Mode decomposition and prediction Kolmogorov-Arnold network power load blind source separation method

By combining empirical modal decomposition and recurrent neural network, the Kolmogorov-Arnold Networks model is used to perform modal decomposition and prediction of power load, and the problem of low decomposition accuracy and insufficient general use of the existing power load blind source separation method is solved. A more efficient power load signal separation is achieved.

CN118885804APending Publication Date: 2024-11-01GUANGXI UNIV
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202411104110.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-01

AI Technical Summary

Technical Problem

The existing power load blind source separation methods have problems such as low decomposition accuracy, high requirements for source signal quality, and insufficient general use. In particular, the methods based on non-negative matrix decomposition and independent component analysis are not effective in actual power systems.

Method used

Combining empirical modal decomposition and recurrent neural network, the Kolmogorov-Arnold Networks (KAN) model is used to perform modal decomposition and prediction of power loads. The power load signals are processed through differential, integral, GRU, LSTM and BiLSTM technologies to improve separation accuracy and General use.

Benefits of technology

It improves the accuracy and generality of blind source separation of power loads, can effectively deal with noise and correlation interference, and improves the separation effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118885804A_ABST
    Figure CN118885804A_ABST
Patent Text Reader

Abstract

The invention provides a Kolmogorov-Arnold network power load blind source separation method based on modal decomposition and prediction. According to the modal decomposition and prediction method, the advantages of EMD (empirical mode decomposition), a differential integral preprocessing technology, a GRU (Gated Recirculation Unit) gating circulation unit, an LSTM (Long Short-Term Memory) network, a BiLSTM (Bidirectional Long Short-Term Memory) network and a KAN (Karan Area Network) deep learning model are combined. Differential, integral and modal decomposition processing is carried out on original time sequence data, the information amount of the data is increased, complex features of the data are more easily captured by a neural network, then multiple recurrent neural network models are utilized to carry out multi-dimensional feature extraction and preliminary prediction, and finally a preliminary prediction result is input into a KAN network to carry out information separation and optimization. According to KAN, namely Kolmogorov-Arnold Networks, data features are captured through a KAN deep learning model for adaptive optimization, and the reliability and precision of a prediction result are further improved. According to the modal decomposition and prediction method, the blind source information processing technology can be optimized, and the power load blind source separation precision can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of non-intrusive electric load blind source separation technology, electric power system load forecasting, and signal processing technology, and in particular to a method for electric load blind source separation using differentiation, integration, empirical mode decomposition, recurrent neural network, and deep learning model, which is suitable for blind source separation of electric power system loads. Background Art

[0002] Existing methods based on non-negative matrix factorization (NMF) have problems such as large uncertainty in decomposition results, easy to fall into local optimality, low decomposition accuracy caused by sensitivity to pathological conditions, and insufficient versatility.

[0003] In addition, the accuracy of existing methods based on independent component analysis (ICA) is highly dependent on the statistical independence assumption of source signals and the quality of source signals. However, in actual power systems, there may be certain correlations and noise interference between loads, which makes the separation effect of ICA poor.

[0004] Therefore, a Kolmogorov-Arnold network power load blind source separation method based on modal decomposition and prediction is proposed to solve the problems of low decomposition accuracy, high requirements on source signal quality and insufficient versatility in existing power load blind source separation methods. Summary of the invention

[0005] The present invention proposes a Kolmogorov-Arnold network power load blind source separation method for mode decomposition and prediction, which combines empirical mode decomposition and recurrent neural network for information processing and prediction in blind source separation problems, and has the function of optimizing the power load blind source separation information processing technology and improving the power load blind source separation accuracy; the steps in the use process are:

[0006] Step (1): Get and store the user's previous The total power signal at the moment to the continuous total power signal at the current moment, the original power load total power signal sequence is obtained for:

[0007] (1)

[0008] in, For users The total power signal at time, For users The total power signal at time, For users The total power signal at time, is the total power signal of the user at the current moment;

[0009] Step (2): Perform first-order differential processing on the original power load total power signal sequence to increase the amount of data information and obtain the differential signal sequence for:

