A photovoltaic load prediction method and device based on artificial intelligence

Abstract

The present application belongs to the field of computer technology, and discloses a photovoltaic load prediction method and device based on artificial intelligence, the method comprising: obtaining historical photovoltaic data of a photovoltaic power station; expanding the historical photovoltaic data based on a data homeomorphic generative adversarial network to obtain a training data set; using the training data set to train a fully connected neural network based on exploratory gradient descent to obtain a feature extraction model; obtaining photovoltaic data to be tested, and inputting the feature extraction model to obtain target feature data; inputting the target feature data into a trained feature dimension reduction model to obtain target dimension reduction data; inputting the target dimension reduction data into a trained classifier model to obtain a load prediction interval. The present application can solve the problem of insufficient training data and improve the generalization ability, stability and model performance of the feature extraction model.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a photovoltaic load prediction method and device based on artificial intelligence. Background Art

[0002] As the world relies more and more on renewable energy, the stability and reliability of photovoltaic power generation as a clean and sustainable form of energy are crucial to the balance of the power grid and the efficient use of energy. However, photovoltaic power generation systems are affected by many factors, such as weather changes, seasonal changes, equipment performance, etc., which lead to significant volatility and uncertainty in photovoltaic power generation. With the development of artificial intelligence technology, especially the remarkable achievements of deep learning in image recognition, natural language processing and other fields, its application potential in photovoltaic power generation prediction is gradually emerging.

[0003] However, existing deep learning models can automatically extract features and process complex nonlinear relationships, but they require a large amount of high-quality training data, resulting in inefficiency and high cost in data collection and labeling. In addition, problems such as overfitting, gradient disappearance or explosion are prone to occur during the training process of deep learning models, affecting the generalization ability and prediction accuracy of the model. Summary of the invention

[0004] The present application provides a photovoltaic load forecasting method and device based on artificial intelligence, which can solve the problem of insufficient training data and improve the generalization ability, stability and model performance of the feature extraction model.

[0005] In a first aspect, an embodiment of the present application provides a photovoltaic load prediction method based on artificial intelligence, comprising:

[0006] Obtain historical photovoltaic data of photovoltaic power plants;

[0007] The historical photovoltaic data is expanded based on the data homeomorphism generative adversarial network to obtain the training data set;

[0008] The training data set is used to train a fully connected neural network based on exploratory gradient descent to obtain a feature extraction model;

[0009] Obtain the photovoltaic data to be tested, and input it into the feature extraction model to obtain the target feature data;

[0010] Input the target feature data into the trained feature dimension reduction model to obtain the target dimension reduction data;

[0011] The target dimension reduction data is input into the trained classifier model to obtain the load forecast interval.

[0012] Furthermore, the training process of the generative adversarial network based on data homeomorphism includes:

[0013] Initialize the parameters of the generator and discriminator of the generative adversarial network;

[0014] Generation step: let the generator extract the periodic trend characteristics of the real photovoltaic data based on the trend extraction function and Fourier transform, and generate a noise vector based on the periodic trend characteristics and random noise;

[0015] The noise vector and the corresponding true PV data are input into the discriminator;

[0016] Return to the generation step until the first preset number of times is reached to obtain a generative adversarial network based on data homeomorphism.

[0017] Furthermore, the above-mentioned data homeomorphism-based generative adversarial network expands the historical photovoltaic data to obtain a training data set, including:

[0018] Input historical photovoltaic data into the generative adversarial network to obtain generated data and corresponding discrimination results;

[0019] If the judgment result is true, the homeomorphic verification score of the generated data is calculated;

[0020] If the homeomorphism verification score is greater than or equal to the preset verification threshold, the generated data is put into the training dataset.

[0021] Furthermore, the above-mentioned training data set is used to train a fully connected neural network based on exploratory gradient descent to obtain a feature extraction model, including:

[0022] Initialize the network parameters of the fully connected neural network;

[0023] Input step: input the training data set into the fully connected neural network, and make each layer of the fully connected neural network undergo linear transformation and nonlinear activation function processing;

[0024] An exploratory gradient descent update strategy is used to predict the update parameters of a fully connected neural network.

[0025] Return to the input step until the second preset number is reached to obtain a feature extraction model.

[0026] Furthermore, the fully connected neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer and an output layer connected in sequence; the number of neurons in the first hidden layer, the second hidden layer, the third hidden layer and the fourth hidden layer increases gradually; the number of neurons in the output layer is 512.

[0027] Furthermore, the training process of the feature dimensionality reduction model includes:

[0028] The training data set is input into the feature extraction model to obtain photovoltaic feature data;

[0029] Initialize the parameters of the autoencoder; the autoencoder includes an encoder and a decoder;

[0030] Encoding step: inputting the photovoltaic characteristic data into an encoder of the autoencoder to obtain first characteristic data;

[0031] Inputting the first feature data into a decoder of the autoencoder to obtain second feature data;

[0032] Calculating a reverse reconstruction loss value based on the first feature data and the second feature data;

[0033] Update the weights and biases of the encoder and decoder according to the inverse reconstruction loss value;

[0034] Return to the encoding step until the third preset number of times is reached to obtain a feature dimensionality reduction model.

