Energy generating capacity prediction method based on evolutionary variational mode decomposition

Abstract

The invention discloses an energy generating capacity prediction method based on evolutionary variational mode decomposition, and aims to improve the accuracy and efficiency of energy generating capacity prediction. According to the method, an energy generating capacity time sequence is decomposed into a high-frequency mode, an intermediate-frequency mode and a low-frequency mode, trend smoothing is conducted on the low-frequency mode through a least square method, trend components of the intermediate-frequency mode are removed through a differential filter so as to enhance the capturing capacity of periodic fluctuation, and stable prediction is conducted on the high-frequency mode through a dendritic learning model. And carrying out adaptive adjustment on the modal weight in combination with a multi-population optimized substance state search algorithm, and fusing prediction results of each modal through a weighted summation method. According to the method, the problem of capturing periodic and trend characteristics in energy generating capacity prediction is effectively solved, so that the power system management and resource allocation efficiency is optimized.

Description

Technical Field

[0001] The present invention relates to the field of energy management and optimization, and in particular to an energy generation prediction method based on evolutionary variational mode decomposition. Background Art

[0002] Energy generation forecasting is a key link in power system management and optimization, especially with the development of smart grids and large-scale application of renewable energy. The accuracy of the forecast is directly related to the stable operation of the power system, the effective utilization of resources, and the efficient integration of renewable energy. However, the accuracy of power generation forecasting is affected by many uncertain factors such as weather conditions, equipment failures, and demand fluctuations, which increase the complexity and uncertainty of the forecast. Therefore, there are multiple challenges in accurately forecasting energy generation.

[0003] In the past few decades, power forecasting technology has undergone a transformation from traditional statistical methods to complex machine learning algorithms. Early forecasting models were mainly based on statistical regression analysis, such as autoregressive models (AR) and moving average models (MA), which performed well when the data series was relatively stable and without strong cyclical fluctuations. However, with the opening of the power market and the increase in the proportion of renewable energy, the load and power generation of the power system have shown more complex nonlinear characteristics and significant time variability.

[0004] In order to capture these complex dynamic relationships, more advanced technologies such as neural networks, support vector machines (SVM) and ensemble learning have been gradually introduced into power generation forecasting. Especially in renewable energy generation, such as photovoltaic and wind farms, the complexity and challenges of forecasting are particularly prominent. Since the output of renewable energy is affected by many uncontrollable factors, such as meteorological conditions (wind speed, solar radiation, temperature, etc.) and equipment performance, the forecasting model must have a high degree of adaptability and accuracy. Typically, some studies use BP neural networks combined with photovoltaic power station output data and meteorological data for forecasting, trying to improve the accuracy of the forecast by learning the complex nonlinear relationship of the input data. Similarly, some studies use segmented time series and Elman neural networks to deal with power attenuation and time changes, and improve the model's ability to capture power fluctuations.

[0005] Patent No. CN202211458203.6 discloses a multi-factor wind farm power generation forecasting method, which optimizes parameters through the beetle swarm algorithm and uses approximate entropy method and principal component analysis for data processing. However, in practical applications, it may still face the complexity of parameter selection and data processing. Patent No. CN202310147820.2 discloses a short-term power generation forecasting method based on ensemble learning, which extracts trend terms and periodic terms of the sequence through the MSTL decomposition method. Although it takes periodicity and trend into account, it may still lack sufficient flexibility and accuracy when dealing with forecasts in complex environments.

[0006] Although existing forecasting technologies have made progress in some aspects, there are still some shortcomings. First, many models face challenges in computational efficiency and stability when processing large-scale data sets. Especially under the demand for real-time or near real-time forecasting, the model must respond quickly and process continuously changing input data. In addition, many existing models fail to fully consider long-term trends and cyclical changes in the data, which is particularly important in the forecasting of renewable energy power generation because the output of these energy sources strongly depends on seasonal and daily variation factors.

[0007] Although neural networks and support vector machines perform well in capturing nonlinear relationships in data, they often require a large amount of training data and a complex parameter tuning process. In addition, the black-box nature of these models also makes it difficult to interpret the results and to intuitively understand the factors affecting model predictions. Therefore, developing an efficient prediction method that can handle large-scale data sets and fully reflect the long-term trends and cyclical changes in data is an important scientific and technological challenge facing the power industry. Summary of the invention

[0008] In order to solve the problems of insufficient prediction accuracy and failure to fully consider the periodicity and trend in the actual prediction scenario in the existing energy generation prediction methods, the present invention proposes an energy generation prediction method based on evolutionary variational mode decomposition, which decomposes the energy generation time series into three modes of high frequency, medium frequency and low frequency to effectively extract the intrinsic characteristics of the sequence, specifically including the following processes:

[0009] S1: Perform variational modal decomposition on the original power generation signal and decompose it into S intrinsic mode functions, that is, solve the following variational problem, and its calculation formula is:

[0010]

[0011] Where k represents the index of the mode, S represents the number of modes, and u k represents the kth mode function, t is the time variable, and represents each time point in the time series of the signal; represents the instantaneous rate of change of each modal signal; wk represents the center frequency of the kth mode, δ(t) represents the impulse response of the Hilbert transform, is the standard expression of the Hilport transform, which represents the impulse response of the filter. is at frequency ω k The complex exponential function at represents the adjustment of the modal function in the frequency domain, f represents the original signal, that is, the sum of all modes is equal to the target signal, and the purpose of the algorithm is to minimize all modal functions u k A collection of .