[0010] (2)

[0011] in, For users The total power signal at time, For users Total power signal at time;

[0012] Step (3): Perform first-order integration processing on the original power load total power signal sequence to increase the amount of data information and make the data features more significant, and obtain the integral signal sequence for:

[0013] (3)

[0014] in, For users The total power signal at the moment The accumulation of the total power signal at time For users The total power signal at the moment The accumulation of the total power signal at time For users The total power signal at the moment The accumulation of the total power signal at time For users The accumulation of the total power signal at time to the total power signal at the current time, for of the moment;

[0015] Step (4): Use the EMD empirical mode decomposition method to perform modal decomposition on the original power load total power signal sequence, remove the noise and interference in the power time series, improve the accuracy and reliability of signal processing, and take the first modal component signal sequence , the second modal component signal sequence , the third modal component signal sequence and the residual component signal sequence for:

[0016] (4)

[0017] (5)

[0018] (6)

[0019] (7)

[0020] in, and The method to obtain is as follows:

[0021] Step (4-1): Find the extreme points of the total power signal sequence of all original power loads, connect the local maximum points into an upper envelope through a cubic spline curve, and connect the local minimum points into a lower envelope. The upper and lower envelopes contain all the data points to obtain the maximum envelope. and minimum envelope ;

[0022] Step (4-2): Average the two envelopes to get the average value for:

[0023] (8)

[0024] Step (4-3): Subtract the average value from the original power load total power signal sequence to obtain the intermediate signal for:

[0025] (9)

[0026] Step (4-4): Here, if If the conditions of the intrinsic mode function are met, then is the first-order intrinsic mode function of the original signal; if If not satisfied, proceed to step (4-5);

[0027] The conditions of the intrinsic mode function are: in the entire data set, the sum of the number of local maxima and local minima must be equal to the number of zero crossings or differ by at most one; at any time point, the mean of the upper envelope defined by the local maximum and the lower envelope defined by the local minimum is zero;

[0028] Step (4-5): As the original signal, perform a second screening:

[0029] (10)

[0030] Where: yes The mean of the upper and lower envelopes;

[0031] Repeat the filter until Satisfying the definition of the intrinsic mode function, the first-order intrinsic mode function is obtained ;

[0032] Step (4-6): Convert the original signal minus Get the remaining amount for:

[0033] (11)

[0034] Steps (4-7): As the new original signal, re-execute steps (4-1) to (4-6) to obtain the new residual , repeat this process n times and when the nth remaining amount When it becomes a monotonic function or a constant and cannot be decomposed into the intrinsic mode function, the entire EMD process ends; the original signal It can be expressed as n intrinsic mode function components and residual components The specific formula is:

[0035] (12)

[0036] in, is the intrinsic mode function component; is the residual component, representing the average trend of the signal, which is a constant sequence or a monotonic sequence. is the accumulation of the first-order intrinsic mode function to the n-th-order intrinsic mode function;

[0037] Step (5): Use the GRU gated recurrent unit architecture to , , , , , , After Predict the total power at each moment, capture data features, and alleviate gradient disappearance;

[0038] The prediction result is recorded as:

[0039] (13)

[0040] There is a logo in the upper right corner The value of is the value predicted by the GRU gated recurrent unit; The prediction result is recorded as:

[0041] (14)

[0042] The prediction result is recorded as:

[0043] (15)

[0044] The prediction result is recorded as:

[0045] (16)

[0046] The prediction result is recorded as:

[0047] (17)

[0048] The prediction result is recorded as:

[0049] (18)

[0050] The prediction result is recorded as:

[0051] (19)

[0052] The formula for GRU is as follows:

[0053] (20)

[0054] (twenty one)

[0055] (twenty two)

[0056] (twenty three)

[0057] in, is the input data at the current moment, is the output of the unit at the current moment, is the output of the unit at the previous moment, To reset the gate, To update the gate, is a candidate hidden state, represents element-wise multiplication, represents the sigmoid function, represents the hyperbolic tangent function, , , , , , and b are parameters to be learned in the deep network, which are the first weight of the reset gate, the second weight of the reset gate, the bias of the reset gate, the first weight of the update gate, the second weight of the update gate, the bias of the update gate, the first weight of the candidate hidden state, the second weight of the candidate hidden state, and the bias of the candidate hidden state respectively;

