Load prediction method and system based on similar day screening

Abstract

The invention relates to a load prediction method and system based on similar day screening, and the method comprises the following steps: obtaining real-time load data (including temperature, air pressure and humidity) and historical load data, splicing the real-time load data and historical load data into a training data set, and carrying out the normalization processing. Afterwards, the normalized data is used for training a BP neural network to obtain a first prediction sequence and a second prediction sequence, and the average change influence value of the first prediction sequence and the second prediction sequence is calculated to reduce the data dimension; next, non-similar day data is filtered by determining real-time and historical similar days to form a feature set. Further, data decomposition is performed on the feature set, and high-frequency and low-frequency components are divided. The high-frequency component is used for training a CNN-BiLSTM model, the low-frequency component is used for training an informer model, and the prediction results of the CNN-BiLSTM model and the low-frequency component are fused to obtain final load prediction. According to the method, the influence of non-similar day data is effectively reduced, and the model is promoted to be more focused on similar historical data with high correlation, so that the prediction accuracy is improved.

Description

Technical Field

[0001] The invention relates to a load forecasting method and system based on similar day screening, belonging to the technical field of power load forecasting. Background Art

[0002] Power load forecasting is an important part of power system operation, planning and dispatching. With the rapid development of smart grids, the collection, storage and processing capabilities of power load data have been significantly improved, providing a rich data source for load forecasting. However, traditional load forecasting methods are often based on a single data source and a simple forecasting model, which makes it difficult to accurately capture the complex changing characteristics of power load. Traditional load forecasting methods usually rely only on historical load data as the basis for forecasting, ignoring other factors that may affect load changes, such as weather conditions, holiday effects, economic activities, etc. This single data source limits the generalization ability and accuracy of the forecasting model.

[0003] The patent document with the patent number "CN118554429A" has developed a short-term power load forecasting method based on multi-source data. The problem with this method is that it mainly relies on the extensive collection and preprocessing of multi-source data. Although the diversity of data is taken into account, it does not screen the data for similarity, which may lead to the introduction of irrelevant noise data during the model training process. The transfer learning method is used to build a target model based on the trained short-term power load forecasting source model. Although transfer learning can accelerate the model training process, the specific structure and parameter adjustment of the model may be limited by the characteristics of the source model, and it is difficult to fully adapt to the characteristics of the new data. Summary of the invention

[0004] In order to solve the above problems existing in the prior art, the present invention proposes a load forecasting method and system based on similar day screening.

[0005] The technical solution of the present invention is as follows:

[0006] In one aspect, the present invention provides a load forecasting method based on similar day screening, comprising the following steps:

[0007] Acquire real-time load data and historical load data, including temperature, air pressure and humidity, and concatenate the real-time load data and historical load data as a training data set, normalize the training data set, and use the normalized training data set to train the BP neural network;

[0008] Input the training data set into the trained BP neural network to obtain a first prediction sequence and a second prediction sequence, calculate the average change impact value of the first prediction sequence and the second prediction sequence, and reduce the dimension of the training data set according to the average change impact value;

[0009] Determine the similar days between the real-time load data and the historical load data, and filter the data corresponding to the non-similar days in the training data set according to the similar days to obtain the feature set;

[0010] Perform data decomposition on the feature set, and divide the decomposed signal into high-frequency components and low-frequency components according to a preset frequency threshold;

[0011] The CNN-BiLSTM model is trained using high-frequency components, and the informer model is trained using low-frequency components. The CNN-BiLSTM model and the informer model are used to predict the load, and the prediction results are fused to obtain the final load result.

[0012] As a preferred implementation, screening the dimensions in the training data set includes the following steps:

[0013] Randomly select a data in each dimension of the training data set and increase it by 10% to obtain the first training set p1:

[0014]

[0015] Randomly select a data from each dimension of the training data set and reduce it by 10% to get the second training set p2:

[0016]

[0017] Among them, s m represents the load that needs to be predicted for the mth dimension of the training data set, x mi represents the i-th data of the m-th dimension of the training data set, and n represents the maximum number of data in one dimension;

[0018] The first training set p1 and the second training set p2 m Input the trained BP neural network to get the first prediction sequence and the second prediction sequence

[0019] in, Represents the prediction result of the data of the mth dimension of the first training set p1, Represents the prediction result of the data of the mth dimension of the second training set p2;

[0020] Calculate the average change impact value of the first prediction series and the second prediction series:

[0021]

[0022] A1, let i=1;

[0023] A2. If If it is greater than or equal to MIV, no operation is performed; otherwise, the i-th dimension of the training data set is deleted;

[0024] A3, i=i+1;

[0025] A4. Execute step A2 until i=m.