[0012] The quadratic penalty factor α and the Lagrange multiplier λ are introduced to transform the constrained variational problem into an unconstrained variational problem. The calculation formula is:

[0013]

[0014] Where L is the Lagrangian objective function, the unconstrained variational problem is solved by the alternating direction multiplier method, the saddle point is found by alternating update calculation, and u is optimized by alternating k , while satisfying the constraints, minimizing the original objective function:

[0015]

[0016] Where f is the original signal, j is the imaginary unit, ω is the angular frequency, jω is used to map the signal from the time domain to the frequency domain, indicating the amplitude and phase of the signal at different frequencies, sgn(ω+ω k ) is to adjust the phase of the mode function in the frequency domain, is the representation of the kth mode function in the frequency domain, The frequency domain representation of the kth mode is frequency-shifted so that its center frequency is aligned with ω k , is the frequency representation of the original signal, and λ(ω) is the frequency domain representation of the Lagrange multiplier.

[0017] S2: For different modes obtained by decomposition, perform the following processing steps respectively:

[0018] S2-1: For low frequency mode Mo 1 , the least square method is used to smooth the trend and fit the long-term trend. The calculation formula is:

[0019] Mo 1 ′=p 0 +p 1 l+p 2 l 2 +p 3 l 3

[0020] S2-2: For the intermediate frequency mode Mo2 , a differential filter is used to remove trend components and enhance the ability to capture periodic fluctuations. The calculation formula is:

[0021]

[0022] S2-3: For high frequency mode Mo 3 , a dendritic learning model is used for stable prediction to accurately capture nonlinear relationships. The calculation formula is:

[0023] Mo 3 ′=DLM(Mo 3 )

[0024] S3: The prediction results of each mode are integrated by weighted summation to obtain the final energy generation prediction result NY'. The calculation formula is:

[0025] NY'=αMo 1 '+βMo 2 '+γMo 3 '

[0026] Among them, Mo 1 ′、Mo 2 'and Mo 3 ' are the prediction results of low-frequency, medium-frequency and high-frequency modes after processing, l is the input data, is the cth data point of the original data sequence, g is the total length of the sequence, and p 0 、p 1 、p 2 and p 4 are polynomial coefficients determined by the least squares method, and α, β, and γ are weight parameters of each mode.

[0027] S4: Optimize the weight parameters α, β and γ using a multi-population improved material state search algorithm, specifically including the following steps:

[0028] S4-1: Initialize the population and set the maximum number of iterations iter, where the population consists of multiple candidate weight parameter combinations, and a candidate weight parameter combination represents a set of weight parameters (α i ,β i ,γ i ), the candidate weight parameter combinations in the population represent individuals;

[0029] S4-2: Each individual’s position update is subject to the direction vector d i The control of this vector determines the individual Q i Moving direction and step length in the solution space, direction vector d i The calculation formula is:

[0030]

[0031] in, represents the direction vector of individual i after the update in the q+1 generation, q is the current generation, iter is the maximum number of iterations, niter is the current number of iterations, Q best is the position of the current global best individual, Q i is the individual's current position.

[0032] S4-3: Calculate velocity vector v i , and its calculation formula is:

[0033]

[0034] in, and are the upper and lower bounds of the j-th dimension, η∈[0,1], and n is the population size.

[0035] S4-4: By updating the direction vector and velocity vector, the new position Q ij The calculation formula is:

[0036]

[0037] Among them, v i,j is the velocity of the ith individual in the jth dimension, λ∈[0,1]; collision behavior occurs during the molecular evolution process. When two molecules d a and d b When the distance between them is less than the collision distance, they exchange direction vectors, i.e., d a =d b , d b =d a ;

[0038] S4-5: Calculate the collision distance r, the calculation formula is:

[0039]

[0040] Among them, η∈[0,1];

[0041] S4-6: There is a certain probability of random change in the process of molecular evolution. In order to simulate this mechanism, the calculation formula of molecular position is:

[0042]

[0043] Among them, rand(0,1) represents a random number between [0,1], and H is a hyperparameter set for each stage.

[0044] As a technical preferred solution of the present invention, the variational problem in step S1 is constructed by: first, using Hilbert transform to calculate the analytical signal of the original signal, and analyzing its spectrum through frequency domain transform; then estimating the center frequency of each modal function and limiting the bandwidth of the modal spectrum; finally, optimizing the spectrum distribution through frequency domain bandwidth calculation.