[0035] Furthermore, the training process of the classifier model includes:

[0036] Inputting photovoltaic feature data into a feature dimension reduction model to obtain dimension-reduced feature data;

[0037] Initialize the weight parameters and bias parameters of the fractional-order neural network;

[0038] Classification step: input the dimension-reduced feature data into the fractional-order neural network to obtain the classification result;

[0039] Calculate the loss error between the classification result and the corresponding labeled interval of the reduced dimension feature data;

[0040] Calculate the update gradient of weight parameters and bias parameters according to the back-propagation algorithm and loss error;

[0041] Adjust the learning rate of the fractional neural network based on each update gradient;

[0042] Return to the classification step until the fourth preset number of times is reached to obtain a classifier model.

[0043] In a second aspect, an embodiment of the present application provides a photovoltaic load prediction device based on artificial intelligence, comprising:

[0044] The first acquisition module is used to acquire historical photovoltaic data of the photovoltaic power station;

[0045] An expansion module is used to expand historical photovoltaic data based on a data homeomorphism generative adversarial network to obtain a training data set;

[0046] A feature extraction training module is used to train a fully connected neural network based on exploratory gradient descent using a training data set to obtain a feature extraction model;

[0047] The second acquisition module is used to acquire the photovoltaic data to be tested and input the feature extraction model to obtain the target feature data;

[0048] A dimension reduction module is used to input the target feature data into the trained feature dimension reduction model to obtain the target dimension reduction data;

[0049] The prediction module is used to input the target dimension reduction data into the trained classifier model to obtain the load prediction interval.

[0050] In a third aspect, an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a photovoltaic load forecasting method based on artificial intelligence as described in any of the above embodiments are performed.

[0051] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a photovoltaic load forecasting method based on artificial intelligence as in any of the above embodiments.

[0052] In summary, compared with the prior art, the technical solution provided in the embodiment of the present application has at least the following beneficial effects:

[0053] An artificial intelligence-based photovoltaic load forecasting method provided in an embodiment of the present application, firstly, expands the acquired historical photovoltaic data through a generative adversarial network based on data homeomorphism, while ensuring that the expanded data is consistent with the real photovoltaic data in statistical characteristics and physical sense, solves the problem of insufficient training data and improves the generalization ability of the feature extraction model; secondly, the feature extraction model of the present application is trained by a fully connected neural network and an exploratory gradient descent method, so that it can converge to a better local minimum faster during the training process, solves the problems of gradient vanishing, gradient exploding and falling into the local optimum that occur in the traditional neural network training process, and improves the stability of feature extraction and model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A flowchart of a photovoltaic load forecasting method based on artificial intelligence is provided as an exemplary embodiment of the present application.

[0055] Figure 2 A structural diagram of a photovoltaic load prediction device based on artificial intelligence is provided as an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0056] The technical solutions 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, rather than all of the embodiments.

[0057] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0058] See also Figure 1 , the embodiment of the present application provides a photovoltaic load prediction method based on artificial intelligence, including:

[0059] Step S11, obtaining historical photovoltaic data of the photovoltaic power station.

[0060] Among them, historical photovoltaic data includes historical power generation data, historical meteorological data and historical equipment operation data. The collected data is acquired in real time through sensors and automation systems and stored in a structured JSON format;

[0061] Specifically, the attributes of historical PV data may include:

[0062] P_a is the instantaneous power generation (kW), T_a is the ambient temperature (℃), I_a is the light intensity (lux), V_a is the wind speed (m / s), H_a is the relative humidity (%), C_a is the equipment status code (discrete value), D_a is the power generation efficiency (%), S_a is the season identifier (discrete value), R_a is the rainfall (mm), and L_a is the load demand (kW).

[0063] It should be noted that the above data attributes are only an example for illustrating the present embodiment. In practical applications, the attributes of data are usually more than 10 attributes, and may even reach dozens or even hundreds.

[0064] Furthermore, the corresponding load labeling of the collected historical photovoltaic data can be achieved by manual input; specifically, the labeled categories include: "within 10KW range", "10KW-20KW range", "20KW-30KW range", "30KW-40KW range", "40KW-50KW range", "50KW and above range", a total of 6 categories.

[0065] Step S12: Expand the historical photovoltaic data based on the data homeomorphism generative adversarial network to obtain a training data set.

[0066] It is understandable that in this application, the acquisition, labeling and preprocessing of the training data set are time-consuming and labor-intensive, and insufficient training data can easily lead to poor model generalization ability and affect the accuracy of the model. Therefore, this application uses a generative adversarial network algorithm based on data homeomorphism to generate samples, thereby achieving the expansion of photovoltaic data. Among them, the generative adversarial network algorithm based on data homeomorphism includes two components: a generator and a discriminator. The task of the generator is to generate high-quality generated data based on the input noise and a small amount of real data, and the discriminator attempts to distinguish between generated data and real data.

