Load potential prediction method based on TTNRBO-VMD algorithm

Abstract

The invention discloses a load potential prediction method based on a TTNRBO-VMD algorithm, and belongs to the technical field of adjustable load resource pool optimization scheduling, and the method comprises the following steps: S1, constructing a TTNRBO-VMD model, introducing an intrinsic component IMF obtained through signal decomposition into a variation model for processing on the basis of empirical mode decomposition through variational mode decomposition VMD, and obtaining a TTNRBO-VMD model; a plurality of K modal components vk (t) with relatively high robustness to noise can be obtained; and S2, establishing a user potential prediction model, and performing optimization solution on each modal component after VMD decomposition by using a long short-term memory (LSTM) network algorithm. The load potential prediction method based on the TTNRBO-VMD algorithm is used for supporting construction and scheduling of an adjustable load resource pool, the convergence speed is increased, the prediction precision and applicability are improved, and the problems that demand potential prediction is long in time consumption, large in workload, high in subjectivity, low in reliability, large in data demand and the like are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of optimized scheduling of adjustable load resource pools, and particularly to a load potential prediction method based on the TTNRBO-VMD algorithm. Background Art

[0002] At present, the methods for evaluating the potential of demand response users can be mainly divided into two types: questionnaire surveys and data analysis. However, for power companies, the number of their power users is huge, and some users do not have a deep understanding of demand response (DR). The questionnaire survey method has problems such as long time consumption, large workload, strong subjectivity, and low reliability. The data analysis-based method can be further divided into two ways: mechanism-driven and data-driven. Although the mechanism-driven method has high reliability, the load modeling process is complex, overly dependent on user load parameters, and difficult to promote on a large scale. Although the data-driven method has a simple modeling process, a large number of data samples are required during the training of machine learning algorithms. If the quantity is limited, the training effect is also difficult to meet the requirements. Summary of the Invention

[0003] The purpose of the present invention is to provide a load potential prediction method based on the TTNRBO-VMD algorithm, which is used to support the construction and scheduling of adjustable load resource pools, can accelerate the convergence speed, can improve the prediction accuracy and applicability, and can solve problems such as long time consumption, large workload, strong subjectivity, low reliability, and large data demand in demand potential prediction.

[0004] To achieve the above purpose, the present invention provides a load potential prediction method based on the TTNRBO-VMD algorithm, including the following steps:

[0005] S1. Construct a TTNRBO-VMD model. Based on empirical mode decomposition, variational mode decomposition (VMD) introduces the intrinsic mode functions (IMFs) obtained by decomposing the signal into a variational model for processing, and obtains K modal components with strong robustness to noise.

[0006] S2. Establish a user potential prediction model, and use the long short-term memory network (LSTM) algorithm to optimize and solve each modal component after VMD decomposition.

[0007] Preferably, the process of obtaining the K modal components v k (t) in S1 is as follows:

[0008] Use the Hilbert transform to obtain the analytical signal of v k (t), and modulate the central band of v k (t) to the corresponding baseband, then demodulate and estimate the bandwidth using the Gaussian smoothness of the demodulated signal. The constraint relationship can be expressed as:

[0009]

[0010] In the formula, {v k} = {v1,..., v k} is the modal component IMF after VMD decomposition; {ω k} = {ω1,..., ω k} is the central frequency corresponding to each IMF; δ(t) is the impulse function; is the gradient calculation; f(t) is the original signal, K is the number of modes after modal decomposition, t is the current iteration number, j is the imaginary unit, s.t. means subject to, represents the integral kernel in the case of K modes;

[0011] Introduce the Lagrange multiplier τ(t) and the penalty factor α, and transform the constrained variational mode problem into an unconstrained variational mode problem. The transformed Lagrangian function is expressed as:

[0012]

[0013] where s(t) is the filtered signal;

[0014] Use the alternating direction multiplier method to iteratively update the modal components v k n+1 , ω k n+1 , τ. The obtained Lagrangian saddle point is the optimal solution of the modal decomposition; v k , ω k , τ's iterative constraint conditions are:

[0015]

