Wind power prediction system and method based on structural contrast evolutionary neural structure search

Abstract

This invention belongs to the field of data prediction technology, specifically disclosing a wind power prediction system and method based on structural comparative evolutionary neural network structure search. The system resamples and performs feature selection on historical wind farm operation data to obtain wind power data; defines and encodes a candidate neural network structure search space to obtain the neural network structure encoding DNN; trains the DNN and verifies the error; and constructs a new training set X using the DNN and the verification error. new And train a structural contrast model model ac Using evolutionary algorithms, model ac Instead of evaluating fitness, the neural network is compared and iteratively optimized to obtain the optimal neural network structure. The optimal neural network structure is then trained and its parameters tuned using a training set to obtain the final optimal structure. Wind power data is input into the optimal neural network structure, and the predicted wind power is output. This invention features high prediction efficiency and accuracy, good stability, low resource consumption, and strong generalization ability.

Description

Technical Field

[0001] This application belongs to the field of data processing technology, specifically relating to a wind power prediction system and method based on structural contrastive evolutionary neural structure search, which has high prediction efficiency and accuracy, good stability, low resource consumption, and strong generalization ability. Background Technology

[0002] As the global energy transition accelerates, wind energy, as a core component of clean and renewable energy, continues to increase its share in the energy mix. Vigorously developing wind energy can not only reduce dependence on fossil fuels and alleviate environmental pollution and greenhouse gas emissions, but also promote energy architecture optimization and sustainable economic development, while creating numerous skilled jobs and possessing significant ecological and social value. However, the inherent intermittent and random characteristics of wind energy lead to drastic fluctuations in wind farm output power, posing serious challenges to the safe and stable operation of the power system, dispatch planning, and the balance of power supply and demand.

[0003] Accurate wind power forecasting is crucial for solving the aforementioned wind power problems. Grid controllers, energy market planners, and regulatory agencies all require reliable forecast data to formulate reasonable power generation plans, dispatch backup power sources, optimize grid operation strategies, ensure a stable power supply, and meet end-user electricity demand. Therefore, the accuracy and efficiency of wind power forecasting technology directly affect the economic benefits of wind farms and the operational safety of the power system, and has become one of the core research topics in the field of new energy.

[0004] With the rapid development of artificial intelligence technology, wind power prediction methods have gradually shifted from traditional statistical time series models (such as autoregressive moving average models and exponential smoothing) to advanced machine learning and deep learning techniques. For example, deep neural networks (DNNs), with their powerful nonlinear fitting and feature extraction capabilities, have achieved breakthroughs in image recognition, natural language processing, and time series prediction, and have been widely used in the construction of wind power prediction models in recent years. However, because wind power is affected by various factors such as meteorological conditions, geographical environment, and equipment operating status, the data exhibits complex nonlinear, time-series correlation, and scene diversity characteristics. Fixed-architecture neural network models are difficult to adapt to the operating conditions of different wind farms and cannot flexibly cope with dynamic changes in data distribution, resulting in limited prediction accuracy. Therefore, developing automated and adaptive neural network architecture optimization algorithms is crucial for generating high-quality, self-adjusting wind power prediction models.

[0005] Neural Architecture Search (NAS) aims to automate the design of neural network architectures, providing an effective way to overcome the limitations of manual design. However, NAS is inherently a complex optimization problem, facing multiple challenges such as complex constraints, discrete representations, and high computational costs. Evolutionary Neural Architecture Search (ENAS), based on evolutionary computation, is a NAS algorithm with unique advantages in achieving full automation of NAS and handling complex constraints and multi-objective optimization. It has been widely applied in image processing, natural language processing, and time series forecasting. However, the application of ENAS algorithms in wind power forecasting still has the following significant limitations: 1. High computational cost: Traditional ENAS algorithms require complete training for each candidate architecture to evaluate performance, which requires a lot of computational resources and time even on small datasets, making it difficult to meet the real-time requirements of engineering applications.

[0006] 2. Insufficient exploration of the search space: Existing coding methods mostly use fixed formats to describe network structures, which makes it difficult to fully depict the diversity of complex network topologies and limits the ability to explore novel and efficient architectures.

[0007] 3. Imbalance between performance evaluation accuracy and efficiency: Fitness evaluation methods that rely on accurate performance predictions increase computational overhead and are susceptible to data noise, making it difficult to balance search efficiency and prediction performance.

[0008] To address the aforementioned issues, there is an urgent need to develop an automated modeling method that can efficiently explore network structures, reduce computational costs, and accurately capture the characteristics of wind power data. This method would solve the shortcomings of existing wind power prediction models, such as complex structural design, high computational costs, and unstable prediction performance, thereby promoting the engineering application and upgrading of wind power prediction technology. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention proposes a wind power prediction system based on structural comparative evolutionary neural structure search, which features high prediction efficiency and accuracy, good stability, low resource consumption, and strong generalization ability. It also provides a wind power prediction method based on structural comparative evolutionary neural structure search.

[0010] The wind power prediction system based on structural contrastive evolutionary neural structure search of this invention is implemented as follows: it includes a data acquisition module, a structure encoding module, a training and verification module, a model building module, an iterative optimization module, a training and tuning module, and a power prediction module. The data acquisition module is used to perform time resampling, feature selection, dataset partitioning, and sliding window sample reconstruction on historical wind farm operation data to obtain wind power data including training set, validation set, and test set. The structure encoding module is used to define the candidate neural network structure search space, encode the structure search space, and obtain the structure encoding DNN of the neural network. The training and verification module is used to train the structure encoding DNN using the training set in the wind power data, and to verify the trained structure encoding DNN using the verification set in the wind power data, so as to obtain the verification error of the sampled structure. The model training module is used to construct a new training set X by using the structure encoding DNN obtained from the structure encoding module and the verification error obtained from the training and verification module as features and labels, respectively. new and using the training set X new Structural comparison model model ac Training is performed to optimize model parameters so that the structural contrast model can establish a mapping relationship between structural encoding features and model validation error; The iterative optimization module is used to compare the trained structure with the model using an evolutionary algorithm. model ac By directly comparing neural network structures and iteratively optimizing them, an optimized neural network structure can be obtained. The training and optimization module is used to train and optimize the neural network structure using the training set in the wind power data to obtain the optimal neural network structure. The power prediction module is used to input the current wind power data into the optimal neural network structure and output the predicted wind power.

[0011] Furthermore, the data acquisition module includes a feature selection unit, a temporal resampling unit, a dataset partitioning unit, and a sliding window sample reconstruction unit. The feature selection unit is used to select input features related to wind power prediction from historical operating data of wind farms. The input features include wind speed, wind direction, and wind power. The time resampling unit is used to resample the historical operation data of the wind farm according to a predetermined sampling interval Δt to obtain multivariate sequence data with uniform time granularity; The dataset partitioning unit is used to partition the historical operation data of wind farms after data cleaning, outlier processing, and time resampling in chronological order to obtain a training set, a validation set, and a test set. First, the preprocessed continuous historical operation data is sorted according to timestamps, and data subsets for different time periods are divided according to a preset ratio. Then, feature selection rules, normalization parameters, and other data transformation parameters are determined using only the training set data, and the obtained parameters are applied to the validation set and the test set respectively. The sliding window sample reconstruction unit is used to construct samples for the training set, validation set, and test set respectively using the sliding window method after the dataset partitioning unit has completed the dataset partitioning, so as to form corresponding supervised learning sample pairs (X,y); where X represents the model input sample, which consists of wind farm operation feature data from multiple consecutive historical moments within the sliding window, including wind speed, wind direction, temperature, and other effective input variables after feature selection; y represents the model prediction target.