[0067] Step S13, using the training data set to train a fully connected neural network based on exploratory gradient descent to obtain a feature extraction model. In this application, a 6-layer fully connected neural network is used for feature extraction, and the fully connected neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer and an output layer connected in sequence; the number of neurons in the first hidden layer, the second hidden layer, the third hidden layer and the fourth hidden layer increases gradually.

[0068] Specifically, the first hidden layer has 128 neurons for preliminary feature extraction; the second hidden layer has 256 neurons for further abstraction of features; the third hidden layer has 512 neurons for further deepening feature extraction; the fourth hidden layer has 1024 neurons for further deepening feature extraction. The output layer has 512 neurons.

[0069] Step S14, obtaining photovoltaic data to be measured and inputting it into a feature extraction model to obtain target feature data.

[0070] Step S15, inputting the target feature data into the trained feature dimension reduction model to obtain the target dimension reduction data.

[0071] Step S16, input the target dimension reduction data into the trained classifier model to obtain the load forecast interval.

[0072] Finally, the load forecast interval is the same as the annotation content of the historical photovoltaic data, including "within 10KW interval", "10KW-20KW interval", "20KW-30KW interval", "30KW-40KW interval", "40KW-50KW interval", and "50KW and above interval", a total of 6 possible load forecast intervals.

[0073] The above embodiment provides a photovoltaic load prediction method based on artificial intelligence. First, the acquired historical photovoltaic data is expanded through a generative adversarial network based on data homeomorphism. While ensuring that the expanded data is consistent with the real photovoltaic data in statistical characteristics and physical sense, the problem of insufficient training data is solved and the generalization ability of the feature extraction model is improved. Secondly, the feature extraction model of the present application is trained by a fully connected neural network and an exploratory gradient descent method, so that it can converge to a better local minimum faster during the training process, solves the problems of gradient vanishing, gradient exploding and falling into the local optimum that occur in the traditional neural network training process, improves the stability of feature extraction and model performance, and thus improves the prediction accuracy and stability of the photovoltaic load interval.

[0074] In some embodiments, the training process of the data homeomorphism-based generative adversarial network includes:

[0075] Step S01, initializing the parameters of the generator and discriminator of the generative adversarial network.

[0076] Specifically, let the generator be , the discriminator is , the weight of the generator is , the weight of the discriminator is , the bias of the generator is , the bias of the discriminator is , the initialization method can be expressed as:

[0077]

[0078]

[0079] in, It means that it obeys a specific distribution. represents normal distribution; Represents the initialization standard deviation, which can be set to 0.01.

[0080] Step S02, generation step: let the generator extract the periodic trend characteristics of the real photovoltaic data based on the trend extraction function and Fourier transform, and generate a noise vector based on the periodic trend characteristics and random noise. It is worth noting that the real photovoltaic data can directly use historical photovoltaic data, and the historical photovoltaic data can be expanded after the training is completed.

[0081] In the generator, by analyzing the time series characteristics of real photovoltaic data, key periodic trend features are extracted to guide the generation of noise vectors. Specifically, the extraction method is expressed as:

[0082]

[0083] in, It is a periodic trend characteristic; is the real photovoltaic data; and To adjust the weights of Fourier transform and trend analysis; represents the fast Fourier transform, Represents the trend extraction function.

[0084] Furthermore, the fast Fourier transform is used to analyze the periodic characteristics in the data, and the calculation method is expressed as:

[0085]

[0086] In the formula, is the number of data points, is an imaginary unit; For the Real photovoltaic data.

[0087] Furthermore, the trend extraction function uses a linear regression method to determine the long-term trend of the data. The calculation method is expressed as:

[0088]

[0089] in, Extract function weights for trends, is the time variable of the real photovoltaic data, and are the slope and intercept of the linear trend line obtained by the least squares method. Preferably, Can be set to 0.3.

[0090] In one embodiment, the calculation method of the slope and intercept of the linear trend line obtained by the least squares method is expressed as:

[0091]

[0092]

[0093] in, for Real photovoltaic data at all times.

[0094] Furthermore, based on the homeomorphic characteristics of the data and combined with external environmental factors such as the influence of meteorological conditions, a noise vector is designed to simulate photovoltaic power generation data under different conditions. The noise vector generation method is expressed as follows:

[0095]

[0096] in, is the noise vector; Represents the random noise added to ensure data diversity, is the identity matrix; It is a dynamic adjustment factor based on the variability of environmental data.

[0097] Specifically, dynamic regulatory factors based on the variability of environmental data It allows the noise vector to be dynamically adjusted according to the real-time environmental data, thereby improving the adaptability of the generated data and the accuracy of the prediction model. The calculation method is expressed as:

[0098]

[0099] in, predefined variability thresholds for environmental data; For the Real-time measurements of environmental indicators; For the Historical averages of environmental indicators; is the total number of categories of environmental indicators.

[0100] Furthermore, in order to make the structured noise more effectively reflect the actual characteristics of photovoltaic data, the weight of the trend extraction function can be adaptively adjusted by considering the seasonal changes of the data. The adjustment method is expressed as:

[0101]

[0102] in, is to adjust the intensity, Represents the total number of days in a year. is the current season day of the year.