[0016] In the formula, v k n+1 (ω), and are the results of v k n+1 , ω k n+1 , τ through the Hilbert transform in the time domain, represents the i-th modal component after the (n + 1)-th iteration, represents the i-th modal component after the n-th iteration, represents the central frequency of the k-th mode after the n-th iteration, λ represents the Lagrange multiplier;

[0017] The selection of the penalty factor α and the decomposition layer number K in VMD decomposition will have a great impact on the decomposition effect. Introduce the tuna optimization TSO to eliminate the problem of inaccurate iteration of the penalty factor and the decomposition layer number, and use the TSO algorithm to optimize and solve the parameters K and α of the VMD algorithm;

[0018] The iterative process of the TSO algorithm is mainly divided into a spiral foraging stage and a parabolic foraging stage according to the value of the random number r1.

[0019] Preferably, in the spiral foraging stage, that is, when r1 < 0.5, there is:

[0020]

[0021] β = e bl ·cos(2πr3);

[0022]

[0023] In the formula: X i (t + 1) is the i-th modal component in the (t + 1)-th iteration; X i (t) is the i-th modal component in the t-th iteration; X σ (t) is the modal component reference point randomly generated in the t-th iteration; X ψ (t) is the optimal solution in the t-th iteration, and α1, α2 are weight coefficients that control the change trend of the modal components, where θ is a constant with a value of 0.7; t and t max are the current iteration number and the maximum iteration number; N P is the total number of modes, β is the adaptive change coefficient, and l is the compensation factor.

[0024] Preferably, in the parabolic foraging stage, that is, when r1 ≥ 0.5, there is:

[0025]

[0026] In the formula: T F is a random number with a value of 1 or -1; r1, r2, r3, r4, r5 are random numbers with a value range of [0, 1].

[0027] Preferably, the Newton - Raphson optimization algorithm NRBO is used to optimize the tuna iterative process; for a set with a fixed number of modes N P assuming that the dataset to be optimized has n dimensions, the initial position of the randomly generated set is:

[0028] x nj = lo + rand(up - lo);

[0029] In the formula, x nj is the position of the m-th modal component in the set in the k-th dimension; k is a natural number with a value range of [1, n], and the value range of m is [1, N P; up and lo are the lower and upper limits of the parameters to be optimized respectively; rand is a random number between (0, 1).

[0030] Calculate the initial fitness value: According to the set fitness function, calculate the fitness value of each modal component, and screen out the best fitness value, the worst fitness value and their corresponding positions x b and x w ; Apply the NRSR rule to explore new solutions; For the nth modal component in the tth iteration, apply NRSR to explore the modal position, and the new position is expressed as:

[0031]

[0032] In the formula, is the new position explored by applying the NRSR rule, t and n are the iteration number and the modal component serial number respectively; r1 and r2 respectively represent random numbers between [0, 1]; are 3 positions updated from the current position by NRBO to enhance the local search performance and the global search performance respectively, and the expressions are:

[0033]

[0034] In the formula, Nr is the value calculated by applying the NRSR rule; x b is the current best position of the modal component, is the position of the nth modal component in the tth iteration; Rho is a step factor that guides each modal component in the correct direction; δ is a coefficient that adaptively changes with the iteration number, which can avoid the occurrence of local optimum problems.

[0035] Preferably, introduce the trap avoidance operation TAO to improve the quality of the solution and avoid local optimum traps. By combining the best position x of the modal component b and its corresponding NRSR vector position to generate a new position with enhanced quality By comparing the random number rand between (0, 1) with the DF value, the DF value is usually 0.6; The new value generated is:

[0036]

[0037] In the formula, θ1 and θ2 are random numbers between (-1, 1) and (-0.5, 0.5) respectively, μ1 and μ2 are random numbers, Mean() represents taking the average value. Compare a random number Δ between (0, 1) with 0.5. If Δ≥0.5, μ1 and μ2 are assigned 1, otherwise calculate according to the following formula, that is:

[0038]

[0039] Preferably, the solution steps in S2 are as follows:

[0040] f t = σ(W f [H t-1 , X t + b f );

[0041] i t = σ(W i [H t-1 , X t + b i );

[0042]

[0043] O t = σ(W o [H t-1 , X t + b o );

[0044] H t = O t tanh(C t );

[0045] Where: W f , W i , W c , W o are the matrices of the forget gate, input gate, cell state, and output gate respectively; X t is the input modal component; H t is the hidden state; f t , i t , C t , O t are the outputs of the forget gate, input gate, cell state, and output gate modal components respectively; b f , b i , b c , b o are the corresponding bias constants, and σ is the sigmoid function; is the temporary state.

