New energy grid connection flexibility resource prediction model construction method based on artificial intelligence

Abstract

The invention relates to the technical field of new energy grid connection, in particular to a new energy grid connection flexibility resource prediction model training method based on artificial intelligence, and aims at solving the problems that a traditional prediction method is insufficient in precision and poor in adaptability when coping with new energy grid connection flexibility resources. By collecting and processing multi-dimensional historical data, adopting machine learning and deep learning algorithms and combining artificial intelligence, effective features can be extracted from the data, and new energy grid-connected flexibility resources are accurately predicted. Through data expansion and optimization, the problems of data scarcity and uneven distribution are effectively solved, the generalization ability and prediction precision of the model are improved, more scientific and accurate decision support is provided for power system scheduling, and stable operation and flexible scheduling of a new energy grid-connected system are promoted.

Description

Technical Field

[0001] The present invention relates to the field of new energy grid-connected technology, and in particular to a method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence. Background Art

[0002] With the transformation of the global energy structure, the proportion of new energy is increasing. The volatility and intermittent characteristics of wind energy, solar energy, etc. make the flexibility and stability of new energy grid connection an important challenge for power system management. Traditional power system dispatching methods are difficult to adapt to these dynamic changes. More advanced technical means, especially artificial intelligence technology, are needed to achieve accurate prediction and optimal dispatch of new energy grid-connected flexibility resources. In the new energy grid-connected system, the real-time power output of wind power, photovoltaic and other power generation units is affected by many factors such as weather, equipment operating status, dispatching instructions, etc., and the dispatching status of energy storage equipment increases the complexity of the system. Therefore, how to effectively collect grid-connected data, accurately predict the flexibility of grid-connected resources, and ensure the stable operation of the power grid has become an urgent problem to be solved. Summary of the invention

[0003] The present invention provides a carbon-based artificial intelligence new energy grid-connected flexibility resource prediction model construction method, which can effectively solve the problems in the background technology.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] A method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence includes the following steps:

[0006] Collect grid-connected data based on real-time operation data of new energy grid-connected systems and perform manual annotation;

[0007] A data expansion model is constructed using a generative adversarial network algorithm based on entropy loss constraints to expand the grid-connected data.

[0008] A feature extraction model is constructed using a neural network algorithm based on dynamic adaptive oscillation to extract features from grid-connected data;

[0009] The feature dimension reduction model is constructed by using an autoencoder based on boundary smoothing strategy to reduce the dimension of the extracted features;

[0010] The extreme learning machine algorithm based on stationarity constraint is used to build a classifier model to classify the features after dimensionality reduction.

[0011] Furthermore, the collection of the grid-connected data includes the following steps:

[0012] Obtain power output data based on sensors deployed in each renewable energy power generation unit;

[0013] Obtain meteorological data of the location area of ​​each renewable energy power generation unit through the meteorological station;

[0014] Obtain the dispatch plan and operating status of each renewable energy power generation unit based on the dispatch platform.

[0015] Furthermore, the training method of the data expansion model comprises the following steps:

[0016] Initialize the generator and discriminator of the generative adversarial network;

[0017] The generator generates data samples under the entropy loss constraint;

[0018] The discriminator evaluates the authenticity of the data sample based on the self-similarity loss of fractal theory and guides the optimization of the generator;

[0019] The training is iterated until a preset stopping condition is met, thereby completing the training of the data expansion model.

[0020] Furthermore, the training method of the feature extraction model comprises the following steps:

[0021] Initialize the parameters of the neural network and the hyperparameters of the dynamic adaptive oscillation;

[0022] Adjust the oscillation frequency of each parameter based on the loss function of the neural network and simultaneously update the phase of the oscillation;

[0023] Adaptively adjust the amplitude according to the parameter update effect;

[0024] Use sine and cosine functions to simulate oscillation behavior and calculate the updated values ​​of the neural network weights and biases;

[0025] The above steps are iterated repeatedly until a preset stop iteration condition is met to complete the training of the feature extraction model.

[0026] Furthermore, the training method of the feature dimensionality reduction model comprises the following steps:

[0027] Initialize the weights and bias parameters of the encoder in the autoencoder, as well as the weights and bias parameters of the decoder;

[0028] The autoencoder is trained by optimizing the loss function and the parameters of the autoencoder are updated synchronously with the optimization operation.

[0029] Automatic boundary smoothing strategy is used to handle outliers and boundary effects in the dimensionality reduction process;

[0030] Optimize the training process of the autoencoder using the convex hull convergence strategy;

[0031] The training is iterated until a preset stopping condition is met, thereby completing the training of the feature dimensionality reduction model.

[0032] Furthermore, the training method of the classifier model comprises the following steps:

[0033] Initialize the network structure of the extreme learning machine, including the input layer, hidden layer and output layer;

[0034] By limiting the change of the connection weights between the hidden layer and the output layer, the neural network is constrained to be stable.

[0035] Adopting adaptive bias correction strategy to adjust the bias term of extreme learning machine algorithm;

[0036] A multi-task learning framework is used to optimize the extreme learning machine algorithm, and the losses of attribute prediction, sample distribution prediction, and uncertainty prediction are comprehensively considered;

[0037] Update the parameters of the extreme learning machine algorithm based on the training error;

[0038] The training is iterated until a preset stopping condition is met, thereby completing the training of the classifier model.

[0039] Furthermore, the generator optimizes the generated data samples by minimizing the change in entropy, which is defined as:

[0040]

[0041] Among them, k B is the Boltzmann constant; P c (i) is the probability distribution of the samples generated by the generator on the i-th feature dimension; na is the feature dimension of the generated samples; S c is the entropy of the generated sample; Diversity loss for generating samples;

[0042] The diversity loss of generated data samples uses the nonlinear Lagrange multiplier method, which is calculated as follows:

[0043]

[0044] in, is the penalty coefficient of diversity constraint; σ c is a scale factor used to normalize the diversity of generated samples and is a training parameter; c To control the weight parameter of diversity enhancement;

[0045] The probability distribution is calculated based on the energy value of the feature, which is specifically expressed as:

[0046]

[0047] Among them, Z c is a normalization constant that ensures that the sum of all probabilities is 1; E c (i) is the energy value of the feature in the i-th dimension; T c is the temperature term, k B is the Boltzmann constant;

[0048] The complexity of features of different dimensions is represented by the energy value of the feature, and the calculation method is expressed as:

[0049]

[0050] in, is the eigenvalue of the jth dimension of the generated sample, X c，j is the eigenvalue of the jth dimension of the real sample; is the feature weight coefficient, which is a training parameter used to quantify the impact of each feature dimension on entropy change.

[0051] Furthermore, by optimizing the total loss function of the discriminator, the authenticity of the generated data samples is evaluated and the fractal self-similarity characteristics of the grid-connected data are captured. The total loss function of the discriminator is the sum of the traditional adversarial loss and the fractal self-similarity loss, and the calculation method is expressed as:

[0052]

[0053] Among them, λ c To control the balance coefficient of the discriminator loss;

[0054] The traditional way of calculating adversarial loss is expressed as:

[0055]

[0056] in, To combat losses; is the expectation; ~ means it obeys a specific distribution; X c Indicates real grid-connected data; D c () is the mapping function of the discriminator, which represents the probability that the input grid-connected data is the real grid-connected data; G c is the generator; P c is the distribution of real grid-connected data;

[0057] The calculation method of fractal self-similarity loss is expressed as:

[0058]

[0059] Among them, X c,i represents the eigenvalue of the real grid-connected data in the i-th dimension, α c is the self-similarity coefficient; β c is the multi-scale analysis coefficient; Mc is the number of scales in fractal multiscale analysis, indicating the self-similarity of grid-connected data generated from multiple scale analyses, enabling the discriminator to identify fractal features of grid-connected data at different levels of detail.

