A big data analysis method for photovoltaic power generation

Abstract

The present invention relates to the field of photovoltaic technology, and discloses a big data analysis method for photovoltaic power generation, including obtaining performance monitoring data and environmental monitoring data of photovoltaic panels; using a distributed federated learning method to build a local model at each local node, and building a federated model on a central server; using the local model to use a generative adversarial network based on quantum feature fusion to perform data expansion on the performance monitoring data and environmental monitoring data of photovoltaic panels, extracting features from the expanded data and classifying the extracted feature graphs, and uploading the model parameters of the local model to the central server; using the federated model to aggregate and update the model parameters uploaded by each local model; using the trained federated model to analyze and process the performance monitoring data and environmental monitoring data of the photovoltaic panels to be analyzed, and obtain the photovoltaic power generation category. The present invention realizes distributed model training and data mining of multi-node joint models, and effectively ensures data security.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic technology, and in particular to a big data analysis method for photovoltaic power generation. Background Art

[0002] With the rapid development and widespread application of photovoltaic technology, a large amount of data related to photovoltaic panel performance and environmental factors has been generated. The effective analysis of this data is of great significance for optimizing the operating efficiency of photovoltaic systems, predicting maintenance needs, and improving energy output. However, due to the large amount of data and the complexity of high-dimensional features, traditional data processing methods face many challenges in processing speed, accuracy, and privacy protection.

[0003] The Chinese invention patent with application number CN202110364884.9 proposes a photovoltaic power generation data monitoring system and method based on blockchain. The scheme includes: the investor uses the monitoring equipment deployed by the monitoring party at the photovoltaic power generation site to obtain the solar irradiance data collected by the monitoring equipment forwarded by the monitoring party, and compares and verifies it with the first fingerprint data uploaded to the blockchain and agreed upon by the monitoring equipment. In this way, it can be determined whether the solar irradiance data sent by the monitoring party obtained by the investor is true and reliable, and the immutable and decentralized characteristics of the blockchain are used to deter malicious tampering and ensure data security and reliability.

[0004] The Chinese invention patent with application number CN202110322549.2 proposes a fault determination method, device, processor and photovoltaic system for a photovoltaic power station. The photovoltaic power station includes a photovoltaic string and a predetermined device, and the predetermined device is directly electrically connected to the photovoltaic string. The fault determination method includes: obtaining predetermined data of the predetermined device within a predetermined time period, and the predetermined data includes DC power and internal temperature; obtaining voltage data of the predetermined device at the initial time, and the initial time is the starting time of the predetermined time period; and determining whether the photovoltaic power station has a fault based on the predetermined data and voltage data. In the fault determination method, whether the photovoltaic power station has a fault is determined based on the DC power, internal temperature and voltage data of the predetermined device obtained, so that whether the photovoltaic power station has a fault can be determined more accurately, thereby solving the problem of low accuracy in determining whether a photovoltaic power station has a fault in the prior art.

[0005] The Chinese invention patent with application number CN202110383624.6 proposes a photovoltaic power prediction method, storage medium and terminal device. The photovoltaic power prediction method includes the following steps: inputting a historical meteorological data sample set into a classification model, and obtaining a first meteorological type training set through the classification model; inputting a real-time meteorological data sample set into the classification model, and the classification model performs a rolling check on the first meteorological type training set according to the real-time meteorological data sample set, and obtains a second meteorological type training set; inputting the second meteorological type training set into a dynamic combination prediction model, and obtaining first prediction data through the dynamic combination prediction model; inputting the first prediction data into a verification model, and obtaining second prediction data through the verification model; the second prediction data is the prediction data of photovoltaic power; classifying the meteorological type according to historical data and real-time data, improving the accuracy of the classification data, and improving the prediction accuracy by cross-validating several prediction values.

[0006] The above technical solution has the following problems that still need to be further solved:

[0007] 1. In traditional model training, data usually needs to be stored and processed centrally, which can easily lead to data privacy leakage and is difficult to solve the limitations of data resource islands.

[0008] 2. Existing data generation methods usually do not use quantum computing technology, which limits the feature representation capabilities of data samples and the sample diversity during model training, resulting in the failure to effectively solve the problem of insufficient data.

[0009] 3. Traditional neural networks are relatively fixed and limited in terms of feature extraction and data adaptability, and lack an effective dynamic adjustment mechanism to adapt to complex and changeable data characteristics.

[0010] 4. Conventional classifiers usually fail to take into account the nonlinearity and dynamic changes of data, which limits the classification effect and accuracy of decision recommendations of the model when processing actual complex data. Summary of the invention

[0011] In view of the above-mentioned deficiencies in the prior art, the present invention provides a big data analysis method for photovoltaic power generation. By adopting a distributed federated learning model training architecture, distributed model training and data mining of multi-node joint models are realized to break the data resource island problem of the traditional centralized training model and aggregate to obtain more effective data. At the same time, the distributed federated learning architecture can prevent the leakage of local data of each node. The data is used locally and the model is trained locally, which effectively ensures data security.