[0012] Furthermore, in the structure encoding module, the candidate neural network structure search space includes the network topology and computational operations of the candidate neural network; in the structure search space, the neural network structure is represented in the form of a directed acyclic graph, where nodes represent computational units and edges represent data transmission relationships; wherein, the network topology is used to describe the connection relationship between each node and the network structure in the candidate neural network structure, and the computational operations are used to characterize the specific computational operations performed by the nodes; The structure encoding of the neural network (DNN) includes connection relationship encoding and computation operation encoding; wherein, the connection relationship encoding is used to represent the connection relationship between nodes in the candidate neural network structure, and the computation operation encoding is used to record the type of computation operation corresponding to each node in the candidate neural network structure.

[0013] Furthermore, the connection relationship encoding adopts the form of an adjacency matrix, and the calculation operation encoding adopts the form of an operation list; the elements in the adjacency matrix take values ​​of 0 or 1, respectively indicating that there is no connection relationship between the corresponding nodes; the operation list records the specific calculation operations of each node in the candidate neural network structure.

[0014] Furthermore, the specific function of the training and verification module is to randomly sample the training set from the wind power data obtained by the data acquisition module. N A sample, and with N The structure encoding DNN obtained by training the structure encoding module with samples is saved as a DNN. n Then, the validation set from the wind power data is used to test the DNN. n Verification was performed to obtain the DNN. n Predicted value ; final calculationN DNN corresponding to each sample n Root mean square error (RMSE): , Where: N val The number of samples in the validation set. y i The first i The predicted value and the corresponding actual value of each sample.

[0015] Furthermore, the model building module includes a new dataset building unit, a structure comparison model training unit, and a neural network structure selection unit; The new dataset construction unit is used to sample the datasets obtained from the training and validation modules. N Each sample and its corresponding RMSE are used as features and labels, respectively, and saved as a new dataset X. new ; The structural contrast model training unit is used to train the dataset X. new Training Structure Comparison Model model ac Specifically, it encodes the DNN using the trained structure. n Using the corresponding RMSE as the supervision signal, the structural comparison model is compared. model ac Structural representations are obtained through comparative learning of supervised learning sample pairs (X,y), and the structural comparison model is jointly optimized with classification cross-entropy loss to complete the structural comparison model. model ac Training; The neural network structure selection unit is used to predefine a set of candidate operators, network connection rules, and hyperparameter value ranges based on the input and output requirements of the wind power prediction task, and construct a neural network structure search space Ω. A structure generator is used to generate candidate network structures (Arch) that satisfy the constraints from the search space Ω according to random sampling rules. The candidate network structure is represented as a directed acyclic graph composed of a set of node operators, a connection matrix, and a set of hyperparameters. The structure encoding module converts the candidate network structure (Arch) into a fixed-length structure encoding vector, which is then input into the trained structure comparison model. model ac The prediction performance of different candidate structures is evaluated; the optimal network structure Arch is selected based on the prediction performance ranking results; then, a wind power prediction network is constructed based on Arch, and the network is trained using supervised learning samples (X,y) generated by a sliding window to obtain the final wind power prediction model.

[0016] Furthermore, the iterative optimization module includes an initialization population unit, an individual structure sorting unit, a structure encoding selection unit, a new population selection unit, and an evolutionary condition unit. The initialization population unit is used to initialize N individual structures representing the structure encoding DNN in the search space Ω using a random sampling method to form an initialization population; wherein, each individual structure is encoded separately before being added to the whole; The individual structure sorting unit is used to map the structure of each individual in the initial population to the corresponding structure encoding DNN, select a binary tournament selection strategy in the evolutionary algorithm, and use a structure comparison model to select the individual structures. model ac The individual structures are directly compared, and then sorted according to the comparison results. The structure coding selection unit is used to select the top-ranked individual structures as parent structure codes based on the aforementioned sorting results, and then generate new child structure codes with the same size as the parent structure codes through crossover and mutation operations. The new population selection unit is used to merge the parent structure code and the new offspring structure code to form a combined population, using a structure comparison model. model ac Individuals in the combined population are compared to obtain ranking scores and sorted in descending order of scores. The top N individuals in the descending order are selected and retained as the next generation of the new population, and the rest are eliminated. The evolutionary condition unit is used to determine whether the current evolutionary iteration number t has reached the preset maximum number of evolutionary generations T. max Among them, the maximum number of generations T max The search space size, the number of candidate network structures, and the computational resource requirements are pre-defined; if the current iteration number t <T max If the current iteration number t ≥ T, then continue to execute the next round of evolution; max If this happens, the evolutionary process will terminate.

[0017] Furthermore, the training and optimization module is used to train the parameters of the optimized neural network structure using the training set and corresponding supervised learning sample pairs (X,y) from the wind power data obtained by the data acquisition module, and to iteratively update the weights of the aforementioned training set during the parameter training process, so that the optimized neural network structure gradually approaches the minimization of the loss function; after the parameter training is completed and a preset convergence condition is reached, the optimal neural network structure is obtained; the preset convergence condition is to verify the optimized neural network structure after parameter training using a validation set, and the loss function of the optimized neural network structure reaches the minimum trigger value or the number of iterations reaches the upper limit within a preset number of consecutive rounds.

[0018] The wind power prediction method based on structural contrastive evolutionary neural structure search of this invention is implemented as follows: it includes data acquisition, structural encoding, training and validation, model building, iterative optimization, training tuning, and power prediction steps. The specific steps are as follows: A. Data Acquisition: Time resampling, feature selection, dataset partitioning, and sliding window sample reconstruction are performed on historical wind farm operation data to obtain wind power data including training set, validation set, and test set; B. Structure Encoding: Define the candidate neural network structure search space, encode the structure search space, and obtain the structure encoding DNN of the neural network; C. Training and Validation: The structure encoding DNN is trained using the training set in the wind power data, and the trained structure encoding DNN is validated using the validation set in the wind power data to obtain the validation error of the sampled structure. D. Model Construction: Using the structured encoding DNN obtained in step B and the validation error obtained in step C as features and labels respectively, a new training set X is constructed. new And using training set X new Structural comparison model model ac Training is performed to optimize model parameters so that the structural contrast model can establish a mapping relationship between structural encoding features and model validation error; E. Iterative Optimization: Using evolutionary algorithms, the trained structure is compared with the model. model ac By directly comparing neural network structures and iteratively optimizing them, an optimized neural network structure can be obtained. F. Training and optimization: Use the training set in the wind power data to train and optimize the neural network structure to obtain the optimal neural network structure; G. Power Prediction: Input the current wind power data into the optimal neural network structure and output the predicted wind power.