[0060] Furthermore, the initialization method of the generative adversarial network generator and discriminator is expressed as:

[0061]

[0062]

[0063] φ 0p =0;

[0064] A 0p =01;

[0065] Among them, ~ indicates compliance with a specific distribution; It means the mean is 0 and the variance is Normal distribution of represents normal distribution; represents the initialization variance of the neural network; W 0p is the initial weight of the neural network, b 0p is the initial bias of the neural network; φ 0p is the initial phase of oscillation; A 0p is the initial amplitude of oscillation.

[0066] Furthermore, the oscillation frequency of each parameter is adjusted based on the local curvature estimation of the loss surface, and the historical gradient information is used for adjustment. The adjustment method is expressed as:

[0067]

[0068] Among them, β p is the historical gradient weight factor, which controls the influence of the historical gradient on the current frequency adjustment; ω p is the oscillation frequency, is the oscillation frequency of the tth iteration; is the weight of the neural network at the t-1th iteration; is the weight of the neural network at the kth iteration; L p is the loss function of the neural network; is the sign of the partial derivative; ω0 is the fundamental oscillation frequency; α p is a hyperparameter that adjusts the oscillation responsiveness;

[0069] Update the phase and amplitude of the oscillation based on the parameter update effect:

[0070] The phase update method of the oscillation is expressed as:

[0071]

[0072] Among them, φ p is the phase of the oscillation, is the phase of the oscillation at the tth iteration; is the phase of the oscillation at the t-1th iteration; δ p is the phase adjustment factor; tanh() is the hyperbolic tangent function, which can limit the adjustment range of the phase to avoid excessive adjustment leading to instability of the algorithm; λ p is a hyperparameter that adjusts phase sensitivity;

[0073] The adaptive adjustment method of amplitude is expressed as:

[0074]

[0075] Among them, A p is the amplitude of the oscillation, is the amplitude of the oscillation at the tth iteration; is the amplitude of the oscillation at the t-1th iteration; γ p is the amplitude adjustment factor.

[0076] Furthermore, the updated values ​​of the weights and biases of the neural network are calculated based on the oscillation behavior, and the update method is expressed as:

[0077]

[0078] in, is the weight of the neural network at the tth iteration; is the bias of the neural network at the tth iteration; is the weight of the neural network at the t-1th iteration; The bias of the neural network at the t-1th iteration; t * is the current iteration number of the neural network.

[0079] Furthermore, the way to initialize the autoencoder parameters is expressed as:

[0080]

[0081] in, represents the weight of the encoder, Indicates the encoder bias; represents the weight of the decoder, represents the bias of the decoder; randn() represents the random function; d r,in is the dimension of the encoder input layer, d r,mid is the dimension of the middle layer of the encoder, d r,out is the dimension of the encoder output layer; zeros() is a function that generates a full 0 vector.

[0082] Furthermore, during the training process of the autoencoder, the parameter update amount for each iteration is calculated as:

[0083]

[0084] in, is the weight parameter update amount of the encoder; is the encoder bias parameter update amount; is the update amount of the decoder weight parameters; is the update amount of the bias parameter of the decoder; is the learning rate of the autoencoder at the tth iteration;

[0085] The update method of the autoencoder parameters is expressed as:

[0086]

[0087] Among them, ← represents the parameter update operation;

[0088] The calculation method of the loss function of the autoencoder is expressed as:

[0089]

[0090] Among them, L r is the loss function of the autoencoder; X r is the input data of the autoencoder, is the output reconstructed by the decoder, Sig() is the Sigmoid activation function; |||| is the L2 norm; is the output of the smoothed encoder; is the weight coefficient of the i-th dimension autoencoder; n r is the vector dimension corresponding to the samples of the current batch input to the autoencoder;

[0091] The calculation method of the weight coefficient of the autoencoder is expressed as:

[0092]

[0093] in, is the characteristic variance of the input data of the i-th dimension autoencoder; is the feature of the input data of the i-th dimension autoencoder; ∈ re is a small constant to avoid division by zero; Var() is the variance function.

[0094] Furthermore, the output of the encoder is smoothed by a smoothing function, and the calculation method is expressed as:

[0095]

[0096] Among them, τ r () is a smooth function, δ r is the smoothing strength parameter; θ r is an automatically adjusted threshold parameter used to control the degree of smoothing;

[0097] The smoothing strength parameter is determined by an adaptive method based on the local density of the data, and the calculation method is expressed as:

[0098]

[0099] in, calculate In its neighborhood The kernel density estimation in can accurately process data in high-density and sparse areas, improving the robustness and adaptability of the overall model; represents the neighborhood; k r () is the kernel density estimation function; mean() is the average function; ∈ rea is a preset small constant.

[0100] Furthermore, the learning rate is dynamically adjusted based on the data distribution characteristics of the training process. The calculation method is expressed as:

[0101]

[0102] in, is the learning rate of the initial autoencoder; is the learning rate of the autoencoder at the tth iteration; γ r is the learning rate decay factor of the autoencoder, is the average diameter of the convex hull in the tth iteration, is the diameter of the initial convex hull.

[0103] Furthermore, in the initialization process of the extreme learning machine network structure, the weights of the hidden layer are randomly initialized, the bias term of the extreme learning machine is set to a constant of 0, and the weights of the output layer are optimized through training;

[0104] After initialization, the stationary function is used to constrain the update amplitude of the weights from the hidden layer to the output layer. The calculation method of the stationary function is expressed as:

[0105]

[0106] Among them, W u,i is the weight between the i-th hidden node and the output layer of the extreme learning machine; f u () is the stationary function; ΔW u,i Represents the weight change of the i-th node; ||W u,i ||2 is the L2 norm of the i-th hidden node weight; |||| is the L2 norm; λ u is the stability regularization parameter, which controls the magnitude of weight updates.

[0107] Furthermore, the bias correction term is calculated by monitoring the training error and validation error of the extreme learning machine. The calculation method is expressed as:

[0108]

[0109] Among them, N u is the number of training samples input to the extreme learning machine; y u,i is the true label of the i-th training sample; is the prediction output of the extreme learning machine for the i-th training sample; η u is the learning rate of the extreme learning machine, which is used to control the correction speed; α u (t) is the dynamic adjustment factor of the tth iteration;

[0110] The calculation method of the dynamic adjustment factor is expressed as:

[0111]

[0112] Among them, β u is a hyperparameter that controls the rate at which the correction factor decreases.

[0113] Furthermore, the total loss function of the extreme learning machine comprehensively considers the losses of attribute prediction, sample distribution prediction, and uncertainty prediction, and the calculation method is expressed as:

[0114] L u =L u,task1 +Y u ·L u,task2 +α u ·L u,task3 ;

[0115] L u,task1 is a loss function based on classification error, specifically using cross entropy loss, and the calculation method is expressed as:

[0116]

[0117] Among them, y ui,j and are the j-th class probability values ​​of the true label of the i-th sample and the label predicted by the extreme learning machine; kes is the number of categories;

[0118] L u,task2 It is the loss function for sample distribution prediction, specifically using KL divergence loss, and the calculation method is expressed as:

[0119]

[0120] Where DKL(||) represents the Kullback-Leibler divergence, P u (X ui )and are the i-th real data distribution and predicted data distribution respectively;

[0121] L u,task3 is the uncertainty prediction loss function, which uses entropy loss to measure the uncertainty of model prediction. The calculation method is expressed as:

[0122]

[0123] in, It represents the prediction entropy of the i-th sample, reflecting the prediction uncertainty of the model for this sample.