[0012] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0013] A big data analysis method for photovoltaic power generation includes the following steps:

[0014] Obtain performance monitoring data and environmental monitoring data of photovoltaic panels;

[0015] Use distributed federated learning to build local models on each local node and build a federated model on the central server;

[0016] The local model is used to expand the performance monitoring data and environmental monitoring data of photovoltaic panels using a generative adversarial network based on quantum feature fusion. After feature extraction of the expanded data, the extracted feature graphs are classified, and the model parameters of the local model are uploaded to the central server.

[0017] The federated model is used to aggregate and update the model parameters uploaded by each local model, and the updated model parameters are sent to each local model to start a new round of iteration until the training converges;

[0018] The trained federated model is used to analyze and process the performance monitoring data and environmental monitoring data of the photovoltaic panels to be analyzed to obtain the photovoltaic power generation category.

[0019] Preferably, the data expansion of the performance monitoring data and environmental monitoring data of photovoltaic panels using a generative adversarial network based on quantum feature fusion includes:

[0020] Initialize the network parameters of the quantum code generator and discriminator, as well as the initial state of the quantum bits;

[0021] Perform quantum state encoding on the input photovoltaic panel performance monitoring data and environmental monitoring data to obtain quantum feature encoding;

[0022] In the adversarial training phase, the generator is used to generate new data samples according to the quantum feature encoding, and the discriminator is used to evaluate the authenticity of the generated data samples and calculate the loss function based on the quantum state fidelity;

[0023] After each round of training, the fidelity of the overall quantum state is evaluated and the quantum state is corrected;

[0024] Repeat the above steps until the preset stop iteration condition is met.

[0025] Preferably, the quantum state encoding of the input photovoltaic panel performance monitoring data and environmental monitoring data is performed as follows:

[0026]

[0027] in, is the quantum feature encoding function; is the feature vector of the input feature vector into the quantum feature encoding function; The first Features Indicates the effect on Pauli on qubits matrix; is the quantum bit number; is the number of quantum bits; It is an element-wise multiplication operation.

[0028] Preferably, the loss function based on quantum state fidelity is calculated as follows:

[0029]

[0030]

[0031] in, is the loss function of the generative adversarial network, is the generator function, is the discriminator function, is the eigenvector of the input eigenvector into the quantum eigencoding function, To counter the loss function, is the regularization parameter, is the fidelity function between quantum states, The real data feature vector is transformed from the initial state through quantum gate operation The converted quantum state is Generate data feature vectors for the generator through quantum gate operations from the initial state The converted quantum state is For expectations, To obey the set distribution; is the characteristic distribution of real data, is the characteristic distribution of the noise vector, is random noise.

[0032] Preferably, extracting features from the expanded data includes:

[0033] A fully connected neural network optimized based on the environment-aware wolfpack algorithm is used to extract features from the augmented data.

[0034] Preferably, the feature extraction of the expanded data using a fully connected neural network optimized based on an environment-aware wolf pack algorithm includes:

[0035] Initialize the structure of the fully connected neural network;

[0036] Dynamically adjust the parameters of the wolf pack algorithm according to the data collection time interval;

[0037] The wolfpack algorithm is used to perform incremental learning on the augmented data;

[0038] Use the influence of leaders to dynamically adjust the behavior strategy of the wolf pack algorithm;

[0039] Repeat the above steps until the preset stop iteration condition is met.

[0040] As a preferred method, the behavior strategy of the wolf pack algorithm is adjusted dynamically by using the influence of the leader:

[0041]

[0042] in, The influence of individual leaders in the wolf pack algorithm, For the greatest leadership influence, For the smallest leader influence, is the total number of iterations of the wolf pack algorithm, is the current iteration number.

[0043] Preferably, classifying the extracted feature graph includes:

[0044] A high-order neural network based on neuron migration is used to classify the extracted feature maps.

[0045] Preferably, using a high-order neural network based on neuron migration to classify the extracted feature graph includes:

[0046] Initialize the parameters of the high-order neural network;

[0047] Using the neuron migration mechanism, the network parameters are dynamically adjusted according to the real-time feedback of the data;

[0048] Using a self-optimizing network architecture search strategy, different high-order neural network architecture variants are tested and evaluated to obtain the optimal high-order neural network structure;

[0049] During the training and iterative optimization phase of the model, weights and biases are adjusted to minimize the loss function;

[0050] Repeat the above steps until the preset stop iteration condition is met.