[0045] As a technical preferred solution of the present invention, the DLM described in step S2 is a dendritic learning model, which is composed of a synaptic layer, a dendritic layer, a membrane layer, a somatic layer and an axon layer; wherein the synaptic layer uses a Sigmoid function to process the input signal, the dendritic layer uses an addition operation to sum the output of the synaptic layer, the membrane layer uses a multiplication operation to perform a nonlinear combination of the results of each dendritic layer, the somatic layer is normalized by a standard score, and the axon layer multiplies the discrete weight of each input corresponding to the synaptic layer by the processed somatic output to obtain the final result of the model.

[0046] The calculation formulas of the components of the DLM dendrite learning model are:

[0047] Synaptic layer:

[0048]

[0049] Dendritic layer:

[0050]

[0051] Film layer:

[0052]

[0053] Somatic layer:

[0054]

[0055] Axon layer:

[0056]

[0057] Among them, x i is the input of power generation, k s is a positive constant, w ij is the weight of the synaptic layer, θ ij is the threshold, Y is the output of the synapse, N is the number of synapses, and Z j is the output of the dendritic layer, M is the number of dendritic layers, V is the output of the membrane layer, num represents the number of neurons, and std represent the mean and standard deviation of the calculated data, σ represents the Sigmoid function, var represents the variance, and k νIt is a hyperparameter of the somatic layer. By continuously adjusting the weights and thresholds of the model through back propagation, DLM can learn various complex functional relationships.

[0058] As a technical preferred solution of the present invention, the multi-population improved material state search algorithm in step S4 is further improved into a multi-population improved material state search algorithm, which divides the population into two sub-populations according to the fitness of the individuals: the first half is the better population Q g (q), the latter part is the inferior population Q b (q) The individuals in the inferior sub-population are updated by the individuals in the superior sub-population.

[0059] A multi-population improved material state search algorithm includes the following steps:

[0060] T1: When the collision step in SMS is completed, sort the population by fitness from smallest to largest to obtain a new population Q r (q), the first half is defined as The second half is defined as First keep Q g (q) Middle front The individual is regarded as an excellent solution, and a new individual is generated for every 4 individuals. The population size generates the optimal subpopulation G g (q), which is calculated as:

[0061]

[0062] in N is the total population size.

[0063] T2: Set the probability p to ensure that the best subpopulation G(q) is randomly selected from the population with a probability of (1-p), which is calculated as follows:

[0064]

[0065] in,

[0066] T3: At the beginning, the optimal subpopulation G(q) is randomly selected; as the number of iterations increases, G(q) tends to select the excellent individuals after sorting; when in the gaseous state, the individuals are made more random; after gradually transforming to the solid state, the population tends to search around the excellent individuals; the optimal subpopulation G(q) is mutated to obtain the subpopulation H(q); the new subpopulation V(q) is composed of the populations H(q) and Q b (q) is generated, and its calculation formula is:

[0067] H i (q) = G r1(q)+F·(G r2 (q)-G r3 (q))

[0068]

[0069] in, r1, r2, r3 are A random integer in the range G r1 (q), G r2 (q) and G r3 (q) represents the selected subpopulation, F is the variation factor, which controls the degree of difference amplification, and CR is the crossover probability;

[0070] T4: Use the greedy selection strategy to select the Q b (q) A better individual, the formula is as follows:

[0071]

[0072] Among them, f represents the fitness function, mutation operation and subpopulation update strategy are introduced, and the generation of better solutions is ensured by greedy selection strategy.

[0073] T5: As the iteration process proceeds, the population gradually transitions from gas to solid, and the search strategy gradually develops from randomness to convergence; by continuously minimizing the value of the objective function during the evolution process, the algorithm gradually converges to the optimal point in the solution space and obtains the final weights of different modes.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] Enhance the prediction accuracy of the prediction model: The present invention effectively decomposes the power generation signal into multiple modes corresponding to different frequency components through evolutionary variational mode decomposition (EVMD). This enables the prediction model to more accurately capture the long-term trends, periodic changes, and short-term fluctuations in the signal. Compared with traditional models, this method improves the model's adaptability to complex dynamic systems and the accuracy of predictions, especially when faced with the high volatility and uncertainty of renewable energy power generation.

[0076] Modality-specific processing strategies: The modality-specific processing strategies adopted by the present invention (such as least squares method to smooth low-frequency modes, differential filter to enhance medium-frequency modes, and dendritic learning model to optimize high-frequency modes) allow the most appropriate analysis and prediction methods to be used for data with different characteristics. The introduction of this method not only improves the processing efficiency of data in each frequency band, but also optimizes the effect of the overall prediction structure, thereby achieving more accurate data fitting and prediction over the entire spectrum.

[0077] Optimized parameter adjustment mechanism: The multi-population optimized material state search algorithm (MPSMS) is used to adaptively adjust the modal weights, providing the optimal weight configuration for each modality. This adaptive adjustment mechanism can dynamically respond to external changes and changes in internal modal characteristics during the prediction process, significantly improving the stability and prediction performance of the model in a changing environment.