[0046] Therefore, the present invention adopts the above-mentioned load potential prediction method based on the TTNRBO-VMD algorithm, which has a more accurate prediction accuracy of the Yongfu adjustable potential, and is effective and practical for excavating the potential on the demand side, providing a technical solution for the construction of the resource pool.

[0047] The technical solution of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings

[0048] Figure 1 is a flowchart of an embodiment of a load potential prediction method based on the TTNRBO-VMD algorithm of the present invention;

[0049] Figure 2 is a structural diagram of the TTNRBO-VMD model of an embodiment of a load potential prediction method based on the TTNRBO-VMD algorithm of the present invention. Detailed implementation manners

[0050] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0051] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs.

[0052] Embodiment 1

[0053] As Figure 1 shown, the present invention provides a load potential prediction method based on the TTNRBO-VMD algorithm, including the following steps:

[0054] S1. Construct a TTNRBO-VMD model;

[0055] Variational mode decomposition (VMD) is an adaptive and completely non-recursive signal decomposition technique that can decompose a multi-component signal into multiple single-component amplitude-modulated and frequency-modulated signals. Based on empirical mode decomposition, VMD introduces the intrinsic mode functions IMF obtained by signal decomposition into a variational model for processing, and K modal components v k (t) with strong robustness to noise can be obtained. The detailed process is as follows:

[0056] (1) Use Hilbert transform to obtain the analytical signal of v k (t), and modulate the center band of v k (t) to the corresponding baseband, then demodulate and estimate the bandwidth using the Gaussian smoothness of the demodulated signal. The constraint relationship can be expressed as:

[0057]

[0058] In the formula, {v k} = {v1,..., v k} are the modal components IMF after VMD decomposition; {ω k} = {ω1,..., ω k} are the center frequencies corresponding to each IMF; δ(t) is the impulse function; For gradient calculation; f(t) is the original signal, K is the number of modes after modal decomposition, t is the current iteration number, j is the imaginary unit, s.t. means subject to, represents the integral kernel when there are K modes;

[0059] Introduce the Lagrange multiplier τ(t) and the penalty factor α, and transform the constrained variational mode problem into an unconstrained variational mode problem. The transformed Lagrangian function is expressed as:

[0060]

[0061] where s(t) is the filtered signal;

[0062] Use the alternating direction multiplier method to iteratively update the modal components v k n+1 and ω k n+1 and τ. The obtained Lagrangian saddle point is the optimal solution of the modal decomposition; v k and ω k The iterative constraint conditions for τ are:

[0063]

[0064] In the formula, v k n+1 (ω), and are the results of v k n+1 and ω k n+1 and τ through the Hilbert transform in the time domain. represents the i-th modal component after the (n + 1)-th iteration, represents the i-th modal component after the n-th iteration, represents the center frequency of the k-th mode after the n-th iteration, and λ represents the Lagrange multiplier;

[0065] The selection of the penalty factor α and the decomposition layer number K in the VMD decomposition will have a greater impact on the decomposition effect. If α is too large, it may lead to modal aliasing and loss of important signals; if α is too small, it will cause too many modes and introduce noise. If K is too large, it will overfit the noise and increase the signal complexity. If K is too small, it will lead to insufficient decomposition.

[0066] To eliminate the problems of inaccurate penalty factors and iteration of decomposition layers, the present invention introduces Tunaswarm optimization (TSO). Tunaswarm optimization is mainly a meta-heuristic optimization algorithm that simulates the cooperative foraging behavior of tuna groups. By simulating the spiral foraging and parabolic foraging strategies of tuna groups to find the optimal solution, the present invention uses the TSO algorithm to optimize and solve the parameters K and α of the VMD algorithm.