[0051] As a preferred embodiment, the adjustment method of using the neuron migration mechanism to dynamically adjust the network parameters according to the real-time feedback of the data is as follows:

[0052]

[0053] in, The incremental calculation function for adjusting the parameters of the neuron migration adjustment strategy, is the weight of the high-order neural network, is the learning rate of the neuron migration mechanism at the tth iteration, is the loss function of the high-order neural network, is the weight of the high-order neural network at the tth iteration, is the element-wise product symbol, is the symbolic function, is the weight of the high-order neural network at the t-1th iteration.

[0054] The present invention has the following beneficial effects: BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of a big data analysis method for photovoltaic power generation;

[0056] Figure 2 This is a schematic diagram of the principle of distributed federated learning method;

[0057] Figure 3 Schematic diagram of the structure of a generative adversarial network based on quantum feature fusion;

[0058] Figure 4 Schematic diagram of the fully connected neural network structure optimized based on the environment-aware wolf pack algorithm. DETAILED DESCRIPTION

[0059] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0060] like Figure 1 As shown, a big data analysis method for photovoltaic power generation provided by an embodiment of the present invention includes the following steps S1 to S5:

[0061] S1. Obtain performance monitoring data and environmental monitoring data of photovoltaic panels;

[0062] S2, using distributed federated learning method to build local models in each local node and build a federated model in the central server;

[0063] S3. Use the local model to expand the performance monitoring data and environmental monitoring data of the photovoltaic panels using a generative adversarial network based on quantum feature fusion, extract features from the expanded data, classify the extracted feature graphs, and upload the model parameters of the local model to the central server;

[0064] S4. Use the federated model to aggregate and update the model parameters uploaded by each local model, and send the updated model parameters to each local model to start a new round of iteration until the training converges;

[0065] S5. Use the trained federated model to analyze and process the performance monitoring data and environmental monitoring data of the photovoltaic panels to be analyzed to obtain the photovoltaic power generation category.

[0066] In an optional embodiment of the present invention, the training data collected in step S1 mainly includes the performance monitoring system of the photovoltaic panel and the environmental monitoring equipment. The collected data types cover various indicators ranging from the instant output of the photovoltaic panel to the surrounding environmental conditions. The collected data is stored in a structured JSON format. In one embodiment, the attributes of the data include: current value , indicating the current output by the photovoltaic panel, in ampere (A); voltage value , represents the voltage output by the photovoltaic panel, in volts (V); temperature , represents the surface temperature of the photovoltaic panel, in degrees Celsius (°C); light intensity , represents the ambient light intensity, the unit is Lumen (Lm); wind speed , indicating ambient wind speed in meters per second (m / s); relative humidity , represents the relative humidity of the environment, in percentage (%); PM2.5 concentration , indicating the concentration of PM2.5 in the air, in micrograms per cubic meter (μg / m³); device status , indicating the operating status of the photovoltaic panel, such as normal, under maintenance, fault, etc.; efficiency , represents the conversion efficiency of photovoltaic panels, the unit is percentage (%); the energy generated , which represents the total amount of electrical energy generated by photovoltaic panels in a specific period of time, in kilowatt-hours (kWh).

[0067] It should be noted that this embodiment is only used to illustrate one data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0068] Furthermore, the collected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeled categories include: high efficiency, medium efficiency, and low efficiency, a total of 3 categories.

[0069] In an optional embodiment of the present invention, in step S2, in the federated learning framework, each node does not need to share the data of its local training model, but trains the local model and sends the updated model to the centralized learning unit for aggregation. The model training architecture of distributed federated learning is as follows: Figure 2 shown.

[0070] In the model training architecture of distributed federated learning, each node performs local model training in each round of selection and uploads the trained model parameters to the central server. The central server completes parameter aggregation and update, and sends the updated parameters to each node to start a new round of iteration until the training converges.

[0071] In the distributed federated learning model training architecture, the federated average algorithm is used as the model aggregation algorithm. The federated average algorithm implements the collaborative training of local models through multiple global iterations. Specifically, for each global iteration, the number of nodes is set to The total number of samples is , and the first The number of samples per node is , then the objective function of federated learning is defined as:

[0072]

[0073]

[0074] In the formula, is the objective function of federated learning, are model parameters, is the model parameter For The loss prediction of sample data is is the training loss, For the The sample characteristics of the sample data are No. labels of sample data. Preferably, the training loss The cross entropy loss function is used for calculation.

[0075] Furthermore, for the For a node, the objective function of the node is defined as:

[0076]

[0077] In the formula, For the The number of samples per node, For the The objective function of each node is For the The data distribution of each node.

