Model training method and device and power load prediction method

Abstract

The invention discloses a model training method and device and a power load prediction method. The model training method comprises the following steps: acquiring power load data; performing decomposition processing on the power load data in the training set to obtain a plurality of IMF components, and performing combination and superposition processing on the IMF components to obtain a plurality of target IMF components; updating model parameters in the neural network model to obtain a target neural network model; predicting the target IMF component by using the target neural network model to obtain a plurality of prediction results, and fusing the plurality of prediction results to obtain a target prediction result corresponding to the power load data; and determining an error between the target prediction result and the test set, and when the error is less than a preset threshold, determining that the neural network model completes training. According to the invention, the technical problem of low prediction precision of the power load in related technologies is solved.

Description

Technical Field

[0001] The present application relates to the field of power load forecasting, and specifically, to a model training method, device and power load forecasting method. Background Art

[0002] In recent years, with the rapid growth trend of "smart home", "new energy vehicles" and "5G technology", the electricity market has also developed rapidly. Forecasting electricity can help companies plan ahead to avoid affecting the normal lives of residents and the operation and production of enterprises when power supply is insufficient, and causing waste of resources when power supply is excessive. Forecasting electricity is an important measure to ensure stable power supply and healthy economic development. Due to the sharp increase in electricity demand and the combined influence of multiple factors, the nonlinear characteristics of electricity data are more obvious, and the difficulty of prediction has increased significantly. It will be difficult to meet the electricity needs of users based on traditional power load analysis and prediction.

[0003] As scholars deepen their research, more and more algorithms are applied to power load forecasting. In the early 20th century, the autoregressive moving average series model attracted much attention and became a research hotspot at the time. For linear stationary power load data, this model can achieve relatively ideal results. It provides an effective way and method to solve problems related to linear stationarity and occupies an important position in the research field at the time. However, power load data in real life often presents non-stationary and nonlinear characteristics and has high time complexity.

[0004] Since the 21st century, as science and technology have gradually developed and matured, neural networks have attracted widespread attention from scholars and have been introduced into the field of power load prediction. Neural networks have outstanding self-learning and adaptive capabilities, and can deal with nonlinear and non-stationary problems of power load data relatively well. For example, the traditional BP neural network predicts power load data, and then the network uses a gradient descent algorithm to update weights and thresholds during training. However, in this prediction method, if the initial value is not selected correctly, the network is prone to fall into local individual extreme points, thereby reducing the prediction accuracy. At the same time, power load data often has a high time complexity, and traditional neural network prediction algorithms usually cannot obtain satisfactory prediction accuracy.

[0005] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0006] The present application provides a model training method, device and power load prediction method to at least solve the technical problem of low power load prediction accuracy in related technologies.

[0007] According to one aspect of the present application, a model training method is provided, including: obtaining power load data, wherein the power load data is divided into a training set and a test set; decomposing the power load data in the training set to obtain multiple intrinsic mode function IMF components, merging and superimposing the IMF components according to the permutation entropy of the IMF components to obtain multiple target IMF components; updating the model parameters in the neural network model through an adaptive dual-center leader collaborative particle swarm algorithm to obtain a target neural network model; predicting the target IMF component using the target neural network model to obtain multiple prediction results, fusing the multiple prediction results to obtain a target prediction result corresponding to the power load data; determining the error between the target prediction result and the test set, and determining that the neural network model has completed training when the error is less than a preset threshold.

[0008] Optionally, the power load data in the training set is decomposed to obtain multiple intrinsic mode function IMF components, including: step S1, adding noise to the power load data in the training set to obtain first power load data; step S2, decomposing the first power load data to obtain multiple first components; calculating the average value of the multiple first components to obtain the first IMF component; step S3, determining the difference between the power load data in the training set and the first IMF component as the first residual signal; adding noise to the first residual signal to obtain the first target residual signal; decomposing the first target residual signal to obtain multiple second mode component; calculating the average value of multiple second modal components to obtain the second IMF component; step S4, determining the difference between the (k-1)th residual signal and the kth IMF component as the kth residual signal; adding noise to the kth residual signal to obtain the kth target residual signal; decomposing the kth target residual signal to obtain multiple (k+1)th modal components; calculating the average value of multiple (k+1)th modal components to obtain the (k+1)th IMF component, wherein k is a positive integer greater than 1; step S5, executing step S4 in ascending order of k until the residual signal cannot be decomposed and multiple IMF components are obtained.

[0009] Optionally, according to the permutation entropy of the IMF components, the IMF components are merged and superimposed to obtain multiple target IMF components, including: reconstructing the IMF components in phase space to obtain multiple sub-vectors; sorting the multiple sub-vectors according to the index value of each sub-vector to obtain a target sequence; determining multiple symbol sequences corresponding to the target sequence, and determining the probability of occurrence of each symbol sequence; determining the initial permutation entropy of the target sequence according to the probability of occurrence of each symbol sequence; normalizing the initial permutation entropy to obtain the target permutation entropy corresponding to the IMF components; and re-merging and superimposing the IMF components whose target permutation entropy is within a preset threshold to obtain multiple target IMF components.

[0010] Optionally, the model parameters in the neural network model are updated by an adaptive dual-center leader collaborative particle swarm algorithm to obtain a target neural network model, including: initializing the model parameters in the neural network model and obtaining the number of input neurons in the neural network model and the number of neurons in the reserve pool; determining the target dimension of particles in the particle swarm according to the number of input neurons in the neural network model and the number of neurons in the reserve pool; initializing the particle swarm in the leader collaborative particle swarm algorithm based on the adaptive dual-center and determining the initial position and initial velocity of each particle; determining the individual extreme value and the group extreme value according to the initial position and the initial velocity; determining the first position of the first center particle in the particle swarm according to the target dimension and the individual extreme value; determining the second position of the second center particle in the particle swarm according to the target dimension and the positions of multiple different particles; updating the first position and the second position by adaptive weights; determining the target group extreme value in the updated first position, the second position and the position of each particle in the particle swarm; taking the target particle corresponding to the target group extreme value as the leader, and updating the position and velocity of each particle in the particle swarm until the preset stop condition is met; setting the model parameters in the neural network model to the parameters of the group extreme value obtained after the preset stop condition is met to obtain the target neural network model.

[0011] Optionally, the target particle corresponding to the extreme value of the target group is used as the leader, and the position and speed of each particle in the particle group are updated, including: detecting the preset life span of the leader, and when the preset life span of the leader is exhausted, randomly generating a challenger in the particle group; obtaining the leadership indicator of the challenger and the leadership indicator of the challenger, and comparing the leadership indicator of the challenger with the leadership indicator of the challenger; if the leadership indicator of the challenger is greater than the leadership indicator of the leader, using the challenger to replace the leader, and updating the position and speed of each particle in the particle group.

[0012] Optionally, if the leadership index of the challenger is greater than the leadership index of the leader, the challenger is used to replace the leader, including: using the challenger to pre-replace the leader within a preset time to perform parameter iterations several times, and if the individual extreme value of any particle is updated during any iteration, the challenger is used to formally replace the leader.

[0013] Optionally, after comparing the leadership indicator of the challenger with the leadership indicator of the leader, the method further includes: if the leadership indicator of the challenger is equal to the leadership indicator of the leader, randomly selecting a dimension in the target dimension, updating the challenger, and comparing the updated leadership indicator of the challenger with the leadership indicator of the leader.

[0014] According to another aspect of the present application, a method for predicting electric power load is also provided, comprising: obtaining electric power load data to be processed; inputting the electric power load data to be processed into a neural network model to obtain a predicted electric power load at a future moment output by the neural network model, wherein the neural network model is obtained by training using the above model training method.

[0015] According to another aspect of the present application, a model training device is also provided, including: an acquisition module for acquiring power load data, wherein the power load data is divided into a training set and a test set; a processing module for decomposing the power load data in the training set to obtain multiple intrinsic mode function IMF components, and merging and superimposing the IMF components according to the permutation entropy of the IMF components to obtain multiple target IMF components; an updating module for updating the model parameters in the neural network model through an adaptive dual-center leader collaborative particle swarm algorithm to obtain a target neural network model; a prediction module for predicting the target IMF component using the target neural network model to obtain multiple prediction results, and fusing the multiple prediction results to obtain a target prediction result corresponding to the power load data; a generation module for determining the error between the target prediction result and the test set, and determining that the neural network model has completed training when the error is less than a preset threshold.

[0016] According to another aspect of the present application, a non-volatile storage medium is also provided, the storage medium includes a stored program, wherein when the program runs, the device where the storage medium is located is controlled to execute the above model training method.

[0017] According to another aspect of the present application, an electronic device is provided, including: a memory and a processor, the processor being used to run a program stored in the memory, wherein the above model training method is executed when the program is running.

[0018] According to another aspect of the present application, a computer program is also provided, wherein the above model training method is implemented when the computer program is executed by a processor.