[0019] Furthermore, the specific process of step C is as follows: randomly sampling the training set from the wind power data obtained in step A to obtain... N A sample, and with N The structure encoding DNN obtained by training step B with samples is saved as a DNN. n Then, the validation set from the wind power data is used to test the DNN. n Verification was performed to obtain the DNN. n Predicted value ; final calculation N DNN corresponding to each sample n Root mean square error (RMSE): , Where: N valThe number of samples in the validation set. y i The first i The predicted value and the corresponding true value of each sample; The specific process of step D is as follows: D10. The sampled data obtained in step C. N Each sample and its corresponding RMSE are used as features and labels, respectively, and saved as a new dataset X. new ; D20, Using dataset X new Training Structure Comparison Model model ac Specifically, it encodes the DNN using the trained structure. n Using the corresponding RMSE as the supervision signal, the structural comparison model is compared. model ac Structural representations are obtained through comparative learning of supervised learning sample pairs (X,y), and the structural comparison model is jointly optimized with classification cross-entropy loss to complete the structural comparison model. model ac Training; D30. Based on the input and output requirements of the wind power prediction task, predefine the candidate operator set, network connection rules, and hyperparameter value range to construct a neural network structure search space Ω. Using a structure generator, candidate network structures (Arch) satisfying the constraints are generated from the search space Ω according to random sampling rules. The candidate network structure is represented as a directed acyclic graph composed of a set of node operators, a connection matrix, and a set of hyperparameters. The candidate network structure (Arch) is then converted into a fixed-length structure encoding vector by the structure encoding module and input into the trained structure comparison model. model ac The prediction performance of different candidate structures is evaluated; the optimal network structure Arch is selected based on the prediction performance ranking results; then, a wind power prediction network is constructed based on Arch, and the network is trained using supervised learning samples (X,y) generated by a sliding window to obtain the final wind power prediction model.

[0020] The beneficial effects of this invention are: 1. This invention enhances the ability of the structure search process to characterize complex network topologies by uniformly encoding neural network structures, enabling different candidate neural network structures to be effectively expressed and manipulated within the same search space. Simultaneously, it introduces a structure comparison evaluation mechanism based on structure encoding, replacing the inefficient "complete training + validation" approach of traditional ENAS with a relative performance comparison between candidate neural network structures (i.e., indirectly evaluating the fitness of candidate structures through a surrogate model). This involves first training candidate neural network structures with a small number of samples to obtain their validation errors (such as RMSE), then using "structure encoding" and "validation errors" as feature-labels to construct a surrogate model. In subsequent evolutionary processes, the surrogate model quickly predicts the performance of candidate neural network structures, eliminating the need for repeated training. This not only reduces the impact of data noise on performance evaluation but also guides evolution towards "better structures" to improve search efficiency. Furthermore, it significantly reduces the computational complexity and time overhead of the evolutionary search phase, effectively solving the problem of high computational resource consumption and low efficiency in existing ENAS algorithms for wind power prediction scenarios. Based on this, this invention integrates the structure comparison model (… model ac By combining it with the ENAS algorithm, the overall efficiency of neural network structure search is further improved while ensuring the validity of search results, enabling the evolutionary process to quickly converge to a high-performance network structure under limited computing resources.

[0021] 2. This invention uses a directed acyclic graph (DAG) to uniformly represent neural network structures, and uses adjacency matrices (connection relationship encoding) and operation lists (computation operation encoding) to finely encode network topology and operation types, thereby improving the ability to characterize complex and diverse network architectures and overcoming the limitations of traditional fixed-format encoding in terms of insufficient search space exploration, thus enhancing the search capability and expression flexibility of network structures.

[0022] 3. The optimal neural network structure obtained in this invention, due to iterative optimization (selection, crossover, mutation) and parameter tuning (small-batch stochastic gradient descent to optimize network weights) specifically for wind power data (such as wind speed, wind direction, historical power, etc.), makes the final optimal network structure more closely fit the characteristics of wind power data (such as the temporal dependence of wind speed, wind direction, and power). It exhibits higher prediction accuracy, stronger robustness, and better generalization ability on real wind farm data (it can adapt to wind power data in different scenarios, such as extreme weather, equipment aging, etc.), and can effectively cope with the challenges of nonlinearity, temporal sequence, and scenario diversity of wind power.

[0023] 4. The entire modeling process of this invention, from data preprocessing and structure search to final prediction, is fully automated. This not only avoids the problems of low efficiency and poor stability caused by relying on manual experience to repeatedly adjust the network structure and input delay parameters in traditional methods, but also balances computational efficiency and prediction performance, thereby meeting the engineering application requirements of power systems for real-time and reliable wind power prediction.

[0024] In summary, this invention addresses the shortcomings of traditional wind power prediction models, such as reliance on manual design of structures, high computational costs, and unstable prediction performance. By encoding and representing neural network structures and introducing a structure comparison and evaluation mechanism, it achieves evolutionary optimization search and efficient screening of neural network structures, thereby constructing a high-performance wind power prediction model. This not only improves the level of automated modeling of wind power prediction models but also reduces model structure search time and computational resource consumption, while achieving a synergistic improvement in prediction performance and engineering practicality. It has good application prospects and promotional value. Attached Figure Description

[0025] Figure 1 This is a flowchart of the prediction method of the present invention; Figure 2 This is an example of the structural encoding of the prediction method of the present invention; Figure 3 This is an example of a structural comparison model for the prediction method of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this does not limit the present invention in any way. Any changes or improvements made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0027] This invention relates to a wind power prediction system based on structural comparative evolutionary neural architecture search, comprising a data acquisition module, a structure encoding module, a training and validation module, a model building module, an iterative optimization module, a training tuning module, and a power prediction module. The data acquisition module is used to perform time resampling, feature selection, dataset partitioning, and sliding window sample reconstruction on historical wind farm operation data to obtain wind power data including training set, validation set, and test set. The structure encoding module is used to define the candidate neural network structure search space, encode the structure search space, and obtain the structure encoding DNN of the neural network. The training and verification module is used to train the structure encoding DNN using the training set in the wind power data, and to verify the trained structure encoding DNN using the verification set in the wind power data, so as to obtain the verification error of the sampled structure. The model training module is used to construct a new training set X by using the structure encoding DNN obtained from the structure encoding module and the verification error obtained from the training and verification module as features and labels, respectively. new and using the training set X new Structural comparison model model ac Training is performed to optimize model parameters so that the structural contrast model can establish a mapping relationship between structural encoding features and model validation error; The iterative optimization module is used to compare the trained structure with the model using an evolutionary algorithm. model ac By directly comparing neural network structures and iteratively optimizing them, an optimized neural network structure can be obtained. The training and optimization module is used to train and optimize the neural network structure using the training set in the wind power data to obtain the optimal neural network structure. The power prediction module is used to input the current wind power data into the optimal neural network structure and output the predicted wind power.