[0026] As a preferred implementation, the method for determining similar days is:

[0027] Calculate the trend similarity R between the historical load data and the real-time load data in the fth historical time period mn (f):

[0028]

[0029] Among them, a n Indicates real-time load data, a f represents the load data of the fth historical time period, E() represents the mathematical expectation, and D() represents the variance;

[0030] Calculate the difference between the historical load data and the real-time load data in the fth historical time period. mn (f):

[0031]

[0032] Where t represents the number of preset load data, T represents the total number of load data, and x f,t represents the load data of the tth historical time period, x n,t Represents t real-time load data;

[0033] The trend similarity R mn (f) and the difference O mn (f) Perform normalization to obtain the similarity coefficient V mn (f):

[0034]

[0035] Calculate the date matching coefficient c between the historical load data and the real-time load data in the fth historical time period i (t):

[0036]

[0037] Among them, a1 and a3 represent the preset attenuation coefficient, β i represents the preset logical variable, f() represents the integer function, r() represents the remainder function, b represents the preset number of days, N1 and M2 represent the days of the week, and N3 represents the date distance;

[0038] The similarity coefficient V mn (f) Date matching coefficient c i (t) are added together to obtain the total matching coefficient, and the average total matching coefficient of f historical time periods is calculated. The historical time period with a total matching coefficient greater than or equal to the average total matching coefficient is regarded as a similar day.

[0039] As a preferred implementation, the method of data decomposition is:

[0040] S1. Let the feature set be represented by x(t);

[0041] S2, preprocess the feature set x(t) to obtain the signal sequence {x1(t),...,x2(t),...,x i (t)}:

[0042] x i (t) = x(t) + εδ i (t);

[0043] Where ε is the noise factor, δ i (t) represents the i-th noise;

[0044] S3, use empirical mode decomposition EMD to decompose the i-th feature signal x in the signal sequence i (t) Decompose to obtain K modal components, calculate the mean of the K modal components to obtain the intrinsic mode function I of the i-th characteristic signal MF,i (t):

[0045]

[0046] in, represents the jth modal component;

[0047] The residual signal r(t) of the i-th characteristic signal is expressed as:

[0048] r(t)=x i (t)-I MF,i (t);

[0049] S4, executing step S2 using the residual signal r(t) as a feature set until S4 reaches a preset maximum number of executions;

[0050] S5, the i-th characteristic signal x i (t) is expressed as:

[0051]

[0052] Among them, H represents the preset maximum number of executions, I MF,p (t) represents i characteristic signals x iThe p-th decomposition eigenmode function I of (t) MF,p (t);

[0053] Before Personal I MF,p (t) is taken as the high frequency component, and the rest is taken as the low frequency component.

[0054] As a preferred implementation, the method for fusing the prediction results of the CNN-BiLSTM model and the informer model is:

[0055]

[0056] Among them, XG represents the final prediction result, Represents the prediction result of CNN-BiLSTM model, Represents the prediction result of the informer model, Indicates splicing.

[0057] On the other hand, the present invention also provides a load forecasting system based on similar day screening, comprising:

[0058] Data acquisition module: obtain real-time load data and historical load data, including temperature, air pressure and humidity, splice the real-time load data and historical load data as a training data set, normalize the training data set, and use the normalized training data set to train the BP neural network;

[0059] Data processing module: input the training data set into the trained BP neural network to obtain the first prediction sequence and the second prediction sequence, calculate the average change impact value of the first prediction sequence and the second prediction sequence, and reduce the dimension of the training data set according to the average change impact value;

[0060] Similar day calculation module: determine the similar days between real-time load data and historical load data, filter the data corresponding to non-similar days in the training data set according to the similar days, and obtain the feature set;

[0061] Data decomposition module: decomposes the feature set and divides the decomposed signal into high-frequency components and low-frequency components according to the preset frequency threshold;

[0062] Load forecasting module: Use high-frequency components to train the CNN-BiLSTM model, use low-frequency components to train the informer model, use the CNN-BiLSTM model and the informer model to predict the load, and fuse the prediction results to obtain the final load result.