[0124] Furthermore, the gradient descent method is used to update the parameters of the extreme learning machine algorithm, and the calculation method is expressed as:

[0125]

[0126] Among them, W u (t) is the weight of the extreme learning machine at the tth iteration; b u (t) is the bias of the extreme learning machine at the tth iteration; W u (t+1) is the weight of the extreme learning machine at the t+1th iteration; b u (t+1) is the bias of the extreme learning machine at the t+1th iteration; and are the gradients of the extreme learning machine's loss function with respect to weights and biases, respectively.

[0127] The beneficial effects of the present invention are:

[0128] The present invention generates more training samples through a generative adversarial network algorithm based on entropy loss constraints, which greatly solves the problem of scarcity of training data, enhances the diversity and representativeness of the training set, improves the generalization ability of the model, and reduces the prediction error caused by uneven data.

[0129] The neural network algorithm based on dynamic adaptive oscillation in the present invention can effectively avoid the gradient vanishing and explosion problems in high-dimensional feature space, enhance the adaptability and stability of the feature extraction model in complex data, optimize the learning process of the neural network, reduce the influence of local optimal solutions, and improve the accuracy of feature extraction.

[0130] The boundary smoothing-based autoencoder of the present invention effectively handles boundary effects and outlier problems in the process of dimensionality reduction, enhances the robustness of the dimensionality reduction result, and can maintain the main feature information of the data, making the reduced dimensionality data more suitable for further processing by the classifier, thereby improving the processing efficiency and accuracy of the model.

[0131] The extreme learning machine based on stationarity constraint in the present invention ensures the stability of network parameters during training through stationarity constraint in classification tasks, avoids the occurrence of overfitting, and improves classification accuracy. At the same time, the combination of adaptive bias correction and multi-task learning framework enables the classifier to adapt to different data distributions, thereby improving the accuracy of classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0132] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0133] Figure 1 A flow chart of a method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence in the present invention;

[0134] Figure 2 It is a training flow chart of the autoencoder algorithm based on boundary smoothing in the present invention;

[0135] Figure 3 This is a training flow chart of the neural network algorithm based on dynamic adaptive oscillation in the present invention. DETAILED DESCRIPTION

[0136] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0137] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only and do not represent the only implementation method.

[0138] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0139] In the existing technology, the prediction model of renewable energy grid-connected flexibility resources usually faces problems such as data scarcity and uneven distribution. Especially when there are insufficient samples or complex data types, the existing training methods often cannot fully explore the potential patterns in the data, thus affecting the prediction accuracy of the model. In addition, the existing data expansion technology mostly relies on simple interpolation or noise generation methods, which cannot effectively increase the diversity and representativeness of the data.

[0140] Traditional feature extraction methods are prone to gradient vanishing or gradient exploding problems in high-dimensional data, which makes the neural network training process unstable and the training effect difficult to guarantee. Existing neural network algorithms often fall into local optimal solutions during training, resulting in insufficient feature extraction, which in turn affects the overall performance of the model.

[0141] Furthermore, existing feature dimensionality reduction methods often ignore the impact of boundary effects and outliers when processing high-dimensional data, resulting in dimensionality reduction results that are not stable and robust enough. When processing complex data, existing methods cannot effectively retain the key feature information of the data, resulting in difficulties in subsequent classifier processing.

[0142] Traditional classifier algorithms often lack effective constraints on the training process when dealing with complex data sets, which can easily lead to overfitting and training instability. Although the existing extreme learning machine method has a fast training speed, it lacks classification accuracy and model convergence when dealing with multi-feature and multi-dimensional data, especially when the data is unevenly distributed, and cannot achieve good classification results.

[0143] Based on the technical problems existing in the prior art, the present invention discloses a method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence, such as Figures 1 to 3 As shown, the following steps are included:

[0144] Based on the real-time operation data of the new energy grid-connected system, the grid-connected data is collected and manually labeled; the generative adversarial network algorithm based on entropy loss constraints is used to build a data expansion model to expand the grid-connected data; a neural network algorithm based on dynamic adaptive oscillation is used to build a feature extraction model to extract features from the grid-connected data; an autoencoder based on a boundary smoothing strategy is used to build a feature dimensionality reduction model to reduce the dimensionality of the extracted features; a classifier model is built using an extreme learning machine algorithm based on stationarity constraints to classify the features after dimensionality reduction.

[0145] Compared with the prior art, the present invention adopts a generative adversarial network algorithm inspired by thermodynamics for data expansion. The generator generates samples under the constraint of entropy loss, and the discriminator evaluates the similarity between the generated samples and the real samples, which solves the problems of data scarcity and uneven distribution in the task of new energy grid-connected flexibility resource prediction. The generative adversarial network simulates the entropy changes in thermodynamics to make the generated samples more representative and accurate in the feature space.

[0146] Furthermore, a neural network algorithm based on dynamic adaptive oscillation is used for feature extraction. By simulating the nonlinear oscillation behavior in physical phenomena, the algorithm can effectively explore and utilize local extreme values ​​in high-dimensional parameter space, improve the stability of the feature extraction process, avoid problems such as gradient vanishing and gradient explosion, and ensure the stability and performance of model training.

[0147] The present invention adopts an autoencoder based on boundary smoothing to perform feature dimensionality reduction. During the dimensionality reduction process, the autoencoder smoothes the boundary effect, reduces the influence of outliers in the dimensionality reduction process, improves the robustness of the dimensionality reduction result, and can effectively retain the main features of the data, providing high-quality input for subsequent classification tasks.

[0148] Finally, the extreme learning machine algorithm based on stationarity constraints is used as the classifier. Through the stationarity constraint mechanism, the change range of the weights between the hidden layer and the output layer is limited, the overfitting phenomenon is reduced, and the classification accuracy and convergence are improved. In addition, through adaptive bias correction and multi-task learning framework, the adaptability and accuracy of the classifier under complex data are further improved.

[0149] In this embodiment, the collection of grid-connected data includes the following steps:

[0150] The power output data is obtained based on the sensors deployed in each renewable energy power generation unit; the meteorological data of the location area of ​​each renewable energy power generation unit is obtained through the meteorological station; the scheduling plan and operation status of each renewable energy power generation unit are obtained based on the scheduling platform.

[0151] Specifically, the data source used to train the new energy grid-connected flexibility resource prediction model mainly comes from the real-time operation data of the new energy grid-connected system, covering the real-time power output, grid-connected voltage, current, meteorological data (such as wind speed, light intensity, etc.) and dispatching instructions of multiple new energy power generation units including wind power, photovoltaics, energy storage systems, etc. Key information.

[0152] Furthermore, the grid-connected data collection methods include but are not limited to: deploying sensors at wind farms, photovoltaic power stations, energy storage facilities and other nodes to collect real-time data such as power output, voltage, current, etc.; obtaining meteorological data of each new energy power generation area through cooperative meteorological stations, such as wind speed, light intensity, temperature and other information; collecting information such as the scheduling plan, operating status and scheduling instructions of new energy power generation from the scheduling platform. The above-mentioned grid-connected data is transmitted to the cloud data center in real time through wireless communication protocols (such as Zigbee, LoRa, 5G, etc.) for unified storage and processing.