[0028] The data acquisition module includes a feature selection unit, a temporal resampling unit, a dataset partitioning unit, and a sliding window sample reconstruction unit. The feature selection unit is used to select input features related to wind power prediction from historical operating data of wind farms. The input features include wind speed, wind direction, and wind power. The time resampling unit is used to resample the historical operation data of the wind farm according to a predetermined sampling interval Δt (e.g., 10 min, 15 min or 1 h) to obtain multivariate sequence data with uniform time granularity. The dataset partitioning unit is used to partition the historical operation data of wind farms after data cleaning, outlier processing, and time resampling in chronological order to obtain a training set, a validation set, and a test set. First, the preprocessed continuous historical operation data is sorted according to timestamps, and data subsets for different time periods are divided according to a preset ratio. Then, only the training set data is used to determine feature selection rules, normalization parameters, and other data transformation parameters, and the obtained parameters are applied to the validation set and the test set respectively to avoid introducing future time information into the model training process. The sliding window sample reconstruction unit is used to construct samples for the training set, validation set, and test set respectively using the sliding window method after the dataset partitioning unit has completed the dataset partitioning, so as to form corresponding supervised learning sample pairs (X,y); where X represents the model input sample, which consists of wind farm operation feature data of multiple consecutive historical moments within the sliding window, including wind speed, wind direction, temperature, and other effective input variables after feature selection; y represents the model prediction target, that is, the wind power value at the corresponding prediction moment.

[0029] In the sliding window sample reconstruction unit, let the length of the window be... L The sliding step size is S The prediction step size (lead time) is H Then for any subset of the multivariate sequence { x t Construct input samples X t = [x t-L+1 ,...,x t ] The corresponding tag is , where superscript ( p ) represents power; thus, a sample set is obtained for subsequent search and prediction model training.

[0030] In the structure encoding module, the candidate neural network structure search space includes the network topology and computational operations of the candidate neural network. In this search space, the neural network structure is represented as a directed acyclic graph, where nodes represent computational units and edges represent data transfer relationships. The network topology describes the connections between nodes and the network structure within the candidate neural network, and the computational operations characterize the specific computational operations performed by the nodes. By designing this search space, a balance between exploration capability and computational complexity can be achieved, providing an effective set of candidate structures for the evolutionary algorithm. The structure encoding of the neural network (DNN) includes connection relationship encoding and computation operation encoding; wherein, the connection relationship encoding is used to represent the connection relationship between nodes in the candidate neural network structure, and the computation operation encoding is used to record the type of computation operation corresponding to each node in the candidate neural network structure.

[0031] The connection relationships are encoded using an adjacency matrix, while the computational operations are encoded using an operation list. Elements in the adjacency matrix take values ​​of 0 or 1, representing either no connection or a connection between corresponding nodes (i.e., 0 indicates no connection between nodes in the adjacency matrix, and 1 indicates a connection). The operation list records the specific computational operations performed on each node in the candidate neural network structure. This encoding scheme accurately represents the network structure, facilitates searching and manipulation, and ensures compatibility between the search space and evolutionary algorithms. The operation list describes the specific structure of each network block, while the adjacency matrix represents the connection methods between different blocks, thus achieving accurate description and effective management of the network structure.

[0032] The specific function of the training and verification module is to randomly sample the training set from the wind power data obtained by the data acquisition module. N A sample, and with N The structure encoding DNN obtained by training the structure encoding module with samples is saved as a DNN. n Then, the validation set from the wind power data is used to test the DNN. n Verification was performed to obtain the DNN. n Predicted value ; final calculation N DNN corresponding to each sample n Root mean square error (RMSE): , Where: N val The number of samples in the validation set. y i The first i The predicted value and the corresponding actual value of each sample.

[0033] The model building module includes a new dataset building unit, a structural comparison model training unit, and a neural network structure selection unit; The new dataset construction unit is used to sample the datasets obtained from the training and validation modules. N Each sample and its corresponding RMSE are used as features and labels, respectively, and saved as a new dataset X. new ; The structural contrast model training unit is used to train the dataset X. new Training Structure Comparison Model model ac Specifically, it encodes the DNN using the trained structure. n Using the corresponding RMSE as the supervision signal, the structural comparison model is compared. model acStructural representations are obtained through comparative learning of supervised learning sample pairs (X,y), and the structural comparison model is jointly optimized with classification cross-entropy loss to complete the structural comparison model. model ac Training; The neural network structure selection unit is used to predefine a set of candidate operators, network connection rules, and hyperparameter value ranges based on the input and output requirements of the wind power prediction task, and construct a neural network structure search space Ω. A structure generator is used to generate candidate network structures (Arch) that satisfy the constraints from the search space Ω according to random sampling rules. The candidate network structure is represented as a directed acyclic graph composed of a set of node operators, a connection matrix, and a set of hyperparameters. The structure encoding module converts the candidate network structure (Arch) into a fixed-length structure encoding vector, which is then input into the trained structure comparison model. model ac The prediction performance of different candidate structures is evaluated; the optimal network structure Arch is selected based on the prediction performance ranking results; then, a wind power prediction network is constructed based on Arch, and the network is trained using supervised learning samples (X,y) generated by a sliding window to obtain the final wind power prediction model.

[0034] Currently, the data pair ( Aα, Xα The target labels are only 0 and 1, generated by comparing the performance metrics of two candidate network structures, Arch1 and Arch2. "0" indicates that "Arch1" performs worse than "Arch2", and "1" indicates the opposite. This method aims to reduce the reliance on precise performance predictions and instead focus on evaluating relative performance, thereby improving efficiency and accuracy in the search for efficient structures.

[0035] The iterative optimization module includes an initialization population unit, an individual structure sorting unit, a structure encoding selection unit, a new population selection unit, and an evolutionary condition unit. The initialization population unit is used to initialize N individual structures representing the structure encoding DNN in the search space Ω using a random sampling method to form an initialization population; wherein, each individual structure is encoded separately before being added to the whole; The individual structure sorting unit is used to map the structure of each individual in the initial population to the corresponding structure encoding DNN, select a binary tournament selection strategy in the evolutionary algorithm, and use a structure comparison model to select the individual structures. model ac The individual structures are directly compared, and then sorted according to the comparison results. The structure coding selection unit is used to select the top-ranked individual structures (e.g., the top 20%) as parent structure codes based on the aforementioned ranking results, and then generate new child structure codes with the same size as the parent structure codes through crossover and mutation operations. The new population selection unit is used to merge the parent structure code and the new offspring structure code to form a combined population, using a structure comparison model. model ac Individuals in the combined population are compared to obtain ranking scores and sorted in descending order of scores. The top N individuals in the descending order are selected and retained as the next generation of the new population, and the rest are eliminated. The evolutionary condition unit is used to determine whether the current evolutionary iteration number t has reached the preset maximum number of evolutionary generations T. max Among them, the maximum number of generations T max The search space size, the number of candidate network structures, and the computational resource requirements are pre-defined; if the current iteration number t <T max If the current iteration number t ≥ T, then continue to execute the next round of evolution; max If this happens, the evolutionary process will terminate.

[0036] It's important to note that the entire process in the iterative optimization module follows the evolutionary algorithm's workflow: population initialization (initializing population units), fitness evaluation (individual structure sorting units), parent selection and offspring reproduction (structure encoding selection units), and environmental selection (new population selection units). Generally, evolutionary algorithms involve four hyperparameters: the maximum number of iterations in the entire evolutionary process, population size, crossover rate, and mutation rate of offspring. A major factor influencing the performance of evolutionary algorithms is environmental selection, as effective environmental selection ensures that individuals in the new population have better fitness than those in previous populations.