[0078] Reduce computing resource consumption: Through precise modal decomposition and targeted processing methods, the present invention can effectively reduce unnecessary computing resource consumption. Compared with traditional prediction models that require a large amount of data input and complex calculations, the present invention achieves a more efficient data processing and parameter optimization process by reducing data redundancy and optimizing algorithm steps. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 A flow chart of the energy generation prediction method based on evolutionary variational mode decomposition provided by the present invention;

[0080] Figure 2 A visualization forecast result diagram of enterprise thermal power generation according to an embodiment of the present invention;

[0081] Figure 3 A visualization forecast result diagram of enterprise hydroelectric power generation according to an embodiment of the present invention;

[0082] Figure 4 This is a visualization forecast result diagram of enterprise wind power generation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0083] The present invention is further described below in conjunction with the examples. However, the present invention can be implemented in many different ways and should not be construed as being limited to the examples shown; on the contrary, these examples provide those skilled in the art with implementation methods that meet applicable legal requirements.

[0084] Example 1: Figure 1 As shown, the energy generation prediction method based on evolutionary variational mode decomposition proposed in the present invention specifically includes the following steps:

[0085] S1: Perform variational modal decomposition on the original power generation signal and decompose it into S intrinsic mode functions, that is, solve the following variational problem, and its calculation formula is:

[0086]

[0087] Where k represents the index of the mode, S represents the number of modes, and u k represents the kth mode function, t is the time variable, and represents each time point in the time series of the signal; represents the instantaneous rate of change of each modal signal; w k represents the center frequency of the kth mode, δ(t) represents the impulse response of the Hilbert transform, is the standard expression of the Hilport transform, which represents the impulse response of the filter. is at frequency ω k The complex exponential function at represents the adjustment of the modal function in the frequency domain, f represents the original signal, that is, the sum of all modes is equal to the target signal, and the purpose of the algorithm is to minimize all modal functions u k A collection of .

[0088] The quadratic penalty factor α and the Lagrange multiplier λ are introduced to transform the constrained variational problem into an unconstrained variational problem. The calculation formula is:

[0089]

[0090] Where L is the Lagrangian objective function, the unconstrained variational problem is solved by the alternating direction multiplier method, the saddle point is found by alternating update calculation, and u is optimized by alternating k , while satisfying the constraints, minimizing the original objective function:

[0091]

[0092] Where f is the original signal, j is the imaginary unit, ω is the angular frequency, jω is used to map the signal from the time domain to the frequency domain, indicating the amplitude and phase of the signal at different frequencies, sgn(ω+ω k ) is to adjust the phase of the mode function in the frequency domain, is the representation of the kth mode function in the frequency domain, The frequency domain representation of the kth mode is frequency-shifted so that its center frequency is aligned with ω k , is the frequency representation of the original signal, and λ(ω) is the frequency domain representation of the Lagrange multiplier.

[0093] S2: For different modes obtained by decomposition, perform the following processing steps respectively:

[0094] S2-1: For low frequency mode Mo 1 , the least square method is used to smooth the trend and fit the long-term trend. The calculation formula is:

[0095] Mo 1 ′=p 0 +p 1 l+p 2 l 2 +p 3 l 3

[0096] S2-2: For the intermediate frequency mode Mo 2 , a differential filter is used to remove trend components and enhance the ability to capture periodic fluctuations. The calculation formula is:

[0097]

[0098] S2-3: For high frequency mode Mo 3 , a dendritic learning model is used for stable prediction to accurately capture nonlinear relationships. The calculation formula is:

[0099] Mo 3 ′=DLM(Mo 3 )

[0100] S3: The prediction results of each mode are integrated by weighted summation to obtain the final energy generation prediction result NY'. The calculation formula is:

[0101] NY'=αMo 1 '+βMo 2 '+γMo 3 '

[0102] Among them, Mo 1 ′、Mo 2 'and Mo 3 ' are the prediction results of low-frequency, medium-frequency and high-frequency modes after processing, l is the input data, is the cth data point of the original data sequence, g is the total length of the sequence, and p 0 、p 1 、p 2 and p 4 are polynomial coefficients, determined by the least squares method, α, β and γ are weight parameters of each mode;

[0103] S4: Optimize the weight parameters α, β and γ using a multi-population improved material state search algorithm, specifically including the following steps:

[0104] S4-1: Initialize the population and set the maximum number of iterations iter, where the population consists of multiple candidate weight parameter combinations, and a candidate weight parameter combination represents a set of weight parameters (α i ,β i ,γ i ), the candidate weight parameter combinations in the population represent individuals;

[0105] S4-2: Each individual’s position update is subject to the direction vector d i The control of this vector determines the individual Q i Moving direction and step length in the solution space, direction vector d i The calculation formula is:

[0106]

[0107] in, represents the direction vector of individual i after the update in the q+1 generation, q is the current generation, iter is the maximum number of iterations, niter is the current number of iterations, Q best is the position of the current global best individual, Q i is the individual's current position;

[0108] S4-3: Calculate velocity vector v i , and its calculation formula is:

[0109]

[0110] in, and are the upper and lower bounds of the j-th dimension, η∈[0,1], and n is the population size.