[0103] Step S03: input the noise vector and the corresponding real photovoltaic data into the discriminator.

[0104] In the adversarial training phase, the generator and the discriminator are trained alternately. The generator tries to produce more and more realistic data, and the discriminator strives to improve its ability to distinguish between true and false. After each training, the parameters of the generator and the discriminator are adjusted to optimize the overall performance. The adversarial training iteratively updates the parameters by minimizing the loss function of the generator and the discriminator, which is expressed as:

[0105]

[0106] in, is the loss function; express expectations; represents the discriminator function; is the input of the discriminator, is the distribution of the input of the discriminator; represents a generator function, is the input of the generator, is the distribution of the input to the generator; and Represent the distribution of real data and noise data respectively.

[0107] Step S04, returning to the generation step until a first preset number of times is reached to obtain a generative adversarial network based on data homeomorphism.

[0108] The first preset number of times may be set to 1000 times.

[0109] The above embodiment optimizes the structure of the generator and the discriminator. In the generator, a structured noise compensation mechanism is used to adjust the noise input by analyzing the periodic trend characteristics in the real data, so as to better simulate the seasonal and daily variation characteristics of the photovoltaic data.

[0110] In some embodiments, the above-mentioned data homeomorphism-based generative adversarial network expands historical photovoltaic data to obtain a training data set, including:

[0111] Step S121, inputting historical photovoltaic data into a generative adversarial network to obtain generated data and corresponding discrimination results.

[0112] Step S122: If the judgment result is true, the homeomorphism verification score of the generated data is calculated.

[0113] Specifically, the calculation method of the homeomorphism verification score is expressed as:

[0114]

[0115] in, is the homeomorphic validation score; To generate the characteristics of the data; is the L2 norm; and are the weights of characteristic differences and maximum mean differences, respectively; Represents the currently generated training dataset; Represents the maximum mean difference function, which is used to measure the statistical distance between two sets of data.

[0116] Step S123: if the homeomorphism verification score is greater than or equal to the preset verification threshold, the generated data is put into the training data set.

[0117] In addition to the generated data output by the generative adversarial network based on historical photovoltaic data, the historical photovoltaic data itself will also be put into the training data set to train the fully connected neural network together with the generated data.

[0118] Specifically, the homeomorphic verification of the above-mentioned embodiment is also performed during the training phase, that is, after each training cycle, a set of preset homeomorphic verification rules are used to evaluate the quality of the generated data; the homeomorphic verification method of the present application can ensure that the generated data is not only statistically indistinguishable from the real photovoltaic data, but also physically consistent.

[0119] In some embodiments, the above-mentioned training data set is used to train a fully connected neural network based on exploratory gradient descent to obtain a feature extraction model, including:

[0120] Step S131, initializing the network parameters of the fully connected neural network.

[0121] Specifically, initialize the parameters of the fully connected neural network, including the weights and bias parameters of the fully connected neural network:

[0122]

[0123]

[0124] in, is the weight matrix of the i-th layer of the fully connected neural network; is the bias vector of the i-th layer; is the scaling factor of the neural network weights, is the scaling factor of the bias of the neural network; represents a normal distribution with a mean of 0 and a standard deviation of 1; is a tiny noise. Preferably, is a noise vector whose elements are all 0.001.

[0125] Step S132, input step: input the training data set into the fully connected neural network, and make each layer of the fully connected neural network perform linear transformation and nonlinear activation function processing.

[0126] Specifically, the training data set passes through each layer of the fully connected neural network, and each layer is processed by linear transformation and nonlinear activation function to finally obtain the feature representation of the output layer. The output calculation method of each layer is:

[0127]

[0128] in, represents the first The activation output of the layer; The fully connected neural network The output of the layer; is the adaptive ReLU activation function.

[0129] Suppose the input of the adaptive ReLU activation function is , the calculation method of the adaptive ReLU activation function is expressed as:

[0130]

[0131] in, and are the adaptive coefficients of the positive and negative parts respectively, To obtain the maximum value function, is the minimum value function. Preferably, Set to 0.3, Set to 0.7.

[0132] Step S133, using an exploratory gradient descent update strategy to predict update parameters of the fully connected neural network.

[0133] Specifically, this application adopts an exploratory gradient descent update strategy. Before each update of the fully connected neural network parameters, the possible gradient changes are predicted, and the step size and direction of the gradient update are adjusted according to the prediction results to optimize the parameter adjustment during the training process to avoid falling into the local optimum too early. The calculation method is expressed as:

[0134]

[0135]

[0136] in, and They are the first Layer weights and The update amount of the layer bias; is the learning rate of the neural network; is the loss function with respect to the parameter The gradient of is the loss function with respect to the parameter The gradient of is the loss function of the neural network; is the exploration intensity coefficient; is the prediction function of the gradient.

[0137] Preferably, Set to 0.03, Cross entropy loss is used.