[0078] In one embodiment, At the iteration, As an example, let the parameter update method of the first node The parameter gradient of each node is , then in The way the model updates parameters at the iteration is expressed as:

[0079]

[0080] In the formula, For the The model parameters of the iteration, For the The model parameters of the iteration, is the learning rate for the current parameter update, For the The number of samples per node.

[0081] Furthermore, the parameter update method of the global model of the central server is expressed as:

[0082]

[0083] In the formula, For the Parameters of the global model of the central server at iteration .

[0084] Repeating the iteration operation indicates that the global model of the central server and the model training of each node are completed. In one embodiment, the preset condition for stopping the iteration is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 5000 times.

[0085] In the present invention, the local model obtains the final trained federated model by aggregating parameters, and the final trained federated model includes a feature extraction model and a classifier model. In one embodiment, the photovoltaic power generation data is input into the federated model for data processing, and finally the categories of photovoltaic power generation data are obtained, including: high efficiency, medium efficiency, and low efficiency, a total of 3 categories.

[0086] The present invention adopts a distributed federated learning model, which avoids centralized data storage and processing through local data processing and model training of multiple nodes, and solves the problems of data privacy leakage and resource islands.

[0087] In an optional embodiment of the present invention, the acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model and affect the accuracy of the model. The present invention uses a generative adversarial network algorithm based on quantum feature fusion to generate samples, thereby achieving data expansion. The generative adversarial network consists of two parts: a generator and a discriminator, such as Figure 3As shown. The generator is responsible for generating new data samples, and the discriminator evaluates the authenticity of the samples. The generator structure of the present invention adopts quantum state representation, and the generator is used as a quantum coding generator to convert traditional feature coding into quantum bit form, so as to utilize the parallelism and super-position properties of quantum computing, provide richer feature representation for the generated data samples, and enhance the anti-interference ability of the model.

[0088] The training process of the generative adversarial network algorithm based on quantum feature fusion is as follows:

[0089] 1. Initialize the network parameters of the quantum code generator and discriminator, as well as the initial state of the quantum bits. Specifically, let the number of quantum bits be , the initialization state is , the initialization expression of the quantum state is:

[0090]

[0091] In the formula, Indicates The polar angle on the Bloch sphere of a quantum bit, Indicates The phase angle of the qubits; It is an element multiplication operation; is the first initial state of the quantum bit, is the second initial state of the quantum bit; Is an imaginary unit.

[0092] 2. Perform quantum state encoding on the input training data to map the high-dimensional data features to quantum bits. The present invention uses quantum gate operations to perform feature conversion to enhance the nonlinear representation capability of data. Specifically, the input data is mapped to the quantum state through quantum encoding for feature conversion. The quantum feature encoding method is expressed as:

[0093]

[0094] In the formula, is the quantum feature encoding function; is the feature vector input into the quantum feature encoding function; The first Features, Indicates the effect on Pauli on qubits matrix.

[0095] In one embodiment, the feature vector input into the quantum feature encoding function is considered to be a composite feature vector, which can be expressed as a linear combination of multiple sub-features, expressed as:

[0096]

[0097] In the formula, is the linear combination weight coefficient of the features, It is The first feature sub-characteristic functions; is the number of decomposed sub-features, is a sub-feature of the feature input into the quantum feature encoding function. Preferably, The first feature The sub-characteristic function can be the result of Fourier transform or the result of wavelet transform. Preferably, Set to 0.1.

[0098] 3. In the adversarial training phase, the generator generates new data samples based on the characteristics of quantum coding, and the discriminator evaluates the authenticity of the generated data and calculates the loss function based on the fidelity of the quantum state. Furthermore, the parameters of the generator and the discriminator are updated through the gradient descent optimization algorithm to reduce the difference between the real data samples and the generated data samples. Specifically, the update method of the parameters of the generator and the discriminator is expressed as:

[0099]

[0100]

[0101] In the formula, are the parameters of the generator, are the parameters of the discriminator; It is The learning rate of the generative adversarial network for the iteration, is the loss function of the generative adversarial network, is the gradient of the loss function of the generative adversarial network with respect to the generator parameters, is the gradient of the loss function of the generative adversarial network with respect to the discriminator parameters; It is Parameters of the generator for the iterations; It is The parameters of the discriminator at iterations; It is Parameters of the generator for the iterations; It is The parameters of the discriminator at iterations; is a generator function, is the discriminator function; It is The learning rate control parameter for each iteration.

[0102] Furthermore, the present invention uses a dynamic parameter adjustment mechanism to monitor the fidelity of the quantum state and the loss of the generator and the discriminator in real time, and uses the adaptive control strategy in feedback control theory to optimize the learning process, thereby enhancing the robustness of the model to unstable training dynamics. Specifically, the learning rate control parameter is adjusted based on the current quantum fidelity and the historical performance of the loss function, and the adjustment method is expressed as:

[0103]

[0104] In the formula, It is The fidelity between quantum states of iterations, and is the fidelity between the quantum states of the current and previous iterations; It is The loss function value of the generative adversarial network at the iteration, It is The loss function value of the generative adversarial network at the iteration. is the first hyperparameter controlling the learning rate, is the second hyperparameter controlling the learning rate, is the third hyperparameter controlling the learning rate. Preferably, Set to 2, Set to 2, Set to 0.3.