[0111] S4-4: By updating the direction vector and velocity vector, the new position Q ij The calculation formula is:

[0112]

[0113] Among them, v i,j is the velocity of the ith individual in the jth dimension, λ∈[0,1]; collision behavior occurs during the molecular evolution process. When two molecules d a and d b When the distance between them is less than the collision distance, they exchange direction vectors, i.e., d a =d b , d b =d a ;

[0114] S4-5: Calculate the collision distance r, the calculation formula is:

[0115]

[0116] Among them, η∈[0,1].

[0117] S4-6: There is a certain probability of random change in the process of molecular evolution. In order to simulate this mechanism, the calculation formula of molecular position is:

[0118]

[0119] Among them, rand(0,1) represents a random number between [0,1], and H is a hyperparameter set for each stage.

[0120] As a technical preferred solution of the present invention, the variational problem in step S1 is constructed by: first, using Hilbert transform to calculate the analytical signal of the original signal, and analyzing its spectrum through frequency domain transform; then estimating the center frequency of each modal function and limiting the bandwidth of the modal spectrum; finally, optimizing the spectrum distribution through frequency domain bandwidth calculation.

[0121] As a technical preferred solution of the present invention, the DLM described in step S2 is a dendritic learning model, which is composed of a synaptic layer, a dendritic layer, a membrane layer, a somatic layer and an axon layer; wherein the synaptic layer uses a Sigmoid function to process input signals, the dendritic layer uses an addition operation to sum the output of the synaptic layer, the membrane layer uses a multiplication operation to perform a nonlinear combination of the results of each dendritic layer, the somatic layer is normalized by a standard score, and the axon layer multiplies the discrete weight of each input corresponding to the synaptic layer by the processed somatic output to obtain the final result of the model. The calculation formula of each component of the DLM as a dendritic learning model is:

[0122] Synaptic layer:

[0123]

[0124] Dendritic layer:

[0125]

[0126] Film layer:

[0127]

[0128] Somatic layer:

[0129]

[0130] Axon layer:

[0131]

[0132] Among them, x i is the input of power generation, k s is a positive constant, w ij is the weight of the synaptic layer, θ ij is the threshold, Y is the output of the synapse, N is the number of synapses, and Z j is the output of the dendritic layer, M is the number of dendritic layers, V is the output of the membrane layer, num represents the number of neurons, and std represent the mean and standard deviation of the calculated data, σ represents the Sigmoid function, var represents the variance, and k ν It is a hyperparameter of the somatic layer. By continuously adjusting the weights and thresholds of the model through back propagation, DLM can learn various complex functional relationships.

[0133] As a technical preferred solution of the present invention, the multi-population improved material state search algorithm in step S4 is further improved into a multi-population improved material state search algorithm, which divides the population into two sub-populations according to the fitness of the individuals: the first half is the better population Q g (q), the latter part is the inferior population Q b (q), update the individuals in the inferior sub-population by the individuals in the superior sub-population. A multi-population improved material state search algorithm includes the following steps:

[0134] T1: When the collision step in SMS is completed, sort the population by fitness from smallest to largest to obtain a new population Q r (q), the first half is defined as The second half is defined as First keep Q g (q) Middle front The individual is regarded as an excellent solution, and a new individual is generated for every 4 individuals. The population size generates the optimal subpopulation G g (q), which is calculated as:

[0135]

[0136] in N is the total population size.

[0137] T2: Set the probability p to ensure that the best subpopulation G(q) is randomly selected from the population with a probability of (1-p), which is calculated as follows:

[0138]

[0139] in,

[0140] T3: At the beginning, the optimal subpopulation G(q) is randomly selected; as the number of iterations increases, G(q) tends to select the excellent individuals after sorting; when in the gaseous state, the individuals are made more random; after gradually transforming to the solid state, the population tends to search around the excellent individuals; the optimal subpopulation G(q) is mutated to obtain the subpopulation H(q); the new subpopulation V(q) is composed of the populations H(q) and Q b (q) is generated, and its calculation formula is:

[0141] H i (q) = G r1 (q)+F·(G r2 (q)-G r3 (q))

[0142]

[0143] in, r1, r2, r3 are A random integer in the range G r1 (q), G r2 (q) and G r3 (q) represents the selected subpopulation, F is the mutation factor that controls the degree of difference amplification, and CR is the crossover probability.

[0144] T4: Use the greedy selection strategy to select the Q b (q) A better individual, the formula is as follows:

[0145]

[0146] Among them, f represents the fitness function, mutation operation and subpopulation update strategy are introduced, and the generation of better solutions is ensured by greedy selection strategy.