[0058] Step (6): Use LSTM long short-term memory network to , , , , , , After Predict the total power at each moment, capture data features, and alleviate gradient explosion;

[0059] The prediction result is recorded as:

[0060] (twenty four)

[0061] There is a logo in the upper right corner The value is the value predicted by the LSTM long short-term memory network; The prediction result is recorded as:

[0062] (25)

[0063] The prediction result is recorded as:

[0064] (26)

[0065] The prediction result is recorded as:

[0066] (27)

[0067] The prediction result is recorded as:

[0068] (28)

[0069] The prediction result is recorded as:

[0070] (29)

[0071] The prediction result is recorded as:

[0072] (30)

[0073] The formula for LSTM is as follows:

[0074] f t = σ ( w f [ h l s t m , t − 1 , x l s t m , t ] + b f ) (31)

[0075] i t = σ ( w i [ h l s t m , t − 1 , x l s t m , t ] + b i ) (32)

[0076] o t = σ ( w o [ h l s t m , t − 1 , x l s t m , t ] + b o ) (33)

[0077] C t ̃ = t a n h ( w c [ h l s t m , t − 1 , x l s t m , t ] + b c ) (34)

[0078] (35)

[0079] (36)

[0080] in, is the input data at the current moment, is the input status at the current moment, is the state of this unit, i.e., the candidate gate, is the state of the unit at the previous moment, is the output of the unit at the current moment, is the output of the unit at the previous moment; and The two matrices are directly concatenated together as input, with the corresponding weights , , , Multiply and add the corresponding bias , , , ; Forget gate for the last unit state Perform selective forgetting and output the historical information of memory; the input gate is for the current input state Selective input is performed, and the output obtained is the current input information; the sum of the memory of historical information and the current input information is the current unit state ; The output gate selectively outputs the current unit state to obtain the final output ; , , and They are the weight of the forget gate, the weight of the input gate, the weight of the output gate, and the weight corresponding to the input state at the current moment; , , , They are the bias of the forget gate, the bias of the input gate, the bias of the output gate, and the bias corresponding to the input state at the current moment; is the output of the forget gate; is the output of the output gate; is the output of the input gate;

[0081] Step (7): Use BiLSTM bidirectional long short-term memory network to , , , , , , After Predict total power at all times to improve noise tolerance;

[0082] The prediction result is recorded as:

[0083] (37)

[0084] There is a logo in the upper right corner The value is the value predicted by the BiLSTM bidirectional long short-term memory network; The prediction result is recorded as:

[0085] (38)

[0086] The prediction result is recorded as:

[0087] (39)

[0088] The prediction result is recorded as:

[0089] (40)

[0090] The prediction result is recorded as:

[0091] (41)

[0092] The prediction result is recorded as:

[0093] (42)

[0094] The prediction result is recorded as:

[0095] (43)

[0096] The formula of BiLSTM is as follows:

[0097] h = [ h f t 1 , h b t 1 ] (44)

[0098] in, is the forward LSTM output, is the reverse LSTM output, is the number of layers;

[0099] Step (8): Take the sequence obtained by modal decomposition in step (4) and the prediction results obtained from steps (5) to (6) as the input of KAN , feature extraction and adaptive optimization combination processing are performed through the KAN network, and finally the output result of the KAN network is obtained ,Right now is the result of blind source separation of power load; KAN is Kolmogorov-Arnold Networks;

[0100] The surface formula of KAN is:

[0101] (45)

[0102] The deep formula of KAN is:

[0103] (46)