[0063] The present invention has the following beneficial effects:

[0064] The present invention filters non-similar day data by determining real-time and historical similar days to form a feature set, so that the model is more focused on similar historical data with high correlation with current load changes, thereby reducing the interference of non-similar day data on the prediction results and improving the accuracy of the prediction. The feature set is decomposed to divide the high-frequency and low-frequency components, and trained using the CNN-BiLSTM model and the informer model respectively. This decomposition and targeted modeling method can better capture the different frequency characteristics in the load data and further improve the prediction accuracy. For example, the CNN-BiLSTM model can effectively extract local features and time series information in the high-frequency components, while the informer model performs well in processing low-frequency components and can mine long-term dependencies and global information. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 The present invention is a flowchart for implementing the method. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] It should be understood that the step numbers used in this document are only for convenience of description and are not intended to limit the order in which the steps are executed.

[0068] It should be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0069] The terms “include” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0070] The term "and / or" means and includes any and all possible combinations of one or more of the associated listed items.

[0071] Embodiment 1:

[0072] See also Figure 1 The present invention provides a load forecasting method based on similar day screening, comprising the following steps:

[0073] Acquire real-time load data and historical load data, including temperature, air pressure and humidity, and concatenate the real-time load data and historical load data as a training data set, normalize the training data set, and use the normalized training data set to train the BP neural network;

[0074] Input the training data set into the trained BP neural network to obtain a first prediction sequence and a second prediction sequence, calculate the average change impact value of the first prediction sequence and the second prediction sequence, and reduce the dimension of the training data set according to the average change impact value;

[0075] Determine the similar days between the real-time load data and the historical load data, and filter the data corresponding to the non-similar days in the training data set according to the similar days to obtain the feature set;

[0076] Perform data decomposition on the feature set, and divide the decomposed signal into high-frequency components and low-frequency components according to a preset frequency threshold;

[0077] The CNN-BiLSTM model is trained using high-frequency components, and the informer model is trained using low-frequency components. The CNN-BiLSTM model and the informer model are used to predict the load, and the prediction results are fused to obtain the final load result.

[0078] As a preferred implementation, screening the dimensions in the training data set includes the following steps:

[0079] Randomly select a data in each dimension of the training data set and increase it by 10% to obtain the first training set p1:

[0080]

[0081] Randomly select a data from each dimension of the training data set and reduce it by 10% to get the second training set p2:

[0082]

[0083] Among them, a m represents the load that needs to be predicted for the mth dimension of the training data set, x mi represents the i-th data of the m-th dimension of the training data set, and n represents the maximum number of data in one dimension;

[0084] The first training set p1 and the second training set p2 m Input the trained BP neural network to get the first prediction sequence and the second prediction sequence

[0085] in, Represents the prediction result of the data of the mth dimension of the first training set p1, Represents the prediction result of the data of the mth dimension of the second training set p2;

[0086] Calculate the average change impact value of the first prediction series and the second prediction series:

[0087]

[0088] A1, let i=1;

[0089] A2. If If it is greater than or equal to MIV, no operation is performed; otherwise, the i-th dimension of the training data set is deleted;

[0090] A3, i=i+1;

[0091] A4. Execute step A2 until i=m.

[0092] for example This indicates the impact value of temperature change. If the impact value of temperature change is less than the average impact value, the data corresponding to the temperature will be deleted from the training data set. If the temperature data is missing, the dimension of the training data set will be -1.

[0093] As a preferred implementation, the method for determining similar days is:

[0094] Calculate the trend similarity R between the historical load data and the real-time load data in the fth historical time period mn (f):

[0095]

[0096] Among them, a n Indicates real-time load data, a f represents the load data of the fth historical time period, E() represents the mathematical expectation, and D() represents the variance;

[0097] Calculate the difference between the historical load data and the real-time load data in the fth historical time period. mn (f):

[0098]

[0099] Where t represents the number of preset load data, T represents the total number of load data, and x f,t represents the load data of the tth historical time period, x n,t Represents t real-time load data;

[0100] The trend similarity R mn (f) and the difference Omn (f) Perform normalization to obtain the similarity coefficient V mn (f):

[0101]

[0102] Calculate the date matching coefficient c between the historical load data and the real-time load data in the fth historical time period i (t):

[0103]

[0104] Among them, a1 and a3 represent the preset attenuation coefficient, β i represents a preset logical variable, takes 1 when the historical date and the date to be predicted belong to the same date type (both are holidays or working days), otherwise takes 0, f() represents an integer function, r() represents a remainder function, b represents a preset number of days (the date difference between the ith historical day and the date to be predicted), N1 and N2 represent the number of days in a week (7 in this embodiment), and N3 represents the date distance (according to the holiday distribution in my country, the number of days between some major holidays is usually less than 30 days (such as Mid-Autumn Festival and National Day), and N3 is taken as 20 in this embodiment);

[0105] The similarity coefficient V mn (f) Date matching coefficient c i (t) are added together to obtain the total matching coefficient, and the average total matching coefficient of f historical time periods is calculated. The historical time period with a total matching coefficient greater than or equal to the average total matching coefficient is regarded as a similar day.