[0147] T5: As the iteration process proceeds, the population gradually transitions from gas to solid, and the search strategy gradually develops from randomness to convergence; by continuously minimizing the value of the objective function during the evolution process, the algorithm gradually converges to the optimal point in the solution space and obtains the final weights of different modes.

[0148] Example 2: In the field of energy, accurate prediction of energy generation is of great significance for optimizing the allocation of power resources, improving the operating efficiency of the power system, and supporting the efficient integration of renewable energy. With the gradual development of smart grids, real-time dynamic analysis of energy production and consumption is particularly important. In particular, for energy forms such as thermal power generation, hydropower generation, and wind power generation, they are subject to greater volatility due to the influence of factors such as weather and equipment operating status. Therefore, accurate prediction of energy generation can not only improve the stability of the power grid, but also help relevant departments make reasonable planning and scheduling. To this end, this embodiment is based on the evolutionary variational mode decomposition (EVMD) algorithm proposed in this patent, and accurately predicts future energy generation through in-depth analysis of historical power generation data.

[0149] This example selects historical monthly power generation data of different types of energy (thermal, hydro, and wind) in a certain region. All data come from enterprises above the scale of the National Bureau of Statistics, and linear interpolation is used to fill missing values ​​caused by environmental or technical reasons. The evaluation data set we use includes 15 years of monthly data from 2000 to 2020, of which 70% of the data is used for training and 30% of the data is used for testing. The following are the detailed steps for processing the data:

[0150] First, the time series data of historical energy generation was collected and preprocessed, including:

[0151] Data cleaning: Remove obvious outliers and noise data.

[0152] Missing value processing: Missing data points are filled by linear interpolation to ensure data continuity and integrity.

[0153] Normalization: In order to accelerate model convergence and reduce computational complexity, the normalization method is used to process data. The specific normalization formula is as follows:

[0154]

[0155] Among them, x(t i ) is the power generation data at a certain moment, x min is the minimum value among all data, x max is the maximum value among all the data.

[0156] In order to deeply analyze the inherent laws of time series, variational mode decomposition (VMD) is used to decompose the energy generation time series into three sub-modes with different frequencies (low-frequency mode, medium-frequency mode and high-frequency mode). Through this decomposition process, the main periodic components, long-term trend changes and short-term fluctuation characteristics in the signal can be effectively identified and isolated, thereby providing valuable information for the establishment of subsequent prediction models.

[0157] For each decomposed sub-mode, different prediction methods are used for processing:

[0158] For low-frequency modes, the least square method is used to smooth the trend and fit the long-term change trend.

[0159] For the intermediate frequency mode, a differential filter is used to remove the trend component and enhance the ability to capture periodic fluctuations.

[0160] For high-frequency modes, a dendritic learning model is used for prediction to capture short-term fluctuation characteristics.

[0161] In order to further improve the prediction accuracy, this embodiment uses the multi-population improved material state search algorithm (MPSMS) to optimize the weight of each sub-modality. In this way, we can ensure that the contribution of each modality to the final prediction result is fully considered, while avoiding over-reliance on a certain modality, thereby improving the prediction stability and reliability of the entire model.

[0162] Material State Search Algorithm (MPSMS): This algorithm simulates the transformation process of matter from gas to solid and dynamically adjusts the weights of each sub-mode, so that the system has greater randomness in the initial exploration stage, and gradually tends to converge to an excellent solution as the iteration proceeds. The core advantage of the MPSMS algorithm is that it can simultaneously consider the balance between global search and local optimization, so that it can achieve efficient solutions to prediction problems under the influence of multiple uncertain factors.

[0163] In order to scientifically and accurately evaluate the prediction performance of each model, this embodiment uses three key evaluation indicators: mean square error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). All evaluation indicators are calculated using de-standardized prediction data and real data.

[0164] Metric 1: Mean Square Error (MSE)

[0165]

[0166] Metric 2: Mean Absolute Error (MAE)

[0167]

[0168] Metric 3: Mean Absolute Percentage Error (MAPE)

[0169]

[0170] Where m is the total number of sequences, T n is the true value, O n is the predicted value.

[0171] In order to verify the effectiveness of the patented method, this embodiment conducted a comparative experiment on the EVMD model with a variety of commonly used prediction models (such as LSTM, MLP, ANFIS, Elman neural network, SVM), and the results are shown in Table 1.