[0104] in, represents the input of KAN, i.e., the sequence obtained by modal decomposition in step (4) and the prediction results obtained from steps (5) to (6); for The pth element of ; and They are the output of the KAN surface layer and the output of KAN respectively; is an internal function, forming a KAN layer, where the input dimension , output dimension ; is the qth layer external function, forming a KAN layer, where the input dimension , output dimension ; Indicates from arrive The accumulation of Indicates from arrive The shape of KAN consists of an integer array [ n 0 , n 1 , . . . , n L ] express, is the number of nodes at layer 0; is the number of nodes in the first layer; is the number of nodes in the Lth layer; T is the bias of KAN; is the accumulation of functions; is the 0th layer external function; is the first layer external function; is the L-2 layer external function; is the L-1th layer external function;

[0105] use represents the i-th neuron in the k-th layer, and express - The activation value of the neuron between the kth layer and the k+1th layer is Activation function, connection and The activation function is , ; ; ;

[0106] Pre-activation express, Post-activation express, The activation value of is the sum of all incoming post-activation values:

[0107] (47)

[0108] The output matrix of the kth layer for:

[0109] (48)

[0110] In matrix form is the B-spline function matrix of the kth layer, is the input matrix of the kth layer, is the output matrix of the kth layer.

[0111] Compared with the prior art, the present invention has the following advantages and effects:

[0112] (1) Existing methods based on non-negative matrix decomposition (NMF) have problems such as large uncertainty in decomposition results, easy to fall into local optimality, low decomposition accuracy caused by sensitivity to pathological conditions, and insufficient versatility. However, the present invention can use the recurrent neural network model to extract multi-dimensional features, and then use the KAN deep learning model to capture data features for adaptive optimization, thereby improving the accuracy and versatility of the separation results.

[0113] (2) The accuracy of existing methods based on independent component analysis (ICA) is highly dependent on the statistical independence assumption of the source signal and the quality of the source signal. However, in actual power systems, there may be certain correlations and noise interference between loads, which makes the separation effect of ICA poor. The present invention can reduce signal noise and improve separation effect by combining recurrent neural network with KAN deep learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 It is a structural framework diagram of an information collection device included in the method of the present invention.

[0115] Figure 2 It is a flow chart of modal decomposition and prediction implementation of the method of the present invention.

[0116] Figure 3 The figure is a schematic diagram of the use process of Kolmogorov-Arnold Networks in the method of the present invention. DETAILED DESCRIPTION

[0117] A Kolmogorov-Arnold Networks power load blind source separation method for modal decomposition and prediction proposed by the present invention is described in detail as follows in conjunction with the accompanying drawings:

[0118] Figure 1 It is a structural framework diagram of an information acquisition device included in the method of the present invention. The external power supply is connected to the information acquisition terminal through the incoming line and the plug. The power is first converted and processed by the power supply module to provide the required voltage for the signal acquisition module, the core processing module, the display module, the wireless communication module, and the switch module respectively: the surge protection circuit works immediately to prevent the surge voltage in the power grid from damaging the terminal; the AC / DC conversion circuit converts the alternating current into direct current to provide stable power support for each module inside the terminal; the filter circuit filters the power supply to remove the clutter and noise in the power supply to ensure the power supply quality; the DC / DC conversion circuit can convert one DC voltage into another DC voltage to meet the voltage requirements of different modules. After the signal acquisition module is activated by the power supply module, information is collected: the sensor used to collect current signals uses the excellent stability and accuracy of the constantan resistance wire to detect changes in current; the resistor voltage divider circuit is responsible for voltage dividing the signal, and after adjusting the signal strength, it is amplified by the amplifier circuit, and then the A / D conversion circuit converts the analog signal into a digital signal for processing by the core processing module. After the core processing module receives the data from the signal acquisition module, the DSP chip analyzes and processes the data and controls the switch module, that is, the switch action of the relay is controlled according to the predetermined logic. The transistor drive circuit is used to drive the relay circuit, and the EEPROM is used to store the terminal configuration information and data to ensure that the data is not lost even if the power is off. The information processed by the core processing module can be displayed in real time through the display module, and can also be sent to the remote server and other terminals through the wireless communication module.

