A wind energy prediction method and system based on deep learning

Abstract

This application relates to a deep learning-based wind energy prediction method and system. The method includes: preprocessing raw wind energy data to generate standard wind energy data, and dividing it into a first wind energy dataset and a second wind energy dataset; generating multiple sets of model initialization parameters and determining the parameter evaluation values ​​of these parameters; iteratively updating the multiple sets of model initialization parameters based on the parameter evaluation values ​​until the target model initialization parameters are obtained; initializing a deep learning prediction model based on the target model initialization parameters and training the initialized deep learning prediction model; and verifying the output accuracy of the trained model based on the second wind energy dataset, determining the model as the target wind energy prediction model if it meets expectations. This method can reduce the possibility of the model getting trapped in local optima during training, improve model training stability and prediction accuracy, and simultaneously enhance the reliability and generalization ability of the model in practical applications.

Description

Technical Field

[0001] This application relates to the field of wind energy forecasting technology, and in particular to a wind energy forecasting method and system based on deep learning. Background Technology

[0002] With the continuous expansion of installed wind power capacity, wind power's share in the power system is constantly increasing. Because wind energy is affected by various factors such as meteorological conditions, geographical environment, and equipment operating status, target wind energy data such as wind power output and effective wind speed typically exhibit intermittent, random fluctuations, and nonlinear characteristics, leading to complex variation patterns and significant prediction difficulties. Existing wind energy prediction methods mainly include physical prediction methods and data-driven prediction methods. Physical prediction methods typically rely on fluid dynamics, thermodynamics, or numerical weather prediction models, which suffer from complex modeling, high computational load, and long time consumption. In contrast, data-driven prediction methods, especially deep learning methods, can learn the nonlinear mapping relationship between input data and target wind energy data through historical data, thus gradually becoming an important technical route for wind energy prediction.

[0003] However, existing deep learning prediction models still have some significant problems when applied to wind energy prediction: On the one hand, deep learning model training is a high-dimensional non-convex optimization process, which is highly sensitive to model initialization parameters. Existing technologies often use random initialization or empirical setting to determine initialization parameters, lacking effective evaluation and optimization of initialization parameters. This can easily lead to slow model convergence, unstable training process, or even getting stuck in local optima, thus affecting the accuracy and stability of prediction results. On the other hand, after the model training is completed, existing technologies usually lack a mechanism to re-evaluate the model's prediction accuracy based on independent datasets. Therefore, it is difficult to effectively determine whether the trained model truly meets the accuracy and generalization requirements of practical applications. Summary of the Invention

[0004] Therefore, it is necessary to provide a deep learning-based wind energy prediction method and system that can optimize model initialization parameters and re-evaluate the prediction accuracy of the trained model, thereby reducing the possibility of the model getting stuck in local optimum solutions during training, improving model training stability and prediction accuracy, and enhancing the reliability and generalization ability of the model in practical applications.

[0005] On the one hand, a deep learning-based wind energy prediction method is provided, the method comprising: Raw wind energy data is acquired and preprocessed to generate standard wind energy data, which is then divided into a first wind energy dataset and a second wind energy dataset. Generate multiple sets of model initialization parameters for initializing a preset deep learning prediction model, and determine the parameter evaluation values ​​of the multiple sets of model initialization parameters based on the first wind energy dataset and the preset parameter evaluation function. Based on the parameter evaluation value, the multiple sets of model initialization parameters are iteratively updated until a preset number of iterations is reached. The model initialization parameter corresponding to the optimal parameter evaluation value among the multiple sets of updated model initialization parameters is determined as the target model initialization parameter. The deep learning prediction model is initialized based on the target model initialization parameters, and the initialized deep learning prediction model is trained based on the first wind energy dataset to obtain the trained deep learning prediction model. Based on the second wind energy dataset, determine whether the output accuracy of the trained deep learning prediction model meets expectations; if so, determine the trained deep learning prediction model as the target wind energy prediction model.

[0006] On the other hand, a deep learning-based wind energy forecasting system is provided, the system comprising: The raw data processing module is used to acquire raw wind energy data and preprocess it to generate standard wind energy data, and divide the standard wind energy data into a first wind energy dataset and a second wind energy dataset. The initialization parameter evaluation module is used to generate multiple sets of model initialization parameters for initializing a preset deep learning prediction model, and to determine the parameter evaluation values ​​of the multiple sets of model initialization parameters based on the first wind energy dataset and the preset parameter evaluation function. The initialization parameter determination module is used to iteratively update multiple sets of model initialization parameters according to the parameter evaluation value until a preset number of iterations is reached, and determine the model initialization parameter corresponding to the optimal parameter evaluation value among the updated multiple sets of model initialization parameters as the target model initialization parameter. The prediction model training module is used to initialize the deep learning prediction model based on the target model initialization parameters, and train the initialized deep learning prediction model based on the first wind energy dataset to obtain the trained deep learning prediction model. The prediction model determination module is used to determine whether the output accuracy of the trained deep learning prediction model meets expectations based on the second wind energy dataset. If so, the trained deep learning prediction model is determined as the target wind energy prediction model.