[0172] Table 1 Comparison of evaluation indicators of prediction models

[0173]

[0174] As can be seen from Table 1, under the comparative experimental results of different models, the error index of the model proposed in this patent is significantly lower than that of the comparative model, and has better prediction performance. In order to more intuitively compare the prediction effects of each model, Figure 2-4 The visualization prediction results of 6 models on three data sets are shown. Original represents the original power generation, the horizontal axis represents time, and the vertical axis represents power generation. The vertical purple dotted line in the figure represents the critical point between the training and prediction stages. As the years increase, the power generation shows a continuous upward trend. Figure 2 and Figure 4 It can be seen that SVM cannot accurately capture the long-term trend of power generation in the prediction of thermal and wind power generation, resulting in prediction failure. Figure 3 It can be seen that the LSTM and MLP models are insufficient in predicting cyclical fluctuations. Although Elman and ANFIS can learn some features to predict the continuous upward trend, they still lack accuracy. In contrast, only EVMD can accurately predict the continuous upward trend and cyclical fluctuations of power generation while maintaining extremely high prediction accuracy, with the lowest error and the prediction result closest to the true value. Experiments have shown that the EVMD model proposed in this patent can more accurately predict energy generation, thereby optimizing resource allocation and improving overall operation utilization, and effectively solving key problems in energy management and power generation forecasting.

[0175] Through this embodiment, it can be concluded that the energy generation prediction method based on evolutionary variational mode decomposition (EVMD) proposed in this patent can effectively improve the prediction accuracy and stability of energy generation. Compared with other traditional prediction models, the EVMD model can better capture the trends, cyclical fluctuations and short-term fluctuations in the data when processing complex power generation time series data, thereby providing more accurate prediction results, helping power system managers to better perform resource scheduling and planning, and optimize the operation efficiency of the power system.

[0176] The above embodiments only express the implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention. It should be pointed out that, for those of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.

Claims

1. An energy generation prediction method based on evolutionary variational mode decomposition, characterized by: The following steps are involved: S1: Perform variational mode decomposition (VMD) on the original power generation signal and decompose it into S intrinsic mode functions, that is, solve the following variational problem, and its calculation formula is: Where k is the index of the mode; S is the number of modes; u k represents the kth mode function; t is the time variable, representing each time point in the time series of the signal; represents the instantaneous rate of change of each modal signal; w k represents the center frequency of the kth mode; δ(t) represents the impulse response of the Hilbert transform; is the standard expression for the Hilport transform, which represents the impulse response of the filter; is at frequency ω k The complex exponential function at represents the adjustment of the modal function in the frequency domain; f represents the original signal, that is, the sum of all modes is equal to the target signal; the purpose of the algorithm is to minimize all modal functions u k A collection of; The quadratic penalty factor α and the Lagrange multiplier λ are introduced to transform the constrained variational problem into an unconstrained variational problem. The calculation formula is: Where L is the Lagrangian objective function, the unconstrained variational problem is solved by the alternating direction multiplier method, the saddle point is found by alternating update calculation, and u is optimized by alternating k , while satisfying the constraints, minimizing the original objective function: Where f is the original signal, j is the imaginary unit, ω is the angular frequency, jω is used to map the signal from the time domain to the frequency domain, indicating the amplitude and phase of the signal at different frequencies, sgn(ω+ω k ) is to adjust the phase of the mode function in the frequency domain, is the representation of the kth mode function in the frequency domain, The frequency domain representation of the kth mode is frequency-shifted so that its center frequency is aligned with ω k , is the frequency representation of the original signal, and λ(ω) is the frequency domain representation of the Lagrange multiplier; S2: For different modes obtained by decomposition, perform the following processing steps respectively: S2-1: For the low-frequency mode Mo1, the least square method is used to smooth the trend and fit the long-term change trend. The calculation formula is: Mo1′=p0+p1l+p2l 2 +p3l 3 S2-2: For the intermediate frequency mode Mo2, a differential filter is used to remove the trend component and enhance the ability to capture periodic fluctuations. The calculation formula is: S2-3: For the high-frequency mode Mo3, the dendritic learning model (DLM) is used for stable prediction to accurately capture the nonlinear relationship. The calculation formula is: Mo3′=DLM(Mo3) S3: The prediction results of each mode are integrated by weighted summation to obtain the final energy generation prediction result NY'. The calculation formula is: NY'=αMo1'+βMo2'+γMo3' Among them, Mo1', Mo2' and Mo3' are the prediction results of low-frequency, medium-frequency and high-frequency modes after processing respectively; l is the input data; is the cth data point of the original data sequence; g is the total length of the sequence; p0, p1, p2 and p3 are polynomial coefficients determined by the least squares method; α, β and γ are weight parameters of each mode; S4: Optimize the weight parameters α, β and γ using a multi-population improved material state search algorithm, specifically including the following steps: S4-1: Initialize the population and set the maximum number of iterations iter, where the population consists of multiple candidate weight parameter combinations, and a candidate weight parameter combination represents a set of weight parameters (α i ,β i ,γ i ), the candidate weight parameter combinations in the population represent individuals; S4-2: Each individual’s position update is subject to the direction vector d i The control of this vector determines the individual Q i Moving direction and step length in the solution space, direction vector d i The calculation formula is: in, represents the direction vector of individual i after the update in the q+1 generation, q is the current generation, iter is the maximum number of iterations, niter is the current number of iterations, Q best is the position of the current global best individual, Q i is the individual's current position; S4-3: Calculate velocity vector v i , and its calculation formula is: in, and are the upper and lower bounds of the jth dimension, η∈[0,1], and n is the population size; S4-4: By updating the direction vector and velocity vector, the new position Q ij The calculation formula is: Among them, v i,j is the velocity of the ith individual in the jth dimension, λ∈[0,1]; collision behavior occurs during the molecular evolution process. When two molecules d a and d b When the distance between them is less than the collision distance, they exchange direction vectors, i.e., d a =d b , d b =d a ; S4-5: Calculate the collision distance r, the calculation formula is: Among them, η∈[0,1]; S4-6: There is a certain probability of random change in the process of molecular evolution. In order to simulate this mechanism, the calculation formula of molecular position is: Among them, rand(0,1) represents a random number between [0,1], and H is a hyperparameter set for each stage.