[0138] The exploratory gradient prediction function uses historical gradient data for linear regression prediction. The calculation method is expressed as:

[0139]

[0140] in, Indicated in The loss function of the training is about the parameters The gradient of is the attenuation factor; To explore the strength coefficient; is the total number of training times. Preferably, Set to 0.98, Set to 2.

[0141] parameter The gradient calculation process and of consistency.

[0142] Step S134, return to the input step until a second preset number of times is reached to obtain a feature extraction model.

[0143] The second preset number of times may be set to 1000 times.

[0144] The exploratory gradient descent method adopted in the above embodiment not only considers the current gradient information, but also adopts a forward exploration mechanism to predict the possible change trend of the gradient before the gradient is updated, and adjust the learning step size and direction to achieve faster convergence speed and better local minimum, thereby enhancing the generalization ability of the feature.

[0145] In some embodiments, the training process of the feature dimensionality reduction model includes:

[0146] Step S21, inputting the training data set into the feature extraction model to obtain photovoltaic feature data.

[0147] Step S22, initializing the parameters of the autoencoder; the autoencoder includes an encoder and a decoder.

[0148] The photovoltaic feature data after feature extraction is input into the feature dimensionality reduction model for training the feature dimensionality reduction model. This application adopts an autoencoder algorithm based on inverse reconstruction as the feature dimensionality reduction model. The autoencoder includes two parts: an encoder and a decoder.

[0149] Specifically, initialize the parameters of the autoencoder, including the weights and bias parameters of the autoencoder:

[0150]

[0151]

[0152] in, is the initial weight matrix of the autoencoder; is the initial bias of the autoencoder; and Represents the number of neurons in the input and output layers respectively; Indicated in A uniformly distributed random number in the interval.

[0153] Step S23, encoding step: inputting the photovoltaic characteristic data into the encoder of the autoencoder to obtain first characteristic data.

[0154] Step S24, inputting the first feature data into the decoder of the autoencoder to obtain second feature data.

[0155] Specifically, the photovoltaic characteristic data is converted into first characteristic data in a low-dimensional space through an encoder, and the first characteristic data is reconstructed into second characteristic data with the same dimension as the original data through a decoder.

[0156] During the forward propagation process, the encoder output is calculated as:

[0157]

[0158] The reconstructed output of the decoder is:

[0159]

[0160] is the first characteristic data; is the Sigmoid activation function; and are the weights of the encoder and decoder respectively; and are the biases of the encoder and decoder respectively; is the photovoltaic characteristic data; It is the second feature data.

[0161] Step S25, calculating a reverse reconstruction loss value based on the first feature data and the second feature data.

[0162] Specifically, the inverse reconstruction loss value of the autoencoder is calculated, and the calculation method of the inverse reconstruction loss value is defined as:

[0163]

[0164] in, is the inverse reconstruction loss function; is the number of samples of the photovoltaic characteristic data input in the current batch; represents the L2 norm; is the regularization parameter; is the KL divergence.

[0165] KL divergence is a measure of the first characteristic data distribution The second characteristic data distribution The difference between them is used to ensure that the key features of the data are not lost during the dimensionality reduction process. The calculation method is expressed as:

[0166]

[0167] in, and The first characteristic data and the second characteristic data are respectively The probability distribution of dimensions, is the feature dimension of the first feature data before dimensionality reduction.

[0168] Step S26, updating the weights and biases of the encoder and decoder according to the inverse reconstruction loss value.

[0169] Specifically, according to the calculated inverse reconstruction loss value, the gradient descent algorithm is used to update the weights and biases of the encoder and decoder. The update method is expressed as:

[0170]

[0171]

[0172]

[0173]

[0174] in, is the learning rate; is the updated encoder weight; is the updated encoder bias; is the updated decoder weight; is the updated decoder bias.

[0175] Further, For example, the calculation method is expressed as:

[0176]

[0177] in, represents the partial derivative of the inverse reconstruction loss function with respect to the second feature data, is the partial derivative during decoding, is the partial derivative during encoding.

[0178] Step S27, return to the encoding step until a third preset number of times is reached to obtain a feature dimensionality reduction model.

[0179] Specifically, the autoencoder that has been trained for a third preset number of times is used as a feature dimensionality reduction model.

[0180] In order to enhance the adaptability of the feature dimensionality reduction of the model, the above embodiment adopts an inverse reconstruction strategy to process the high-dimensional data after input feature extraction, so as to minimize the dependency between different features, thereby improving the encoding quality and decoding accuracy.

[0181] In some embodiments, the training process of the classifier model includes:

[0182] Step S31, inputting photovoltaic feature data into a feature dimension reduction model to obtain dimension-reduced feature data.

[0183] Step S32, initializing the weight parameters and bias parameters of the fractional-order neural network.

[0184] Specifically, all weights and bias parameters in the fractional-order neural network are randomly initialized, and the weights of the fractional-order neural network model are set to , the bias of the fractional-order neural network is , the initialization method is expressed as:

[0185]

[0186]

[0187] In the formula, is the initial weight matrix of the fractional-order neural network, is the initial bias vector of the fractional-order neural network, is the initialization standard deviation of the fractional-order neural network, Indicates the generation of normally distributed random numbers. and are the dimensions of the input and output layers of the fractional-order neural network, respectively.