[0019] According to another aspect of the present application, a computer program product is also provided, which includes a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above model training method is implemented.

[0020] In the present application, power load data is obtained, wherein the power load data is divided into a training set and a test set; the power load data in the training set is decomposed to obtain multiple intrinsic mode function IMF components, and the IMF components are merged and superimposed according to the permutation entropy of the IMF components to obtain multiple target IMF components; the model parameters in the neural network model are updated by an adaptive dual-center leader collaborative particle swarm algorithm to obtain a target neural network model; the target IMF component is predicted using the target neural network model to obtain multiple prediction results, and the multiple prediction results are fused to obtain a target prediction result corresponding to the power load data; the error between the target prediction result and the test set is determined, and when the error is less than a preset threshold, the way in which the neural network model completes training is determined, thereby achieving the purpose of avoiding the traditional neural network from using a gradient descent algorithm to update weights and thresholds, and the network is prone to fall into local extreme points, thereby achieving the technical effect of improving the prediction accuracy of power load data with high time complexity, and thus solving the technical problem of low prediction accuracy of power load in related technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0022] Figure 1 is a flow chart of a model training method according to an embodiment of the present application;

[0023] Figure 2 is a flow chart of another model training method according to an embodiment of the present application;

[0024] Figure 3 A schematic diagram of a relative error absolute value comparison according to an embodiment of the present application;

[0025] Figure 4 A schematic diagram of a comparison between a real value and four neural network algorithms according to an embodiment of the present application;

[0026] Figure 5 Another schematic diagram of the comparison of absolute values ​​of relative errors according to an embodiment of the present application;

[0027] Figure 6 A schematic diagram of another comparison between a true value and four neural network algorithms according to an embodiment of the present application;

[0028] Figure 7 Another schematic diagram of the comparison of absolute values ​​of relative errors according to an embodiment of the present application;

[0029] Figure 8A schematic diagram of another comparison between a true value and four neural network algorithms according to an embodiment of the present application;

[0030] Fig. 9 Another schematic diagram of the comparison of absolute values ​​of relative errors in an embodiment of the present application;

[0031] Fig.10 A schematic diagram of another comparison between a true value and four neural network algorithms in an embodiment of the present application;

[0032] Fig.11 Another schematic diagram of the comparison of absolute values ​​of relative errors in an embodiment of the present application;

[0033] Fig.12 A schematic diagram of another comparison between a true value and four neural network algorithms in an embodiment of the present application;

[0034] Fig.13 Another schematic diagram of the comparison of absolute values ​​of relative errors in an embodiment of the present application;

[0035] Fig.14 A schematic diagram of another comparison between a true value and four neural network algorithms in an embodiment of the present application;

[0036] Fig.15 is a flow chart of a method for predicting power load according to an embodiment of the present application;

[0037] Fig.16 is a structural diagram of a model training device according to an embodiment of the present application;

[0038] Fig.17 It is a hardware structure block diagram of a computer terminal according to a model training method of an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0040] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0041] In order to better understand the embodiments of the present application, the technical terms involved in the embodiments of the present application are explained as follows: CEEMDAN-LDPSO-ESN: a combined prediction method for optimizing echo state neural network with complete integrated empirical mode decomposition with adaptive noise and adaptive dual-center leader collaborative particle swarm algorithm CEEMDAN: complete integrated empirical mode decomposition with adaptive noise; EEMD: integrated empirical mode decomposition; EMD: empirical mode decomposition; ESN: echo state neural network; LDPSO: adaptive dual-center leader collaborative particle swarm algorithm; PSO: particle swarm algorithm; DCPSO: dual-center particle swarm algorithm; ADPSO: adaptive dual-center particle swarm algorithm; PE: column entropy; ACL-PSO: leader particle swarm algorithm; pBest: local optimal solution (individual extreme value); gBest: global optimal solution (group extreme value); MAE: mean absolute error; MAPE: mean absolute percentage error; EMD-PE-ESN: echo state neural network based on empirical mode decomposition and permutation entropy.

[0042] According to an embodiment of the present application, a method embodiment of a model training method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0043] Figure 1 is a flow chart of a model training method according to an embodiment of the present application, such as Figure 1 As shown, the method comprises the following steps:

[0044] Step S101, obtaining power load data, wherein the power load data is divided into a training set and a test set.

[0045] Collect power load data. The data sources include but are not limited to Nanjing's 15-minute load power supply data, the hourly power load demand data of various countries in the European power transmission system, and the Australian electricity price and power load data set. Clean the power load data to remove possible outliers or missing values ​​to ensure data quality. Divide the processed power load data into a training set and a test set. The training set is used to train the prediction model, while the test set is used to evaluate the prediction performance of the model. The division ratio is set according to specific needs. For example, the first 80% of the data can be used as the training set and the remaining 20% ​​of the data can be used as the test set.

[0046] Step S102, decomposing the power load data in the training set to obtain multiple IMF components, merging and superimposing the IMF components according to the permutation entropy of the IMF components to obtain multiple target IMF components.

[0047] For example, the CEEMDAN algorithm is used to decompose the power load data in the training set to obtain a series of IMF components. The CEEMDAN algorithm avoids modal aliasing by adding adaptive noise during the decomposition process, making the decomposed IMF components purer and able to more accurately reflect the inherent structure of the power load data. Among them, the IMF components are a series of signal components decomposed from the original power load data by the CEEMDAN algorithm. Each IMF component has the following characteristics: it is an approximate orthogonal decomposition of the power load data; each IMF component represents a different frequency or scale feature in the power load data; the IMF component has an instantaneous frequency characteristic, that is, at any point in time, the frequency of the IMF signal is definable, which makes the IMF component very suitable for analyzing nonlinear and non-stationary signals.

[0048] The permutation entropy (PE) is calculated for each IMF component to evaluate its complexity and randomness. Permutation entropy is a measure of time series complexity and can be used to evaluate the strength of the nonlinear characteristics of each IMF component. According to the set threshold, the IMF components with permutation entropy values ​​within the threshold are screened out. These components are considered to be the target IMF components containing key information. These target IMF components are merged and superimposed to generate a new sequence for subsequent modeling and prediction.

[0049] Step S103, updating the model parameters in the neural network model through the adaptive dual-center leader collaborative particle swarm algorithm to obtain the target neural network model.

[0050] Among them, neural network models include: Echo State Network (ESN), which is a special recurrent neural network (RNN) mainly used for time series prediction and pattern recognition. The main feature of ESN is that the evolution of its internal state follows the echo state principle, which enables the network to remember the historical information of the input sequence, but avoids the common gradient vanishing or explosion problems in traditional RNN, so that it performs better when processing long sequence data. The workflow of ESN is as follows: 1. Initialization: Set network parameters, including the connection weights, input weights and output weights of the reservoir. Among them, the connection weights and input weights of the reservoir are usually kept fixed, and the output weights are the parts that need to be adjusted during the training process. 2. Training phase: The input sequence enters the network through the input layer. The sequence information is passed to the reservoir through the input weights, interacts with the internal state of the reservoir, and generates a new reservoir state. The new reservoir state is connected to the output layer, and the predicted output is generated through the output weights. Through linear regression or other optimization algorithms, the output weights are adjusted so that the predicted output is as close to the output of the real sequence as possible. 3. Prediction phase: Use the output weights obtained in the training phase to predict the new input sequence. The input sequence enters the network and undergoes dynamic evolution in the reserve pool to generate a prediction result.

[0051] In step S103, the model parameters of ESN are set, including input weight, reserve pool weight, feedback weight, etc. The parameters of the particle swarm algorithm (PSO) are initialized, including particle position, speed, learning factor, etc. The ESN model parameters are optimized using the LDPSO algorithm. The LDPSO algorithm is an improved PSO algorithm that enhances the search capability of the particle swarm by adaptively adjusting the positions of the two central particles and introducing the mechanism of the leader particle swarm algorithm, thereby avoiding local optimal solutions and improving the globality of model parameter optimization. Through iterative optimization, the LDPSO algorithm adjusts the ESN model parameters to the optimal state and obtains the target neural network model.

[0052] Step S104, using the target neural network model to predict the target IMF component to obtain multiple prediction results, and fusing the multiple prediction results to obtain the target prediction result corresponding to the power load data.

[0053] In step S104, the neural network model (here, the ESN model optimized by LDPSO) uses its internally learned weights and connection parameters to independently predict each target IMF component. When predicting, the model predicts the load value at the next moment or in the future through the echo state mechanism based on the historical information of the input sequence. The prediction results of each target IMF component are obtained. These prediction results are component prediction values ​​of the power load, and each component prediction result reflects the information of a specific frequency or time scale in the power load data.