[0106] As a preferred implementation, the method of data decomposition is:

[0107] S1. Let the feature set be represented by x(t);

[0108] S2, preprocess the feature set x(t) to obtain the signal sequence {x1(t),...,x2(t),...,x i (t)}:

[0109] x i (t) = x(t) + εδ i (t);

[0110] Where ε is the noise factor, δ i (t) represents the i-th noise;

[0111] S3, use empirical mode decomposition EMD to decompose the i-th feature signal x in the signal sequence i (t) Decompose to obtain K modal components, calculate the mean of the K modal components to obtain the intrinsic mode function I of the i-th characteristic signal MF,i(t):

[0112]

[0113] in, represents the jth modal component;

[0114] The residual signal r(t) of the i-th characteristic signal is expressed as:

[0115] r(t)=x i (t)-I MF,i (t);

[0116] S4, executing step S2 using the residual signal r(t) as a feature set until S4 reaches a preset maximum number of executions;

[0117] S5, the i-th characteristic signal x i (t) is expressed as:

[0118]

[0119] Among them, H represents the preset maximum number of executions, I MF,p (t) represents i characteristic signals x i The p-th decomposition eigenmode function I of (t) MF,p (t);

[0120] Before Personal I MF,p (t) is taken as the high frequency component, and the rest is taken as the low frequency component.

[0121] As a preferred implementation, the method for fusing the prediction results of the CNN-BiLSTM model and the informer model is:

[0122]

[0123] Among them, XG represents the final prediction result, Represents the prediction result of CNN-BiLSTM model, Indicates the prediction result of the informer model. Indicates splicing.

[0124] The CNN-BiLSTM model is a hybrid deep learning architecture that combines the advantages of convolutional neural networks (CNNs) and bidirectional long short-term memory networks (BiLSTMs).

[0125] The Informer model is a time series prediction model based on the Transformer architecture. It combines the characteristics of Transformer, self-attention mechanism and CNN, and is suitable for time series prediction with long-term dependencies and multiple time scales.

[0126] Embodiment 2:

[0127] The present invention also provides a load forecasting system based on similar day screening, comprising:

[0128] Data acquisition module: obtain real-time load data and historical load data, including temperature, air pressure and humidity, splice the real-time load data and historical load data as training data set, normalize the training data set, and use the normalized training data set to train the BP neural network;

[0129] Data processing module: input the training data set into the trained BP neural network to obtain the first prediction sequence and the second prediction sequence, calculate the average change impact value of the first prediction sequence and the second prediction sequence, and reduce the dimension of the training data set according to the average change impact value;

[0130] Similar day calculation module: determine the similar days between real-time load data and historical load data, filter the data corresponding to non-similar days in the training data set according to the similar days, and obtain the feature set;

[0131] Data decomposition module: decomposes the feature set and divides the decomposed signal into high-frequency components and low-frequency components according to the preset frequency threshold;

[0132] Load forecasting module: Use high-frequency components to train the CNN-BiLSTM model, use low-frequency components to train the informer model, use the CNN-BiLSTM model and the informer model to predict the load, and fuse the prediction results to obtain the final load result.

[0133] The system is used to implement the method in Example 1, which will not be described in detail here.

[0134] In the embodiments of the present application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. Among them, A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b and c can be represented by: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, c can be single or multiple.

[0135] Those of ordinary skill in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented in a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0136] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0137] In several embodiments provided in the present application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), disk or optical disk, and other media that can store program codes.

[0138] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A load forecasting method based on similar day screening, characterized in that: The following steps are involved: Acquire real-time load data and historical load data, including temperature, air pressure and humidity, and concatenate the real-time load data and historical load data as a training data set, normalize the training data set, and use the normalized training data set to train the BP neural network; Input the training data set into the trained BP neural network to obtain a first prediction sequence and a second prediction sequence, calculate the average change impact value of the first prediction sequence and the second prediction sequence, and reduce the dimension of the training data set according to the average change impact value; Determine the similar days between the real-time load data and the historical load data, and filter the data corresponding to the non-similar days in the training data set according to the similar days to obtain the feature set; Perform data decomposition on the feature set, and divide the decomposed signal into high-frequency components and low-frequency components according to a preset frequency threshold; The CNN-BiLSTM model is trained using high-frequency components, and the informer model is trained using low-frequency components. The CNN-BiLSTM model and the informer model are used to predict the load, and the prediction results are fused to obtain the final load result.

