Multi-scene deep learning source-load prediction method based on VMD and RIME
Through VMD decomposition and RIME optimization algorithm, a multi-scene deep learning source-load prediction method based on VMD and RIME is constructed, which solves the problems of insufficient timing feature capture and local optimization in the prior art, and improves prediction accuracy and applicability.
Patent Information
- Application Number
- CN202411289427.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-14
- Publication Date
- 2025-03-04
AI Technical Summary
In the prior art, prediction models fail to effectively capture timing features and are prone to fall into a local optimal state.
Using a multi-scene deep learning source-load prediction method based on VMD and RIME, the original time data is decomposed through VMD to obtain K eigenmodal components, and the LSTM neural network structure is optimized using RIME optimization algorithm.
Effectively capture timing features, avoiding the solution process from falling into local optimization, and improving the accuracy of the prediction model and multi-scene applicability.
Smart Images
Figure CN119561004A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power system source-load prediction, and in particular to a multi-scenario deep learning source-load prediction method based on VMD and RIME. Background Art
[0002] In recent years, with the continuous intensification of the energy crisis, countries have been vigorously developing renewable energy power generation technology. However, the increasing proportion of renewable energy power generation and the large-scale access of various flexible loads have further increased the uncertainty in modern power systems. Traditional prediction methods have large errors for highly complex and non-stationary data. Some scholars have proposed that deep learning algorithms can be introduced to fully explore data characteristics and time series rules, thereby improving the accuracy of prediction models and their applicability in multiple scenarios. Deep learning prediction models can effectively extract the intrinsic connections of time series data through the analysis of massive data, and achieve higher prediction accuracy through iterative training of model hyperparameters. The theory and application of deep learning have become one of the important development directions of contemporary system and control science. In the research process of deep learning prediction methods, the construction of prediction models, the setting of parameters, and the training of hyperparameters are several key issues.
[0003] However, the current prediction models do not effectively capture the time series characteristics and are prone to fall into local optimal states during the solution process.
[0004] In view of this, a multi-scenario deep learning source-load prediction method based on VMD and RIME is needed. Summary of the invention
[0005] In view of the problem that the prediction model in the prior art does not effectively capture the time series characteristics and is prone to fall into a local optimal state during the solution process, the present invention provides a multi-scenario deep learning source-load prediction method based on VMD and RIME, which can use the VMD method to decompose the original time data and use the RIME optimization algorithm to optimize the LSTM neural network structure. The specific technical solution is as follows:
[0006] A multi-scenario deep learning source-load prediction method based on VMD and RIME includes the following steps:
[0007] Collect power supply side data and load side data based on the set sampling time interval, generate a data set, and use it as the original signal;
[0008] Set the number of modal functions K and the penalty factor α, and use VMD to decompose the original signal into a combination of K intrinsic mode components;
[0009] Determine the network structure parameters, including the number of input layer units, the number of output layer units, the number of hidden layer units, the learning rate and the maximum number of iterations, and use the LSTM neural network to establish a prediction model;
[0010] The data set is divided into a training set and a validation set, and the IMF components are predicted to obtain the prediction results.
[0011] Preferably, the process of setting the number K of modal functions is as follows:
[0012] Establishing a K value library, wherein the K value library includes all available K values;
[0013] Starting from K=1, gradually increase the number of modes and calculate the center frequency of each mode;
[0014] When the change degree of the center frequency falls into the set stable state threshold range or repetition occurs after the number of modes is increased, the current K value is used as the set number K of mode functions.
[0015] Preferably, the penalty factor α is determined as follows:
[0016] Select an initial penalty factor value and determine an iteration termination condition, wherein the number of iterations and the iteration termination condition are: when the change of the penalty factor is less than a set threshold, the iteration is stopped;
[0017] Perform VMD decomposition step: Perform VMD decomposition using the current penalty factor value to obtain a set of mode functions;
[0018] Perform the decomposition result evaluation step: check whether the modal function contains the center frequency, and increase or decrease the penalty factor based on the evaluation result and use it as the new penalty factor;
[0019] The VMD decomposition and evaluation steps are repeated with new penalty factor values until the termination condition is met or the iteration limit is reached.
[0020] The structure of the LSTM neural network described above includes two hidden layers, two Dropout layers and one fully connected layer, where the number of hidden layer units is given by the RIME optimization algorithm; the absolute mean error is selected as the loss value function; and the Dropout layer parameter is set to 0.4.