[0105] Furthermore, the dynamic adjustment method of the learning rate of the generative adversarial network is expressed as:

[0106]

[0107] In the formula, It is a hyperparameter that controls the sensitivity of learning rate changes. When the loss function fluctuates greatly, the learning rate is reduced to increase the stability of training. Preferably, Set to 2.

[0108] In one embodiment, the loss function of the generative adversarial network is calculated based on quantum fidelity and is expressed as:

[0109]

[0110]

[0111] In the formula, is the fidelity function between quantum states, is the regularization parameter; is the adversarial loss function; is the real data feature vector from the initial state through quantum gate operation The quantum state obtained by transformation; The generator generates the data feature vector from the initial state through quantum gate operation The quantum state obtained by transformation; Express expectations, is subject to a specific distribution; is the characteristic distribution of real data; is the characteristic distribution of the noise vector. Preferably, Set to 0.3.

[0112] Furthermore, the calculation method of the fidelity function between quantum states is expressed as:

[0113]

[0114] Furthermore, due to and All of them are operated from the initial state through quantum gates The transformation process of the expanded state is expressed as:

[0115]

[0116]

[0117] In the formula, is the quantum characteristic encoding function.

[0118] 4. After each round of training, the fidelity of the overall quantum state is evaluated and the quantum state is corrected to ensure the accurate retention of quantum properties. The fidelity correction method of the quantum state is expressed as:

[0119]

[0120] In the formula, is the real data feature vector from the initial state through quantum gate operation The quantum state obtained by transformation; The generator generates the data feature vector from the initial state through quantum gate operation The quantum obtained by conversion; yes and Fidelity between.

[0121] 5. Repeat the above steps 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.

[0122] After the data expansion model training is completed, the trained data expansion model is used to increase the number of samples. In one embodiment, assuming that the original sample of node k1 is 800, and the data expansion model generates 200 samples through expansion, the expanded data set contains 1000 samples.

[0123] The present invention adopts a generative adversarial network combined with a quantum feature fusion strategy, and enhances the feature representation capability of data samples through a generator and a discriminator represented by quantum bits, expands the training data set, and alleviates data scarcity.

[0124] In an optional embodiment of the present invention, the training data of each node is used to train a feature extraction model to fully mine the features of big data analysis of photovoltaic power generation. The present invention uses a 3-layer fully connected neural network for feature extraction, and the structure is as follows: Figure 4 shown.

[0125] In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, they may encounter problems such as gradient vanishing, gradient exploding, or falling into local optimal solutions, which will affect the stability of training and the performance of the model. The present invention uses a neural network optimized based on an environment-aware wolf pack algorithm for feature extraction. The wolf pack algorithm is a group intelligence optimization algorithm that simulates the hunting behavior and leadership hierarchy of wolves to achieve the search for the optimal solution. Based on the traditional wolf pack algorithm, the present invention adopts a progressive learning framework, allowing the model to gradually increase its complexity and depth during the training process to better adapt to the characteristics of photovoltaic power generation data.

[0126] The training process of the neural network algorithm optimized based on the environment-aware wolf pack algorithm is as follows:

[0127] 1. Initialize the structure of the neural network. In one embodiment, the initialization method is expressed as:

[0128]

[0129]

[0130] In the formula, is the number of nodes in the input layer of the neural network; is the dimension of the data features input to the neural network; is the number of nodes in the hidden layer of the neural network; is a scaling factor adjusted according to data complexity. Preferably, The value ranges from 2 to 10.

[0131] Furthermore, the parameters of the neural network are initialized by random initialization, and the distribution of the initialized parameters obeys a normal distribution with a mean of 0 and a variance of 0.01.

[0132] 2. Dynamically adjust the parameters of the wolf pack algorithm to better adapt to changes in the data environment. In one embodiment, the learning rate and exploration rate of the wolf pack algorithm are dynamically adjusted according to the data collection time interval. The adjustment method is expressed as follows:

[0133]

[0134]

[0135] In the formula, is the learning rate of the wolfpack algorithm; is the exploration rate of the wolf pack algorithm; is the initial learning rate of the wolf pack algorithm; is the initial exploration rate of the wolf pack algorithm; For the The learning rate of the wolfpack algorithm for each iteration; For the The exploration rate of the wolfpack algorithm in iterations; Adjustment coefficient for the first wolf pack algorithm; Adjust coefficients for the second wolf pack algorithm; Indicates the time interval for collecting photovoltaic power generation data, in hours. Preferably, Set to 3, Set to 4. Preferably, Set to 0.3, Set to 0.2, Set to 1, Set to 1.