[0054] The prediction results of all target IMF components are linearly superimposed. Since the purpose of CEEMDAN decomposition and PE screening is to decompose complex signals into multiple simpler signal components, these components are recombined after prediction to restore the overall change trend of the original signal. Considering the possible errors in the prediction process, the superimposed prediction results can be error corrected. For example, by introducing additional correction terms or smoothing the prediction results to ensure the accuracy and stability of the final prediction results. After superposition and possible error correction, the final target prediction results of the power load data are obtained. This step integrates all the information extracted by the neural network model from multiple IMF components to provide a complete load forecast value.

[0055] Step S105, determining the error between the target prediction result and the test set, and determining that the neural network model has completed training when the error is less than a preset threshold.

[0056] Compare the target prediction result obtained in step S104 with the actual power load data in the test set, and calculate the prediction error, for example, using MAE (Mean Absolute Error) or MAPE (Mean Absolute Percentage Error) to evaluate the prediction performance. If the prediction error is less than the preset threshold, it indicates that the model prediction result matches the actual data in the test set, and the model training is completed. If the error is greater than the threshold, it is necessary to continue training and adjust the model parameters until the error meets the requirements.

[0057] Preferably, in step S105, the neural network model can be loaded into the memory, for example, the raw data of the neural network model can be loaded from the non-volatile memory into the volatile memory so that the processor runs the neural network model. The raw data of the neural network model refers to unprocessed data, which may include parameters and structural data of the neural network model. The structural data may be a calculation relationship based on the parameters, such as the forward propagation calculation relationship between the intermediate layers and the neurons. Specifically, the structural data may include the structure-related code of the neural network model, such as the code for performing related calculations between the intermediate layers and the neurons. In one embodiment, an area for loading the neural network model can be divided in the memory, which may include a structural data storage area and a parameter storage area. The structural data storage area is used to store structure-related codes, and the parameters referenced by it can point to the address of specific parameters in the parameter storage area through a pointer. During the training process of the neural network model, it may be necessary to update the parameters frequently, and the parameter values ​​in the parameter storage area can be updated.

[0058] According to the above steps, power load data is obtained, wherein the power load data is divided into a training set and a test set; the power load data in the training set is decomposed to obtain multiple intrinsic mode function IMF components, and the IMF components are merged and superimposed according to the permutation entropy of the IMF components to obtain multiple target IMF components; the model parameters in the neural network model are updated by the adaptive dual-center leader collaborative particle swarm algorithm to obtain a target neural network model; the target IMF component is predicted using the target neural network model to obtain multiple prediction results, and the multiple prediction results are fused to obtain the target prediction result corresponding to the power load data; the error between the target prediction result and the test set is determined, and when the error is less than a preset threshold, the way in which the neural network model completes the training is determined, thereby achieving the purpose of avoiding the traditional neural network from being easily trapped in local extreme points when using the gradient descent algorithm to update the weights and thresholds, thereby achieving the technical effect of improving the prediction accuracy of power load data with high time complexity.

[0059] The following Figure 1 The steps shown are exemplary and explanatory.

[0060] According to some optional embodiments of the present application, decomposition processing is performed on the power load data in the training set to obtain multiple intrinsic mode function IMF components, which can be achieved through the following steps.

[0061] Step S1, adding noise to the power load data in the training set to obtain the first power load data; step S2, decomposing the first power load data to obtain multiple first components; calculating the average value of the multiple first components to obtain the first IMF component; step S3, determining the difference between the power load data in the training set and the first IMF component as the first residual signal; adding noise to the first residual signal to obtain the first target residual signal; decomposing the first target residual signal to obtain multiple second modal components; calculating the average value of the multiple second modal components to obtain the first 2IMF component; step S4, determining the difference between the (k-1)th residual signal and the kth IMF component as the kth residual signal; adding noise to the kth residual signal to obtain the kth target residual signal; decomposing the kth target residual signal to obtain multiple (k+1)th modal components; calculating the average value of multiple (k+1)th modal components to obtain the (k+1)th IMF component, wherein k is a positive integer greater than 1; step S5, executing step S4 in order of k from small to large until the residual signal cannot be decomposed and multiple IMF components are obtained.

[0062] The above steps describe the specific process of CEEMDAN algorithm to decompose power load data, which is to extract the intrinsic mode of the signal and reduce the complexity of subsequent prediction model processing. The implementation methods of these steps are explained in detail below: Step S1: Add adaptive white noise to the original power load data set. The addition of adaptive noise is to improve the modal aliasing problem in the decomposition process and ensure the purity of each intrinsic mode function (IMF). After the noise is added, the first power load data obtained is the original data set with noise, which will be used as the input of the next decomposition process. Step S2: Use the CEEMDAN algorithm to decompose the first power load data with adaptive noise to obtain a series of first components. The CEEMDAN algorithm is improved on the basis of EMD (empirical mode decomposition) and aims to improve the accuracy and stability of decomposition through multiple integration and noise addition. All the first components obtained (i.e., the results of multiple decompositions) are averaged to obtain a relatively pure first IMF component. This is a key difference between CEEMDAN and EEMD algorithms, which aims to further reduce modal aliasing. Step S3: Calculate the difference between the original power load data and the first IMF component to obtain the first residual signal. This represents the part of the signal that has not yet been decomposed. Add noise to the first residual signal to generate the first target residual signal, and prepare for the next round of decomposition. Use the CEEMDAN algorithm to decompose the first target residual signal to obtain multiple second modal components. Calculate the average value of these components to finally obtain a pure second IMF component. Step S4 is a loop process for generating subsequent IMF components. The difference between the (k-1)th residual signal and the kth IMF component is determined as the kth residual signal, and then noise is added to the kth residual signal to obtain the kth target residual signal. Perform CEEMDAN decomposition on the kth target residual signal to obtain multiple (k+1)th modal components. Average these components to obtain the (k+1)th IMF component. This process is repeated until the residual signal cannot be further decomposed. Step S5: When the residual signal cannot be further decomposed, the loop terminates. For example, it occurs when the IMF component becomes very slow-changing or close to a trend signal. After performing the above steps, multiple IMF components are obtained, which represent the intrinsic characteristics of different frequencies and time scales of the original power load data in the training set.

[0063] According to some other optional embodiments of the present application, according to the permutation entropy of the IMF components, the IMF components are merged and superimposed to obtain multiple target IMF components, which can be achieved by the following method: reconstructing the IMF components in phase space to obtain multiple sub-vectors; sorting the multiple sub-vectors according to the index value of each sub-vector to obtain a target sequence; determining multiple symbol sequences corresponding to the target sequence, and determining the probability of occurrence of each symbol sequence; determining the initial permutation entropy of the target sequence according to the probability of occurrence of each symbol sequence; normalizing the initial permutation entropy to obtain the target permutation entropy corresponding to the IMF components; re-merging and superimposing the IMF components whose target permutation entropy is within a preset threshold to obtain multiple target IMF components.

[0064] In the above embodiment, each IMF component is reconstructed in phase space, which is a process for capturing the dynamic characteristics of the time series, especially the nonlinear and non-stationary characteristics. By selecting an appropriate embedding dimension m and delay time, a one-dimensional time series can be converted into points in an m-dimensional vector space to form a trajectory in the phase space. From the reconstructed phase space, a series of sub-vectors are extracted. Each sub-vector contains signal values ​​arranged by time delay,

[0065] The generated sub-vectors are sorted according to their positions in the phase space (i.e., index values) to generate an ordered sequence, which helps the subsequent symbolization process. The sorted sub-vector sequence is used as the target sequence, which contains the dynamic characteristics of the signal in the reconstructed phase space and provides a basis for subsequent symbolization and permutation entropy calculation. Each sub-vector in the target sequence is converted into a symbol sequence. The elements of the sub-vector can be sorted and then mapped to a specific symbol or pattern based on the sorting result, so that the continuous value signal can be converted into discrete symbols, which is convenient for analyzing its complexity. The number of occurrences of each different symbol sequence in all generated symbol sequences is counted, and the probability of its occurrence is calculated.

[0066] The permutation entropy is calculated based on the probability of occurrence of the symbol sequence, which measures the complexity and randomness of the signal. In order to make the permutation entropy values ​​comparable between different IMF components, normalization processing is required. A permutation entropy threshold is set, and the IMF components with higher complexity are screened out according to the size of the normalized permutation entropy. The IMF components with larger permutation entropy values ​​contain the main dynamic information of the signal. The IMF components with permutation entropy within the preset threshold are re-merged and superimposed to generate a new signal or component, namely the target IMF component.