[0021] The above also includes the following steps:
[0022] After the data set is divided into a training set and a validation set, each IMF component is normalized after the data set is divided and used as the input of the LSTM neural network.
[0023] The process of predicting the IMF components described above and obtaining the prediction results is as follows:
[0024] Input the training set data into the LSTM neural network to train the hyperparameters of the prediction model;
[0025] The validation set data is input into the LSTM neural network to obtain the forecast data of each IMF component, and the final forecast result is obtained by superposition.
[0026] The above-mentioned power supply side data includes the active power of photovoltaic and wind power.
[0027] The above-mentioned load side data include four types of loads: residential, agricultural, industrial and commercial.
[0028] A computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the multi-scenario deep learning source-load prediction method based on VMD and RIME as described above.
[0029] A processor is used to run a program, wherein when the program is running, the multi-scenario deep learning source-load prediction method based on VMD and RIME as described above is executed.
[0030] Compared with the prior art, the present invention has the following beneficial effects:
[0031] The present invention collects power supply side data and load side data based on a set sampling time interval, generates a data set, and uses it as the original signal; then sets the number of modal functions K and the penalty factor α, and uses VMD to decompose the original signal into a combination of K intrinsic modal components; secondly, determines the network structure parameters, including the number of input layer units, the number of output layer units, the number of hidden layer units, the learning rate and the maximum number of iterations, and uses the LSTM neural network to establish a prediction model; finally, divides the data set into a training set and a validation set, predicts the IMF component, and obtains the prediction result. The present invention uses the VMD method to decompose the original time data, and uses the RIME optimization algorithm to optimize the LSTM neural network structure, so that the prediction model can more effectively capture the time series characteristics while effectively avoiding the solution process from falling into the local optimum. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.
[0033] Figure 1 This is the overall process of the combined prediction method proposed in the present invention.
[0034] Figure 2 This is a flow chart of the LSTM network structure parameter optimization proposed in the present invention.
[0035] Figure 3 Schematic diagram of the decomposition results using the VMD method.
[0036] Figure 4 The power supply side prediction results under different scenarios are obtained using the combined prediction model.
[0037] Figure 5 The load side prediction results under different scenarios are obtained using the combined prediction model.
[0038] Figure 6 This is a comparison chart of the prediction effects of the combined prediction method proposed in the present invention and other prediction methods. DETAILED DESCRIPTION
[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0040] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.
[0041] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.
[0042] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0043] Combine the following Figure 1-Figure 6 , further explain the implementation methods of the present invention.
[0044] In one embodiment of the present invention, a multi-scenario deep learning source-load prediction method based on VMD and RIME is provided. The following first briefly introduces variational mode decomposition and frost optimization algorithm:
[0045] Variational mode decomposition (VMD) is an adaptive signal decomposition method that decomposes time series signals into a series of intrinsic mode functions (IMFs) with specific center frequencies and frequency bands, iteratively updates IMFs using the alternating direction multiplier method, and adaptively searches for the optimal center frequency and frequency band of IMFs. The core of the VMD method is to construct a variational model and solve variational problems.
[0046] (1) Constructing a variational problem
[0047] The sum of IMFs is used as the input signal, and the minimum sum of IMF frequency bands is used as the objective function. Let the original time series signal be f(t), and use Hilbert transform to solve K IMFs. The calculation formula is as follows:
[0048]
[0049] Where: u k (t) is IMF, k=1,2,…,K; ω k is the IMF center frequency; δ(t) is the unit pulse function; is the derivative of the function with respect to time.
[0050] (2) Solving variational problems
[0051] To solve the variational model, the augmented Lagrangian function can be used to transform the above model into an equivalent unconstrained optimization problem. The transformation expression is as follows:
[0052]
[0053] Where α is the quadratic penalty factor, which is used to limit the frequency band; λ is the Lagrange multiplier. The following is the convergence condition:
[0054]
[0055] Where ε is the precision control value, which is used to limit the relative error.
[0056] The above equation is solved using the Alternating Direction Method of Multipliers (ADMM), and each modal component and its center frequency are continuously updated during the solution process.