2. The load forecasting method based on similar day screening according to claim 1 is characterized in that: Filtering the dimensions in the training dataset includes the following steps: Randomly select a data in each dimension of the training data set and increase it by 10% to obtain the first training set p1: Randomly select a data from each dimension of the training data set and reduce it by 10% to get the second training set p2: Among them, s m represents the load that needs to be predicted for the mth dimension of the training data set, x mi represents the i-th data of the m-th dimension of the training data set, and n represents the maximum number of data in one dimension; The first training set p1 and the second training set p2 m Input the trained BP neural network to get the first prediction sequence and the second prediction sequence in, Represents the prediction result of the data of the mth dimension of the first training set p1, Represents the prediction result of the data of the mth dimension of the second training set p2; Calculate the average change impact value of the first prediction series and the second prediction series: A1, let i=1; A2. If If it is greater than or equal to MIV, no operation is performed; otherwise, the i-th dimension of the training data set is deleted; A3, i=i+1; A4. Execute step A2 until i=m.

3. The load forecasting method based on similar day screening according to claim 1 is characterized in that: The method for determining similar days is: Calculate the trend similarity R between the historical load data and the real-time load data in the fth historical time period mn (f): Among them, a n Indicates real-time load data, a f represents the load data of the fth historical time period, E() represents the mathematical expectation, and D() represents the variance; Calculate the difference between the historical load data and the real-time load data in the fth historical time period. mn (f): Where t represents the number of preset load data, T represents the total number of load data, and x f,t represents the load data of the tth historical time period, x n,t Represents t real-time load data; The trend similarity R mn (f) and the difference O mn (f) Perform normalization to obtain the similarity coefficient V mn (f): Calculate the date matching coefficient c between the historical load data and the real-time load data in the fth historical time period i (t): Among them, a1 and a3 represent the preset attenuation coefficient, β i represents the preset logical variable, f() represents the integer function, r() represents the remainder function, b represents the preset number of days, N1 and N2 represent the days of the week, and N3 represents the date distance; The similarity coefficient V mn (f) Date matching coefficient c i (t) are added together to obtain the total matching coefficient, and the average total matching coefficient of f historical time periods is calculated. The historical time period with a total matching coefficient greater than or equal to the average total matching coefficient is regarded as a similar day.

4. The load forecasting method based on similar day screening according to claim 1 is characterized in that: The method of data decomposition is: S1. Let the feature set be represented by x(t); S2, preprocess the feature set x(t) to obtain the signal sequence {x1(t),...,x2(t),...,x i (t)}: x i (t)=x(t)+εδ i (t); Where ε is the noise factor, δ i (t) represents the i-th noise; S3, use empirical mode decomposition EMD to decompose the i-th feature signal x in the signal sequence i (t) Decompose to obtain K modal components, calculate the mean of the K modal components to obtain the intrinsic mode function I of the i-th characteristic signal MF,i (t): in, represents the jth modal component; The residual signal r(t) of the i-th characteristic signal is expressed as: r(t)=x i (t)-I MF,i (t); S4, executing step S2 using the residual signal r(t) as a feature set until S4 reaches a preset maximum number of executions; S5, the i-th characteristic signal x i (t) is expressed as: Among them, H represents the preset maximum number of executions, I MF,p (t) represents i characteristic signals x i The p-th decomposition eigenmode function I of (t) MF,p (t); Before Personal I MF,p (t) is taken as the high frequency component, and the rest is taken as the low frequency component.

5. The load forecasting method based on similar day screening according to claim 1 is characterized in that: The method for fusing the prediction results of the CNN-BiLSTM model and the informer model is: Among them, XG represents the final prediction result, Represents the prediction result of CNN-BiLSTM model, Indicates the prediction result of the informer model. Indicates splicing.