[0119] Figure 2 : is a flow chart of the modal decomposition and prediction implementation of the method of the present invention. The implementation steps of the modal decomposition and prediction method are:

[0120] Step (1): Get and store the user's previous The total power signal at the moment to the continuous total power signal at the current moment, the original power load total power signal sequence is obtained ;

[0121] Step (2): Perform first-order differential processing on the original power load total power signal sequence to increase the amount of data information and make the data features more significant, and obtain the differential signal sequence ;

[0122] Step (3): Perform first-order integration processing on the original power load total power signal sequence to increase the amount of data information and make the data features more significant, and obtain the integral signal sequence ;

[0123] Step (4): Use the EMD empirical mode decomposition method to perform modal decomposition on the original power load total power signal sequence, remove the noise and interference in the power time series, improve the accuracy and reliability of signal processing, and take the first modal component signal sequence , the second modal component signal sequence , the third modal component signal sequence and the residual component signal sequence ;

[0124] Step (5): Use the GRU gated recurrent unit architecture to , , , , , , After The prediction of the total power at each moment captures the data features and alleviates the gradient disappearance; the prediction result of GRU prediction is: , , , , , , ;

[0125] Step (6): Use LSTM long short-term memory network to , , , , , , After The prediction of total power at each moment captures data features and alleviates gradient explosion. The prediction result of LSTM prediction is: , , , '、 , , ;

[0126] Step (7): Use BiLSTM bidirectional long short-term memory network to , , , , , , After The total power prediction at each moment improves noise tolerance. The prediction result of BiLSTM prediction is: , , , '''、 , , ;

[0127] Figure 3 The following is a flow chart of the Kolmogorov-Arnold Networks method of the present invention. The sequence obtained by modal decomposition in step (4) and the prediction results obtained in steps (5) to (6) are used as the input of KAN. , after feature extraction and adaptive optimization combination processing by KAN network, the output result of KAN network can be obtained ,Right now It is the result of blind source separation of power load;

[0128] The surface formula of KAN is:

[0129] (45)

[0130] The deep formula of KAN is:

[0131] (46)

[0132] in, represents the input of KAN, i.e., the sequence obtained by modal decomposition in step (4) and the prediction results obtained from steps (5) to (6); for The pth element of ; and They are the output of the KAN surface layer and the output of KAN respectively; is an internal function, forming a KAN layer, where the input dimension , output dimension ; is the qth layer external function, forming a KAN layer, where the input dimension , output dimension ; Indicates from arrive The accumulation of Indicates from arrive The shape of KAN consists of an integer array [ n 0 , n 1 , . . . , n L ] express, is the number of nodes at layer 0; is the number of nodes in the first layer; is the number of nodes in the Lth layer; T is the bias of KAN; is the accumulation of functions; is the 0th layer external function; is the first layer external function; is the L-2 layer external function; is the L-1th layer external function;

[0133] use represents the i-th neuron in the k-th layer, and express - The activation value of the neuron between the kth layer and the k+1th layer is Activation function, connection and The activation function is , ; ; ;

[0134] Pre-activation express, Post-activation express, The activation value of is the sum of all incoming post-activation values:

[0135] (47)

[0136] The output matrix of the kth layer for:

[0137] (48)

[0138] In matrix form is the B-spline function matrix of the kth layer, is the input matrix of the kth layer, is the output matrix of the kth layer.