[0067] The iterative process of the TSO algorithm is mainly divided into a spiral foraging stage and a parabolic foraging stage according to the value of the random number r1.

[0068] In the spiral foraging stage, that is, when r1 < 0.5, there is:

[0069]

[0070] β = e bl ·cos(2πr3);

[0071]

[0072] Where: X i (t + 1) is the i-th modal component of the (t + 1)-th iteration; X i (t) is the i-th modal component of the t-th iteration; X σ (t) is the modal component reference point randomly generated in the t-th iteration; X ψ (t) is the optimal solution of the t-th iteration, and α1 and α2 are weight coefficients that control the change trend of the modal component, where θ is a constant with a value of 0.7; t and t max are the current iteration number and the maximum iteration number; N P is the total number of modes, β is the adaptive change coefficient, and l is the compensation factor.

[0073] In the parabolic foraging stage, that is, when r1 ≥ 0.5, there is:

[0074]

[0075] Where: T F is a random number with a value of 1 or -1; r1, r2, r3, r4, and r5 are random numbers with a value range of [0, 1].

[0076] To further explore the optimal results, the Newton-Raphson (NRBO) algorithm is used to optimize the tuna iteration process. NRBO mainly uses the search rule (NRSR) and the trap avoidance operator (TAO) to explore the entire search process and further explores the optimal results using multiple groups of matrices. Aiming at the problems of slow convergence speed, easy to fall into local optimum, and large power fluctuations in the later stage of the TSO algorithm, the present invention uses the NRBO optimization algorithm to optimize the tuna iteration process.

[0077] For a set with a fixed number of modes N P assuming that the dataset to be optimized has n dimensions, the initial position of the randomly generated set is:

[0078] x nj = lo + rand(up - lo);

[0079] where x nj is the position of the m-th mode component in the set at the k-th dimension; k is a natural number, with a value range of [1, n], and the value range of m is [1, N P ; up and lo are respectively the lower and upper limits of the parameters to be optimized; rand is a random number between (0, 1).

[0080] Calculate the initial fitness value: According to the set fitness function, calculate the fitness value of each mode component, and screen out the best fitness value, the worst fitness value, and their corresponding positions x b and x w ; Apply the NRSR rule to explore new solutions; For the n-th mode component in the t-th iteration, apply the NRSR to explore the mode position, and the new position is expressed as:

[0081]

[0082] where, is the new position explored by applying the NRSR rule, t and n are respectively the iteration number and the mode component serial number; r1 and r2 respectively represent random numbers between [0, 1]; are respectively 3 positions updated from the current position by NRBO to enhance the local search performance and the global search performance, and the expressions are:

[0083]

[0084] where Nr is the value calculated by applying the NRSR rule; x b is the current best position of the mode component, is the position of the nth modal component in the t-th iteration; Rho is a step factor that guides each modal component in the correct direction; δ is a coefficient that adaptively changes with the number of iterations, which can avoid the occurrence of local optimal problems.

[0085] To improve the quality of the solution and avoid local optimal traps, a trap avoidance operation (TAO) is introduced to improve the quality of the solution and avoid local optimal traps. By combining the best position x of the modal components b and its corresponding NRSR vector position to generate a new position with enhanced quality By comparing the (0,1) random number rand with the DF value, the DF value is usually 0.6; the new value generated is:

[0086]

[0087] In the formula, θ1 and θ2 are random numbers between (-1,1) and (-0.5,0.5) respectively, μ1 and μ2 are random numbers, Mean() represents taking the average value. Using a random number Δ between (0,1) to compare with 0.5, if Δ≥0.5, μ1 and μ2 are assigned 1, otherwise they are calculated according to the following formula, that is:

[0088]

[0089] The algorithm calculation process is as Figure 2 shown.