[0037] The training and optimization module is used to train the parameters of the optimized neural network structure using the training set and corresponding supervised learning sample pairs (X,y) from the wind power data obtained by the data acquisition module, and to iteratively update the weights of the aforementioned training set during the parameter training process, so that the optimized neural network structure gradually approaches the minimization of the loss function. After the parameter training is completed and a preset convergence condition is reached, the optimal neural network structure is obtained. The preset convergence condition is to verify the optimized neural network structure after parameter training using a validation set, and the loss function of the optimized neural network structure reaches the minimum trigger value or the maximum number of iterations within a preset number of consecutive rounds.

[0038] like Figure 1 , 2As shown in Figure 3, the wind power prediction method based on structural contrastive evolutionary neural structure search of the present invention includes the steps of data acquisition, structural encoding, training and verification, model construction, iterative optimization, training tuning, and power prediction. The specific steps are as follows: A. Data Acquisition: Time resampling, feature selection, dataset partitioning, and sliding window sample reconstruction are performed on historical wind farm operation data to obtain wind power data including training set, validation set, and test set; B. Structure Encoding: Define the candidate neural network structure search space, encode the structure search space, and obtain the structure encoding DNN of the neural network; C. Training and Validation: The structure encoding DNN is trained using the training set in the wind power data, and the trained structure encoding DNN is validated using the validation set in the wind power data to obtain the validation error of the sampled structure. D. Model Construction: Using the structured encoding DNN obtained in step B and the validation error obtained in step C as features and labels respectively, a new training set X is constructed. new And using training set X new Structural comparison model model ac Training is performed to optimize model parameters so that the structural contrast model can establish a mapping relationship between structural encoding features and model validation error; E. Iterative Optimization: Using evolutionary algorithms, the trained structure is compared with the model. model ac By directly comparing neural network structures and iteratively optimizing them, an optimized neural network structure can be obtained. F. Training and optimization: Use the training set in the wind power data to train and optimize the neural network structure to obtain the optimal neural network structure; G. Power Prediction: Input the current wind power data into the optimal neural network structure and output the predicted wind power.

[0039] The specific steps of step A are as follows: A10. Feature Selection: Select input features related to wind power prediction from historical wind farm operation data. The input features include wind speed, wind direction, and wind power. A20. Time resampling: Resample the historical operating data of the wind farm according to a predetermined sampling interval Δt (e.g., 10 min, 15 min or 1 h) to obtain multivariate sequence data with uniform time granularity; A30. Dataset Partitioning: The historical operation data of wind farms, after data cleaning, outlier handling, and time resampling, is partitioned according to time order to obtain training, validation, and test sets. First, the preprocessed continuous historical operation data is sorted according to timestamps, and data subsets for different time periods are divided according to preset proportions. Then, feature selection rules, normalization parameters, and other data transformation parameters are determined using only the training set data, and the obtained parameters are applied to the validation and test sets respectively. A40. Sliding Window Sample Reconstruction: After completing the dataset partitioning in step A30, the sliding window method is used to construct samples for the training set, validation set, and test set to form corresponding supervised learning sample pairs (X,y); where X represents the model input sample, which consists of wind farm operation feature data from multiple consecutive historical moments within the sliding window, including wind speed, wind direction, temperature, and other effective input variables after feature selection; y represents the model prediction target.

[0040] In the A40 sub-step, let the window length be L, the sliding step size be S, and the prediction step size (lead time) be H. Then, for any subset of multivariate sequences { x t Construct input samples X t = [x t-L+1 ,...,x t ] The corresponding tag is , where superscript ( p ) represents power; thus, a sample set is obtained for subsequent search and prediction model training.

[0041] In step B, the candidate neural network structure search space includes the network topology and computational operations of the candidate neural network. In the structure search space, the neural network structure is represented in the form of a directed acyclic graph, where nodes represent computational units and edges represent data transmission relationships. The network topology is used to describe the connection relationships between each node and the network structure in the candidate neural network structure, and the computational operations are used to characterize the specific computational operations performed by the nodes. The structural encoding of a neural network (DNN) includes connection relationship encoding and computation operation encoding. Connection relationship encoding is used to represent the connection relationships between nodes in the candidate neural network structure, while computation operation encoding is used to record the type of computation operation corresponding to each node in the candidate neural network structure.

[0042] The specific process of step C is as follows: randomly sampling the training set from the wind power data obtained in step A to obtain... N A sample, and with NThe structure encoding DNN obtained by training step B with samples is saved as a DNN. n Then, the validation set from the wind power data is used to test the DNN. n Verification was performed to obtain the DNN. n Predicted value ; final calculation N DNN corresponding to each sample n Root mean square error (RMSE): , Where: N val The number of samples in the validation set. y i The first i The predicted value and the corresponding actual value of each sample.

[0043] The specific process of step D is as follows: D10. The sampled data obtained in step C. N Each sample and its corresponding RMSE are used as features and labels, respectively, and saved as a new dataset X. new ; D20, Using dataset X new Training Structure Comparison Model model ac Specifically, it encodes the DNN using the trained structure. n Using the corresponding RMSE as the supervision signal, the structural comparison model is compared. model ac Structural representations are obtained through comparative learning of supervised learning sample pairs (X,y), and the structural comparison model is jointly optimized with classification cross-entropy loss to complete the structural comparison model. model ac Training; D30. Based on the input and output requirements of the wind power prediction task, predefine the candidate operator set, network connection rules, and hyperparameter value range to construct a neural network structure search space Ω. Using a structure generator, candidate network structures (Arch) satisfying the constraints are generated from the search space Ω according to random sampling rules. The candidate network structure is represented as a directed acyclic graph composed of a set of node operators, a connection matrix, and a set of hyperparameters. The candidate network structure (Arch) is then converted into a fixed-length structure encoding vector by the structure encoding module and input into the trained structure comparison model. model ac The prediction performance of different candidate structures is evaluated; the optimal network structure Arch is selected based on the prediction performance ranking results; then, a wind power prediction network is constructed based on Arch, and the network is trained using supervised learning samples (X,y) generated by a sliding window to obtain the final wind power prediction model.

[0044] The specific process of step E is as follows: E10. Initialize N individual structures representing the structure encoding DNN within the search space Ω using random sampling to form an initial population; each individual structure is encoded separately before being added to the population. E20. Map the structure of each individual in the initial population to the corresponding structure encoding DNN. Select the binary tournament selection strategy in the evolutionary algorithm and use the structure comparison model to select the individual structures. model ac The individual structures are directly compared, and then sorted according to the comparison results. E30. Based on the aforementioned ranking results, select the top-ranked individual structures (e.g., the top 20%) as parent structure codes, and then generate new child structure codes with the same size as the parent structure codes through crossover and mutation operations. E40. Merge the parent generation structure code with the new offspring structure code to form a combined population, and use a structure comparison model. model ac Individuals in the combined population are compared to obtain ranking scores and sorted in descending order of scores. The top N individuals in the descending order are selected and retained as the next generation of the new population, and the rest are eliminated. E50. Determine whether the current evolution iteration number t has reached the preset maximum number of evolution generations T. max Among them, the maximum number of generations T max The search space size, the number of candidate network structures, and the computational resource requirements are pre-defined; if t <T max If the current iteration number t ≥ T, then continue to execute the next round of evolution; max If this happens, the evolutionary process will terminate.