[0007] The aforementioned deep learning-based wind energy prediction method and system optimizes key initial parameters that affect training performance before model training. This allows the deep learning prediction model to enter the training process with a better initial state, thereby reducing training fluctuations caused by random initialization or empirical settings, lowering the possibility of the model getting stuck in local optima, improving the stability and convergence of model training, and enhancing the accuracy of wind energy prediction results. Furthermore, by further validating the model's predictive ability after training, it avoids judging model performance solely based on the fitting results during the training phase, thus more effectively evaluating the model's actual prediction accuracy and improving the generalization ability, prediction reliability, and application value of the target wind energy prediction model. Attached Figure Description

[0008] Figure 1 This is an application environment diagram of a deep learning-based wind energy prediction method in one embodiment; Figure 2 This is a flowchart illustrating a deep learning-based wind energy prediction method in one embodiment. Figure 3 This is a schematic diagram of a deep learning-based wind energy prediction system architecture in one embodiment; Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0010] This application provides a deep learning-based wind energy prediction method that can be applied to, for example... Figure 1 The application environment shown may include a wind energy data acquisition terminal 101 and a server 102, which communicate via a network. The server 102 can be implemented as a standalone server, a server cluster consisting of multiple servers, a cloud server, a distributed computing platform, or other computing devices with data processing, model training, and predictive analysis capabilities. The wind energy data acquisition terminal 101 can be, but is not limited to, a sensor node, a data acquisition device, an industrial control computer, an edge computing device, a smart gateway, or other devices with data acquisition and data transmission functions. The wind energy data acquisition terminal 101 can be deployed in scenarios such as wind farms, wind turbine generators, wind measurement towers, booster stations, or meteorological monitoring stations to collect raw data related to wind energy status and power generation operation status, and send the raw data to the server 102.

[0011] Server 102 is used to uniformly process the received raw data and execute a deep learning-based wind energy prediction process. Specifically, server 102 can clean, align, normalize, and construct samples from the collected raw data to form standard data that can be used for model processing. Based on this, training samples and validation samples are constructed around the target data, and the initialization parameters of the preset deep learning prediction model are searched and optimized to obtain a better initial state of the model. Subsequently, the model is trained using the training samples, enabling the model to extract the nonlinear mapping relationship between the raw data and the target data, and the prediction performance of the model is verified by the validation samples. When the model's prediction performance meets the preset requirements, the model is determined as the target wind energy prediction model for subsequent real-time prediction applications.

[0012] It should be noted that the raw data may include meteorological environmental data and equipment operation data. Meteorological environmental data may include, for example, wind speed, wind direction, temperature, air pressure, and humidity. Equipment operation data may include, for example, historical power generation, rotor speed, nacelle position angle, and blade pitch angle. Not all of the above raw data are used as the final prediction target, but rather as basic data to characterize wind energy variation patterns and equipment operating status. Server 102 can identify target data with predictive significance and application value from the raw data, and use the target data as the output of the wind energy prediction task, while using other data related to the target data as model input features. Target data may be wind power, effective wind speed, and available wind energy intensity at future times or within future time periods—data that reflects the effectiveness of wind energy utilization and is suitable for trend prediction.

[0013] In some embodiments, after the target wind energy prediction model is determined, the server 102 can also receive real-time data before the time to be predicted, process the real-time data according to the data processing method corresponding to the training phase, and input the processed data into the target wind energy prediction model to output the prediction result corresponding to the target data at the time to be predicted.

[0014] In one embodiment, such as Figure 2 As shown, a deep learning-based wind energy prediction method is provided, which is then applied to... Figure 1 Taking server 102 as an example, the explanation includes the following steps: Step 201: Obtain raw wind energy data and preprocess it to generate standard wind energy data. Divide the standard wind energy data into a first wind energy dataset and a second wind energy dataset. Step 202: Generate multiple sets of model initialization parameters for initializing the preset deep learning prediction model, and determine the parameter evaluation values ​​of the multiple sets of model initialization parameters based on the first wind energy dataset and the preset parameter evaluation function. Step 203: Iteratively update multiple sets of model initialization parameters based on parameter evaluation values ​​until a preset number of iterations is reached. Then, determine the model initialization parameter corresponding to the optimal parameter evaluation value among the updated multiple sets of model initialization parameters as the target model initialization parameter. Step 204: Initialize the deep learning prediction model based on the target model initialization parameters, and train the initialized deep learning prediction model based on the first wind energy dataset to obtain the trained deep learning prediction model. Step 205: Based on the second wind energy dataset, determine whether the output accuracy of the trained deep learning prediction model meets expectations. If so, determine the trained deep learning prediction model as the target wind energy prediction model.

[0015] In the aforementioned deep learning-based wind energy prediction method, by optimizing the key initial parameters that affect the training effect before model training, the deep learning prediction model can enter the training process with a better initial state. This reduces training fluctuations caused by random initialization or empirical settings, lowers the possibility of the model getting stuck in local optima, improves the stability and convergence effect of model training, and helps to improve the accuracy of wind energy prediction results. At the same time, by further verifying the prediction ability of the model after training, it is possible to avoid judging the model performance solely based on the fitting results during the training phase. This allows for a more effective evaluation of the model's actual prediction accuracy, improving the generalization ability, prediction reliability, and application value of the target wind energy prediction model.