[0090] S2. Establish a user potential prediction model;

[0091] LSTM (Long Short-Term Memory Network) is a special type of Recurrent Neural Network (RNN) that can learn long-term dependency information. It performs well in tasks such as time series prediction, natural language processing, and speech recognition that require capturing long-term dependencies. The invention uses this algorithm to optimize and solve each modal component after VMD decomposition. The solution steps are:

[0092] f t = σ(W f [H t-1 , X t +b f );

[0093] i t = σ(W i [H t-1 , X t +b i );

[0094]

[0095] O t= σ(W o [H t-1 , X t + b o );

[0096] H t = O t tanh(C t );

[0097] Where: W f , W i , W c , W o are the matrices of the forget gate, input gate, cell state, and output gate respectively; X t is the input modal component; H t is the hidden state; f t , i t , C t , O t are the outputs of the forget gate, input gate, cell state, and output gate modal components respectively; b f , b i , b c , b o are the corresponding bias constants, σ is the sigmoid function; C~ t is the temporary state.

[0098] Therefore, the present invention adopts the above-mentioned load potential prediction method based on the TTNRBO-VMD algorithm, which can be applied to the construction project of the demand-side resource pool, used to support the construction and scheduling of the adjustable load resource pool, speeds up the convergence rate, improves the prediction accuracy and applicability, and solves the problems of long time-consuming, large workload, strong subjectivity, low reliability, and large data demand in demand potential prediction.

[0099] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that: they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A load potential prediction method based on the TTNRBO-VMD algorithm, characterized in that: It includes the following steps: S1. Construct the TTNRBO-VMD model. Based on empirical mode decomposition, variational mode decomposition (VMD) introduces the intrinsic mode functions (IMFs) obtained by decomposing the signal into a variational model for processing, and obtains K modal components v k (t) that have strong robustness to noise; S2. Establish a user potential prediction model, and use the long short-term memory network LSTM algorithm to optimize and solve each modal component after VMD decomposition.

2. The load potential prediction method based on the TTNRBO-VMD algorithm according to claim 1, characterized in that: The process of obtaining K modal components v k (t) in S1 is as follows: k (t) is as follows: Obtain \(v\) using Hilbert transform k (t)'s analytic signal, and modulate the center band of \(v\) k (t) to the corresponding baseband, then demodulate and estimate the bandwidth using the Gaussian smoothness of the demodulated signal. The constraint relationship can be expressed as: where {v k} = {v1,..., v k} are the mode components IMF after VMD decomposition; {ω k} = {ω1,..., ω k} are the central frequencies corresponding to each IMF; δ(t) is the impulse function; is the gradient calculation; f(t) is the original signal, K is the number of modes after mode decomposition, t is the current iteration number, j is the imaginary unit, s.t. means subject to, represents the integral kernel at K modes; Introduce the Lagrange multiplier τ(t) and the penalty factor α to transform the constrained variational mode problem into an unconstrained variational mode problem. The transformed Lagrangian function is expressed as: Among them, s(t) is the filtered signal; Iteratively update the modal component v using the alternating direction method of multipliers k n+1 , ω k n+1 , τ, and the obtained Lagrangian saddle point is the optimal solution of the modal decomposition; v k , ω k , The iterative constraint conditions for τ are as follows: where \(v\) k n+1 \((\omega)\) and are the results of the Hilbert transform of \(v\) k n+1 \(\omega\) k n+1 \(\tau\) in the time domain, and \(\hat{v}_{i}^{n + 1}\) represents the \(i\)-th modal component after the \((n + 1)\)-th iteration, \(\hat{v}_{i}^{n}\) represents the \(i\)-th modal component after the \(n\)-th iteration, \(\omega_{k}^{n}\) represents the center frequency of the \(k\)-th mode after the \(n\)-th iteration, and \(\lambda\) represents the Lagrange multiplier; The selection of the penalty factor α and the decomposition layer number K in VMD decomposition will have a great impact on the decomposition effect. Introduce the tuna optimization TSO to eliminate the problem of inaccurate iteration of the penalty factor and the decomposition layer number, and use the TSO algorithm to optimize and solve the parameters K and α of the VMD algorithm; The iterative process of the TSO algorithm is mainly divided into a spiral foraging stage and a parabolic foraging stage according to the value of the random number r1.