[0045] In step F, the training set and corresponding supervised learning sample pairs (X,y) from the wind power data in step A are used to train the parameters of the optimized neural network structure using mini-batch stochastic gradient descent. During the parameter training process, the weights of the aforementioned training set are iteratively updated so that the optimized neural network structure gradually approaches the minimization of the loss function. After the parameter training is completed and the preset convergence condition is met, the optimal neural network structure is obtained. The preset convergence condition is that the optimized neural network structure after parameter training is verified using a validation set, and the loss function of the optimized neural network structure reaches the minimum trigger value or the number of iterations reaches the upper limit within a preset number of consecutive iterations.

[0046] Example This embodiment uses publicly available wind farm operation data from the International New Energy Laboratory as the verification object. Input features are constructed using historical wind speed, direction, and power information, with the prediction target being the wind power value at a future time, thus forming the data pairs required for the wind power prediction task.

[0047] S100: Data from the International New Energy Laboratory for 2011 and 2012 were selected. Data from 16MW wind turbines was collected in both years with a sampling time of 5 minutes. After downloading the sampled data, 210,242 samples were obtained. Then, three important features—wind speed, wind direction, and wind power—were selected as input variables. The downloaded data was resampled at 15-minute intervals as needed, resulting in 68,880 samples. Before the experiment, these 68,880 samples were divided into three sets: a training set, a validation set, and a test set, as shown in Table 1.

[0048] Table 1. Dataset partitioning used in the experiment

[0049] S200: Define the candidate neural network structure search space, encode the structure search space, and obtain the neural network structure encoding (DNN). The specific process is as follows: S210. Define the search space: The candidate neural network structure search space comprises the network topology and computational operations of the candidate neural networks, used to limit the range of candidate neural network structures. The network topology describes the connection relationships between nodes or network modules, and the computational operations characterize the specific types of operations performed by the nodes. In the candidate neural network structure search space, the neural network structure is represented in the form of a directed acyclic graph (DAG), where nodes represent computational units and edges represent data transmission relationships; the computational operations of each node are selected from a preset set of operations. By combining the topology and computational operations, effective expression and search of various candidate neural network structures are achieved, providing a foundation for subsequent structure encoding, structure comparison and evaluation, and evolutionary optimization.

[0050] S220. Structure Encoding: To facilitate a unified representation and manipulation of neural network structures during evolutionary search, candidate neural network structures are structure-encoded. The structure encoding (DNN) of the neural network comprises two parts: connection relationship encoding and computational operation encoding. The connection relationship encoding represents the connections between nodes in the network and uses an adjacency matrix. The computational operation encoding records the computational operation types corresponding to each node in the candidate neural network structure and uses an operation list. Elements in the adjacency matrix take values ​​of 0 or 1, where 0 indicates no connection between corresponding nodes and 1 indicates a connection between corresponding nodes. By combining the adjacency matrix and the operation list, a unified description of the neural network's topology and computational features is achieved, thereby supporting the evolutionary algorithm's search, comparison, and optimization of different candidate structures.

[0051] S300: As shown in Table 1, using the data from the entire year of 2011 as the training set, 500 samples were randomly sampled from the training set. These 500 samples were used to train the structure encoding DNN obtained from S200, and the trained structure encoding DNN was saved as a DNN. n Subsequently, the DNN was validated using data from the first quarter of 2012. n Validation was performed, and the DNN corresponding to 500 structural samples was obtained. n Predicted value Finally, calculate N DNN corresponding to each sample n Root mean square error (RMSE): , Where: N val The number of samples in the validation set. y i The first i The predicted value and the corresponding actual value of each sample.

[0052] S400: Construct a new training set X using the structured encoding DNN obtained in step S200 and the root mean square error (RMSE) obtained in step S300, respectively, with features and labels. new And using training set X new Training Structure Comparison Model model ac The specific process is as follows: S410. Take the 500 samples obtained from S300 and their corresponding RMSEs as features and labels respectively, and save them as a new dataset X. new .

[0053] S420, Using dataset X new Training Structure Comparison Model model acSpecifically, it encodes the DNN using the trained structure. n Using the corresponding RMSE as the supervision signal, the structural comparison model is compared. model ac Structural representations are obtained through comparative learning of supervised learning sample pairs (X,y), and the structural comparison model is jointly optimized with classification cross-entropy loss to complete the structural comparison model. model ac Training.

[0054] S430. Based on the input and output requirements of the wind power prediction task, a candidate set, connection rules, and hyperparameter value range are pre-defined to construct a neural network structure search space Ω. A structure generator generates candidate network structures (Arch) that satisfy the constraints from the search space Ω according to random sampling rules. The candidate network structure is represented as a directed acyclic graph composed of a set of node operators, a connection relation matrix, and a hyperparameter set. The structure encoding module converts the candidate network structure (Arch) into a fixed-length structure encoding vector, which is then input into the trained structure comparison model. model ac The prediction performance of different candidate structures is evaluated; the optimal network structure Arch is selected based on the prediction performance ranking results; then, a wind power prediction network is constructed based on Arch, and the network is trained using supervised learning samples (X,y) generated by a sliding window to obtain the final wind power prediction model.

[0055] Currently, the data pair ( Aα, Xα The target labels are only 0 and 1, generated by comparing the performance metrics of two candidate network structures, Arch1 and Arch2. "0" indicates that "Arch1" performs worse than "Arch2", and "1" indicates the opposite. This method aims to reduce the reliance on precise performance predictions and instead focus on evaluating relative performance, thereby improving efficiency and accuracy in the search for efficient structures.

[0056] S500: Utilizes an evolutionary algorithm to compare the trained structure with the model. model ac Instead of evaluating fitness, a direct comparison of neural network structures is performed, and iterative optimization yields an optimized neural network structure. The specific process is as follows: S510. In the search space Ω, N individual structures representing the structure encoding DNN are initialized using random sampling to form an initial population; each individual structure is encoded separately before being added to the whole.

[0057] S520. Map the structure of each individual in the initial population to the corresponding structure encoding DNN. In the evolutionary algorithm, select the binary tournament selection strategy, i.e., select two individuals from the population each time (sampling with replacement). Then, use the selected individual structures... modelac The individual structures are directly compared and then sorted according to the comparison structure.

[0058] S530. Based on the aforementioned ranking results, select the top-ranked individual structures (e.g., the top 20%) as parent structure codes, and then generate new child structure codes with the same size as the parent structure codes through crossover and mutation operations.

[0059] S540. Merge the parent generation structure code with the new offspring structure code to form a combined population, and use a structure comparison model. model ac Individuals in the combined population are compared to obtain ranking scores and sorted in descending order of scores. The top N individuals in the descending order are selected and retained as the next generation of the new population, and the rest are eliminated.

[0060] S550. Determine whether the current evolution iteration number t has reached the preset maximum evolution generation number T. max Among them, the maximum number of generations T max The search space size, the number of candidate network structures, and the computational resource requirements are pre-defined; if t <T max Then proceed to step S530 to continue the next round of evolution; if the current iteration number t≥T max If this happens, the evolutionary process will terminate.