[0057] The rime-ice (RIME) optimization algorithm is an efficient optimization algorithm proposed by Su et al. in February 2023. Inspired by the growth mechanism of rime-ice in nature, the algorithm simulates the formation process of soft frost and hard frost, proposes a soft frost search strategy and a hard frost piercing mechanism, continuously updates the population to achieve optimization, and thus finds the optimal solution.
[0058] The optimization process of the RIME algorithm is as follows:
[0059] 1) Initialize the population: randomly generate a group of frost particles, each of which represents a potential solution.
[0060] 2) A soft frost search strategy is proposed. According to the five motion characteristics of frost particles, the condensation process of each particle is simulated and the position of the frost particles is calculated:
[0061]
[0062] In the formula, represents the new position of the particle after the update, i and j represent the jth particle of the i-th frost factor; R best,j is the position of the jth particle of the optimal factor in the frost population R; parameter r1 is a random number in the range of (-1,1) used to control the direction of particle movement; at the same time, cosθ changes with the increase of the number of iterations; β is an environmental factor used to ensure the convergence of the algorithm and changes with the number of iterations; h is the degree of adhesion, a random number in the range of (0,1) used to control the center distance between two frost particles; Ub ij and Lb ij are the upper and lower bounds of the search space, respectively, limiting the effective area of particle movement; E is the adhesion coefficient, which affects the condensation probability of frost particles and increases with the number of iterations; parameter r2 is a random number in the range of (0,1).
[0063] 3) Inspired by the hard frost penetration phenomenon, a hard frost penetration mechanism is proposed for the exchange between frost particles, which improves the convergence of the algorithm and the ability to escape from the local optimum. The position replacement formula between frost particles is shown in formula (6).
[0064]
[0065] Where: is the updated particle position; R best,j is the position of the jth particle of the optimal factor in the frost population R; F normr (S i ) is the normalized value of the current frost factor fitness; r3 is a random number in the range of (-1,1).
[0066] Based on the above description, in this embodiment, a multi-scenario deep learning source-load prediction method based on VMD and RIME includes the following steps:
[0067] S1: Collect power supply side data and load side data based on the set sampling time interval, generate a data set, and use it as the original signal;
[0068] S2: Set the number of modal functions K and the penalty factor α, and use VMD to decompose the original signal into a combination of K intrinsic mode components;
[0069] This step involves decomposing the raw data using the VMD method.
[0070] Variational mode decomposition (VMD) can decompose the original signal into a combination of K intrinsic mode components. Each IMF has a certain center frequency and bandwidth, but there will be no mode aliasing between them.
[0071] Before using VMD to process the original signal, the number of mode functions K and the penalty factor α must be set.
[0072] Among them, the value range of K directly affects the decomposition effect of VMD. If K is too large, the signal of the same mode will be decomposed into multiple center frequencies; if K is too small, the noise cannot be effectively removed from the original signal. Therefore, most scholars will perform multiple decompositions when selecting parameter K, and determine the value of parameter K by observing the center frequencies of all IMFs after decomposition.
[0073] Some scholars have proposed that the center frequency method can be used to determine the optimal number of IMFs. Once the center frequencies of the components of the final intrinsic mode function reach a stable state, the K value is considered to be the optimal value. Therefore, for different time series data, we use different K values for decomposition and determine the most appropriate value through frequency analysis of the decomposition results.
[0074] The penalty factor α will also affect the decomposition effect of the signal, because different penalty factors represent different bandwidths. The smaller the penalty factor, the larger the bandwidth of each component. Excessive bandwidth will cause some components to contain other component signals; the larger the penalty factor, the smaller the bandwidth of each component, which may cause excessive decomposition of the signal. When the K value is determined, the more reasonable the penalty factor α is set, the better the separation effect of noise and valid signals. The common value range of this coefficient is [1000,3000].
[0075] S3: Determine the network structure parameters, including the number of input layer units, the number of output layer units, the number of hidden layer units, the learning rate and the maximum number of iterations, and use the LSTM neural network to establish a prediction model;
[0076] This step involves optimizing the LSTM network parameters using the RIME algorithm.
[0077] When using LSTM neural network to establish a prediction model, the network structure parameters should be determined first, such as the number of input layer units, the number of output layer units, the number of hidden layer units, the learning rate, and the maximum number of iterations. The number of hidden layer units determines the fitting ability of the network, the learning rate determines the update speed of each parameter in the network, and the maximum number of iterations should be selected within a reasonable range. Therefore, these three parameters should be reasonably selected, and relying on subjective experience to adjust the parameters may lead to unreasonable parameters, which directly affects the prediction accuracy of the model. Therefore, using optimization algorithms to optimize the structural parameters of neural networks is an efficient method.