[0067] In some optional embodiments of the present application, updating the model parameters in the neural network model by an adaptive dual-center leader collaborative particle swarm algorithm to obtain a target neural network model can be achieved by the following method: initializing the model parameters in the neural network model and obtaining the number of input neurons in the neural network model and the number of neurons in the reserve pool; determining the target dimension of particles in the particle swarm according to the number of input neurons in the neural network model and the number of neurons in the reserve pool; initializing the particle swarm in the leader collaborative particle swarm algorithm based on the adaptive dual-center, and determining the initial position and initial velocity of each particle; determining the individual extreme value and the group extreme value according to the initial position and the initial velocity; determining the first position of the first center particle in the particle swarm according to the target dimension and the individual extreme value; determining the second position of the second center particle in the particle swarm according to the target dimension and the positions of multiple different particles; updating the first position and the second position by adaptive weights; determining the target group extreme value in the updated first position, the second position and the position of each particle in the particle swarm; taking the target particle corresponding to the target group extreme value as the leader, and updating the position and velocity of each particle in the particle swarm until the preset stop condition is met; setting the model parameters in the neural network model to the parameters of the group extreme value obtained after the preset stop condition is met to obtain the target neural network model.

[0068] In the above embodiment, first, the weights in the ESN are randomly initialized. These weights include input weights, reserve pool weights, and output weights. During initialization, the spectral radius of the reserve pool weights is ensured to be less than 1 to ensure the stability of the echo state. The number of input neurons K and the number of neurons N in the reserve pool are determined. K is related to the dimension of the input data. Based on the number of input neurons K of the ESN model and the number of neurons N in the reserve pool, the target dimension of the particles in the particle swarm is determined to be n=K+N. Each particle represents a set of possible model parameters. N is set according to the complexity of the model and the prediction requirements. A larger N can improve the expressiveness of the model.

[0069] Furthermore, the particle swarm is initialized, and the position and velocity of each particle are randomly generated to ensure that the particles cover a wide range of parameter space, which is conducive to subsequent search and optimization. The initial position of each particle is regarded as its individual extreme value, and the position of the best individual in the current particle swarm is determined as the group extreme value. The group extreme value represents the optimal parameter combination in the particle swarm search process.

[0070] Furthermore, based on the target dimension and individual extreme values, the first position of the first center particle is calculated using adaptive weights. The position of the first center particle usually reflects the center of the entire particle swarm. Considering the positions of multiple different particles, the second position of the second center particle is determined by adaptive weights. The position of the second center particle helps to search a wider parameter space. In each iteration, by evaluating the search effect of the particle swarm, the weights of the first position and the second position are dynamically adjusted to guide the particle swarm to more effectively approach the global optimal solution.

[0071] Furthermore, after updating the first position, the second position, and the position of each particle in the particle swarm, the group extreme value is re-evaluated to determine the optimal parameter combination under the current search state. The target particle corresponding to the current group extreme value is used as the leader. The leader particle has the function of guiding the overall search direction of the particle swarm to prevent the particle swarm from falling into the local optimal solution too early. According to the position of the leader particle, as well as the positions of the first center particle and the second center particle, the position and speed of each particle in the particle swarm are updated. The update formula takes into account the current position of the particle, the individual extreme value, the group extreme value, and the position of the two center particles. Repeat the above process until the preset stop condition is met, such as the maximum number of iterations, the minimum error threshold met, or the convergence condition of the particle swarm.

[0072] Furthermore, after the optimization process is completed, the parameters corresponding to the group extreme values ​​that meet the preset stop conditions are set as model parameters in the neural network model, which include input weights, reserve pool weights, and output weights. Through the above steps, the target neural network model optimized by the LDPSO algorithm is obtained. The parameters of the model are adjusted to improve the prediction performance, especially when processing power load data with complex nonlinear and periodic characteristics.

[0073] Optionally, taking the target particle corresponding to the extreme value of the target group as the leader, and updating the position and speed of each particle in the particle group can be achieved by the following method: detecting the preset life span of the leader, and when the preset life span of the leader is exhausted, randomly generating a challenger in the particle group; obtaining the leadership index of the challenger and the leadership index of the challenger, and comparing the leadership index of the challenger with the leadership index of the challenger; if the leadership index of the challenger is greater than the leadership index of the leader, using the challenger to replace the leader, and updating the position and speed of each particle in the particle group.

[0074] At the beginning of the algorithm, a particle is randomly selected or determined according to a certain strategy as a leader and is given a preset lifespan value. This lifespan value indicates how many iterations after which the position of the leader will be re-evaluated, or the duration of the leader as the optimal solution. After each iteration, the lifespan value of the leader decreases, which reflects the "aging" of the leader during the particle swarm search process. As the algorithm progresses, if the leader fails to continuously improve or maintain its optimal solution, its lifespan will gradually run out. When the leader's lifespan value drops to 0 or below a certain threshold, it indicates that the leader may no longer be the optimal solution, or its guiding role in the search process is weakened. When the leader's lifespan is exhausted, a particle is randomly selected from the particle swarm as a challenger. This particle is usually not directly selected based on the current optimal solution of the particle swarm (i.e., the group extreme value gBest), but is randomly selected to increase the diversity of the search. For the challenger, its leadership index is calculated. The leadership index can be the fitness value corresponding to the challenger's position, the update frequency of pBest in the most recent iteration, or its contribution to the overall performance improvement of the particle swarm. The challenger's leadership index is compared with the leadership index of the current leader. If the challenger's leadership index is higher, it means that the challenger's position may guide the particle swarm to search for the global optimal solution more efficiently. If the challenger's leadership index is indeed greater than the leader's, the challenger is promoted to the new leader, and its position and related properties become the new benchmark for guiding the future movement of the particle swarm. Subsequently, the position and velocity of each particle in the particle swarm are updated according to the new leader's position to continue searching for a better solution.

[0075] Optionally, if the leadership index of the challenger is greater than the leadership index of the leader, the challenger is used to replace the leader, which can be achieved by the following method: the challenger is used to pre-replace the leader within a preset time to perform parameter iterations several times. If the individual extreme value of any particle is updated during any iteration, the challenger is used to formally replace the leader.

[0076] Optionally, after comparing the leadership indicator of the challenger with the leadership indicator of the leader, the following steps may be performed: if the leadership indicator of the challenger is equal to the leadership indicator of the leader, randomly select a dimension in the target dimension, update the challenger, and compare the updated leadership indicator of the challenger with the leadership indicator of the leader.

[0077] It is understandable that when the leadership indicator of the challenger is equal to that of the current leader, in order to break this tie and avoid the algorithm from stagnating, a common strategy is to randomly select a dimension in the target dimension for update. This process can introduce new randomness, help the particle swarm algorithm jump out of the possible local optimum, and continue to explore a better solution space. The following are the specific implementation steps: In the target dimension range ([1,D]) ((D) is the dimension of the particle, that is, the number of parameters to be optimized, the algorithm randomly selects a dimension (d). This dimension (d) will be used for the local update of the challenger. The value of the challenger particle in dimension (d) is randomly changed, which can be achieved by adding a random perturbation or sampling from a predefined distribution. The fitness value or leadership indicator of the updated challenger particle is calculated, which can be obtained by mapping the parameters of the particle into a neural network and then evaluating the model performance. The leadership indicator of the updated challenger is compared with the leadership indicator of the current leader. If the challenger If the challenger's leadership indicator is better or not worse than the leader, the challenger may be selected as the new leader. If the updated challenger's leadership indicator is indeed better than the current leader, the challenger replaces the leader. On the other hand, if the challenger's leadership indicator is worse than the leader after the update, the challenger fails and the leader remains unchanged. Regardless of whether the leader is updated, all particles in the particle swarm will update their positions and velocities according to the new leader's position and velocity update rules to continue the iterative search process. The above process is repeated until the preset stop condition is met, such as reaching the maximum number of iterations or the change in fitness value is less than a set threshold.

[0078] Figure 2 is a flowchart of another model training method according to an embodiment of the present application. Figure 2 As shown, the method comprises the following steps:

[0079] First, the power load data is obtained and preprocessed.

[0080] Secondly, the preprocessed power load data set is fed into the CEEMDAN algorithm for decomposition. The CEEMDAN algorithm decomposes complex time series data into a series of intrinsic mode functions (IMFs) and a trend residual signal. This can separate the different frequency components in the power load data, reduce the complexity of data processing, and improve the prediction accuracy.

[0081] Then, the decomposed IMF components are analyzed, especially the complexity of each component is evaluated by calculating the permutation entropy. By setting a threshold, the IMF components with high complexity (i.e., the permutation entropy is lower than the threshold) are screened out as the focus of the subsequent prediction model. These screened IMF components are considered to be signal components containing the key characteristics of the power load data.

[0082] Then, the screened IMF components are fed into the ESN model optimized by LDPSO. The LDPSO algorithm is improved on the basis of particle swarm optimization (PSO). By introducing dual-center particles and leader mechanism, the search efficiency and the ability to avoid local optimal solutions are improved, thereby optimizing the parameters of the ESN model, such as input weight, reserve pool weight and output weight.

[0083] Again, the optimized LDPSO-ESN model is trained using the screened IMF components. During the training process, the model can predict new test data and output prediction results. Since the model is trained based on the screened IMF components, it can capture the dynamic characteristics of power load data more accurately.