2. The energy generation prediction method based on evolutionary variational mode decomposition according to claim 1 is characterized by: The variational problem in step S1 is constructed by: firstly using Hilbert transform to calculate the analytical signal of the original signal, and analyzing its spectrum through frequency domain transform; then estimating the center frequency of each modal function and limiting the bandwidth of the modal spectrum; and finally optimizing the spectrum distribution through frequency domain bandwidth calculation.

3. The energy generation prediction method based on evolutionary variational mode decomposition according to claim 1 is characterized by: In step S2, the DLM is a dendritic learning model, which consists of a synaptic layer, a dendritic layer, a membrane layer, a somatic layer and an axon layer; wherein the synaptic layer uses a Sigmoid function to process input signals, the dendritic layer uses an addition operation to sum the outputs of the synaptic layer, the membrane layer uses a multiplication operation to perform a nonlinear combination of the results of each dendritic layer, the somatic layer is normalized by a standard score, and the axon layer multiplies the discrete weight of each input corresponding to the synaptic layer by the processed somatic output to obtain the final result of the model.

4. The energy generation prediction method based on evolutionary variational mode decomposition according to claim 3 is characterized by: The calculation formulas of the components of the DLM dendrite learning model are: Synaptic layer: Dendritic layer: Film layer: Somatic layer: Axon layer: Among them, x i is the input of power generation, k s is a positive constant, w ij is the weight of the synaptic layer, θ ij is the threshold, Y is the output of the synapse, N is the number of synapses, and Z j is the output of the dendritic layer, M is the number of dendritic layers, V is the output of the membrane layer, num represents the number of neurons, and std represent the mean and standard deviation of the calculated data, σ represents the Sigmoid function, var represents the variance, and k ν is a hyperparameter of the somatic layer; by continuously adjusting the weights and thresholds of the model through back propagation, DLM can learn various complex functional relationships.

5. The energy generation prediction method based on evolutionary variational mode decomposition according to claim 1 is characterized by: The multi-population improved material state search algorithm in step S4 is further improved into a multi-population improved material state search algorithm, which divides the population into two sub-populations according to the fitness of the individuals: the first half is the better population Q g (q), the latter part is the inferior population Q b (q) The individuals in the inferior sub-population are updated by the individuals in the superior sub-population.

6. The energy generation prediction method based on evolutionary variational mode decomposition according to claim 5 is characterized by: The multi-population improved material state search algorithm comprises the following steps: T1: When the collision step in SMS is completed, sort the population by fitness from smallest to largest to obtain a new population Q r (q), the first half is defined as The second half is defined as First keep Q g (q) Middle front The individual is regarded as an excellent solution, and a new individual is generated for every 4 individuals. The population size generates the optimal subpopulation G g (q), which is calculated as: in N is the total population size; T2: Set the probability p to ensure that the best subpopulation G(q) is randomly selected from the population with a probability of (1-p), which is calculated as follows: in, r∈[1,N]; T3: At the beginning, the optimal subpopulation G(q) is randomly selected; as the number of iterations increases, G(q) tends to select the excellent individuals after sorting; when in the gaseous state, the individuals are made more random; after gradually transforming to the solid state, the population tends to search around the excellent individuals; the optimal subpopulation G(q) is mutated to obtain the subpopulation H(q); the new subpopulation V(q) is composed of the populations H(q) and Q b (q) is generated, and its calculation formula is: H i (q)=G r1 (q)+F·(G r2 (q)-G r3 (q)) in, j=1,2,...,D,r∈[1,N];r1,r2,r3 are A random integer in the range G r1 (q), G r2 (q) and G r3 (q) represents the selected subpopulation, F is the variation factor, which controls the degree of difference amplification, and CR is the crossover probability; T4: Use the greedy selection strategy to select the Q b (q) A better individual, the formula is as follows: Among them, f represents the fitness function, which introduces mutation operation and subpopulation update strategy, and ensures the generation of better solutions through greedy selection strategy; T5: As the iteration process proceeds, the population gradually transitions from gas to solid, and the search strategy gradually develops from randomness to convergence; by continuously minimizing the value of the objective function during the evolution process, the algorithm gradually converges to the optimal point in the solution space and obtains the final weights of different modes.

Images