[0078] The structure of the LSTM neural network includes two hidden layers, two Dropout layers and one fully connected layer. The number of hidden layer units is given by the RIME optimization algorithm. The mean absolute error (MAE) is selected as the error loss function. The Dropout layer parameter is set to 0.4 to prevent overfitting.
[0079] S4: Divide the data set into a training set and a validation set, predict the IMF components, and obtain the prediction results.
[0080] This step is divided into the data set partitioning part and the IMF component prediction part, as follows:
[0081] 1) Dataset division
[0082] 15 minutes is selected as a time interval. Each data set includes 7 days of historical data, totaling 672 data points. The data of the first 6 days are used as training sets. The data of the 7th day are used as validation sets.
[0083] Each IMF component is normalized after the data set is divided and used as the input of the LSTM neural network.
[0084] 2) Forecast of IMF components
[0085] The training set data is input into the LSTM neural network to train the hyperparameters of the prediction model. After inputting the validation set data, the prediction data of each IMF component is obtained, and the final prediction result can be obtained by superposition.
[0086] The prediction method proposed in this embodiment is verified by data from Guangxi Zhuang Autonomous Region, China. The power supply side data includes the active power of photovoltaic and wind power, and the load side data involves four typical loads: residential, agricultural, industrial, and commercial. We take 15 minutes as a time interval, and each data set includes 7 days of historical data, totaling 672 data points. The data of the first 6 days are used as the training set for the training of the LSTM network; the data of the 7th day are used as the test set to compare the deviation between the actual value and the predicted value. The mean absolute percentage error (MAPE) and the root mean square error (RMSE) are selected as error evaluation indicators to analyze the prediction results.
[0087] In order to verify the effectiveness and accuracy of the combined prediction model proposed in this invention for a variety of different scenarios, we selected four scenarios with different meteorological conditions and predicted the power output and load level in the study area. The scenario description is as follows:
[0088] S1: hot weather, taken from historical data in August 2023;
[0089] S2: wind and hail weather, taken from historical data in April 2023;
[0090] S3: Heavy rain weather, taken from historical data in March 2023;
[0091] S4: Cold wave weather, taken from historical data in December 2022.
[0092] The prediction results of power supply measurement under different scenarios are shown in the attached figure, and the prediction results of load side are shown in the attached figure. The error evaluation indicators of the prediction results of the above scenarios are shown in Table 1.
[0093] Table 1 Prediction error analysis of each scenario
[0094]
[0095] From the data of the power supply side, the MAPE of the prediction method proposed in this paper does not exceed 8.5%, and the RMSE is below 2.5MW; from the data of the load side, the MAPE of the prediction method proposed in this paper does not exceed 10%, and the RMSE is below 1.7MW. It can be seen that the prediction method proposed in this paper can be used for multi-scenario and multi-type source-load prediction, and can maintain a high prediction accuracy.
[0096] Comparative analysis of combined prediction models
[0097] In order to verify the accuracy of the prediction method proposed in this paper, three control groups are set up, namely LSTM prediction model, VMD-LSTM prediction model and VMD-PSO-LSTM prediction model. Rainy days are selected as the test scene, and the data is taken from May 2023.
[0098] Comparison of prediction results of different prediction methods Figure 5 The error evaluation indicators are shown in Table 2.
[0099] Table 2 Error analysis of different prediction methods
[0100]
[0101] From the data on the power supply side, compared with the single LSTM prediction model, the MAPE of the VMD-LSTM prediction model decreased by 7.8%, and the RMSE decreased by 13.92%; on this basis, the introduction of the PSO optimization algorithm reduced the MAPE of the prediction model by 24.38%, and the RMSE decreased by 16.18%; compared with the VMD-PSO-LSTM prediction model, the prediction model proposed in this paper reduced the MAPE by 14.36%, and the RMSE by 19.30%. Similar conclusions can be drawn from the load side data. It can be seen that the prediction model proposed in this paper has higher prediction accuracy.