[0153] In one embodiment, the attributes of the grid-connected data include:

[0154] Ra is the timestamp, indicating the time when the grid-connected data is collected, in seconds (s);

[0155] Da is the wind speed, which indicates the wind speed during data collection, in meters per second (m / s);

[0156] Ea is the light intensity, which indicates the light intensity during the acquisition, in watts per square meter (W / m 2 );

[0157] Fa is the system output power, which indicates the power output of the power generation system during acquisition, in kilowatts (kW);

[0158] Ga is the system voltage, which indicates the voltage value of the system during acquisition, in volts (V);

[0159] Ha is the system current, which indicates the current value of the system during acquisition, in ampere (A);

[0160] Ia is the energy storage device status, indicating whether the energy storage device is turned on, and the value is 0 (off) or 1 (on);

[0161] Ja is the ambient temperature, which indicates the ambient temperature during the acquisition, in degrees Celsius (℃);

[0162] Ka is the device fault status, indicating whether there is a fault in the system, and its value is 0 (no fault) or 1 (fault);

[0163] La is the expected dispatch instruction, which indicates the expected power output instruction given by the dispatch system, in kilowatts (kW).

[0164] It should be noted that this embodiment is only intended to illustrate a format and type of grid-connected data of the present invention. In actual applications, the attributes of grid-connected data are usually more than 10 attributes, and the number of attributes of grid-connected data may reach dozens or even hundreds.

[0165] Furthermore, the collected grid-connected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeled categories include stable new energy grid-connected flexibility resources, unstable new energy grid-connected flexibility resources, and new energy grid-connected fault status, a total of 3 categories.

[0166] The present invention aims at the prediction task of new energy grid-connected flexibility resources. The grid-connected training data often have problems such as scarcity and uneven distribution. Consider how to generate richer and more diverse training samples to improve the prediction accuracy.

[0167] In this embodiment, the training method of the data expansion model includes the following steps:

[0168] Initialize the generator and discriminator of the generative adversarial network; the generator generates data samples under the constraint of entropy loss; the discriminator evaluates the authenticity of the data samples based on the self-similarity loss of fractal theory and guides the optimization of the generator; iterate the training until the preset stopping condition is met, and complete the training of the data augmentation model.

[0169] This embodiment adopts a generative adversarial network algorithm based on thermodynamic inspiration to generate samples, thereby realizing grid-connected data expansion. Thermodynamic inspiration means that the thermal equilibrium state in thermodynamics is adopted as the generator part of the generative adversarial network. Through the entropy in thermodynamics, the generator can more finely adjust the characteristic distribution of the grid-connected data during the grid-connected data generation process, so that the generated virtual samples have higher representativeness and accuracy in the grid-connected data space.

[0170] Among them, the generative adversarial network model consists of two parts: a generator and a discriminator. The generator generates new grid-connected data samples under the constraint of entropy loss; the discriminator adopts the design of traditional generative adversarial networks, and guides the optimization of the generator by evaluating the similarity between the generated grid-connected data and the real grid-connected data.

[0171] Specifically, the training process of the thermodynamically inspired generative adversarial network algorithm is as follows:

[0172] First, initialize the generator and discriminator of the generative adversarial network, assuming that the generator is G c , whose network structure includes L c layer, the discriminator is D c , whose network structure includes M cIn one embodiment, the parameters of the generator and the discriminator are randomly initialized, and the initialized values ​​obey a normal distribution with a mean of 0 and a variance of a unit matrix.

[0173] Furthermore, the goal of the generator of the generative adversarial network is to improve the quality of the generated samples by minimizing the change in entropy. The sample output by the generator is expressed as:

[0174] In the formula, The sample generated by the generator; Z c is the random noise input; θ c is the parameter of the generator, is the training parameter; G c () is the mapping function of the generator. The generator continuously optimizes the generated samples to make them more consistent with the distribution of target data in the feature space.

[0175] Different from the traditional generative adversarial network, the traditional generative adversarial network uses the generator loss function to constrain the iterative process of the generator to generate samples, which easily leads to the generated samples being relatively single and having little difference, and the diversity and accuracy of the generated grid-connected data are poor; in the training process of the generator, this embodiment optimizes the generated samples with the help of the entropy change inspired by thermodynamics. The entropy change in thermodynamics is an important indicator for describing the distribution of system states. The generator optimizes the generated data samples by minimizing the entropy change during the training process to improve the diversity and accuracy of the generated grid-connected data. The entropy change is defined as:

[0176]

[0177] Among them, k B is the Boltzmann constant; P c (i) is the probability distribution of the samples generated by the generator on the i-th feature dimension; na is the feature dimension of the generated samples; S c is the entropy of the generated sample; is the diversity loss of generated samples; preferably, k B Set to 2.

[0178] The diversity loss of generated data samples adopts the nonlinear Lagrange multiplier method. By penalizing and rewarding the diversity of generated samples, the generator can find the best balance between accuracy and diversity to avoid overfitting or sample homogeneity. The calculation method is expressed as:

[0179]

[0180] in, is the penalty coefficient of diversity constraint; σ c is a scale factor used to normalize the diversity of generated samples and is a training parameter; cTo control the weight parameter for diversity enhancement; preferably, Set to 0.3, γ c Set to 2.

[0181] Furthermore, the probability distribution is calculated based on the energy value of the feature, which is specifically expressed as:

[0182]

[0183] Among them, Z c is a normalization constant that ensures that the sum of all probabilities is 1; E c (i) is the energy value of the feature in the i-th dimension; T c is the temperature term, k B is the Boltzmann constant.

[0184] Furthermore, the complexity of features of different dimensions is represented by the energy value of the feature, and the calculation method is expressed as:

[0185]

[0186] in, is the eigenvalue of the jth dimension of the generated sample, X c，j is the eigenvalue of the jth dimension of the real sample; is the feature weight coefficient, which is a training parameter used to quantify the impact of each feature dimension on entropy change.

[0187] Furthermore, the task of the discriminator of the generative adversarial network is to distinguish between generated grid-connected data and real grid-connected data. On the basis of the adversarial loss used by the traditional discriminator, the discriminator used in this embodiment evaluates the authenticity of the generated grid-connected data by learning the fractal self-similarity of the sample. The discriminator compares the local features of the input grid-connected data with the global features and analyzes the fractal dimension of the grid-connected data. The total loss function of the discriminator is the sum of the adversarial loss and the fractal loss. By optimizing the total loss function of the discriminator, the authenticity of the generated data sample is evaluated and the fractal self-similarity characteristics of the grid-connected data are captured. The total loss function of the discriminator is the sum of the traditional adversarial loss and the fractal self-similarity loss. The calculation method is expressed as:

[0188]

[0189] Among them, λ c To control the balance coefficient of the discriminator loss;

[0190] The loss function of the discriminator in this embodiment includes two parts, the traditional adversarial loss and the self-similarity loss based on fractal theory. The calculation method of the traditional adversarial loss is expressed as:

[0191]

[0192] in, To combat losses; is the expectation; ~ means it obeys a specific distribution; X c Indicates real grid-connected data; D c () is the mapping function of the discriminator, which represents the probability that the input grid-connected data is the real grid-connected data; G c is the generator; P c This is the distribution of real grid-connected data.