[0139] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by 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 Kolmogorov-Arnold network power load blind source separation method based on modal decomposition and prediction, characterized in that: Combining empirical mode decomposition and recurrent neural network for information processing and prediction in blind source separation problems has the function of optimizing the information processing technology of power load blind source separation and improving the accuracy of power load blind source separation; The steps in use are: Step (1): Get and store the user's previous The total power signal at the moment to the continuous total power signal at the current moment, the original power load total power signal sequence is obtained for: (1) in, For users The total power signal at time, For users The total power signal at time, For users The total power signal at time, is the total power signal of the user at the current moment; Step (2): Perform first-order differential processing on the original power load total power signal sequence to increase the amount of data information and obtain the differential signal sequence for: (2) in, For users The total power signal at time, For users Total power signal at time; Step (3): Perform first-order integration processing on the original power load total power signal sequence to increase the amount of data information and make the data features more significant, and obtain the integral signal sequence for: (3) in, For users The total power signal at the moment The accumulation of the total power signal at time For users The total power signal at the moment The accumulation of the total power signal at time For users The total power signal at the moment The accumulation of the total power signal at time For users The accumulation of the total power signal at time to the total power signal at the current time, for of the moment; Step (4): Use the EMD empirical mode decomposition method to perform modal decomposition on the original power load total power signal sequence, remove the noise and interference in the power time series, improve the accuracy and reliability of signal processing, and take the first modal component signal sequence , the second modal component signal sequence , the third modal component signal sequence and the residual component signal sequence for: (4) (5) (6) (7) in, and The method to obtain is as follows: Step (4-1): Find the extreme points of the total power signal sequence of all original power loads, connect the local maximum points into an upper envelope through a cubic spline curve, and connect the local minimum points into a lower envelope. The upper and lower envelopes contain all the data points to obtain the maximum envelope. and minimum envelope ; Step (4-2): Average the two envelopes to get the average value for: (8) Step (4-3): Subtract the average value from the original power load total power signal sequence to obtain the intermediate signal for: (9) Step (4-4): Here, if If the conditions of the intrinsic mode function are met, then is the first-order intrinsic mode function of the original signal; if If not satisfied, proceed to step (4-5); in, The conditions for the intrinsic mode function are: in the entire data set, the sum of the number of local maxima and local minima must be equal to the number of zero crossings or differ by at most one; at any time point, the upper envelope defined by the local maximum and the lower envelope defined by the local minimum have a mean of zero; Step (4-5): As the original signal, perform a second screening: (10) Where: yes The mean of the upper and lower envelopes; Repeat the filter until Satisfying the definition of the intrinsic mode function, the first-order intrinsic mode function is obtained ; Step (4-6): Convert the original signal minus Get the remaining amount for: (11) Steps (4-7): As the new original signal, re-execute steps (4-1) to (4-6) to obtain the new residual , repeat this process n times and when the nth remaining amount When it becomes a monotonic function or a constant and cannot be decomposed into the intrinsic mode function, the entire EMD process ends; the original signal It can be expressed as n intrinsic mode function components and residual components The specific formula is: (12) in, is the intrinsic mode function component; is the residual component, representing the average trend of the signal, which is a constant sequence or a monotonic sequence. is the accumulation of the first-order intrinsic mode function to the n-th-order intrinsic mode function; Step (5): Use the GRU gated recurrent unit architecture to , , , , , , After Predict the total power at each moment, capture data features, and alleviate gradient disappearance; The prediction result is recorded as: (13) There is a logo in the upper right corner The value of is the value predicted by the GRU gated recurrent unit; The prediction result is recorded as: (14) The prediction result is recorded as: (15) The prediction result is recorded as: (16) The prediction result is recorded as: (17) The prediction result is recorded as: (18) The prediction result is recorded as: (19) The formula for GRU is as follows: (20) (21) (22) (23) in, is the input data at the current moment, is the output of the unit at the current moment, is the output of the unit at the previous moment, To reset the gate, To update the gate, is a candidate hidden state, represents element-wise multiplication, represents the sigmoid function, represents the hyperbolic tangent function, , , , , , and b are parameters to be learned in the deep network, which are the first weight of the reset gate, the second weight of the reset gate, the bias of the reset gate, the first weight of the update gate, the second weight of the