3. A load potential prediction method based on the TTNRBO-VMD algorithm according to claim 2, characterized in that: In the spiral foraging stage, that is, when r1 < 0.5, there is: β = e bl ·cos(2πr3); Where: X i (t + 1) is the i-th modal component at the (t + 1)-th iteration; X i (t) is the i-th modal component at the t-th iteration; X σ (t) is the reference point of the randomly generated modal component at the t-th iteration; X ψ (t) is the optimal solution at the t-th iteration, α1 and α2 are weight coefficients that control the change trend of the modal component, where θ is a constant with a value of 0.7; t and t max are the current iteration number and the maximum iteration number; N P is the total number of modes, β is the adaptive change coefficient, and l is the compensation factor.

4. A load potential prediction method based on the TTNRBO-VMD algorithm according to claim 2, characterized in that: In the parabolic foraging stage, that is, when r1 ≥ 0.5, there is: Where: T F is a random number, taking a value of 1 or -1; r1, r2, r3, r4, and r5 are random numbers, with a value range of [0, 1].

5. A load potential prediction method based on the TTNRBO-VMD algorithm according to claim 4, characterized in that: Use the Newton-Raphson optimization algorithm NRBO to optimize the tuna iteration process; for a fixed number of modes N P in the set, assuming that the dataset to be optimized has n dimensions, the initial position of the randomly generated set is: x nj = lo + rand(up - lo); where x nj is the position of the m-th modal component in the set at the k-th dimension; k is a natural number, and its value range is [1, n], and the value range of m is [1, N P ; up and lo are the lower and upper limits of the parameters to be optimized respectively; rand is a random number between (0, 1). Calculate the initial fitness value: According to the set fitness function, calculate the fitness value of each modal component, and screen out the best fitness value, the worst fitness value, and their corresponding positions x b and x w ; Apply the NRSR rule to explore new solutions; For the nth modal component in the tth iteration, apply NRSR to explore the modal position, and the new position is expressed as: wherein, is the new position explored by applying the NRSR rule, t and n are the iteration number and the mode component number respectively; r1 and r2 respectively represent random numbers between [0, 1]; are respectively three positions updated from the current position by NRBO to enhance the local search performance and the global search performance, and the expression is: where Nr is the value calculated by applying the NRSR rule; x b is the current optimal position of the modal component, is the position of the n-th modal component at the t-th iteration; Rho is a step factor that guides each modal component in the correct direction; δ is a coefficient that adaptively changes with the number of iterations and can avoid the occurrence of local optimal problems.

6. The load potential prediction method based on the TTNRBO-VMD algorithm according to claim 5, wherein: Introduce the trap avoidance operation TAO to improve the quality of the solution and avoid local optimal traps by combining the best positions x of the modal components b and their corresponding NRSR vector positions to generate new positions with enhanced quality By comparing the (0,1) random number rand with the DF value, the DF value is usually 0.6; the new value generated is: In the formula, θ1 and θ2 are random numbers between (-1, 1) and (-0.5, 0.5) respectively, μ1 and μ2 are random numbers, Mean() represents taking the average value. Compare a random number Δ between (0, 1) with 0.

5. If Δ ≥ 0.5, μ1 and μ2 are assigned 1, otherwise they are calculated according to the following formula, that is:

7. A load potential prediction method based on the TTNRBO-VMD algorithm according to claim 1, characterized in that: The solution steps in S2 are: f t = σ(W f [H t-1 , X t + b f ); i t = σ(W i [H t-1 , X t + b i ); O t = σ(W o [H t-1 , X t + b o ); H t = O t tanh(C t ) Where: W f 、W i 、W c 、W o are the matrices of the forget gate, input gate, cell state, and output gate respectively; X t is the input modal component; H t is the hidden state; f t 、i t 、C t 、O t are the outputs of the forget gate, input gate, cell state, and output gate modal components respectively; b f 、b i 、b c 、b o are the corresponding bias constants, and σ is the sigmoid function; is the temporary state.

Images