[0193] Furthermore, the self-similarity loss based on fractal theory is used to measure the self-similarity of the generated grid-connected data. The diversity and quality of the generated samples are improved by measuring the self-similarity of the samples. At the same time, multi-scale analysis is used to improve the recognition ability of the discriminator at different scales. The calculation method of the fractal self-similarity loss is expressed as:

[0194]

[0195] Among them, X c,i represents the eigenvalue of the real grid-connected data in the i-th dimension, α c is the self-similarity coefficient; β c is the multi-scale analysis coefficient; M c is the number of scales in the fractal multi-scale analysis, indicating the self-similarity of the grid-connected data generated from multiple scale analyses, so that the discriminator can identify the fractal features of the grid-connected data at different levels of detail. Preferably, α c Set to 2, β c Set to 3.

[0196] Finally, the above steps are repeated until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0197] After the grid-connected data expansion model training is completed, the trained grid-connected data expansion model is used to increase the number of samples. In one embodiment, assuming that the original collected samples are 800, and the grid-connected data expansion model expands and generates 200 samples, the expanded grid-connected data set contains 1000 samples.

[0198] The present invention aims at the prior art of using neural networks for feature extraction. In some neural network structures, problems such as gradient vanishing, gradient exploding or falling into local optimal solutions may be encountered. A neural network algorithm based on dynamic adaptive oscillation is adopted as a feature extraction model. By simulating nonlinear oscillation behaviors in physical phenomena, local extreme values ​​can be effectively explored and utilized in high-dimensional parameter space, thereby achieving the purpose of optimizing the neural network.

[0199] In this embodiment, the training method of the feature extraction model includes the following steps:

[0200] Initialize the parameters of the neural network and the hyperparameters of the dynamic adaptive oscillation; adjust the oscillation frequency of each parameter based on the loss function of the neural network, and synchronously update the phase of the oscillation; adaptively adjust the amplitude according to the parameter update effect; use sine and cosine functions to simulate the oscillation behavior, and calculate the updated values ​​of the weights and biases of the neural network; repeat the above steps until the preset stop iteration condition is met to complete the training of the feature extraction model.

[0201] Specifically, Figure 2 As shown in the figure, the training process of the neural network algorithm based on dynamic adaptive oscillation is as follows:

[0202] First, the parameters of the neural network are initialized, and at the same time, the hyperparameters of the dynamic adaptive oscillation are initialized, including the initial phase of the oscillation and the initial amplitude of the oscillation. In one embodiment, the initialization method of the generator and the discriminator of the generative adversarial network is expressed as follows:

[0203]

[0204] φ 0p =0;

[0205] A 0p =41;

[0206] Among them, ~ indicates compliance with a specific distribution; It means the mean is 0 and the variance is Normal distribution of represents normal distribution; represents the initialization variance of the neural network; W 0p is the initial weight of the neural network, b 0p is the initial bias of the neural network; φ 0p is the initial phase of oscillation; A 0p is the initial amplitude of oscillation.

[0207] Furthermore, the oscillation frequency of each parameter is calculated according to the current loss function, and the oscillation frequency of each parameter is adjusted based on the local curvature estimation of the loss surface. The historical gradient information is used for adjustment, so that the parameter update is smoother and more efficient search can be performed in the parameter space. The adjustment method is expressed as:

[0208]

[0209] Among them, β p is the historical gradient weight factor, which controls the influence of the historical gradient on the current frequency adjustment; ω p is the oscillation frequency, is the oscillation frequency of the tth iteration; is the weight of the neural network at the t-1th iteration; is the weight of the neural network at the kth iteration; L p is the loss function of the neural network; is the sign of the partial derivative; ω0 is the fundamental oscillation frequency; α p is a hyperparameter that adjusts the oscillation responsiveness; preferably, α p Set to 5, β p Set to 0.3; the loss function of the neural network uses cross entropy loss, which is calculated by the preset Softmax on the output of the last layer of the neural network.

[0210] Furthermore, the update of each parameter is also affected by the phase difference, which is determined by the success rate of the previous update and the interaction strength between the parameters. The phase update depends on the effect of the previous parameter update, which is used to simulate the delayed effect of causality. The oscillating phase update method is expressed as:

[0211]

[0212] Among them, φ p is the phase of the oscillation, is the phase of the oscillation at the tth iteration; is the phase of the oscillation at the t-1th iteration; δ p is the phase adjustment factor; tanh() is the hyperbolic tangent function, which can limit the adjustment range of the phase to avoid excessive adjustment leading to instability of the algorithm; λ p is a hyperparameter that adjusts the phase sensitivity; preferably, δ p Set to 0.1, λ p Set to 0.3.

[0213] Furthermore, the amplitude is automatically adjusted according to the effect of parameter updates in the past few iterations. If the update of a certain parameter continues to reduce the loss, its amplitude is increased; otherwise, the amplitude is reduced. The adaptive adjustment method of the amplitude is expressed as:

[0214]

[0215] Among them, A p is the amplitude of the oscillation, is the amplitude of the oscillation at the tth iteration; is the amplitude of the oscillation at the t-1th iteration; γ p is the amplitude adjustment factor; preferably, γ p Set to 0.95.

[0216] Furthermore, the updated values ​​of the parameters of each neural network are calculated. Specifically, the updated values ​​of the weights and biases of the neural network are calculated based on the oscillation behavior. The sine and cosine functions are used to simulate the oscillation behavior, allowing the parameters to be periodically explored in the gradient direction of the loss function. The update method is expressed as:

[0217]

[0218] in, is the weight of the neural network at the tth iteration; is the bias of the neural network at the tth iteration; is the weight of the neural network at the t-1th iteration; The bias of the neural network at the t-1th iteration; t * is the current iteration number of the neural network.

[0219] Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0220] In the present invention, the grid-connected data after feature extraction is input into a feature dimensionality reduction model for training the feature dimensionality reduction model. This embodiment adopts an autoencoder based on boundary smoothing as the feature dimensionality reduction model. The autoencoder includes two parts, an encoder and a decoder. The encoder compresses the input data to a low-dimensional representation, and the decoder reconstructs the low-dimensional representation into a feature vector that is as similar as possible to the input data.

[0221] In this embodiment, the training method of the feature dimensionality reduction model includes the following steps:

[0222] Initialize the weights and bias parameters of the encoder in the autoencoder, as well as the weights and bias parameters of the decoder; train the autoencoder by optimizing the loss function, and update the parameters of the autoencoder synchronously with the optimization operation; use an automatic boundary smoothing strategy to handle outliers and boundary effects in the dimensionality reduction process; use a convex hull convergence strategy to optimize the training process of the autoencoder; iterate the training until the preset stop condition is met, and complete the training of the feature dimensionality reduction model.

[0223] See Figure 3 The training process of the boundary smoothing based autoencoder algorithm shown is as follows:

[0224] First, the parameters of the autoencoder are initialized. In one embodiment, the method of initializing the parameters of the autoencoder is expressed as:

[0225]

[0226] in, represents the weight of the encoder, Indicates the encoder bias; represents the weight of the decoder, represents the bias of the decoder; randn() represents the random function; d r,in is the dimension of the encoder input layer, d r,mid is the dimension of the middle layer of the encoder, d r,out is the dimension of the encoder output layer; zeros() is a function that generates a full 0 vector.

[0227] Furthermore, in the training process of the autoencoder, the loss function is optimized by continuous iteration, and the parameter update amount of each iteration is calculated as follows:

[0228]

[0229] in, is the weight parameter update amount of the encoder; is the encoder bias parameter update amount; is the update amount of the decoder weight parameters; is the update amount of the bias parameter of the decoder; is the learning rate of the autoencoder at the tth iteration;

[0230] Furthermore, the update method of the autoencoder parameters is expressed as:

[0231]

[0232] Among them, ← represents the parameter update operation.