[0016] In one embodiment, acquiring raw wind energy data and preprocessing it to generate standard wind energy data includes: Acquire raw wind energy data and arrange it in chronological order of sampling time to generate raw wind energy time series data; In the original wind energy time series data, the window center sampling point is determined, and a moving data window centered on the wind energy data corresponding to the window center sampling point is determined according to the preset half window length; Determine the median value of the moving data window, and determine the deviation between the wind energy data corresponding to the sampling point at the center of the window and the median value; Based on the absolute deviation of wind energy data within the mobile data window relative to the median value, the deviation scale estimate is determined, and the anomaly judgment threshold is determined based on the preset threshold coefficient and the deviation scale estimate. When the deviation value is greater than the anomaly judgment threshold, the wind energy data corresponding to the sampling point at the center of the window is identified as an anomaly and replaced with the median value. When the deviation value is less than or equal to the anomaly detection threshold, the wind energy data corresponding to the sampling point at the center of the window is retained; Move the data window along the time direction of the original wind energy time series data and repeatedly perform outlier detection and data replacement processing until the preprocessing of the original wind energy time series data is completed and standard wind energy data is obtained.

[0017] Specifically, in this embodiment, original wind energy time-series data is constructed according to the sampling time sequence, and a moving data window is set around the central sampling point. The median value, deviation scale estimate, and anomaly detection threshold of the wind energy data within the window are used to identify and replace anomalies in the wind energy data corresponding to the central sampling point. This is essentially a correction of anomaly data based on local time-series distribution characteristics. Since the median and absolute deviation are less susceptible to extreme values ​​than the mean, this processing method can preserve the overall trend of the original wind energy time-series changes while reducing the interference of isolated anomalies and sudden abnormal fluctuations on data quality. This is beneficial for improving the stability and reliability of standard wind energy data, thereby improving the accuracy of subsequent deep learning prediction model training and prediction results.

[0018] In one specific embodiment, acquiring raw wind energy data and preprocessing it to generate standard wind energy data includes: The collected raw wind energy data are sorted in ascending order of sampling time to construct a raw wind energy time series dataset X, denoted as X={x1,x2,x...} p ,...,x n}, where x p This represents the wind energy data corresponding to the p-th sampling time, i.e., the central sampling point of the window, and n represents the total number of sampling times, i.e., the total number of time series data samples. The preset half-window length is set to a positive integer 'a', typically 3-5, corresponding to a total window length of 7-11. The moving data window is constructed with the p-th sampling time as the window center. Its definition is ; Calculate the moving data window Median m of all wind energy data p Its definition is m p =median{ }, where median{·} is the median operation function; Further calculate the absolute deviation between the wind energy data at the center sampling point of the window and the median value within the window. Its definition is ; Calculate the median scale estimate S of the absolute deviation p It is used to characterize the degree of dispersion of data within a window, and is defined as follows: Where 1.4826 is the conversion coefficient between the median and standard deviation of the absolute deviation under normal distribution, and e is the offset of the sampling time within the window relative to the center of the window; Set a preset threshold coefficient h, typically 3, corresponding to the 3σ criterion, and the anomaly detection threshold is hS. p Outlier correction is performed on the wind energy data at the center sampling point of the window, and the corrected standard wind energy data y p for: The data window is moved along the time sequence with a step size of 1, and the above calculation process is repeated until outlier corrections for all sampling points are completed, finally yielding a complete standard wind energy dataset Y, defined as Y={y1,y2,...,y...}. p ,..,y n}

[0019] In one embodiment, standard wind energy data is divided into a first wind energy dataset and a second wind energy dataset, including: Extract the data quality characteristic parameters and time distribution characteristic parameters corresponding to the standard wind energy data. Among them, the data quality characteristic parameters shall include at least one or more of the following: abnormal data ratio parameter, missing data ratio parameter, and the time distribution characteristic parameters shall include at least the preset time period coverage parameter. The first sample size is determined based on the total sample size, data quality characteristic parameters, and time distribution characteristic parameters of the standard wind energy data; Wind energy data that meets the first sample size is selected from the standard wind energy data as the first wind energy dataset, and the remaining standard wind energy data other than the first wind energy dataset are determined as the second wind energy dataset.

[0020] Specifically, in this embodiment, by extracting the data quality characteristic parameters and time distribution characteristic parameters corresponding to the standard wind energy data, and determining the first sample size based on the total sample size, the standard wind energy data is then divided. This ensures that the formation of the first and second wind energy datasets is no longer a simple random division, but rather takes into account both sample quality and time coverage. This avoids the excessive concentration of outlier, missing, or specific time-period data in a single dataset, which helps improve the representativeness of the data used for model training. It also enhances the objectivity and reliability of subsequent prediction accuracy verification based on the second wind energy dataset.