[0084] Finally, the prediction results obtained for each IMF component are fused, which can be a linear superposition of the prediction results of all components. The purpose of this step is to reconstruct the original signal and obtain the final power load forecast result. The fusion process can improve the robustness and accuracy of the forecast because the prediction results of each component take into account the different frequency and time scale characteristics of the signal.

[0085] Preferably, the power load data may be decomposed and processed by the following method, wherein the power load data is represented as s(t).

[0086] Step 1.1, adding noise to s(t), can be achieved by the following formula:

[0087] s i (t)=s(t)+ε0w i (t),i=1,2,3...I

[0088] Step 1.2, EMD pair s i (t) is decomposed for the first time, and then its average is calculated, which can be achieved by the following formula:

[0089]

[0090] Step 1.3, calculate the first residual signal, which can be implemented by the following formula:

[0091] r1(t)=s(t)-Avg_IMF1(t) (2)

[0092] Step 1.4, EMD for r1(t)+ε1E1(w i (t)) is decomposed to obtain the second modal component, and then the mean is calculated:

[0093]

[0094] Step 1.5, calculate the kth residual signal and the k+1th modal component obtained by CEEMDAN decomposition, which can be implemented by the following formula:

[0095] r k (t) = r k-1 (t)-Avg_IMF k (t),k=2,3,...,K (4)

[0096]

[0097] Step 1.6, repeat step 1.5 until the residual signal can no longer be decomposed, and the final residual signal is:

[0098]

[0099] Finally, s(t) can be expressed as follows after CEEMDAN decomposition:

[0100]

[0101] Preferably, the IMF components may be merged and superimposed by the following method to obtain a plurality of new (target) IMF components.

[0102] Step 2.1, regard the IMF component as a time series, reconstruct the phase space of the time series {s(i), i=1,2,...,N}, and obtain the following vector:

[0103]

[0104] Where M is the embedding dimension, λ represents the delay time, and k = 1, 2, …, N-(M-1)λ.

[0105] Step 2.2, reorder the reconstructed S(k) in ascending order, and the updated S(k) is as follows:

[0106] S(k)={s(k+(j1-1)λ)≤s(k+(j2-1)λ)≤...≤s(k+(j M -1)λ)} (9)

[0107] Among them, j1,j2,...,j m is the index value, if s(k+(j i1-1)λ)=s(k+(j i2 -1)λ), then sort according to the size of j.

[0108] Step 2.3: According to any S(k) in formula (9), a set of Sign(g) can be obtained:

[0109] Sign(g)=(j1,j2,...,j M ) (10)

[0110] Among them, g=1,2,…,l,l≤M! Since [j1,j2,…,j M ] has M! permutations, so l≤M!. The probability of Sign(g) appearing is P1, P2, ..., P l ,

[0111] Step 2.4, the PE of {s(i),i,2,…,N} is defined based on the Shannon entropy as follows:

[0112]

[0113] when When H p (M) is the largest, while H p (M) = In (M!). For the convenience of description, H p (M) is normalized, and the specific processing is as follows:

[0114]

[0115] Step 2.5, we know that 0≤H p ≤1.H p is the complexity of {s(i),i,2,…,N}, H p The larger it is, the more random {s(i),i,2,…,N} is.

[0116] Step 2.6, according to the size of PE, set a threshold (δ), recombine and superimpose the IMFs with PE values ​​within δ to obtain a new IMF.

[0117] Preferably, the model parameters may also be updated by the following method.

[0118] Step 3.1, initialize ESN and randomly generate W inOnce generated, W cannot be changed, and the initial state x(0) is set to 0. The new IMF of the training data is input to the input u(t) and the output y(t). According to the ESN reserve pool and the input neurons, the dimension n of the particles in the particle swarm is determined, n=K+N, K is the number of ESN input neurons, and N is the number of reserve pool neurons. Initialize the position, speed and related parameters of the particles in the particle swarm, initialize the number of iterations Iter, and the maximum number of iterations IterMax.

[0119] Step 3.2, update x(t) according to formula (13). Calculate the value of fitness, as shown in formula (14):

[0120] x(t+1)=f(W in ×u(t+1)+W×x(t)+W back ×y(t)) (13)

[0121]

[0122] in, is the predicted value, y s is the true value, s means considering its influence on y(t) starting from the M0th value, and s is the influence on y(t) starting from the M0th value.

[0123] Step 3.3, determine if Iter <= IterMax, if so, execute step (1.4), otherwise output gBest, i.e. W out .

[0124] Step 3.4: Update the particles according to the optimization method of the LDPSO algorithm and update Iter. Return to step (1.3).

[0125] Step 3.5, load the new IMF of the test data into the ESN neural network optimized by the LDPSO algorithm, obtain the results of the test data, and superimpose the results to obtain the final prediction data.

[0126] The following are the optimization steps of the LDPSO algorithm: Step 1: Initialize the particle swarm, set the position and speed of the particle, find pBest and gBest, the number of iterations Iter, the maximum number of iterations IterMAx, the dimension of the particle swarm is n, the learning factors c1, c2, select the best particle as the Leader, set the life of the Leader to Θ, initialize the Leader, and calculate fitness. Step 2: Update Iter, Iter = Iter + 1, determine Iter>IterMax, if it is output gBest, otherwise continue to execute Step 3. Step 3: Calculate the individual extreme value of each particle according to formula (15), where the position A is the particle i along vi Direction of flight wv i The position reached after the distance, position B is the particle i along the parallel to pBest i -x i The direction of flight is c1×r1×(pBest i -x i ) distance, the positions of the two center particles are determined according to formulas (16) to (17), and the positions of the two center particles are determined by formulas (18) and (19). When the population size is inappropriate, if the two center particles do not fly, the search ability of the particle group will be reduced and it will be more susceptible to the influence of extreme values. Therefore, the positions of the two center particles are controlled by adaptive weights, so that the center particles are closer to the optimal position and have search capabilities.

[0127] f(pBest i+1 )=min(f(A),f(B),f(x i+1 ),f(pbest i )) (15)

[0128]

[0129]

[0130] Step 4: Determine gBest according to formula (20) and update the Leader. The current gBest is the Leader.

[0131] f(gBest)=min(f(pBest1),f(pBest2),...,f(pBest n ),f(x GCP ),f(x GCPNEW ),f(x scp ),f(x scpnew )) (20)

[0132] Step 5: Update the particle's V according to formula (21) and formula (22) i and X i .

[0133] X i (k+1)=X i (k)+V i (k+1) (21)

[0134] V i (k) = V i +c1·rand·(pBest i (k)-X i (k))+c2·rand·(gBest(k)-Xi (k))(22)

[0135] Update gBest and pBest.

[0136] Step 6: Adjust Θ and θ, θ = θ + 1, evaluate the leadership of the Leader, and update Θ according to the adjustment strategy. If θ ≥ Θ, the lifespan is exhausted and continue to Step 7, otherwise go to Step 2.

[0137] Step 7: Generate a competitor Challenger. When the Leader's Θ is exhausted, a Challenger is generated to challenge and evaluate the Leader.

[0138] Step 8: Evaluate the leadership ability of the competitor Challenger. If the leadership ability of the Challenger can change the current stagnant state of the particle swarm, the Challenger will become the new Leader. The life span of the new Leader Θ=Θ0, and the age θ=0. Otherwise, the old Leader continues to lead the particle swarm, the particle swarm state regresses, and the age θ is updated to Θ-1.

[0139] It is worth noting that the key to the LDPSO algorithm lies in the following three aspects.

[0140] 1. Adaptive central particles: When the population size is not appropriate, if the two central particles do not fly, the search ability of the particle group will be reduced and it will be more susceptible to the influence of extreme values. Therefore, the positions of the two central particles are controlled by adaptive weights, so that the central particles are closer to the optimal position and have search capabilities.

[0141] 2. Generation of Challenger: When θ≥Θ, the Leader's Θ is exhausted and the particle swarm stagnates. To avoid falling into a local extreme value, a Challenger is generated to challenge the Leader or even replace the Leader. In order to protect the Leader's structure, the parameter That is, only one dimension of the Leader is regenerated. If the Leader and Challenger are exactly the same, a dimension h is randomly selected and the Challenger is updated. h , the value is [L j ,U j ]’s random number.

[0142] 3. Evaluation of Challenger's leadership ability: Challenger updates randomly and its leadership ability cannot be directly judged. In order to judge whether Challenger can replace Leader and evaluate Challenger's leadership, Challenger is used to replace Leader for a short period of time for T iterations. iThe update rule is shown in formula (23):

[0143]

[0144] During the iteration, if the pBest position of any particle is updated, the deadlock of the particle group is broken, and the Challenger replaces the Leader, and the age θ and life cycle Θ of the new Leader are set to 0 and Θ0. Otherwise, the particle group returns to the state before the iteration. The Leader is still the original Leader, θ = Θ-1.

