-03-30 Hohai University Method for multi-task-based predicting massiveuser loads based on multi-channel convolutional neural network

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110084424A (en) * 2019-04-25 2019-08-02 国网浙江省电力有限公司 A kind of Methods of electric load forecasting based on LSTM and LGBM
CN110245801A (en) * 2019-06-19 2019-09-17 中国电力科学研究院有限公司 A kind of Methods of electric load forecasting and system based on combination mining model
US20230095676A1 (en) * 2021-09-27 2023-03-30 Hohai University Method for multi-task-based predicting massiveuser loads based on multi-channel convolutional neural network
CN113935244A (en) * 2021-10-22 2022-01-14 广东电网有限责任公司广州供电局 Method and system for predicting short-term power load of urban rural distribution transformer
CN115600500A (en) * 2022-10-31 2023-01-13 南昌工程学院（Cn） Ultrashort-term probability wind power prediction method based on space-time multiscale and K-SDW

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
JUAN REN ET.: "A CNN-LSTM-LightGBM based short-term wind power prediction method based on attention mechanism", 《ENERGY REPORTS》, 7 March 2022 (2022-03-07), pages 437 - 443 *
庄家懿 等: "基于多模型融合的 CNN-LSTM-XGBoost 短期电力负荷预测方法", 《中国电力》, vol. 54, no. 5, 31 May 2021 (2021-05-31), pages 46 - 55 *

Similar Documents

Publication Publication Date Title
CN110610280B (en) Short-term prediction method, model, device and system for power load
CN115640914B (en) Intelligent gas storage optimization method, internet of things system, device and medium
Du et al. Power load forecasting using BiLSTM-attention
CN110675275B (en) Electric load regulation and control method and system for demand side response application of virtual power plant
CN111079989B (en) DWT-PCA-LSTM-based water supply amount prediction device for water supply company
CN112365029A (en) Missing value processing method for air conditioner load prediction and air conditioner load prediction system
CN111861023A (en) Statistical-based hybrid wind power prediction method and device
CN112149890A (en) Comprehensive energy load prediction method and system based on user energy label
CN112288140A (en) Keras-based short-term power load prediction method, storage medium and equipment
CN114330934A (en) Model parameter self-adaptive GRU new energy short-term power generation power prediction method
CN112365056A (en) Electrical load joint prediction method and device, terminal and storage medium
CN112990587A (en) Method, system, equipment and medium for accurately predicting power consumption of transformer area
CN113065715A (en) Multi-load ultra-short-term prediction method for comprehensive energy system
CN115470862A (en) Dynamic self-adaptive load prediction model combination method
CN116599050A (en) Photovoltaic prediction method and related device based on self-attention mechanism
CN116014722A (en) Sub-solar photovoltaic power generation prediction method and system based on seasonal decomposition and convolution network
CN114091776A (en) K-means-based multi-branch AGCNN short-term power load prediction method
CN117494906A (en) Natural gas daily load prediction method based on multivariate time series
CN117113086A (en) Energy storage unit load prediction method, system, electronic equipment and medium
CN113111592A (en) Short-term wind power prediction method based on EMD-LSTM
CN117474151A (en) Method for short-term load actual prediction based on CNN-LSTM and lightGBM dynamic weighted fusion model
CN113780644A (en) Photovoltaic output prediction method based on online learning
Dong et al. Image-based processing mechanism for peak load forecasting in smart grids
CN112614006A (en) Load prediction method, device, computer readable storage medium and processor
CN112330009A (en) Power grid load prediction 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