[0061] S600: Using the training set and corresponding supervised learning sample pairs (X,y) from the wind power data obtained in S100, mini-batch stochastic gradient descent is used to train the parameters of the optimized neural network structure. During the parameter training process, the weights of the aforementioned training set are iteratively updated so that the optimized neural network structure gradually approaches the minimization of the loss function. After the parameter training is completed and the preset convergence condition is met, the optimal neural network structure is obtained. The preset convergence condition is that the optimized neural network structure after parameter training is verified using a validation set, and the loss function of the optimized neural network structure reaches the minimum trigger value or the number of iterations reaches the upper limit within a preset number of consecutive iterations.

[0062] S700: Input the current wind power data into the optimal neural network structure and output the predicted wind power.

[0063] This invention compares the following methods: (1) Sample best : Sample a series of structures, train and select the optimal structure.

[0064] (2) High-fidelity estimation ENAS: High-fidelity estimation first calculates the similarity between each data sample and other data samples using the Euclidean distance (ED) method, selects the 5 farthest points as the starting point, and uses them as the starting point for the search.

[0065] (3) Evolutionary Neural Architecture Search (ENAS): A traditional algorithm for searching evolutionary neural network architectures.

[0066] (4) Agent-Assisted Evolutionary Neural Architecture Search (SA-ENAS): Fitness values ​​are calculated using an agent model.

[0067] (5) AC-ENAS: Method of the embodiment.

[0068] This embodiment uses time T as the search efficiency evaluation index and the coefficient of determination as the metric. R 2 The root mean square error (RMSE) is a performance evaluation index, and its calculation formula is: , , In the formula: N test The number of samples in the test set. y i The first i The predicted value and the corresponding actual value of each sample.

[0069] Tables 2 and 3 respectively show the differences in training time, optimal structure phenotype, and performance comparison of the method in this embodiment (AC-ENAS) and the other three ENAS methods, with the dataset divided by season.

[0070] Table 2 Comparison of prediction performance of different prediction methods

[0071] Table 3 Training Time

[0072] Tables 2 and 3 compare the performance of AC-ENAS and the three ENAS methods at different prediction levels in terms of optimal structural phenotype and training time. The following conclusions can be drawn: (1) For the same modeling dataset, different model prediction methods show significant performance differences. In comparison, AC-ENAS has obvious advantages on different test sets.

[0073] (2) AC-ENAS is more computationally efficient than the other three ENAS on different datasets, meaning it can obtain the best neural network structure in a shorter time. Compared with SA-ENAS, although the time consumption is the same, AC-ENAS has better prediction performance.

[0074] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A wind power prediction system based on structural contrastive evolutionary neural structure search, characterized in that: It includes a data acquisition module, a structure encoding module, a training and validation module, a model building module, an iterative optimization module, a training tuning module, and a power prediction module. The data acquisition module is used to perform time resampling, feature selection, dataset partitioning, and sliding window sample reconstruction on historical wind farm operation data to obtain wind power data including training set, validation set, and test set. The structure encoding module is used to define the candidate neural network structure search space, encode the structure search space, and obtain the structure encoding DNN of the neural network. The training and verification module is used to train the structure encoding DNN using the training set in the wind power data, and to verify the trained structure encoding DNN using the verification set in the wind power data, so as to obtain the verification error of the sampled structure. The model training module is used to construct a new training set X by using the structure encoding DNN obtained from the structure encoding module and the verification error obtained from the training and verification module as features and labels, respectively. new and using the training set X new Structural comparison model model ac Training is performed to optimize model parameters so that the structural contrast model can establish a mapping relationship between structural encoding features and model validation error; The iterative optimization module is used to compare the trained structure with the model using an evolutionary algorithm. model ac By directly comparing neural network structures and iteratively optimizing them, an optimized neural network structure can be obtained. The training and optimization module is used to train and optimize the neural network structure using the training set in the wind power data to obtain the optimal neural network structure. The power prediction module is used to input the current wind power data into the optimal neural network structure and output the predicted wind power.

2. The wind power prediction system based on structural contrastive evolutionary neural structure search according to claim 1, characterized in that: The data acquisition module includes a feature selection unit, a temporal resampling unit, a dataset partitioning unit, and a sliding window sample reconstruction unit. The feature selection unit is used to select input features related to wind power prediction from historical operating data of wind farms. The input features include wind speed, wind direction, and wind power. The time resampling unit is used to resample the historical operation data of the wind farm according to a predetermined sampling interval Δt to obtain multivariate sequence data with uniform time granularity; The dataset partitioning unit is used to partition the historical operation data of wind farms after data cleaning, outlier processing, and time resampling in chronological order to obtain a training set, a validation set, and a test set. First, the preprocessed continuous historical operation data is sorted according to timestamps, and data subsets for different time periods are divided according to a preset ratio. Then, feature selection rules, normalization parameters, and other data transformation parameters are determined using only the training set data, and the obtained parameters are applied to the validation set and the test set respectively. The sliding window sample reconstruction unit is used to construct samples for the training set, validation set, and test set respectively using the sliding window method after the dataset partitioning unit has completed the dataset partitioning, so as to form corresponding supervised learning sample pairs (X,y); where X represents the model input sample, which consists of wind farm operation feature data from multiple consecutive historical moments within the sliding window, including wind speed, wind direction, temperature, and other effective input variables after feature selection; y represents the model prediction target.

3. The wind power prediction system based on structural contrastive evolutionary neural structure search according to claim 1, characterized in that: In the structure encoding module, the candidate neural network structure search space includes the network topology and computational operations of the candidate neural network; in the structure search space, the neural network structure is represented in the form of a directed acyclic graph, where nodes represent computational units and edges represent data transmission relationships; wherein, the network topology is used to describe the connection relationship between each node and the network structure in the candidate neural network structure, and the computational operations are used to characterize the specific computational operations performed by the nodes; The structure encoding of the neural network (DNN) includes connection relationship encoding and computation operation encoding; wherein, the connection relationship encoding is used to represent the connection relationship between nodes in the candidate neural network structure, and the computation operation encoding is used to record the type of computation operation corresponding to each node in the candidate neural network structure.

4. The wind power prediction system based on structural contrastive evolutionary neural structure search according to claim 3, characterized in that: The connection relationship is encoded in the form of an adjacency matrix, and the computation operation is encoded in the form of an operation list. The elements in the adjacency matrix take values ​​of 0 or 1, which respectively indicate that there is no connection relationship between the corresponding nodes or that there is a connection relationship. The operation list records the specific computation operations of each node in the candidate neural network structure.

5. The wind power prediction system based on structural contrastive evolutionary neural structure search according to claim 2, 3 or 4, characterized in that: The specific function of the training and verification module is to randomly sample the training set from the wind power data obtained by the data acquisition module. N A sample, and with N The structure encoding DNN obtained by training the structure encoding module with samples is saved as a DNN. n Then, the validation set from the wind power data is used to test the DNN. n Verification was performed to obtain the DNN. n Predicted value ; final calculation N DNN corresponding to each sample n Root mean square error (RMSE): ， Where: N val The number of samples in the validation set. y i The first i The predicted value and the corresponding actual value of each sample.