[0145] According to the above steps, the present application improves the traditional neural network through two steps. First, CEEMDAN is used to decompose the power load data with high time complexity to obtain several subsequences IMF. According to the PE value of the subsequence, several subsequences are re-merged and superimposed to generate new subsequences. Secondly, the LDPSO-ESN neural network is used to predict the new subsequences. Finally, the prediction results are superimposed to obtain the actual prediction results. This invention effectively improves the prediction accuracy of power load data with high time complexity, and avoids the problem that the traditional neural network is prone to fall into local extreme points when using the gradient descent algorithm to update weights and thresholds.

[0146] In order to solve the problem of modal aliasing, traditional EEMD adds Gaussian white noise to s(t), which can make the extreme points more evenly distributed and cover abnormal signals that easily cause modal aliasing. However, EEMD adds white noise to s(t), which will affect the decomposition of s(t) after a finite number of averaging. The reconstruction error will be reduced as the number of averaging increases, but the amount of calculation will increase to a certain extent. For this reason, CEEMDAN is proposed. The CEEMDAN algorithm adds adaptive white noise to each stage of EMD decomposition. When the number of integrations is small, its reconstruction error is almost zero. CEEMDAN not only overcomes the defects of modal aliasing, but also solves the problem of incomplete decomposition of the EMD algorithm and the problem of increasing the amount of calculation by relying on multiple integrations. The LDPSO-ESN neural network includes: W of ESN out To optimize, the traditional ESN W out The Wiener-Hopf method is usually used to determine the power load data with high time complexity, and the expected prediction results cannot be obtained. In order to further improve the adaptability of ESN, the LDPSO algorithm is used to determine the W of ESN. out Training is performed to improve the prediction accuracy of ESN. The LDPSO algorithm (adaptive dual-center leader collaborative particle swarm algorithm) is an improvement on the traditional PSO algorithm. The improvement ideas are as follows:

[0147] 1. Introducing the dual-center particle swarm algorithm (DCPSO), in view of the fact that the particle shrinkage speed of the DCPSO algorithm slows down in the later stage and it is easy to fall into the local extreme value, an adaptive dual-center particle swarm algorithm is proposed. The traditional particle swarm algorithm approaches the global optimal solution under the guidance of gBest and pBest, and all particle groups will be attracted to the vicinity of gBest, f (pBest) Regardless of size, it will not have much impact on the global extreme value. In order to obtain the global optimal solution and make all particles approach the global optimal solution, a central particle is introduced, which is formed by the center position of the particle swarm. However, the DCPSO algorithm may be affected by the extreme value. When the population size is small, if the two central particles do not fly, the search speed and space of the particle swarm may be reduced, so that all particles cannot approach the global optimal solution. In view of the above problems, an adaptive dual-center particle swarm algorithm (ADPSO) is proposed in this embodiment, so that particles can search in a wider space and reduce the possibility of particles falling into local extreme values. In the DCPSO algorithm, particles run along the discounted direction, which is different from running in the straight line direction. The running in the discounted direction is more complicated than the straight line, and particles can search in a wider space. When considering the update method of individual extreme values, two "center" positions are introduced. The central particle is composed of the center position of the particle. In the entire search process, it does not have speed. In addition, it has the same properties as other particles. To a certain extent, it reduces the possibility of the particle swarm falling into the local extreme value. However, when the population size is not appropriate, if the two central particles do not fly, the search ability of the particle swarm will be reduced, and it is more susceptible to the influence of extreme values. Therefore, the positions of the two central particles are controlled by adaptive weights, so that the central particles are closer to the optimal position and have search capabilities.

[0148] 2. The idea of ​​leader and challenger in the leader particle swarm algorithm (ACL-PSO) is introduced to prevent the particle swarm from falling into the local optimum too early while maintaining the characteristic of rapid convergence of the particle swarm. In the traditional PSO algorithm, all particles are attracted by gBest, and gBest can be regarded as the leader of the particle swarm. The leadership role played by the local optimum is the key to the premature convergence of the particle swarm. For this reason, the ACL-PSO algorithm assigns appropriate lifespan Θ and age growth mode to the Leader. At the same time, in order to obtain the optimal Θ, it can be adjusted according to the leadership ability of the Leader. If the Leader can lead the particle swarm to learn to a better position, the Leader's Θ will be extended. On the contrary, if the Leader cannot lead the particle swarm well, causing the particles to stagnate, the Leader's Θ will decrease. When the lifespan of Θ is exhausted, the competitor Challenger will be generated and even replace the Leader. The Leader in the ACL-PSO algorithm is not necessarily gBest, but any particle with Θ that can change the stagnation of particles. In the ACL-PSO algorithm, the Leader is always generated from a one-dimensional randomly mutated Challenger (the new Leader and the old Leader are different in one dimension), which expands the search range and reduces the probability of particles falling into local extremes.

[0149] 3. The dual-center particle swarm algorithm (DCPSO) was improved and optimized into the ADPSO algorithm. The positions of the two center particles were controlled by adaptive weights, making the center particles closer to the optimal position and having search capabilities. The leader and challenger ideas in the leader particle swarm algorithm (ACL-PSO) were introduced to expand the search range and reduce the probability of particles falling into local extreme values. The ADPSO algorithm and the ACL-PSO algorithm cooperate with each other to improve the algorithm's optimization ability.

[0150] It should be noted that the training set and test set used for model training in this embodiment both include: power load data set. Due to the time series of power load data, the order of data cannot be disrupted, only continuous data can be extracted, and only the power load data of the previous period can be used to predict the power load data of the next period. In this embodiment, the data selection includes the 15-minute granularity load power supply data and daily weather data set of Nanjing, China from November 1, 2002 to August 31, 2003, the Australian electricity price and power load data set from January 1, 2006 to January 1, 2011, and part of the hourly power load demand data of various countries in the European power transmission system from January 1, 2006 to December 31, 2015 as the data set.

[0151] Training data: 1. 720 data in sequence from the 15-minute load power supply data of Nanjing, China from November 1, 2002 to August 31, 2003; 2. 1,100 data in sequence from the hourly power load demand data of various countries in the European transmission system from January 1, 2006 to December 31, 2015; 3. 21,600 data in sequence every half hour from the Australian electricity price and power load data from January 1, 2006 to January 1, 2011.

[0152] Test data: 1. 120 and 60 data in sequence from the 15-minute granularity load and power supply data of Nanjing, China from November 1, 2002 to August 31, 2003; 2. 360 and 72 data in sequence from the hourly power load demand data of various countries in the European transmission system from January 1, 2006 to December 31, 2015 at 0 o'clock; 3. 1,140 and 720 data in sequence from the Australian electricity price and power load data from January 1, 2006 to January 1, 2011 every half hour.

[0153] It should be noted that: 1. From November 1, 2002 to August 31, 2003, the 15-minute granularity load and power supply data of Nanjing, China, uses the first 12 data to predict the 13th data, that is, the data of 15 minutes in the fourth hour is predicted in the first four hours; 2. From January 1, 2006 to December 31, 2015, the hourly power load demand data of various countries in the European transmission system uses the first 5 data to predict the 6th data, that is, the data of the sixth hour is predicted in the first 5 hours; 3. From January 1, 2006 to January 1, 2011, the data of every half hour in the Australian electricity price and power load data are sequentially used. The first 12 data are used to predict the 13th data, that is, the data of the sixth and a half hour is predicted in the first six hours.

[0154] Furthermore, the mean absolute error and mean absolute percentage error between the model prediction value and the true value in the test set are calculated to determine whether the model has converged. Different algorithms are used to predict the same data set, and the MAE and MAPE are used to evaluate the algorithm. The smaller the MAE and MAPE values, the higher the prediction accuracy.

[0155] In order to detect the performance of the neural network model (CEEMDA-LDPSO-ESN) trained by the model training method proposed in this application, three neural networks, BP, ESN and EMD-PE-ESN, are selected for comparison. The number of input neurons K is set according to different data sets (1. The number of neurons for the 15-minute granularity load power supply data in Nanjing, China from November 1, 2002 to August 31, 2003 is 12, 2. The number of neurons for the hourly power load demand data of various countries in the European transmission system from January 1, 2006 to December 31, 2015 is 5, 3. The number of neurons for the data every half hour in the Australian electricity price and power load data from January 1, 2006 to January 1, 2011 is 12), the number of output neurons is 1, the number of reserve pool nodes of ESN, EMD-PE-ESN and CEEMDAN-LDPSO-ESN is set to 500, and the spectral radius is set to 0.75. MAPE and MAE are used to evaluate different algorithms, and the results are shown in Table 1.

[0156] Table 1 Experimental results of four neural networks on three datasets

[0157]

[0158]

[0159] From the performance evaluation of MAPE and MAE in Table 1, the prediction accuracy of ESN neural network is significantly better than that of BP neural network, mainly because ESN has better short-term memory capacity. The MAE and MAPE values ​​of CEEMDA-LDPSO-ESN neural network prediction method are lower than those of BP, ESN and EMD-PE-ESN neural networks, which shows that the prediction effect of CEEMDAN-LDPSO-ESN power load forecasting method is better than the first three algorithms.