[0021] In one specific embodiment, standard wind energy data is divided into a first wind energy dataset and a second wind energy dataset, including: Let the total sample size of standard wind energy data be n, and define the first sample size as n. train The second sample size is n test Satisfying the constraint relationship n train +n test =n; Extract data quality feature parameters, which should include at least the outlier data percentage parameter and the missing data percentage parameter. The outlier data percentage parameter η... err η represents the ratio of the number of outliers replaced during preprocessing to the total sample size, and is the parameter representing the proportion of missing data. miss This is the ratio of the number of missing values ​​in the original data to the total sample size. The data quality correction coefficient γ is calculated based on the outlier data percentage parameter and the missing data percentage parameter, and is defined as follows: ; The value of γ ranges from (0,1], and the better the data quality, the closer γ is to 1; Extract time distribution characteristic parameters, namely the preset time period coverage parameter δ. Preferably, for the daily and annual period characteristics of wind energy data, if the standard wind energy data fully covers at least one complete daily period, δ is 0.95; if it fully covers at least one complete monthly period, δ is 0.98; if it fully covers at least one complete annual period, δ is 1. The basic splitting ratio coefficient β is set, with a default value of 0.75. The final formula for calculating the first sample size is as follows: ; in, This indicates the floor function, ensuring that the first sample size is a positive integer; Based on the time sequence, the first n elements of the standard wind energy dataset Y are selected. train The first wind energy dataset Y consists of several continuous time-series data points. train ; Remaining n test =n n train The second wind energy dataset Y consists of several continuous time-series data. test .

[0022] In one embodiment, multiple sets of model initialization parameters are generated for initializing a preset deep learning prediction model, and parameter evaluation values ​​for each set of model initialization parameters are determined based on a first wind energy dataset and a preset parameter evaluation function, including: Determine the initialization parameters of multiple models to be optimized in the deep learning prediction model, as well as the range of values ​​for the parameters corresponding to the multiple initialization parameters; Select parameter values ​​from the range of values ​​corresponding to multiple model initialization parameters, and combine them according to the correspondence between the multiple model initialization parameters to generate candidate parameter groups; Multiple sets of candidate parameter groups are determined as multiple sets of model initialization parameters, and the multiple sets of model initialization parameters are loaded into the deep learning prediction model respectively to perform pseudo-initialization of the deep learning prediction model and generate multiple pseudo-prediction models. Select at least one time point in the first wind energy dataset as the time point to be predicted, and input the wind energy data in the first wind energy dataset that are before the time point to be predicted into multiple prediction models to obtain the wind energy prediction values ​​for multiple time points to be predicted. The prediction bias of the corresponding prediction model is determined by the difference between the predicted wind energy value at the time of prediction and the actual wind energy value at the corresponding time of prediction in the first wind energy dataset. Based on the preset parameter evaluation function and multiple prediction biases, the parameter evaluation values ​​of multiple sets of model initialization parameters corresponding to multiple proposed prediction models are determined.

[0023] Specifically, in this embodiment, multiple model initialization parameters to be optimized in the deep learning prediction model and their value ranges are first determined. Then, multiple candidate parameter groups are formed within the corresponding parameter value ranges, and each candidate parameter group is loaded into the deep learning prediction model to generate multiple simulated prediction models. Furthermore, the parameter evaluation values ​​of each group of model initialization parameters are determined by combining the prediction deviation of the first wind energy dataset at the simulated prediction time. This allows the quality of the model initialization parameters to directly correspond to its prediction performance on actual wind energy data. Thus, the initialization parameter selection process, which originally relied on randomness or experience, can be transformed into a quantitative evaluation process based on data feedback, reducing the blindness of initialization parameter selection and providing a more reasonable initial parameter foundation for subsequent model training.

[0024] In one embodiment, multiple sets of model initialization parameters are iteratively updated based on parameter evaluation values ​​until a preset number of iterations is reached. The model initialization parameter corresponding to the optimal parameter evaluation value among the updated multiple sets of model initialization parameters is determined as the target model initialization parameter, including: For multiple sets of model initialization parameters, set the corresponding parameter update step size, local search trigger threshold, and parameter acceptance threshold respectively; Based on the parameter update step size, local search trigger threshold, parameter acceptance threshold, and parameter evaluation value, the initialization parameters of multiple sets of models are iteratively updated multiple times. The iterative update includes at least the following steps: Based on the current parameter evaluation values ​​corresponding to multiple sets of model initialization parameters, the model initialization parameter corresponding to the smallest current parameter evaluation value is determined as the current optimal model initialization parameter; Initialize parameters for multiple sets of models and generate corresponding random update factors for each; The model initialization parameters are updated based on the parameter difference between the model initialization parameters and the current optimal model initialization parameters, the parameter update step size corresponding to the model initialization parameters, and the random update factor, to obtain the updated model initialization parameters. When the updated model initialization parameters exceed the corresponding parameter value range, the excess part will be adjusted to the corresponding parameter value range; A first random number is generated based on a preset randomization strategy; In response to the first random number being greater than the local search trigger threshold corresponding to the updated model initialization parameters, the updated model initialization parameters are locally adjusted within a preset perturbation range, using the current optimal model initialization parameters as a perturbation reference, to generate locally adjusted model initialization parameters, and the locally adjusted model initialization parameters are determined as candidate model initialization parameters. Determine the parameter evaluation values ​​corresponding to the updated model initialization parameters and the parameter evaluation values ​​corresponding to the candidate model initialization parameters respectively; A second random number is generated based on a randomization strategy; If the second random number is less than the parameter acceptance threshold corresponding to the updated model initialization parameters, and the parameter evaluation value corresponding to the candidate model initialization parameters is less than the parameter evaluation value corresponding to the updated model initialization parameters, then the updated model initialization parameters are replaced with the candidate model initialization parameters. Update the parameter update step size, local search trigger threshold, and parameter acceptance threshold corresponding to the initialization parameters of multiple sets of models, so as to execute the next iteration update, until the preset number of iterations is reached.