[0102] In summary, the present invention collects power supply side data and load side data based on the set sampling time interval, generates a data set, and uses it as the original signal; then sets the number of modal functions K and the penalty factor α, and uses VMD to decompose the original signal into a combination of K intrinsic modal components; secondly, determines the network structure parameters, including the number of input layer units, the number of output layer units, the number of hidden layer units, the learning rate and the maximum number of iterations, and uses the LSTM neural network to establish a prediction model; finally, divides the data set into a training set and a validation set, predicts the IMF component, and obtains the prediction result. The present invention uses the VMD method to decompose the original time data, and uses the RIME optimization algorithm to optimize the LSTM neural network structure, so that the prediction model can more effectively capture the time series characteristics while effectively avoiding the solution process from falling into the local optimum.
[0103] Those of ordinary skill in the art will appreciate that the units of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0104] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored, etc.
[0105] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0106] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nlyMemory), random access memory (RAM, RandomAccessMemory), mobile hard disk, magnetic disk or optical disk, etc., which can store program code.
[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.
Claims
1. A multi-scenario deep learning source-load prediction method based on VMD and RIME, characterized in that: The following steps are involved: Collect power supply side data and load side data based on the set sampling time interval, generate a data set, and use it as the original signal; Set the number of modal functions K and the penalty factor α, and use VMD to decompose the original signal into a combination of K intrinsic mode components; Determine the network structure parameters, including the number of input layer units, the number of output layer units, the number of hidden layer units, the learning rate and the maximum number of iterations, and use the LSTM neural network to establish a prediction model; The data set is divided into a training set and a validation set, and the IMF components are predicted to obtain the prediction results.
2. According to claim 1, a multi-scenario deep learning source-load prediction method based on VMD and RIME is characterized in that: The process of setting the number K of modal functions is as follows: Establish a K value library, which includes all available K values; Starting from K=1, gradually increase the number of modes and calculate the center frequency of each mode; When the change degree of the center frequency falls into the set stable state threshold range or repetition occurs after the number of modes is increased, the current K value is used as the set number K of mode functions.
3. The multi-scenario deep learning source-load prediction method based on VMD and RIME according to claim 2, characterized in that: The specific process of determining the penalty factor α is as follows: Select an initial penalty factor value and determine an iteration termination condition, wherein the number of iterations and the iteration termination condition are: when the change of the penalty factor is less than a set threshold, the iteration is stopped; Perform VMD decomposition step: Perform VMD decomposition using the current penalty factor value to obtain a set of mode functions; Perform the decomposition result evaluation step: check whether the modal function contains the center frequency, and increase or decrease the penalty factor based on the evaluation result and use it as the new penalty factor; The VMD decomposition and evaluation steps are repeated with new penalty factor values until the termination condition is met or the iteration limit is reached.
4. The multi-scenario deep learning source-load prediction method based on VMD and RIME according to claim 1, characterized in that: The structure of the LSTM neural network includes two hidden layers, two Dropout layers and one fully connected layer, where the number of hidden layer units is given by the RIME optimization algorithm; the absolute mean error is selected as the loss value function; and the Dropout layer parameter is set to 0.
4.
5. The multi-scenario deep learning source-load prediction method based on VMD and RIME according to claim 1, characterized in that: The following steps are also included: After the data set is divided into a training set and a validation set, each IMF component is normalized after the data set is divided and used as the input of the LSTM neural network.
6. The multi-scenario deep learning source-load prediction method based on VMD and RIME according to claim 5, characterized in that: The process of forecasting the IMF components and obtaining the forecast results is as follows: Input the training set data into the LSTM neural network to train the hyperparameters of the prediction model; The validation set data is input into the LSTM neural network to obtain the forecast data of each IMF component, and the final forecast result is obtained by superposition.
7. The multi-scenario deep learning source-load prediction method based on VMD and RIME according to claim 1, characterized in that: The power supply side data includes the active power of photovoltaic and wind power.
8. The multi-scenario deep learning source-load prediction method based on VMD and RIME according to claim 1, characterized in that: The load side data includes four types of loads: residential, agricultural, industrial and commercial.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the multi-scenario deep learning source-load prediction method based on VMD and RIME as described in any one of claims 1 to 8.
10. A processor, characterized in that: The processor is used to run a program, wherein when the program is run, the multi-scenario deep learning source-load prediction method based on VMD and RIME described in any one of claims 1 to 8 is executed.