[0160] Figures 2 to 14 The running results of four neural networks on the test set are shown respectively, and the experimental results are analyzed: Figures 2 to 6 The paper shows the prediction results of four neural networks on the 15-minute granularity load power supply data set in Nanjing, China from November 1, 2002 to August 31, 2003, as well as the relative error absolute value experimental results. Figures 7 to 9 The prediction results of four neural networks on the hourly power load demand data set of European power transmission system countries from January 1, 2006 to December 31, 2015, as well as the absolute value of relative error experimental results are compared. Figures 11 to 14The prediction results of four neural networks on the Australian electricity price and power load data set from January 1, 2006 to January 1, 2011, as well as the relative error absolute value experimental results are compared. From the relative error absolute value comparison chart, it can be seen that the CEEMDA-LDPSO-ESN neural network prediction method is better, BP has the worst prediction effect, and ESN has significantly improved prediction results compared with BP.

[0161] Fig.15 is a flow chart of a method for predicting power load according to an embodiment of the present application, such as Fig.15 As shown, the method comprises the following steps:

[0162] Step S1501: input the collected power load data into the power load forecasting system based on CEEMDAN-LDPSO-ESN.

[0163] In step S1502, the system preprocesses the input power load data, including data cleaning (removing noise or outliers), data standardization (ensuring that all values ​​are in a similar range), and data segmentation (dividing the data set into a training set and a test set). Preprocessing ensures data quality and helps improve the accuracy and stability of the prediction model.

[0164] Step S1503, the pre-processed power load data is decomposed by the CEEMDAN algorithm. The CEEMDAN algorithm decomposes the time series data into several IMF subsequences, each of which represents signal characteristics of different frequencies and time scales, which helps to identify and process the multi-scale complexity in the data.

[0165] Step S1504, calculate the PE (Permutation Entropy) value of each IMF subsequence, which is an indicator of sequence complexity. A threshold is set according to the PE value, and the IMF subsequences with PE values ​​within the threshold are recombined and superimposed to generate a new subsequence. This step helps to remove noise and simplify signals, so that the subsequent prediction model can learn the intrinsic pattern of power load data more effectively.

[0166] Step S1505, use the ESN optimized by LDPSO to predict the newly generated subsequence. The LDPSO algorithm introduces dual-center particles and leader mechanism on the basis of the traditional particle swarm optimization algorithm, which can effectively prevent particles from falling into the local optimal solution and enhance the global search capability, thereby improving the prediction accuracy of the ESN neural network. The prediction results of each subsequence by the LDPSO-ESN neural network are superimposed to obtain the final power load prediction value. This is because the subsequence obtained by CEEMDAN decomposition is a decomposition of the original signal. The prediction value of the original signal can be reconstructed by superimposing the prediction results. This process helps to improve the accuracy of the prediction.

[0167] Step S1506, the final prediction results are displayed on the interface in the form of visualization or data tables through the output module of the system, which is convenient for users to interpret and use. In addition, the prediction results can also be exported to file formats for further data analysis or power system management decisions.

[0168] Fig.16 is a structural diagram of a model training device according to an embodiment of the present application, such as Fig.16 As shown, the device comprises:

[0169] The acquisition module 1601 is used to acquire power load data, wherein the power load data is divided into a training set and a test set.

[0170] The processing module 1602 is used to decompose the power load data in the training set to obtain multiple intrinsic mode function IMF components, and merge and superimpose the IMF components according to the permutation entropy of the IMF components to obtain multiple target IMF components.

[0171] The updating module 1603 is used to update the model parameters in the neural network model through the adaptive dual-center leader collaborative particle swarm algorithm to obtain the target neural network model.

[0172] The prediction module 1604 is used to predict the target IMF component using the target neural network model to obtain multiple prediction results, and fuse the multiple prediction results to obtain the target prediction result corresponding to the power load data.

[0173] The generation module 1605 is used to determine the error between the target prediction result and the test set, and when the error is less than a preset threshold, determine that the neural network model has completed training.

[0174] Optionally, the processing module 1602 is further used to perform the following steps: step S1, adding noise to the power load data in the training set to obtain first power load data; step S2, decomposing the first power load data to obtain multiple first components; calculating the average of the multiple first components to obtain the first IMF component; step S3, determining the difference between the power load data in the training set and the first IMF component as the first residual signal; adding noise to the first residual signal to obtain a first target residual signal; decomposing the first target residual signal to obtain multiple second modal components; calculating multiple second The average value of the modal components is used to obtain the second IMF component; step S4, determining the difference between the (k-1)th residual signal and the kth IMF component as the kth residual signal; adding noise to the kth residual signal to obtain the kth target residual signal; decomposing the kth target residual signal to obtain multiple (k+1)th modal components; calculating the average value of multiple (k+1)th modal components to obtain the (k+1)th IMF component, wherein k is a positive integer greater than 1; step S5, executing step S4 in order from small to large of k until the residual signal cannot be decomposed and multiple IMF components are obtained.

[0175] Optionally, the processing module 1602 is further used to perform the following steps: reconstructing the IMF components in phase space to obtain multiple sub-vectors; sorting the multiple sub-vectors according to the index value of each sub-vector to obtain a target sequence; determining multiple symbol sequences corresponding to the target sequence, and determining the probability of occurrence of each symbol sequence; determining the initial permutation entropy of the target sequence according to the probability of occurrence of each symbol sequence; normalizing the initial permutation entropy to obtain a target permutation entropy corresponding to the IMF components; and re-merging and superimposing the IMF components whose target permutation entropy is within a preset threshold to obtain multiple target IMF components.

[0176] Optionally, the update module 1603 is also used to perform the following steps: initialize the model parameters in the neural network model, and obtain the number of input neurons in the neural network model and the number of neurons in the reserve pool; determine the target dimension of particles in the particle swarm according to the number of input neurons in the neural network model and the number of neurons in the reserve pool; initialize the particle swarm in the leader collaborative particle swarm algorithm based on adaptive dual centers, and determine the initial position and initial velocity of each particle; determine the individual extreme value and the group extreme value according to the initial position and the initial velocity; determine the first position of the first center particle in the particle swarm according to the target dimension and the individual extreme value; determine the second position of the second center particle in the particle swarm according to the target dimension and the positions of multiple different particles; update the first position and the second position through adaptive weights; determine the target group extreme value in the updated first position, second position and the position of each particle in the particle swarm; use the target particle corresponding to the target group extreme value as the leader, and update the position and velocity of each particle in the particle swarm until the preset stop condition is met; set the model parameters in the neural network model to the parameters of the group extreme value obtained after the preset stop condition is met, and obtain the target neural network model.

[0177] Optionally, the update module 1603 is also used to perform the following steps: detecting the preset lifespan of the leader, and when the preset lifespan of the leader is exhausted, randomly generating a challenger in the particle swarm; obtaining the leadership indicator of the challenger and the leadership indicator of the challenger, and comparing the leadership indicator of the challenger with the leadership indicator of the challenger; if the leadership indicator of the challenger is greater than the leadership indicator of the leader, using the challenger to replace the leader, and updating the position and speed of each particle in the particle swarm.

[0178] Optionally, the update module 1603 is also used to perform the following steps: including: using the challenger to pre-replace the leader within a preset time to perform parameter iteration several times, if the individual extreme value of any particle is updated during any iteration, using the challenger to formally replace the leader.

[0179] Optionally, the update module 1603 is also used to perform the following steps after comparing the leadership indicator of the challenger with the leadership indicator of the leader: if the leadership indicator of the challenger is equal to the leadership indicator of the leader, randomly select a dimension in the target dimension, update the challenger, and compare the updated leadership indicator of the challenger with the leadership indicator of the leader.

[0180] It should be noted that the above Fig.16The modules in the embodiment may be program modules (e.g., a set of program instructions for implementing a specific function) or hardware modules. For the latter, they may be expressed in the following forms, but are not limited thereto: the expression form of the above modules is a processor, or the functions of the above modules are implemented by a processor.

[0181] It should be noted that Fig.16 The preferred implementation of the illustrated embodiment can be found in Figure 1 The relevant description of the illustrated embodiment will not be repeated here.

[0182] Fig.17 FIG. 1 shows a hardware structure block diagram of a computer terminal for implementing a model training method. Fig.17 As shown, the computer terminal 170 may include one or more (shown as 1702a, 1702b, ..., 1702n in the figure) processors 1702 (the processor 1702 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 1704 for storing data, and a transmission module 1706 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Fig.17 The structure shown is only for illustration and does not limit the structure of the above electronic device. Fig.17 More or fewer components as shown, or with Fig.17 Different configurations are shown.