[0025] Specifically, in this embodiment, parameter update step size, local search trigger threshold, and parameter acceptance threshold are set for multiple sets of model initialization parameters. During the iteration process, directional updates are performed based on the current optimal model initialization parameters. Simultaneously, a random update factor, local adjustment strategy, candidate parameter acceptance mechanism, and parameter value range constraints are combined to form an initialization parameter optimization process that balances global search and local refinement. Therefore, on the one hand, it guides the initialization parameters of each set of models to gradually converge towards a better parameter region; on the other hand, it retains necessary random perturbations and local exploration capabilities, reducing the possibility of the parameter search process prematurely getting trapped in a local optimum. This is beneficial for improving the efficiency of initialization parameter optimization and the stability and convergence effect of subsequent model training.

[0026] In one specific embodiment, the initialization parameters of the model to be optimized are the initial weight matrices of the restricted Boltzmann machines of each layer of the deep belief network. The swarm intelligence algorithm used for parameter optimization is preferably the bat algorithm, which generates and evaluates multiple sets of initialization parameters, specifically including: The total number of parameters to be optimized in the initial weight matrix of the deep belief network is defined as D, i.e., the dimension of the parameter search space is D. The total number of initial parameter sets, i.e., the population size, is set to K, typically ranging from 20 to 50. The spatial location vector of the individual corresponding to the i-th set of model initialization parameters is G. iD where i=1,2, K, the range of values ​​for each group of parameters is constrained as follows: Among them, Gmin =-1、G max =1 represents the lower and upper limits of the values ​​of the weight parameters to be optimized, which conforms to the conventional value range for neural network weight initialization; Within the range of values, K sets of initial spatial location vectors are randomly generated, which are K sets of model initialization parameters. Each set of parameters is mapped to the initial weight matrix of the deep belief network to complete the model initialization and generate K simulated prediction models. In the first wind energy dataset, m consecutive time points are selected as the proposed prediction time points, where m is typically 100. Historical wind energy data before the proposed prediction time points are input into each proposed prediction model to obtain the proposed wind energy value for each proposed prediction time point. Where i is the proposed prediction model number and j is the proposed prediction time number; The preset parameter evaluation function adopts the fitness function in the form of mean squared error. The parameter evaluation value of the i-th group of model initialization parameters is Fitness(G). iD ), defined as: ; in, The actual standard wind energy data corresponding to the j-th predicted time in the first wind energy dataset is denoted as . The smaller the parameter evaluation value, the lower the prediction deviation of the predicted model corresponding to the initial parameters of this set, and the better the parameter performance.

[0027] In one specific embodiment, multiple sets of model initialization parameters are iteratively updated based on parameter evaluation values ​​until a preset number of iterations is reached. The model initialization parameter corresponding to the optimal parameter evaluation value among the updated multiple sets of model initialization parameters is determined as the target model initialization parameter, including: First, initialize the iteration hyperparameters and set the maximum number of iterations (Iter). max The typical value is 100-500, and the search precision ε is set, with a typical value of 10. 6 Set the search pulse frequency range [f] min ,f max ], typically taking the value f min =0、f max =2, set the initial local search trigger threshold. The typical value is 0.5, which sets the initial parameter acceptance threshold. The typical value is 0.9. The pulse frequency increase coefficient λ and the sound intensity attenuation coefficient α are both set to a constant of 0.9. The initial update step size of the i-th group of parameters is initialized. =0, and its dimension is consistent with the dimension D of the parameter search space; Let the current iteration number be t, where t = 1, 2, Iter maxDuring the iteration process, the current parameter evaluation values ​​of all parameter groups are first calculated, and the model initialization parameters corresponding to the minimum evaluation value are determined as the current optimal model initialization parameters G. best Its formula is expressed as: Where argmin represents the independent variable when the function reaches its minimum value; For the i-th set of parameters, a random update factor rand uniformly distributed in the interval [0,1] is generated to update the search pulse frequency, parameter update step size, and parameter spatial location. The calculation formula is as follows: in, , These represent the spatial positions of the i-th set of parameters in the t-1 and t-1 iterations, respectively. , These are the parameter update step sizes for the corresponding iteration numbers; The updated parameters are subject to value range constraints to prevent them from exceeding reasonable ranges. The constraint formula is as follows: ; Generate a first random number rand1 uniformly distributed in the interval [0,1]. The local search trigger threshold for the t-th iteration is... When rand1 > At that time, a local search is performed to generate candidate model initialization parameters G. candidate Its definition is: ; in, ∈[-1,1] represents a uniformly distributed random perturbation factor. This is the average of the acceptance thresholds for all parameter groups in the t-th iteration; Calculate the parameter evaluation values ​​of the updated parameters and candidate parameters respectively, generate a second random number rand2 uniformly distributed in the interval [0,1], and set the parameter acceptance threshold for the t-th iteration as follows. When rand2 < and When the parameter is selected, the replacement is performed: = ; After completing this round of parameter updates, update the local search trigger threshold and parameter acceptance threshold, calculated using the following formula: The iteration termination condition is reaching the preset maximum number of iterations (Iter). max Or the current optimal parameter evaluation value satisfies Fitness(G) best If ε < ε, after the iteration terminates, initialize the final current optimal model parameters G. bestThese parameters are determined as the initialization parameters for the target model and are used for the initialization of the subsequent deep belief network model.