[0188] Step S33, classification step: input the dimension-reduced feature data into a fractional-order neural network to obtain a classification result.

[0189] Specifically, the reduced-dimensional feature data after feature dimensionality reduction is input into the fractional-order neural network, and the output result is calculated through each layer. In each layer, the reduced-dimensional feature data is multiplied by the weight matrix, and a bias is added, and then an activation function is used to adopt nonlinearity so that the neural network can learn more complex data patterns. Specifically, the reduced-dimensional feature data input into the fractional-order neural network is , the data is propagated through the network layer to calculate the classification results. The calculation process is expressed as:

[0190]

[0191]

[0192]

[0193] in, is the fractional order neural network The linear transformation result of the layer, is the result after passing the activation function, is the final output of the fractional-order neural network, is the activation function improved by fractional calculus, is the number of network layers.

[0194] Furthermore, the calculation method of the activation function improved by fractional calculus is expressed as:

[0195]

[0196] in, is the fractional derivative calculation function, is the order of the fraction, is the coefficient that adjusts the fractional order effect, is the Sigmoid activation function.

[0197] Step S34, calculating the loss error between the classification result and the corresponding labeled interval of the reduced dimension feature data.

[0198] Specifically, the loss function is combined with fuzzy logic rules to evaluate the error between the classification result output by the model and the true labeled interval. Specifically, the loss function calculation method of the fractional-order neural network is expressed as:

[0199]

[0200] In the formula, is the number of samples of the dimension-reduced feature data, is the number of categories, For the The sample A real annotation, is the fractional order neural network for the The predicted probability of samples; is the temperature parameter, which controls the strength of the fuzzy logic; is a normalization index used to emphasize the differences in predicted probabilities, making the model more sensitive to the correct class.

[0201] Preferably, Set to 0.1, Set to 0.3.

[0202] Step S35, calculating the update gradients of the weight parameters and bias parameters according to the back propagation algorithm and the loss error.

[0203] Specifically, the calculation formula for the parameter update gradient of the fractional-order neural network is expressed as:

[0204]

[0205]

[0206]

[0207]

[0208]

[0209]

[0210] in, is the loss function About parameters The gradient of is the loss function About parameters The gradient of is the learning rate of the fractional-order neural network; is the adaptive adjustment function; Update operation for parameters.

[0211] Preferably, Set to 0.01. After calculating the update gradient, the weight parameters and bias parameters will also be updated.

[0212] Step S36, adjusting the learning rate of the fractional-order neural network based on each updated gradient.

[0213] Specifically, the adaptive adjustment function dynamically adjusts the learning rate according to the size of the updated gradient. The adjustment method is expressed as:

[0214]

[0215] in, is the basic learning rate, is the regulating factor, is the L2 norm.

[0216] Preferably, Set to 0.01, Set to 1.

[0217] Furthermore, the adjustment of the basic learning rate depends on the change in model performance, allowing for more detailed adjustment of the learning rate based on the relative change in loss, allowing the model to more flexibly adapt to fluctuations during training. The adjustment method is expressed as:

[0218]

[0219] In the formula, To adjust the intensity parameters; is the sensitivity parameter; is the loss function of the fractional-order neural network trained last time, is the loss function of the fractional-order neural network for this training. Preferably, Set to 0.3, Set to 0.0001.

[0220] Step S37, return to the classification step until the fourth preset number of times is reached to obtain a classifier model.

[0221] The above embodiment adopts a fractional-order neural network based on an adaptive gradient control strategy as a classifier model, and uses fractional-order calculus theory to adjust the connection weights between neurons to enhance the model's ability to process nonlinear data. In traditional fractional-order neural networks, the gradient descent algorithm usually uses a fixed learning rate to update the weights, which limits the performance of the model on complex data sets. The present application adopts an adaptive gradient control mechanism to dynamically adjust the learning rate of each neuron according to its error rate, thereby optimizing the learning process and accelerating the convergence speed, thereby improving the classification accuracy.

[0222] See also Figure 2 Another embodiment of the present application provides a photovoltaic load prediction device based on artificial intelligence, comprising:

[0223] The first acquisition module 101 is used to acquire historical photovoltaic data of a photovoltaic power station.

[0224] The expansion module 102 is used to expand the historical photovoltaic data based on the data homeomorphism generation adversarial network to obtain a training data set.

[0225] The feature extraction training module 103 is used to train a fully connected neural network based on exploratory gradient descent using a training data set to obtain a feature extraction model.

[0226] The second acquisition module 104 is used to acquire photovoltaic data to be tested and input the data into a feature extraction model to obtain target feature data.

[0227] The dimension reduction module 105 is used to input the target feature data into the trained feature dimension reduction model to obtain the target dimension reduction data.

[0228] The prediction module 106 is used to input the target dimension reduction data into the trained classifier model to obtain the load prediction interval.