6. A load forecasting system based on similar day screening, characterized in that: include: Data acquisition module: obtain real-time load data and historical load data, including temperature, air pressure and humidity, splice the real-time load data and historical load data as training data set, normalize the training data set, and use the normalized training data set to train the BP neural network; Data processing module: input the training data set into the trained BP neural network to obtain the first prediction sequence and the second prediction sequence, calculate the average change impact value of the first prediction sequence and the second prediction sequence, and reduce the dimension of the training data set according to the average change impact value; Similar day calculation module: determine the similar days between real-time load data and historical load data, filter the data corresponding to non-similar days in the training data set according to the similar days, and obtain the feature set; Data decomposition module: decomposes the feature set and divides the decomposed signal into high-frequency components and low-frequency components according to the preset frequency threshold; Load forecasting module: Use high-frequency components to train the CNN-BiLSTM model, use low-frequency components to train the informer model, use the CNN-BiLSTM model and the informer model to predict the load, and fuse the prediction results to obtain the final load result.

7. The load forecasting system based on similar day screening according to claim 6 is characterized in that: The data processing module randomly selects a data from each dimension of the training data set and increases it by 10% to obtain the first training set p1: Randomly select a data from each dimension of the training data set and reduce it by 10% to get the second training set p2: Among them, a m represents the load that needs to be predicted for the mth dimension of the training data set, x mi represents the i-th data of the m-th dimension of the training data set, and n represents the maximum number of data in one dimension; The first training set p1 and the second training set p2 m Input the trained BP neural network to get the first prediction sequence and the second prediction sequence in, Represents the prediction result of the data of the mth dimension of the first training set p1, Represents the prediction result of the data of the mth dimension of the second training set p2; Calculate the average change impact value of the first prediction series and the second prediction series: A1, let i=1; A2. If If it is greater than or equal to MIV, no operation is performed; otherwise, the i-th dimension of the training data set is deleted; A3, i=i+1; A4. Execute step A2 until i=m.

8. The load forecasting system based on similar day screening according to claim 6 is characterized in that: The similar day calculation module determines similar days by: Calculate the trend similarity R between the historical load data and the real-time load data in the fth historical time period mn (f): Among them, a n Indicates real-time load data, a f represents the load data of the fth historical time period, E() represents the mathematical expectation, and D() represents the variance; Calculate the difference between the historical load data and the real-time load data in the fth historical time period. mn (f): Where t represents the number of preset load data, T represents the total number of load data, and x f,t represents the load data of the tth historical time period, x n,t Represents t real-time load data; The trend similarity R mn (f) and the difference O mn (f) Perform normalization to obtain the similarity coefficient V mn (f): Calculate the date matching coefficient c between the historical load data and the real-time load data in the fth historical time period i (t): Among them, a1 and a3 represent the preset attenuation coefficient, β i represents the preset logical variable, f() represents the integer function, r() represents the remainder function, b represents the preset number of days, N1 and N2 represent the days of the week, and N3 represents the date distance; The similarity coefficient V mn (f) Date matching coefficient c i (t) are added together to obtain the total matching coefficient, and the average total matching coefficient of f historical time periods is calculated. The historical time period with a total matching coefficient greater than or equal to the average total matching coefficient is regarded as a similar day.

9. The load forecasting system based on similar day screening according to claim 6 is characterized in that: Data decomposition module, the method of data decomposition is: S1. Let the feature set be represented by x(t); S2, preprocess the feature set x(t) to obtain the signal sequence {x1(t),...,x2(t),...,x i (t)}: x i (t)=x(t)+εδ i (t); Where ε is the noise factor, δ i (t) represents the i-th noise; S3, use empirical mode decomposition EMD to decompose the i-th feature signal x in the signal sequence i (t) Decompose to obtain K modal components, calculate the mean of the K modal components to obtain the intrinsic mode function I of the i-th characteristic signal MF,i (t): in, represents the jth modal component; The residual signal r(t) of the i-th characteristic signal is expressed as: r(t)=x i (t)-I MF,i (t); S4, executing step S2 using the residual signal r(t) as a feature set until S4 reaches a preset maximum number of executions; S5, the i-th characteristic signal x i (t) is expressed as: Among them, H represents the preset maximum number of executions, I MF,p (t) represents i characteristic signals x i The p-th decomposition eigenmode function I of (t) MF,p (t); Before Personal I MF,p (t) is taken as the high frequency component, and the rest is taken as the low frequency component.

10. The load forecasting system based on similar day screening according to claim 6, characterized in that: The load forecasting module, the method of fusing the forecast results of the CNN-BiLSTM model and the informer model is: in, Indicates the prediction result of the informer model, XG indicates the final prediction result, Represents the prediction result of CNN-BiLSTM model, Indicates splicing.

Images