[0028] In one embodiment, initializing a preset deep learning prediction model based on target model initialization parameters includes: Construct a pre-defined deep learning prediction model, which includes at least multiple input layer nodes and multiple hidden layer nodes; Input the first wind energy dataset into the input layer to determine the node status of multiple input layer nodes; The node states of multiple hidden layer nodes are determined based on the node states of multiple input layer nodes and the initialization parameters of the target model. Update the node states of multiple input layer nodes based on the node states of multiple hidden layer nodes; Based on the updated node states of multiple input layer nodes, the node states of multiple hidden layer nodes are updated again. Based on the updated node states of multiple hidden layer nodes, the preset deep learning prediction model is initialized.

[0029] Specifically, in this embodiment, by inputting the first wind energy dataset into the input layer and interactively updating the node states of the input layer and hidden layer based on the initialization parameters of the target model, the initialization of the preset deep learning prediction model is not only manifested as parameter loading but also as a gradual adjustment process of node states driven by data. This allows each layer of nodes to form an initial state more adapted to the distribution characteristics of the first wind energy dataset before training begins, thereby improving the starting conditions for model training and enhancing convergence consistency and training stability in subsequent training processes.

[0030] In one specific embodiment, the preset deep learning prediction model is a deep belief network basic unit composed of a single-layer Restricted Boltzmann Machine (RBM). The target model initialization parameters are mapped to the initial weight matrix between the input layer and hidden layer of the RBM, where the visible layer of the RBM is the input layer of the deep learning prediction model. The preset deep learning prediction model is then initialized based on the target model initialization parameters, including: First, a restricted Boltzmann machine network structure is constructed. The number of input layer nodes is defined as I, which is consistent with the temporal feature dimension of the input wind energy data. The state of the i-th input layer node is S. i Where i = 1, 2, ..., I, the number of hidden layer nodes is defined as J, typically 2 to 3 times the number of input layer nodes, and the state of the j-th hidden layer node is S. j Where j=1,2, J, the initial parameters of the target model are mapped to the initial weight matrix ω between the input layer and the hidden layer. ij , where ωij Let be the connection weight between the i-th input layer node and the j-th hidden layer node; The standardized sample data from the first wind energy dataset is input into the input layer to initialize the node states of the input layer, i.e., S. i Based on the time-series wind energy sample values ​​in the first wind energy dataset, and using the input layer node states and the initial weight matrix, the activation states of the hidden layer nodes are calculated using the following formula: in, For a sigmoid activation function with upper and lower bounds, N j (0,1) represents a Gaussian random variable unit with a mean of 0 and a variance of 1, used to introduce random noise to enhance the robustness of the model. σ∈(0,1) is the noise intensity coefficient, typically taking a value of 0.1. θ H =1、θ L =0 represent the upper and lower bounds of the activation function, respectively, and α j =1 is a noise control parameter used to control the slope of the Sigmoid function, e is the base of the natural logarithm, and t j The activation input value for the hidden layer node; Based on the hidden layer node states, the input layer node states are reconstructed and updated. The calculation formula is as follows: ; Where, N i (0,1) represents the Gaussian random variable unit corresponding to the input layer node; Based on the reconstructed input layer node states, the hidden layer node states are updated again, and the calculation formula is as follows: ; in, To reconstruct and update the state of the input layer nodes; After completing the bidirectional state update of the input layer and hidden layer, the initialization of the deep belief network is completed, so that the state of the network nodes is adapted to the distribution characteristics of the input wind energy data, laying the foundation for subsequent unsupervised training.