6. The wind power prediction system based on structural contrastive evolutionary neural structure search according to claim 5, characterized in that: The model building module includes a new dataset building unit, a structure comparison model training unit, and a neural network structure selection unit; the new dataset building unit is used to sample the data obtained from the training and validation module. N Each sample and its corresponding RMSE are used as features and labels, respectively, and saved as a new dataset X. new ; The structural contrast model training unit is used to train the dataset X. new Training Structure Comparison Model model ac Specifically, it encodes the DNN using the trained structure. n Using the corresponding RMSE as the supervision signal, the structural comparison model is compared. model ac Structural representations are obtained through comparative learning of supervised learning sample pairs (X,y), and the structural comparison model is jointly optimized with classification cross-entropy loss to complete the structural comparison model. model ac Training; The neural network structure selection unit is used to predefine a set of candidate operators, network connection rules, and hyperparameter value ranges based on the input and output requirements of the wind power prediction task, and construct a neural network structure search space Ω. A structure generator is used to generate candidate network structures (Arch) that satisfy the constraints from the search space Ω according to random sampling rules. The candidate network structure is represented as a directed acyclic graph composed of a set of node operators, a connection matrix, and a set of hyperparameters. The structure encoding module converts the candidate network structure (Arch) into a fixed-length structure encoding vector, which is then input into the trained structure comparison model. model ac The prediction performance of different candidate structures is evaluated; the optimal network structure Arch is selected based on the prediction performance ranking results; then, a wind power prediction network is constructed based on Arch, and the network is trained using supervised learning samples (X,y) generated by a sliding window to obtain the final wind power prediction model.

7. The wind power prediction system based on structural contrastive evolutionary neural structure search according to claim 6, characterized in that: The iterative optimization module includes an initialization population unit, an individual structure sorting unit, a structure encoding selection unit, a new population selection unit, and an evolutionary condition unit. The initialization population unit is used to initialize N individual structures representing the structure encoding DNN in the search space Ω using a random sampling method to form an initialization population; wherein, each individual structure is encoded separately before being added to the whole; The individual structure sorting unit is used to map the structure of each individual in the initial population to the corresponding structure encoding DNN, select a binary tournament selection strategy in the evolutionary algorithm, and use a structure comparison model to select the individual structures. model ac The individual structures are directly compared, and then sorted according to the comparison results. The structure coding selection unit is used to select the top-ranked individual structures as parent structure codes based on the aforementioned sorting results, and then generate new child structure codes with the same size as the parent structure codes through crossover and mutation operations. The new population selection unit is used to merge the parent structure code and the new offspring structure code to form a combined population, using a structure comparison model. model ac Individuals in the combined population are compared to obtain ranking scores and sorted in descending order of scores. The top N individuals in the descending order are selected and retained as the next generation of the new population, and the rest are eliminated. The evolutionary condition unit is used to determine whether the current evolutionary iteration number t has reached the preset maximum number of evolutionary generations T. max Among them, the maximum number of generations T max The search space size, the number of candidate network structures, and the computational resource requirements are pre-defined; if the current iteration number t <T max If the current iteration number t ≥ T, then continue to execute the next round of evolution; max If this happens, the evolutionary process will terminate.

8. The wind power prediction system based on structural contrastive evolutionary neural structure search according to claim 5, characterized in that: The training and optimization module is used to train the parameters of the optimized neural network structure using the training set and corresponding supervised learning sample pairs (X,y) in the wind power data obtained by the data acquisition module, and iteratively update the weights of the aforementioned training set during the parameter training process, so that the optimized neural network structure gradually approaches the minimization of the loss function. After the parameters are trained and the preset convergence condition is met, the optimal neural network structure is obtained. The preset convergence condition is to verify the optimized neural network structure after parameter training using a validation set, and the loss function of the optimized neural network structure reaches the minimum trigger value or the number of iterations reaches the upper limit within a preset number of consecutive iterations.

9. A wind power prediction method based on structural contrastive evolutionary neural structure search, characterized in that: The process includes data acquisition, structure encoding, training and validation, model building, iterative optimization, training tuning, and power prediction. The specific steps are as follows: A. Data Acquisition: Time resampling, feature selection, dataset partitioning, and sliding window sample reconstruction are performed on historical wind farm operation data to obtain wind power data including training set, validation set, and test set; B. Structure Encoding: Define the candidate neural network structure search space, encode the structure search space, and obtain the structure encoding DNN of the neural network; C. Training and Validation: The structure encoding DNN is trained using the training set in the wind power data, and the trained structure encoding DNN is validated using the validation set in the wind power data to obtain the validation error of the sampled structure. D. Model Construction: Using the structured encoding DNN obtained in step B and the validation error obtained in step C as features and labels respectively, a new training set X is constructed. new And using training set X new Structural comparison model model ac Training is performed to optimize model parameters so that the structural contrast model can establish a mapping relationship between structural encoding features and model validation error; E. Iterative Optimization: Using evolutionary algorithms, the trained structure is compared with the model. model ac By directly comparing neural network structures and iteratively optimizing them, an optimized neural network structure can be obtained. F. Training and optimization: Use the training set in the wind power data to train and optimize the neural network structure to obtain the optimal neural network structure; G. Power Prediction: Input the current wind power data into the optimal neural network structure and output the predicted wind power.

10. The wind power prediction method based on structural contrastive evolutionary neural structure search according to claim 9, characterized in that: The specific process of step C is as follows: randomly sampling the training set from the wind power data obtained in step A to obtain... N A sample, and with N The structure encoding DNN obtained by training step B with samples is saved as a DNN. n Then, the validation set from the wind power data is used to test the DNN. n Verification was performed to obtain the DNN. n Predicted value ; final calculation N DNN corresponding to each sample n Root mean square error (RMSE): ， Where: N val The number of samples in the validation set. y i The first i The predicted value and the corresponding true value of each sample; The specific process of step D is as follows: D10. The sampled data obtained in step C. N Each sample and its corresponding RMSE are used as features and labels, respectively, and saved as a new dataset X. new ; D20, Using dataset X new Training Structure Comparison Model model ac Specifically, it encodes the DNN using the trained structure. n Using the corresponding RMSE as the supervision signal, the structural comparison model is compared. model ac Structural representations are obtained through comparative learning of supervised learning sample pairs (X,y), and the structural comparison model is jointly optimized with classification cross-entropy loss to complete the structural comparison model. model ac Training; D30. Based on the input and output requirements of the wind power prediction task, predefine the candidate operator set, network connection rules, and hyperparameter value range to construct a neural network structure search space Ω. Using a structure generator, candidate network structures (Arch) satisfying the constraints are generated from the search space Ω according to random sampling rules. The candidate network structure is represented as a directed acyclic graph composed of a set of node operators, a connection matrix, and a set of hyperparameters. The candidate network structure (Arch) is then converted into a fixed-length structure encoding vector by the structure encoding module and input into the trained structure comparison model. model ac The prediction performance of different candidate structures is evaluated; the optimal network structure Arch is selected based on the prediction performance ranking results; then, a wind power prediction network is constructed based on Arch, and the network is trained using supervised learning samples (X,y) generated by a sliding window to obtain the final wind power prediction model.

Images