[0136] 3. Carry out the progressive learning stage of the wolf pack algorithm, including the primary stage, intermediate stage and advanced stage.

[0137] In the primary stage, a smaller learning rate and a simple network structure are used for preliminary training to quickly converge to a reasonable solution space; in the intermediate stage, the number of layers and nodes of the network is gradually increased, and more complex activation functions and connection methods are used to improve the expressiveness of the model; in the advanced stage, while ensuring the stability of the model, the depth and width of the network are maximized, and complex nonlinear relationships are used to extract high-order features. Specifically, in the progressive learning stage, the adjustment method of the weights of the neural network is expressed as:

[0138]

[0139] In the formula, is the weight of the neural network, For the The weights of the neural network at the iteration, For the The weights of the neural network at the iteration; The parameters are , input is sample , the sample label is The gradient of the loss function when ; is the loss function of the neural network; is the input sample of the neural network, For sample The corresponding true label.

[0140] In one embodiment, the calculation method of the loss function of the neural network is expressed as:

[0141]

[0142] In the formula, The total number of samples entered for the current batch; For the The true labels of samples, No. The predicted label of each sample is obtained by classifying the feature vector after feature extraction using the preset Softmax function; is the regularization parameter of the neural network; A complexity control parameter that aims to penalize high-order powers of weights to avoid overfitting; is the total number of elements in the weight matrix; Represents the weight matrix The elements, and the complexity of the model is controlled by taking the fourth power of the weight, thereby increasing its generalization ability while ensuring the complexity of the model. Preferably, Set to 0.3, Set to 0.2.

[0143] 4. By monitoring the performance of the model during training, the behavior strategy of the wolf pack algorithm is dynamically adjusted to optimize the adaptability of the wolf pack algorithm. Specifically, the leader's influence is used to adjust the balance between global search and local search. The adjustment method is expressed as:

[0144]

[0145] In the formula, The influence of individual leaders in the wolf pack algorithm; For maximum leadership influence; Influence for the smallest leaders; is the current iteration number; is the total number of iterations of the wolf pack algorithm.

[0146] 5. Repeat the above steps 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.

[0147] The present invention adopts the environment-aware wolf pack algorithm to optimize the neural network, dynamically adjusts the network structure and parameters, adapts to changes in data characteristics, and improves the expression ability and accuracy of the feature extraction model.

[0148] In an optional embodiment of the present invention, each node also needs to build and train a decision model to realize the analysis of photovoltaic power generation big data and intelligent decision recommendation. The present invention adopts a high-order neural network algorithm based on neuron migration as a classifier to classify the data after neural network feature extraction. The high-order neural network algorithm based on neuron migration dynamically adjusts the connection and weight distribution of neurons in the neural network on the basis of a high-order neural network to adapt to the nonlinear and dynamic changes in photovoltaic power generation data.

[0149] The training process of the high-order neural network algorithm based on neuron migration is as follows:

[0150] 1. Initialize the parameters of the high-order neural network. In one embodiment, the initialization method is expressed as:

[0151]

[0152]

[0153] In the formula, is the weight of the high-order neural network, is the bias of the high-order neural network, is the initial weight of the high-order neural network, is the initial bias of the high-order neural network, is a normal distribution, is the standard deviation of weight initialization, is the initial value of the bias. Preferably, Set to , Set to .

[0154] 2. Using the neuron migration mechanism, the network parameters are dynamically adjusted according to the real-time feedback of the data to adapt to the changing data characteristics, which can be expressed as:

[0155]

[0156] In the formula, The incremental calculation function for adjusting the parameters of the neuron migration adjustment strategy, For the The learning rate of the neuron migration mechanism for each iteration, is the loss function of the high-order neural network; For the The weights of the high-order neural network in the iteration, For the The weights of the high-order neural network at the iteration; is a sign function used to determine the direction of weight adjustment; is the element-wise product sign. Preferably, Set to 0.01.

[0157] 3. Using the self-optimizing network architecture search strategy, we continuously test and evaluate different high-order neural network architecture variants to find the model structure that best suits the current data characteristics. Specifically, the self-optimizing network architecture search considers the dynamic adjustment of network depth and width, and realizes the search of the optimal network structure through optimization problems. The optimization problem is expressed as:

[0158]

[0159] In the formula, is the complexity penalty coefficient, which controls the balance between network size and overfitting; represents the depth of the high-order neural network, Represents the width of the high-order neural network; For input parameters The loss function of the high-order neural network is preferably: Set to .