[0183] It should be noted that the one or more processors 1702 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 170. As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0184] The memory 1704 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the model training method in the embodiment of the present application. The processor 1702 executes various functional applications and data processing by running the software programs and modules stored in the memory 1704, that is, the above-mentioned model training method is realized. The memory 1704 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1704 may further include a memory remotely arranged relative to the processor 1702, and these remote memories may be connected to the computer terminal 170 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0185] The transmission module 1706 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 170. In one example, the transmission module 1706 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission module 1706 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0186] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 170 .

[0187] It should be noted that, in some optional embodiments, the above Fig.17 The computer terminal shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of hardware elements and software elements. It should be noted that Fig.17 This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the computer terminal described above.

[0188] It should be noted that Fig.17 The computer terminal shown is used to execute Figure 1 The model training method shown, therefore the relevant explanations in the execution method of the above commands are also applicable to the electronic device and will not be repeated here.

[0189] An embodiment of the present application also provides a non-volatile storage medium, which includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the above model training method.

[0190] A program for a non-volatile storage medium to perform the following functions: obtaining power load data, wherein the power load data is divided into a training set and a test set; decomposing the power load data in the training set to obtain a plurality of intrinsic mode function IMF components, merging and superimposing the IMF components according to the permutation entropy of the IMF components to obtain a plurality of target IMF components; updating the model parameters in the neural network model through an adaptive dual-center leader collaborative particle swarm algorithm to obtain a target neural network model; predicting the target IMF component using the target neural network model to obtain a plurality of prediction results, fusing the plurality of prediction results to obtain a target prediction result corresponding to the power load data; determining the error between the target prediction result and the test set, and determining that the neural network model has completed training when the error is less than a preset threshold.

[0191] An embodiment of the present application also provides an electronic device, including: a memory and a processor, the processor being used to run a program stored in the memory, wherein the above model training method is executed when the program is running.

[0192] The processor is used to run a program that performs the following functions: obtaining power load data, wherein the power load data is divided into a training set and a test set; decomposing the power load data in the training set to obtain multiple intrinsic mode function IMF components, merging and superimposing the IMF components according to the permutation entropy of the IMF components to obtain multiple target IMF components; updating the model parameters in the neural network model through an adaptive dual-center leader collaborative particle swarm algorithm to obtain a target neural network model; predicting the target IMF component using the target neural network model to obtain multiple prediction results, fusing the multiple prediction results to obtain a target prediction result corresponding to the power load data; determining the error between the target prediction result and the test set, and determining that the neural network model has completed training when the error is less than a preset threshold.

[0193] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0194] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0195] In the above-mentioned embodiments of the present application, the collected information is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary protection measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0196] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0197] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0198] In addition, each functional unit in each embodiment of the present application 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.

[0199] 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 application, or the part that contributes to the relevant technology 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 application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk, etc. Various media that can store program codes.

[0200] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A model training method, characterized in that: include: Acquiring power load data, wherein the power load data is divided into a training set and a test set; Decomposing the power load data in the training set to obtain a plurality of intrinsic mode function IMF components, and merging and superimposing the IMF components according to the permutation entropy of the IMF components to obtain a plurality of target IMF components; The model parameters in the neural network model are updated through the adaptive dual-center leader cooperative particle swarm algorithm to obtain the target neural network model; Using the target neural network model to predict the target IMF component to obtain multiple prediction results, and fusing the multiple prediction results to obtain a target prediction result corresponding to the power load data; Determine the error between the target prediction result and the test set, and when the error is less than a preset threshold, determine that the neural network model has completed training.

2. The method according to claim 1, characterized in that The power load data in the training set is decomposed to obtain multiple intrinsic mode function IMF components, including: Step S1, adding noise to the power load data in the training set to obtain first power load data; Step S2, decomposing the first power load data to obtain a plurality of first components; calculating an average value of the plurality of first components to obtain a first IMF component; Step S3, determining the difference between the power load data in the training set and the first IMF component as the first margin signal; adding noise to the first margin signal to obtain a first target margin signal; decomposing the first target margin signal to obtain a plurality of second modal components; calculating the average value of the plurality of second modal components to obtain a second IMF component; Step S4, determining the difference between the (k-1)th residual signal and the kth IMF component as the kth residual signal; adding noise to the kth residual signal to obtain the kth target residual signal; decomposing the kth target residual signal to obtain a plurality of (k+1)th modal components; calculating the average value of the plurality of (k+1)th modal components to obtain the (k+1)th IMF component, wherein k is a positive integer greater than 1; Step S5, executing step S4 in order from small to large k until the residual signal cannot be decomposed and a plurality of IMF components are obtained.

3. The method according to claim 1 or 2, characterized in that: According to the permutation entropy of the IMF components, the IMF components are merged and superimposed to obtain multiple target IMF components, including: Reconstructing the IMF components in phase space to obtain a plurality of sub-vectors; sorting the plurality of sub-vectors according to an index value of each sub-vector to obtain a target sequence; Determine multiple symbol sequences corresponding to the target sequence, and determine the probability of each of the symbol sequences occurring; determine the initial permutation entropy of the target sequence according to the probability of each of the symbol sequences occurring; perform normalization processing on the initial permutation entropy to obtain the target permutation entropy corresponding to the IMF component; The IMF components whose target permutation entropy is within a preset threshold are re-merged and superimposed to obtain a plurality of the target IMF components.

4. The method according to claim 1, characterized in that: The model parameters in the neural network model are updated through the adaptive dual-center leader collaborative particle swarm algorithm to obtain the target neural network model, including: Initializing the model parameters in the neural network model, and obtaining the number of input neurons in the neural network model and the number of neurons in the reserve pool; determining the target dimension of the particles in the particle swarm according to the number of input neurons in the neural network model and the number of neurons in the reserve pool; Initializing a particle swarm in a leader cooperative particle swarm algorithm based on an adaptive dual-center, and determining an initial position and an initial velocity of each particle; determining an individual extreme value and a group extreme value according to the initial position and the initial velocity; Determine a first position of a first central particle in the particle group according to the target dimension and the individual extreme value; determine a second position of a second central particle in the particle group according to the target dimension and the positions of a plurality of different particles; updating the first position and the second position by adaptive weights; Determine a target group extreme value among the updated first position, the second position, and the position of each particle in the particle group; take the target particle corresponding to the target group extreme value as a leader, and update the position and speed of each particle in the particle group until a preset stop condition is met; The model parameters in the neural network model are set to the parameters of the population extreme value obtained after the preset stop condition is met, so as to obtain the target neural network model.

5. The method according to claim 4, characterized in that The target particle corresponding to the extreme value of the target group is used as a leader, and the position and speed of each particle in the particle group are updated, including: detecting a preset life span of the leader, and randomly generating a challenger in the particle group when the preset life span of the leader is exhausted; Acquire the leadership indicator of the challenger and the leadership indicator of the challenger, and compare the leadership indicator of the challenger with the leadership indicator of the challenger; If the leadership index of the challenger is greater than the leadership index of the leader, the challenger is used to replace the leader, and the position and speed of each particle in the particle swarm are updated.

6. The method according to claim 5, characterized in that If the leadership index of the challenger is greater than the leadership index of the leader, using the challenger to replace the leader includes: The challenger is used to replace the leader in advance within a preset time to perform parameter iteration several times. If the individual extreme value of any particle is updated during any iteration, the challenger is used to formally replace the leader.

7. The method according to claim 5, characterized in that After comparing the leadership indicator of the challenger with the leadership indicator of the challenger, the method further includes: If the leadership indicator of the challenger is equal to the leadership indicator of the leader, a dimension is randomly selected from the target dimension, the challenger is updated, and the leadership indicator of the updated challenger is compared with the leadership indicator of the leader.

8. A method for predicting power load, characterized in that: include: Obtaining power load data to be processed; The power load data to be processed is input into a neural network model to obtain a predicted power load at a future moment output by the neural network model, wherein the neural network model is obtained by training using the model training method described in any one of claims 1 to 7.

9. A model training device, characterized in that: include: An acquisition module, used for acquiring power load data, wherein the power load data is divided into a training set and a test set; A processing module, used for decomposing the power load data in the training set to obtain a plurality of intrinsic mode function IMF components, and merging and superimposing the IMF components according to the permutation entropy of the IMF components to obtain a plurality of target IMF components; An updating module is used to update the model parameters in the neural network model through an adaptive dual-center leader collaborative particle swarm algorithm to obtain a target neural network model; A prediction module, used to predict the target IMF component using the target neural network model to obtain multiple prediction results, and to fuse the multiple prediction results to obtain a target prediction result corresponding to the power load data; A generation module is used to determine the error between the target prediction result and the test set, and when the error is less than a preset threshold, determine that the neural network model has completed training.

10. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored program, wherein when the program is running, the device where the non-volatile storage medium is located is controlled to execute the model training method described in any one of claims 1 to 7.

11. An electronic device, characterized in that: include: A memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program executes the model training method described in any one of claims 1 to 7 when running.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the model training method described in any one of claims 1 to 7 is implemented.

Images