[0031] In one embodiment, the initialized deep learning prediction model is trained based on a first wind energy dataset to generate a target wind energy prediction model, including: A preset number of candidate hidden layers for a deep learning prediction model is set, and the initialized deep learning prediction model is trained in multiple rounds based on the number of candidate hidden layers. The training includes at least the following steps: under the number of candidate hidden layers, the node state of the previous hidden layer node is used as the input of the next hidden layer node to update the node state of multiple hidden layer nodes of the preset deep learning prediction model layer by layer, so as to complete one round of training for multiple hidden layer nodes under the number of candidate hidden layers; after completing one round of training, the weight parameters of the preset deep learning prediction model are updated based on the training samples selected from the first wind energy dataset. Repeat the layer-by-layer update and weight parameter update until the preset training termination condition is met, and obtain the prediction result data corresponding to the candidate hidden layer number; Based on the prediction results data corresponding to the candidate hidden layer number, the reconstruction error corresponding to the candidate hidden layer number is determined, and based on the reconstruction error, the number of hidden layers of the preset deep learning prediction model is increased to obtain the increased candidate hidden layer number. Based on the increased number of candidate hidden layers, the initial preset deep learning prediction model is repeatedly updated layer by layer, weight parameters are updated, prediction result data is determined, and reconstruction error is determined to obtain the reconstruction error corresponding to the increased number of candidate hidden layers. Compare the reconstruction error corresponding to the increased number of candidate hidden layers with the reconstruction error corresponding to the number of candidate hidden layers before the increase. If the reconstruction error corresponding to the increased number of candidate hidden layers is greater than the reconstruction error corresponding to the number of candidate hidden layers before the increase, then stop increasing the number of hidden layers and determine the number of candidate hidden layers before the increase as the target number of hidden layers. If the reconstruction error corresponding to the increased number of candidate hidden layers is less than or equal to the reconstruction error corresponding to the number of candidate hidden layers before the increase, then the increased number of candidate hidden layers is determined as the new number of candidate hidden layers, and the steps of increasing the number of hidden layers and determining the corresponding reconstruction error are repeated until the target number of hidden layers is determined. Based on the training results corresponding to the target hidden layer number, a target wind energy prediction model is generated.

[0032] Specifically, in this embodiment, by setting a candidate number of hidden layers and conducting multiple rounds of training at different numbers of hidden layers to determine the prediction results and reconstruction errors, the decision to continue adding layers is made based on the change in reconstruction error before and after increasing the number of hidden layers. This ensures that the determination of the number of hidden layers is based on a comparison of actual training results, rather than relying solely on empirical presets. This allows the model complexity to better match the nonlinear characteristics and variation patterns of wind energy data itself, reducing the risk of insufficient model representation ability due to an insufficient number of hidden layers, and overfitting or training redundancy due to an excessive number of hidden layers. This is beneficial for improving the model's ability to extract features from wind energy data and the final prediction accuracy.

[0033] In one specific embodiment, the model training process employs a layer-by-layer unsupervised pre-training combined with adaptive layer determination based on reconstruction error. First, the training parameters are initialized, setting the initial number of candidate hidden layers w=1, meaning the initial deep belief network consists of a single restricted Boltzmann layer. The model learning rate is then set. The typical value is 0.01~0.1. The search precision ε is set to the preset training termination condition, which is consistent with the precision in the parameter optimization stage. With candidate hidden layer number w, the hidden layer node state of the previous restricted Boltzmann machine is used as the input layer input of the next restricted Boltzmann machine. The input layer-hidden layer state of all restricted Boltzmann machine layers is updated layer by layer in sequence to complete one round of training. After one round of training, the weight matrix of each restricted Boltzmann machine layer is updated, with the weight update amount Δω. ij The calculation formula is: ; in, This represents the joint expectation of the input layer and hidden layer node states under the original input data. This is the joint expectation value of the input layer and hidden layer node states after reconstruction; Repeat the layer-by-layer state update and weight update until the preset training termination condition is met: ,in, For the Frobenius norm operation of a matrix; After training, output the prediction result dataset Z of the deep belief network model for the first wind energy dataset, given the current candidate hidden layer number w. Z is defined as follows: , where z k Let n be the model prediction result value corresponding to the k-th training sample. train This represents the total sample size of the first wind energy dataset; The reconstruction error RE(w) corresponding to the current candidate hidden layer number is calculated using the mean squared error form. The calculation formula is as follows: ; Among them, y k This represents the actual standard wind energy data of the kth sample in the first wind energy dataset. Increase the number of candidate hidden layers by 1, i.e., w = w + 1, and repeat the above training and reconstruction error calculation process to obtain the reconstruction error RE(w) corresponding to the new layer number. Compare the reconstruction error after adding the new layer number with the reconstruction error of the previous layer number. If the following conditions are met: If the target number of hidden layers is w-1, then stop increasing the number of hidden layers and set the target number of hidden layers to w-1. Otherwise, continue to increase the number of layers and repeat the above process until the above stopping condition is met. At the same time, set the maximum number of hidden layers to 5 to avoid training overfitting and computational redundancy caused by too many layers. Based on the weight parameters and network structure corresponding to the number of hidden layers of the target, the final target wind energy prediction model is generated.