[0233] Furthermore, in the training process of the autoencoder, the grid-connected data after feature extraction is processed by forward propagation, and then the autoencoder parameters are updated by the gradient descent method through error back propagation. In this process, the calculation method of the loss function of the autoencoder is expressed as:

[0234]

[0235]

[0236] Among them, L r is the loss function of the autoencoder; X r is the input data of the autoencoder, is the output reconstructed by the decoder, Sig() is the Sigmoid activation function; |||| is the L2 norm; is the output of the smoothed encoder; is the weight coefficient of the i-th dimension autoencoder; n rIt is the vector dimension corresponding to the samples of the current batch input to the autoencoder.

[0237] The weight coefficient of the autoencoder is used to adjust the contribution of each dimension feature in the error. These weights are dynamically allocated according to the importance of the feature. The calculation method is expressed as:

[0238]

[0239] in, is the characteristic variance of the input data of the i-th dimension autoencoder; is the feature of the input data of the i-th dimension autoencoder; ∈ re is a small constant to avoid division by zero; Var() is the variance function.

[0240] Furthermore, in order to deal with outliers and boundary effects in the dimensionality reduction process, an automatic boundary smoothing strategy is adopted to enhance the robustness of the model by dynamically adjusting the boundary value. Specifically, the output of the encoder is smoothed by a smoothing function. The calculation method is expressed as:

[0241]

[0242] Among them, τ r () is a smooth function, δ r is the smoothing strength parameter; θ r It is an automatically adjusted threshold parameter used to control the degree of smoothing and improve adaptability to the boundaries of input data.

[0243] The smoothing strength parameter is determined by an adaptive method based on the local density of the data, and the calculation method is expressed as:

[0244]

[0245] in, calculate In its neighborhood The kernel density estimation in can accurately process data in high-density and sparse areas, improving the robustness and adaptability of the overall model; represents the neighborhood; k r () is the kernel density estimation function; mean() is the average function; ∈ rea is a preset small constant. Preferably, ∈ rea Set to 0.001.

[0246] Furthermore, the convex hull convergence strategy is used to optimize the training process of the autoencoder. The network converges quickly by adjusting the learning step size and weight update strategy. The learning rate is dynamically adjusted based on the data distribution characteristics of the training process. The calculation method is expressed as:

[0247]

[0248] in, is the learning rate of the initial autoencoder; is the learning rate of the autoencoder at the tth iteration; γ r is the learning rate decay factor of the autoencoder, is the average diameter of the convex hull in the tth iteration, is the diameter of the initial convex hull. The learning rate decreases as the convex hull diameter decreases to prevent instability caused by too large a step size in the later stages of learning.

[0249] Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0250] In the present invention, the grid-connected data after dimensionality reduction is input into a classifier to train the classifier model. This embodiment adopts an extreme learning machine algorithm based on stationarity constraints as the classifier model. The extreme learning machine is a single-hidden-layer neural network structure. In a traditional extreme learning machine, the hidden layer weights of the network are usually randomly initialized and fixed, and the weights of the output layer are optimized by the least squares method. Although the calculation is efficient, when processing complex grid-connected data, especially when the grid-connected data set contains multiple features and high-dimensional information, the classification accuracy may be insufficient.

[0251] In order to solve the above problems, this implementation adopts a stationary constraint mechanism to enable the training process of the extreme learning machine to achieve a better balance between classification accuracy and convergence. The stationary constraint mainly adjusts the update strategy of the hidden layer weights and the output layer weights during the learning process of the extreme learning machine to reduce overfitting and improve classification performance. The training method of the classifier model includes the following steps:

[0252] Initialize the network structure of the extreme learning machine, including the input layer, the hidden layer and the output layer; perform stationarity constraint processing on the neural network by limiting the change of the connection weights between the hidden layer and the output layer; use an adaptive bias correction strategy to adjust the bias term of the extreme learning machine algorithm; use a multi-task learning framework to optimize the extreme learning machine algorithm, and comprehensively consider the losses of attribute prediction, sample distribution prediction and uncertainty prediction; update the parameters of the extreme learning machine algorithm according to the training error; iterate the training until the preset stop condition is met, and complete the training of the classifier model.

[0253] Specifically, the training process of the extreme learning machine algorithm based on stationarity constraints is as follows:

[0254] First, the network structure of the extreme learning machine is initialized, including an input layer, a hidden layer and an output layer. The input layer receives the grid-connected training data after feature dimension reduction, and the output layer outputs the classification results. Specifically, assume:

[0255] The input layer data feature of the extreme learning machine is X u =[x u1 , x u2 , ..., x um ], where m is the characteristic dimension of the input grid-connected data;

[0256] The number of hidden layer nodes of the extreme learning machine is h u ;

[0257] The target category label of the output layer of the extreme learning machine is y u =[y u1 ,y u2 , ..., y uk ], where k is the number of categories.

[0258] In the initialization process of the extreme learning machine network structure, the weight W of the hidden layer is u Random initialization, set the bias term b of the extreme learning machine u is a constant of 0, and the weight β of the output layer is optimized through training u ;

[0259] After initialization, the network is subjected to stationarity constraints. The stationarity constraint limits the change of the connection weights between the hidden layer and the output layer, making the weight update more stable during the training process and avoiding the drastic fluctuation of the network parameters. The stationarity function is used to constrain the update amplitude of the weights from the hidden layer to the output layer during each training process. The calculation method of the stationarity function is expressed as:

[0260]

[0261] Among them, W u,i is the weight between the i-th hidden node and the output layer of the extreme learning machine; f u () is the stationary function; ΔW u,i Represents the weight change of the i-th node; ||W u,i || 2 is the L2 norm of the i-th hidden node weight; |||| is the L2 norm; λ u is a stability regularization parameter that controls the magnitude of weight updates. Preferably, λ u Set to 0.2.

[0262] Furthermore, an adaptive bias correction strategy is used to adjust the bias term of the model. The bias correction term is calculated by monitoring the training error and verification error of the extreme learning machine, so that the model can adapt to different grid-connected data distributions and sample characteristics at different training stages. Specifically, in each round of training, the bias term is adjusted according to the distribution of the current grid-connected training data, and the bias correction term is calculated according to the current training error. The calculation method is expressed as:

[0263]

[0264] Among them, N u is the number of training samples input to the extreme learning machine; y u,i is the true label of the i-th training sample; is the prediction output of the extreme learning machine for the i-th training sample; η u is the learning rate of the extreme learning machine, which is used to control the correction speed; α u (t) is the dynamic adjustment factor of the t-th iteration; preferably, η u Set to 0.001.

[0265] Furthermore, the dynamic adjustment factor is dynamically adjusted as the training rounds change, and is specifically set as a decreasing function, which gradually decreases as the training progresses, thereby avoiding overcorrection; the calculation method of the dynamic adjustment factor is expressed as:

[0266]

[0267] Among them, β u is a hyperparameter that controls the rate at which the correction factor decreases.