[0229] In some embodiments, the device further includes a feature dimension reduction training module, configured to perform the following steps:

[0230] The training data set is input into the feature extraction model to obtain photovoltaic feature data.

[0231] Initialize the parameters of the autoencoder; the autoencoder includes an encoder and a decoder.

[0232] Encoding step: inputting the photovoltaic characteristic data into the encoder of the autoencoder to obtain the first characteristic data.

[0233] The first feature data is input into the decoder of the autoencoder to obtain the second feature data.

[0234] A reverse reconstruction loss value is calculated based on the first feature data and the second feature data.

[0235] Update the encoder and decoder weights and biases according to the inverse reconstruction loss value.

[0236] Return to the encoding step until the third preset number of times is reached to obtain a feature dimensionality reduction model.

[0237] Furthermore, the device also includes a classification training module, which is used to perform the following steps:

[0238] The photovoltaic feature data is input into the feature dimension reduction model to obtain the dimension-reduced feature data.

[0239] Initialize the weight parameters and bias parameters of the fractional-order neural network.

[0240] Classification step: Input the reduced dimension feature data into the fractional-order neural network to obtain the classification result.

[0241] Calculate the loss error between the classification result and the labeled interval of the corresponding dimensionality reduction feature data.

[0242] The updated gradients of weight parameters and bias parameters are calculated based on the back-propagation algorithm and loss error.

[0243] Adjust the learning rate of fractional neural networks based on the gradients of each update.

[0244] Return to the classification step until the fourth preset number of times is reached to obtain a classifier model.

[0245] The specific limitations of a photovoltaic load prediction device based on artificial intelligence provided in this embodiment can be found in the above embodiment of a photovoltaic load prediction method based on artificial intelligence, which will not be repeated here. Each module in the above-mentioned photovoltaic load prediction device based on artificial intelligence can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0246] The embodiment of the present application provides a computer device, which may include a processor, a memory, a network interface and a database connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the processor executes the steps of a photovoltaic load forecasting method based on artificial intelligence as in any of the above embodiments.

[0247] The working process, working details and technical effects of the computer device provided in this embodiment can be found in the above embodiment of a photovoltaic load prediction method based on artificial intelligence, which will not be described in detail here.

[0248] The embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of a photovoltaic load forecasting method based on artificial intelligence as in any of the above embodiments are implemented. The computer-readable storage medium refers to a carrier for storing data, which may include but is not limited to a floppy disk, an optical disk, a hard disk, a flash memory, a USB flash drive and / or a memory stick, etc. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The working process, working details and technical effects of the computer-readable storage medium provided in this embodiment can be found in the above embodiment of a photovoltaic load forecasting method based on artificial intelligence, which will not be repeated here.

[0249] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0250] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0251] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A photovoltaic load forecasting method based on artificial intelligence, characterized in that: include: Obtain historical photovoltaic data of photovoltaic power plants; Initialize the parameters of the generator and discriminator of the generative adversarial network; Generating step: the generator is made to extract the periodic trend characteristics of the real photovoltaic data based on the trend extraction function and Fourier transform, and generate a noise vector based on the periodic trend characteristics and random noise; specifically, the extraction method is expressed as: in, It is a periodic trend characteristic; is the real photovoltaic data; and To adjust the weights of Fourier transform and trend analysis; represents fast Fourier transform; represents the trend extraction function; in, Extract function weights for trends, is the time variable of the real photovoltaic data, and is the slope and intercept of the linear trend line obtained by the least squares method; the noise vector generation method is expressed as: in, is the noise vector; represents the added random noise, is the identity matrix; is a dynamic adjustment factor based on the variability of environmental data; a dynamic adjustment factor based on the variability of environmental data The calculation method is expressed as: in, predefined variability thresholds for environmental data; For the Real-time measurements of environmental indicators; For the Historical averages of environmental indicators; is the total number of categories of environmental indicators; Inputting the noise vector and the corresponding real photovoltaic data into a discriminator; Returning to the generation step until a first preset number of times is reached, thereby obtaining a generative adversarial network based on data homeomorphism; The historical photovoltaic data are expanded based on the generative adversarial network of data homeomorphism to obtain a training data set; specifically, the historical photovoltaic data are input into the generative adversarial network to obtain generated data and corresponding discrimination results; if the discrimination result is true, the homeomorphism verification score of the generated data is calculated; the calculation method of the homeomorphism verification score is expressed as: in, is the homeomorphic validation score; To generate the characteristics of the data; is the L2 norm; and are the weights of characteristic differences and maximum mean differences, respectively; represents the training data set; represents the maximum mean difference function; If the homeomorphism verification score is greater than or equal to a preset verification threshold, the generated data is placed into the training data set; Using the training data set to train a fully connected neural network based on exploratory gradient descent to obtain a feature extraction model; Acquire photovoltaic data to be tested, and input the feature extraction model to obtain target feature data; Inputting the target feature data into a trained feature dimension reduction model to obtain target dimension reduction data; The target dimension reduction data is input into a trained classifier model to obtain a load forecast interval.