update gate, the bias of the update gate, the first weight of the candidate hidden state, the second weight of the candidate hidden state, and the bias of the candidate hidden state respectively; Step (6): Use LSTM long short-term memory network to , , , , , , After Predict the total power at each moment, capture data features, and alleviate gradient explosion; The prediction result is recorded as: (24) There is a logo in the upper right corner The value is the value predicted by the LSTM long short-term memory network; The prediction result is recorded as: (25) The prediction result is recorded as: (26) The prediction result is recorded as: (27) The prediction result is recorded as: (28) The prediction result is recorded as: (29) The prediction result is recorded as: (30) The formula for LSTM is as follows: f t = σ ( w f [ h l s t m , t − 1 , x l s t m , t ] + b f ) (31) i t = σ ( w i [ h l s t m , t − 1 , x l s t m , t ] + b i ) (32) o t = σ ( w o [ h l s t m , t − 1 , x l s t m , t ] + b o ) (33) C t ̃ = t a n h ( w c [ h l s t m , t − 1 , x l s t m , t ] + b c ) (34) (35) (36) in, is the input data at the current moment, is the input status at the current moment, is the state of this unit, i.e., the candidate gate, is the state of the unit at the previous moment, is the output of the unit at the current moment, is the output of the unit at the previous moment; and The two matrices are directly concatenated together as input, with the corresponding weights , , , Multiply and add the corresponding bias , , , ; The forget gate is used to calculate the last unit state Perform selective forgetting and output the historical information of memory; the input gate is for the current input state Selective input is performed, and the output obtained is the current input information; the sum of the memory of historical information and the current input information is the current unit state ; The output gate selectively outputs the current unit state to obtain the final output ; , , and They are the weight of the forget gate, the weight of the input gate, the weight of the output gate, and the weight corresponding to the input state at the current moment; , , , They are the bias of the forget gate, the bias of the input gate, the bias of the output gate, and the bias corresponding to the input state at the current moment; is the output of the forget gate; is the output of the output gate; is the output of the input gate; Step (7): Use BiLSTM bidirectional long short-term memory network to , , , , , , After Predict total power at all times to improve noise tolerance; The prediction result is recorded as: (37) There is a logo in the upper right corner The value is the value predicted by the BiLSTM bidirectional long short-term memory network; The prediction result is recorded as: (38) The prediction result is recorded as: (39) The prediction result is recorded as: (40) The prediction result is recorded as: (41) The prediction result is recorded as: (42) The prediction result is recorded as: (43) The formula of BiLSTM is as follows: h = [ h f t 1 , h b t 1 ] (44) in, is the forward LSTM output, is the reverse LSTM output, is the number of layers; Step (8): Take the sequence obtained by modal decomposition in step (4) and the prediction results obtained from steps (5) to (6) as the input of KAN , feature extraction and adaptive optimization combination processing are performed through the KAN network, and finally the output result of the KAN network is obtained ,Right now is the result of blind source separation of power load; KAN is Kolmogorov-Arnold Networks; The surface formula of KAN is: (45) The deep formula of KAN is: (46) in, represents the input of KAN, i.e., the sequence obtained by modal decomposition in step (4) and the prediction results obtained from steps (5) to (6); for The pth element of ; and They are the output of the KAN surface layer and the output of KAN respectively; is an internal function, forming a KAN layer, where the input dimension , output dimension ; is the qth layer external function, forming a KAN layer, where the input dimension , output dimension ; Indicates from arrive The accumulation of Indicates from arrive The shape of KAN consists of an integer array [ n 0 , n 1 , . . . , n L ] express, is the number of nodes at layer 0; is the number of nodes in the first layer; is the number of nodes in the Lth layer; T is the bias of KAN; is the accumulation of functions; is the 0th layer external function; is the first layer external function; is the L-2 layer external function; is the L-1th layer external function; use represents the i-th neuron in the k-th layer, and express - The activation value of the neuron between the kth layer and the k+1th layer is Activation function, connection and The activation function is , ; ; ; Pre-activation express, Post-activation express, The activation value of is the sum of all incoming post-activation values: (47) The output matrix of the kth layer for: (48) In matrix form is the B-spline function matrix of the kth layer, is the input matrix of the kth layer, is the output matrix of the kth layer.

Citation Information

Cited By

  • Multilevel network intrusion detection method and system based on interpretive analysis and empirical knowledge optimization, and medium

    CN119094245A

  • Multi-level network intrusion detection method, system and medium based on explanatory analysis and experience knowledge optimization

    CN119094245B

  • Glass hardness prediction method based on hierarchical prediction neural network

    CN119314585A

  • Low-altitude low-speed small target threat assessment method and system based on LSTM model in combination with KAN network

    CN119919869A

  • Fourier series time sequence prediction method based on KoImogorov-ArnoId theory

    CN120448822A