[0160] In one embodiment, the loss function of the high-order neural network constrains the balance between model complexity and data fitting, and the calculation method is expressed as:

[0161]

[0162] In the formula, is the number of training samples input into the high-order neural network in the current batch, For the The true labels of the input samples; The input data is When , the weight parameter of the high-order neural network is , the predicted label of the sample; It is the input data of the high-order neural network; is the regularization coefficient of the high-order neural network loss function; is the number of weight elements of the high-order neural network.

[0163] Furthermore, considering the changes in historical gradient information and training errors, the learning rate update method of the neuron migration mechanism is expressed as:

[0164]

[0165] Where: For the The learning rate of the neuron migration mechanism for each iteration, For the The learning rate of the neuron migration mechanism for each iteration; is the sensitivity parameter of the learning rate adjustment of the neuron migration mechanism. Preferably, Set to 0.8.

[0166] 4. During the training and iterative optimization phase of the model, the weights and biases are adjusted to minimize the loss function. The adjustment method is expressed as:

[0167]

[0168]

[0169] In the formula, For the The weights of the high-order neural network in the iteration, For the The bias of the high-order neural network in the iteration; For the The weights of the high-order neural network in the iteration, For the The bias of the high-order neural network in the iteration; For the The learning rate of the high-order neural network in the iteration, For the The normalized gradient of the high-order neural network at iterations; It is the input data of the high-order neural network; Labels for input data to high-order neural networks; is the gradient of the loss function of the high-order neural network with respect to the weight parameters, is the gradient of the loss function of the high-order neural network with respect to the bias parameter.

[0170] In one embodiment, the stability of gradient descent is enhanced by adopting an adaptive gradient normalization strategy, thereby improving the learning efficiency and accuracy of the classifier. Specifically, The calculation method of the normalized gradient of the high-order neural network of iterations is expressed as:

[0171]

[0172] In the formula, To prevent division by zero, is the L2 norm. Preferably, Set to 0.0001.

[0173] Furthermore, the learning rate of the high-order neural network is dynamically adjusted according to the gradient changes during the training process, and the calculation method is expressed as:

[0174]

[0175] In the formula, is the basic learning rate, is the learning rate adjustment factor of the high-order neural network, represents the variance of the gradient of the current high-order neural network. Preferably, Set to 0.01, Set to 2.

[0176] 5. Repeat the above steps 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.

[0177] The present invention uses a high-order neural network algorithm of neuron migration as a classifier to achieve efficient classification of dynamic and nonlinear changing data, and optimizes the accuracy and stability of model decision-making.

[0178] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0179] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0180] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0181] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0182] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A big data analysis method for photovoltaic power generation, characterized in that: The following steps are involved: Obtain performance monitoring data and environmental monitoring data of photovoltaic panels; A distributed federated learning method is used to build a local model in each local node, and a federated model is built on the central server. The local model is used to expand the performance monitoring data and environmental monitoring data of photovoltaic panels using a generative adversarial network based on quantum feature fusion. After feature extraction of the expanded data, the extracted feature graphs are classified, and the model parameters of the local model are uploaded to the central server. The federated model is used to aggregate and update the model parameters uploaded by each local model, and the updated model parameters are sent to each local model to start a new round of iteration until the training converges; The trained federated model is used to analyze and process the performance monitoring data and environmental monitoring data of the photovoltaic panels to be analyzed, and the photovoltaic power generation category is obtained; Feature extraction of augmented data includes: A fully connected neural network optimized based on the environment-aware wolfpack algorithm is used to extract features from the augmented data; Initialize the structure of the fully connected neural network; Dynamically adjust the parameters of the wolf pack algorithm according to the data collection time interval; The wolf pack algorithm is used to gradually learn the expanded data. The learning rate and exploration rate of the wolf pack algorithm are dynamically adjusted according to the data collection time interval. The adjustment method is expressed as: p (t) = λ p，0 exp(-α p Δt en ), γ p (t) = γ p，0 (1-exp(-β p Δt en ))In the formula, λ p is the learning rate of the wolf pack algorithm; γ p is the exploration rate of the wolf pack algorithm; p，0 is the initial learning rate of the wolf pack algorithm; γ p，0 is the initial exploration rate of the wolf pack algorithm; p (t) is the learning rate of the wolf pack algorithm at the tth iteration; γ p (t) is the exploration rate of the wolf pack algorithm at the tth iteration; α p is the adjustment coefficient of the first wolf pack algorithm; β p is the adjustment coefficient of the second wolf pack algorithm; Δt en Indicates the collection time interval of photovoltaic power generation data, in hours; The weight adjustment method of the neural network is expressed as: Where W p is the weight of the neural network, is the weight of the neural network at the t+1th iteration, is the weight of the neural network at the yth iteration; The parameters are The input is sample X p , the sample label is Y p The gradient of the loss function when L p ( ) is the loss function of the neural network; X p is the input sample of the neural network, Y p For sample X p The corresponding true label; The calculation method of the loss function of the neural network is expressed as: Where mcs is the total number of samples input in the current batch; y i is the true label of the i-th sample, The predicted label of the i-th sample is obtained by classifying the feature vector after feature extraction using the preset Softmax function; reg，p is the regularization parameter of the neural network; ρ p is a complexity control parameter, the purpose of which is to penalize high-order powers of weights to avoid overfitting; dsc is the total number of elements in the weight matrix; W p，j Represents the weight matrix The jth element in ; Use the influence of leaders to dynamically adjust the behavior strategy of the wolf pack algorithm; Repeat the above steps until the preset stop iteration condition is met; The adjustment method of using the leader's influence to dynamically adjust the behavior strategy of the wolf pack algorithm is: in, is the leader influence of individuals in the wolf pack algorithm, φ p，max For the maximum leader influence, φ p，min is the minimum leader influence, Tax is the total number of iterations of the wolf pack algorithm, and int(t) is the current number of iterations.