2. The photovoltaic load forecasting method based on artificial intelligence according to claim 1 is characterized in that: The method of using the training data set to train a fully connected neural network based on exploratory gradient descent to obtain a feature extraction model includes: Initializing network parameters of the fully connected neural network; Input step: inputting the training data set into the fully connected neural network, and subjecting each layer of the fully connected neural network to linear transformation and nonlinear activation function processing; Using an exploratory gradient descent update strategy to predict update parameters of the fully connected neural network; Adjusting the step size and direction of the fully connected neural network gradient update according to the prediction result; Return to the input step until a second preset number of times is reached to obtain the feature extraction model.

3. The photovoltaic load forecasting method based on artificial intelligence according to claim 2 is characterized in that: The fully connected neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer and an output layer connected in sequence; the number of neurons in the first hidden layer, the second hidden layer, the third hidden layer and the fourth hidden layer increases gradually; the number of neurons in the output layer is 512.

4. The photovoltaic load forecasting method based on artificial intelligence according to claim 2 is characterized in that: The training process of the feature dimensionality reduction model includes: Inputting the training data set into the feature extraction model to obtain photovoltaic feature data; Initializing parameters of an autoencoder; the autoencoder includes an encoder and a decoder; Encoding step: inputting the photovoltaic characteristic data into the encoder of the autoencoder to obtain first characteristic data; Inputting the first feature data into a decoder of the autoencoder to obtain second feature data; Calculating a reverse reconstruction loss value based on the first feature data and the second feature data; Updating weights and biases of the encoder and the decoder according to the inverse reconstruction loss value; Return to the encoding step until a third preset number of times is reached to obtain the feature dimensionality reduction model.

5. The photovoltaic load forecasting method based on artificial intelligence according to claim 4 is characterized in that: The training process of the classifier model includes: Inputting the photovoltaic characteristic data into the characteristic dimension reduction model to obtain dimension-reduced characteristic data; Initialize the weight parameters and bias parameters of the fractional-order neural network; Classification step: inputting the dimension reduction feature data into the fractional-order neural network to obtain a classification result; Calculating the loss error between the classification result and the corresponding labeled interval of the reduced dimension feature data; Calculating the update gradients of the weight parameters and the bias parameters according to the back propagation algorithm and the loss error; Adjusting the learning rate of the fractional-order neural network based on each of the update gradients; Return to the classification step until a fourth preset number of times is reached to obtain the classifier model.

6. A photovoltaic load forecasting device based on artificial intelligence, characterized in that: include: The first acquisition module is used to acquire historical photovoltaic data of the photovoltaic power station; The expansion module is used to initialize the parameters of the generator and discriminator of the generative adversarial network; perform the generation step, wherein the generation step is to enable the generator to extract the periodic trend characteristics of the real photovoltaic data based on the trend extraction function and Fourier transform, and generate a noise vector based on the periodic trend characteristics and random noise; specifically, the extraction method is expressed as: in, It is a periodic trend characteristic; is the real photovoltaic data; and To adjust the weights of Fourier transform and trend analysis; represents fast Fourier transform; represents the trend extraction function; in, Extract function weights for trends, is the time variable of the real photovoltaic data, and is the slope and intercept of the linear trend line obtained by the least squares method; the noise vector generation method is expressed as: in, is the noise vector; represents the added random noise, is the identity matrix; is a dynamic adjustment factor based on the variability of environmental data; a dynamic adjustment factor based on the variability of environmental data The calculation method is expressed as: in, predefined variability thresholds for environmental data; For the Real-time measurements of environmental indicators; For the Historical averages of environmental indicators; is the total number of categories of environmental indicators; Inputting the noise vector and the corresponding real photovoltaic data into a discriminator; Repeat the generation step for a first preset number of times to obtain a generative adversarial network based on data homeomorphism; expand the historical photovoltaic data based on the generative adversarial network of data homeomorphism to obtain a training data set; specifically, input the historical photovoltaic data into the generative adversarial network to obtain generated data and a corresponding discrimination result; if the discrimination result is true, calculate the homeomorphism verification score of the generated data; the calculation method of the homeomorphism verification score is expressed as: in, is the homeomorphic validation score; To generate the characteristics of the data; is the L2 norm; and are the weights of characteristic differences and maximum mean differences, respectively; represents the training data set; represents the maximum mean difference function; If the homeomorphism verification score is greater than or equal to a preset verification threshold, the generated data is placed into the training data set; A feature extraction training module, used to train a fully connected neural network based on exploratory gradient descent using the training data set to obtain a feature extraction model; A second acquisition module is used to acquire photovoltaic data to be tested and input the feature extraction model to obtain target feature data; A dimension reduction module, used for inputting the target feature data into a trained feature dimension reduction model to obtain target dimension reduction data; The prediction module is used to input the target dimension reduction data into the trained classifier model to obtain the load prediction interval.

7. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the photovoltaic load forecasting method based on artificial intelligence as described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the photovoltaic load forecasting method based on artificial intelligence as described in any one of claims 1 to 5 are implemented.

Images