[0268] Furthermore, a multi-task learning framework is used to optimize the model during the training process. The goal of each task is to optimize different aspects of the model through different loss functions. For example, in addition to the standard classification loss function, the model's training objectives are increased by using sample attribute prediction tasks, sample distribution prediction tasks, and uncertainty prediction tasks, thereby improving the model's adaptability to complex data. The total loss function of the extreme learning machine comprehensively considers the losses of attribute prediction, sample distribution prediction, and uncertainty prediction, and the calculation method is expressed as:

[0269] L u =L u,task1 +γ u ·L u,task2 +α u ·L u,task3 ;

[0270] L u,task1 is a loss function based on classification error, specifically using cross entropy loss, and the calculation method is expressed as:

[0271]

[0272] Among them, y ui,j and are the j-th class probability values ​​of the true label of the i-th sample and the label predicted by the extreme learning machine; kes is the number of categories;

[0273] L u,task2 is the loss function for sample distribution prediction, specifically using KL divergence loss, which represents the difference between the grid-connected training data distribution and the predicted distribution after feature dimensionality reduction. The calculation method is expressed as:

[0274]

[0275] Where DKL(||) represents the Kullback-Leibler divergence, P u (X ui )and are the i-th real data distribution and predicted data distribution respectively;

[0276] L u,task3 is the uncertainty prediction loss function, which uses entropy loss to measure the uncertainty of model prediction. The calculation method is expressed as:

[0277]

[0278] in, Represents the predicted entropy of the i-th sample, reflecting the model's prediction uncertainty for the sample

[0279] Furthermore, in each round of training, the parameters of the extreme learning machine model are updated according to the current training error, and the parameters of the extreme learning machine algorithm are updated using the gradient descent method. The calculation method is expressed as:

[0280]

[0281] Among them, W u (t) is the weight of the extreme learning machine at the tth iteration; b u (t) is the bias of the extreme learning machine at the tth iteration; W u (t+1) is the weight of the extreme learning machine at the t+1th iteration; b u (t+1) is the bias of the extreme learning machine at the t+1th iteration; Wu L u and bu L u are the gradients of the extreme learning machine's loss function with respect to weights and biases, respectively.

[0282] Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0283] Use the trained new energy grid-connected flexibility resource prediction model to process new samples, and implement the prediction task of new energy grid-connected flexibility resources in practical applications. In one embodiment, the collected original grid-connected data is input into the trained feature extraction and feature dimension reduction model for feature processing. Further, the processed features are input into the classifier model for classification to obtain the classification results. In this embodiment, the classification categories include: stable new energy grid-connected flexibility resources, unstable new energy grid-connected flexibility resources, and new energy grid-connected fault status, a total of 3 categories.

[0284] Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence, characterized in that: The following steps are involved: Collect grid-connected data based on real-time operation data of new energy grid-connected systems and perform manual annotation; A data expansion model is constructed using a generative adversarial network algorithm based on entropy loss constraints to expand the grid-connected data. A feature extraction model is constructed using a neural network algorithm based on dynamic adaptive oscillation to extract features from grid-connected data; The feature dimension reduction model is constructed by using an autoencoder based on boundary smoothing strategy to reduce the dimension of the extracted features; The extreme learning machine algorithm based on stationarity constraint is used to build a classifier model to classify the features after dimensionality reduction.

2. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 1 is characterized in that: The collection of the grid-connected data comprises the following steps: Obtain power output data based on sensors deployed in each renewable energy power generation unit; Obtain meteorological data of the location area of ​​each renewable energy power generation unit through the meteorological station; Obtain the dispatch plan and operating status of each renewable energy power generation unit based on the dispatch platform.

3. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 1 is characterized in that: The training method of the data expansion model comprises the following steps: Initialize the generator and discriminator of the generative adversarial network; The generator generates data samples under the entropy loss constraint; The discriminator evaluates the authenticity of the data sample based on the self-similarity loss of fractal theory and guides the optimization of the generator; The training is iterated until a preset stopping condition is met, thereby completing the training of the data expansion model.

4. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 1 is characterized in that: The training method of the feature extraction model comprises the following steps: Initialize the parameters of the neural network and the hyperparameters of the dynamic adaptive oscillation; Adjust the oscillation frequency of each parameter based on the loss function of the neural network and simultaneously update the phase of the oscillation; Adaptively adjust the amplitude according to the parameter update effect; Use sine and cosine functions to simulate oscillation behavior and calculate the updated values ​​of the neural network weights and biases; The above steps are iterated repeatedly until a preset stop iteration condition is met to complete the training of the feature extraction model.

5. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 1 is characterized in that: The training method of the feature dimensionality reduction model comprises the following steps: Initialize the weights and bias parameters of the encoder in the autoencoder, as well as the weights and bias parameters of the decoder; The autoencoder is trained by optimizing the loss function and the parameters of the autoencoder are updated synchronously with the optimization operation. Automatic boundary smoothing strategy is used to handle outliers and boundary effects in the dimensionality reduction process; Optimize the training process of the autoencoder using the convex hull convergence strategy; The training is iterated until a preset stopping condition is met, thereby completing the training of the feature dimensionality reduction model.

6. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 1, characterized in that: The training method of the classifier model comprises the following steps: Initialize the network structure of the extreme learning machine, including the input layer, hidden layer and output layer; By limiting the change of the connection weights between the hidden layer and the output layer, the neural network is constrained to be stable. Adopting adaptive bias correction strategy to adjust the bias term of extreme learning machine algorithm; A multi-task learning framework is used to optimize the extreme learning machine algorithm, and the losses of attribute prediction, sample distribution prediction, and uncertainty prediction are comprehensively considered; Update the parameters of the extreme learning machine algorithm based on the training error; The training is iterated until a preset stopping condition is met, thereby completing the training of the classifier model.

7. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 3 is characterized in that: The generator optimizes the generated data samples by minimizing the change in entropy, which is defined as: Among them, k B is the Boltzmann constant; P c (i) is the probability distribution of the samples generated by the generator on the i-th feature dimension; na is the feature dimension of the generated samples; S c is the entropy of the generated sample; Diversity loss for generating samples; The diversity loss of generated data samples uses the nonlinear Lagrange multiplier method, which is calculated as follows: in, is the penalty coefficient of diversity constraint; σ c is a scale factor used to normalize the diversity of generated samples and is a training parameter; c To control the weight parameter of diversity enhancement; The probability distribution is calculated based on the energy value of the feature, which is specifically expressed as: Among them, Z c is a normalization constant that ensures that the sum of all probabilities is 1; E c (i) is the energy value of the feature in the i-th dimension; T c is the temperature term, k B is the Boltzmann constant; The complexity of features of different dimensions is represented by the energy value of the feature, and the calculation method is expressed as: in, is the eigenvalue of the jth dimension of the generated sample, X c，j is the eigenvalue of the jth dimension of the real sample; is the feature weight coefficient, which is a training parameter used to quantify the impact of each feature dimension on entropy change.

8. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 3 is characterized in that: By optimizing the total loss function of the discriminator, the authenticity of the generated data samples is evaluated and the fractal self-similarity characteristics of the grid-connected data are captured. The total loss function of the discriminator is the sum of the traditional adversarial loss and the fractal self-similarity loss. The calculation method is expressed as: Among them, λ c To control the balance coefficient of the discriminator loss; The traditional way of calculating adversarial loss is expressed as: in, To combat losses; is the expectation; ~ means it obeys a specific distribution; X c Indicates real grid-connected data; D c () is the mapping function of the discriminator, which represents the probability that the input grid-connected data is the real grid-connected data; G c is the generator; P c is the distribution of real grid-connected data; The calculation method of fractal self-similarity loss is expressed as: Among them, X c,i represents the eigenvalue of the real grid-connected data in the i-th dimension, α c is the self-similarity coefficient; β c is the multi-scale analysis coefficient; M c is the number of scales in fractal multiscale analysis, indicating the self-similarity of grid-connected data generated from multiple scale analyses, enabling the discriminator to identify fractal features of grid-connected data at different levels of detail.

9. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 4 is characterized in that: The initialization method of the generator and discriminator of the generative adversarial network is expressed as: f 0p =0; A 0p ＝1； Among them, ~ indicates compliance with a specific distribution; It means the mean is 0 and the variance is Normal distribution of represents normal distribution; represents the initialization variance of the neural network; W 0p is the initial weight of the neural network, b 0p is the initial bias of the neural network; φ 0p is the initial phase of oscillation; A 0p is the initial amplitude of oscillation.

10. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 4, characterized in that: The oscillation frequency of each parameter is adjusted based on the local curvature estimation of the loss surface, and the historical gradient information is used for adjustment. The adjustment method is expressed as: Among them, β p is the historical gradient weight factor, which controls the influence of the historical gradient on the current frequency adjustment; ω p is the oscillation frequency, is the oscillation frequency of the tth iteration; is the weight of the neural network at the t-1th iteration; is the weight of the neural network at the kth iteration; L p is the loss function of the neural network; is the sign of the partial derivative; ω0 is the fundamental oscillation frequency; α p is a hyperparameter that adjusts the oscillation responsiveness; Update the phase and amplitude of the oscillation based on the parameter update effect: The phase update method of the oscillation is expressed as: Among them, φ p is the phase of the oscillation, is the phase of the oscillation at the tth iteration; is the phase of the oscillation at the t-1th iteration; δ p is the phase adjustment factor; tanh() is the hyperbolic tangent function, which can limit the adjustment range of the phase to avoid excessive adjustment leading to instability of the algorithm; λ p is a hyperparameter that adjusts phase sensitivity; The adaptive adjustment method of amplitude is expressed as: Among them, A p is the amplitude of the oscillation, is the amplitude of the oscillation at the tth iteration; is the amplitude of the oscillation at the t-1th iteration; γ p is the amplitude adjustment factor.

11. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 10, characterized in that: The updated values ​​of the weights and biases of the neural network are calculated based on the oscillation behavior. The update method is expressed as: in, is the weight of the neural network at the tth iteration; is the bias of the neural network at the tth iteration; is the weight of the neural network at the t-1th iteration; The bias of the neural network at the t-1th iteration; t * is the current iteration number of the neural network.

12. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 5, characterized in that: The way to initialize the autoencoder parameters is expressed as: in, represents the weight of the encoder, Indicates the encoder bias; represents the weight of the decoder, represents the bias of the decoder; randn() represents the random function; d r,in is the dimension of the encoder input layer, d r,mid is the dimension of the middle layer of the encoder, d r,out is the dimension of the encoder output layer; zeros() is a function that generates a full 0 vector.

13. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 12, characterized in that: During the training process of the autoencoder, the parameter update amount for each iteration is calculated as: in, is the weight parameter update amount of the encoder; is the encoder bias parameter update amount; is the update amount of the decoder weight parameters; is the update amount of the bias parameter of the decoder; is the learning rate of the autoencoder at the tth iteration; The update method of the autoencoder parameters is expressed as: Among them, ← represents the parameter update operation; The calculation method of the loss function of the autoencoder is expressed as: Among them, L r is the loss function of the autoencoder; X r is the input data of the autoencoder, is the output reconstructed by the decoder, Sig() is the Sigmoid activation function; |||| is the L2 norm; is the output of the smoothed encoder; is the weight coefficient of the i-th dimension autoencoder; n r is the vector dimension corresponding to the samples of the current batch input to the autoencoder; The calculation method of the weight coefficient of the autoencoder is expressed as: in, is the characteristic variance of the input data of the i-th dimension autoencoder; is the feature of the input data of the i-th dimension autoencoder; ∈ re is a small constant to avoid division by zero; Var() is the variance function.

14. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 13, characterized in that: The output of the encoder is smoothed by a smoothing function, and the calculation method is expressed as: Among them, τ r () is a smooth function, δ r is the smoothing strength parameter; θ r is an automatically adjusted threshold parameter used to control the degree of smoothing; The smoothing strength parameter is determined by an adaptive method based on the local density of the data, and the calculation method is expressed as: in, calculate In its neighborhood The kernel density estimation in can accurately process data in high-density and sparse areas, improving the robustness and adaptability of the overall model; represents the neighborhood; k r () is the kernel density estimation function; mean() is the average function; ∈ rea is a preset small constant.

15. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 5, characterized in that: The learning rate is dynamically adjusted based on the data distribution characteristics of the training process. The calculation method is expressed as: in, is the learning rate of the initial autoencoder; is the learning rate of the autoencoder at the tth iteration; γ r is the learning rate decay factor of the autoencoder, is the average diameter of the convex hull in the tth iteration, is the diameter of the initial convex hull.

16. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 6, characterized in that: In the initialization process of the extreme learning machine network structure, the weights of the hidden layer are randomly initialized, the bias term of the extreme learning machine is set to a constant of 0, and the weights of the output layer are optimized through training; After initialization, the stationary function is used to constrain the update amplitude of the weights from the hidden layer to the output layer. The calculation method of the stationary function is expressed as: Among them, W u,i is the weight between the i-th hidden node and the output layer of the extreme learning machine; f u () is the stationary function; ΔW u,i Represents the weight change of the i-th node; ||W u,i || 2 is the L2 norm of the i-th hidden node weight; |||| is the L2 norm; λ u is the stability regularization parameter, which controls the magnitude of weight updates.

17. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 16, characterized in that: The bias correction term is calculated by monitoring the training error and validation error of the extreme learning machine. The calculation method is expressed as: Among them, N u is the number of training samples input to the extreme learning machine; y u,i is the true label of the i-th training sample; is the prediction output of the extreme learning machine for the i-th training sample; η u is the learning rate of the extreme learning machine, which is used to control the correction speed; α u (t) is the dynamic adjustment factor of the tth iteration; The calculation method of the dynamic adjustment factor is expressed as: Among them, β u is a hyperparameter that controls the rate at which the correction factor decreases.

18. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 17, characterized in that: The total loss function of the extreme learning machine comprehensively considers the losses of attribute prediction, sample distribution prediction, and uncertainty prediction, and is calculated as follows: L u =L u,task1 +g u ·L u,task2 +a u ·L u,task3 ; L u,task1 is a loss function based on classification error, specifically using cross entropy loss, and the calculation method is expressed as: Among them, yui,j and are the j-th class probability values ​​of the true label of the i-th sample and the label predicted by the extreme learning machine; kes is the number of categories; L u,task2 It is the loss function for sample distribution prediction, specifically using KL divergence loss, and the calculation method is expressed as: Where DKL(||) represents the Kullback-Leibler divergence, P u (X ui )and are the i-th real data distribution and predicted data distribution respectively; L u,task3 is the uncertainty prediction loss function, which uses entropy loss to measure the uncertainty of model prediction. The calculation method is expressed as: in, It represents the prediction entropy of the i-th sample, reflecting the prediction uncertainty of the model for this sample.

19. The method for constructing a new energy grid-connected flexibility resource prediction model based on artificial intelligence according to claim 18, characterized in that: The gradient descent method is used to update the parameters of the extreme learning machine algorithm. The calculation method is expressed as: Among them, W u (t) is the weight of the extreme learning machine at the tth iteration; b u (t) is the bias of the extreme learning machine at the tth iteration; W u (t+1) is the weight of the extreme learning machine at the t+1th iteration; b u (t+1) is the bias of the extreme learning machine at the t+1th iteration; and are the gradients of the extreme learning machine's loss function with respect to weights and biases, respectively.

Images