2. A big data analysis method for photovoltaic power generation according to claim 1, characterized in that: The data expansion of photovoltaic panel performance monitoring data and environmental monitoring data using a generative adversarial network based on quantum feature fusion includes: Initialize the network parameters of the quantum code generator and discriminator, as well as the initial state of the quantum bits; Perform quantum state encoding on the input photovoltaic panel performance monitoring data and environmental monitoring data to obtain quantum feature encoding; In the adversarial training phase, the generator is used to generate new data samples according to the quantum feature encoding, and the discriminator is used to evaluate the authenticity of the generated data samples and calculate the loss function based on the quantum state fidelity; After each round of training, the fidelity of the overall quantum state is evaluated and the quantum state is corrected; Repeat the above steps until the preset stop iteration condition is met.

3. The big data analysis method for photovoltaic power generation according to claim 1, characterized in that: The quantum state encoding of the input photovoltaic panel performance monitoring data and environmental monitoring data is expressed as: Among them, U c () is the quantum characteristic encoding function; X c is the feature vector of the input feature vector into the quantum feature encoding function; X c，j represents the jth feature of the input feature vector; σ z，j represents the Pauli Z matrix acting on the jth quantum bit; i is the quantum bit number; n c is the number of quantum bits; It is an element-wise multiplication operation.

4. The big data analysis method for photovoltaic power generation according to claim 1, characterized in that: The loss function based on quantum state fidelity is calculated as: Among them, L c G is the loss function of the generative adversarial network. c is the generator function, D c is the discriminator function, X c is the eigenvector of the input eigenvector into the quantum eigencoding function, L GAN is the adversarial loss function, λ c is the regularization parameter, Fe is the fidelity function between quantum states, |ψ c，real > is the real data feature vector from the initial state |ψ through quantum gate operation c >The quantum state obtained by transformation, |ψ c，gen >Generate data feature vectors for the generator through quantum gate operations from the initial state |ψ c >The quantum state obtained by transformation, is the expectation, ~ is subject to the set distribution; p data (X c ) is the characteristic distribution of real data, p z (z) is the characteristic distribution of the noise vector, G c (z) is the random noise generator function, and z is the random noise.

5. The big data analysis method for photovoltaic power generation according to claim 1, characterized in that: Classification of the extracted feature maps includes: A high-order neural network based on neuron migration is used to classify the extracted feature maps.

6. A big data analysis method for photovoltaic power generation according to claim 5, characterized in that: The extracted feature maps are classified using a high-order neural network based on neuron migration, including: Initialize the parameters of the high-order neural network; Using the neuron migration mechanism, the network parameters are dynamically adjusted according to the real-time feedback of the data; Using a self-optimizing network architecture search strategy, different high-order neural network architecture variants are tested and evaluated to obtain the optimal high-order neural network structure; During the training and iterative optimization phase of the model, weights and biases are adjusted to minimize the loss function; Repeat the above steps until the preset stop iteration condition is met.

7. A big data analysis method for photovoltaic power generation according to claim 6, characterized in that: Using the neuron migration mechanism, the network parameters are adjusted dynamically according to the real-time feedback of the data as follows: Among them, Δ u The parameter adjustment increment calculation function for the neuron migration adjustment strategy, W u is the weight of the high-order neural network, is the learning rate of the neuron migration mechanism at the tth iteration, is the loss function of the high-order neural network, W u (t) is the weight of the high-order neural network at the t-th iteration, ⊙ is the element-wise product sign, sgn is the sign function, W u (t-1) is the weight of the high-order neural network at the t-1th iteration.

Images