[0034] In one embodiment, based on a second wind energy dataset, determining whether the output accuracy of the trained deep learning prediction model meets expectations includes: Multiple times are selected from the second wind energy dataset as prediction times. Wind energy data in the second wind energy dataset that are before the multiple prediction times are input into the trained deep learning prediction model to obtain the wind energy prediction values ​​corresponding to the multiple prediction times. The wind energy prediction values ​​at multiple prediction times are compared with the actual wind energy values ​​in the second wind energy dataset to determine the wind energy prediction difference at multiple prediction times. Based on the preset accuracy evaluation rules and the wind energy prediction difference at multiple prediction times, the prediction accuracy evaluation result of the trained deep learning prediction model is determined. The prediction accuracy evaluation results are compared with the expected accuracy evaluation results to determine whether the prediction accuracy of the trained deep learning prediction model meets expectations.

[0035] Specifically, in this embodiment, multiple prediction times are selected from the second wind energy dataset. Wind energy data prior to the corresponding prediction times are input into the trained deep learning prediction model. The obtained wind energy prediction values ​​are compared with the actual wind energy values. Furthermore, a prediction accuracy evaluation result is generated based on preset accuracy evaluation rules, and then compared with the expected accuracy evaluation result. This effectively constructs a prediction performance verification step independent of the training process. Since this verification is based on data not involved in model training and covers multiple prediction times, it can more objectively reflect the prediction ability of the trained model for new samples, which is beneficial to improving the reliability of the target wind energy prediction model selection results and its usability in practical applications.

[0036] In one specific embodiment, prediction accuracy verification is achieved using three quantitative evaluation metrics. Based on the second wind energy dataset, determining whether the output accuracy of the trained deep learning prediction model meets expectations includes: Second wind energy dataset Y test All consecutive moments are selected as prediction moments, with a total number of moments m=n. test Historical wind energy data prior to each prediction time is input into the trained deep learning prediction model to obtain the wind energy prediction value corresponding to each prediction time. , where b=1,2,...,m are the prediction time numbers; Based on the difference between the predicted and actual values, three accuracy evaluation indicators are calculated, namely, the mean absolute error percentage e. MAPE Coefficient of determination e R square Root mean square error e RMSE The calculation formula is as follows: Among them, y b Let b be the actual standard wind energy value corresponding to the b-th prediction time. is the global average of all actual wind energy values ​​in the second wind energy dataset. The mean absolute error percentage is used to characterize the relative deviation of the prediction. The coefficient of determination is used to characterize the model's ability to fit the data, with a maximum value of 1. The root mean square error is used to characterize the absolute deviation of the prediction. The specific evaluation criteria for the expected accuracy of the three indicators are as follows: The mean absolute error percentage satisfies e MAPE <10% is considered an excellent prediction result, 10% ≤ e MAPE <20% is considered an acceptable prediction result. MAPE ≥20% indicates the model needs improvement; the coefficients of determination satisfy e R square A value >0.9 indicates an excellent model fit, and 0.7 ≤ e R square ≤0.9 indicates a good prediction result, e R square A value <0.5 indicates the model needs optimization; a root mean square error (RMSE) less than 20% of the mean actual wind energy value in the second wind energy dataset indicates a usable prediction result, while a value greater than this threshold indicates the model needs optimization. When all three indicators meet the expected accuracy requirements, the output accuracy of the trained deep learning prediction model is deemed satisfactory, and it can be identified as the target wind energy prediction model. If any indicator fails to meet the requirements, the process returns to the parameter optimization step, and parameter initialization and model training are repeated.

[0037] In one embodiment, after determining the trained deep learning prediction model as the target wind energy prediction model, the method further includes: The real-time wind energy data before the time to be predicted is obtained and preprocessed to obtain the predicted input wind energy data; The predicted wind energy data is input into the input layer of the target wind energy prediction model to determine the node status of multiple input layer nodes; Based on the node states of multiple input layer nodes and the weight parameters between the input layer and the first hidden layer in the target wind energy prediction model, the node states of multiple hidden layer nodes in the first hidden layer are determined. If the target wind energy prediction model includes multiple hidden layers, then based on the node states of multiple hidden layer nodes in the previous hidden layer and the weight parameters between adjacent layers, the node states of multiple hidden layer nodes in the next hidden layer are determined layer by layer until the node states of multiple hidden layer nodes in the last hidden layer are obtained. Output mapping is performed on the node states of multiple hidden layer nodes in the last hidden layer to obtain the wind energy prediction data corresponding to the time to be predicted. The output mapping includes multiplying the node states of multiple hidden layer nodes in the last hidden layer by their corresponding output weight parameters and summing them, and then combining them with the output bias parameters to obtain the wind energy prediction value corresponding to the time to be predicted.

[0038] Specifically, in this embodiment, after determining the target wind energy prediction model, real-time wind energy data prior to the prediction time is acquired and preprocessed. Then, the predicted input wind energy data is input into the input layer of the target wind energy prediction model. The state of each node is determined layer by layer based on the weight parameters between the input layer and each hidden layer. Finally, the wind energy prediction value corresponding to the prediction time is obtained by combining the output weight parameters and output bias parameters, forming a complete application chain from real-time data acquisition to prediction result output. Therefore, the trained target wind energy prediction model can be directly applied to actual wind energy prediction scenarios. While ensuring the connection between the input data and the model calculation process, it achieves the quantitative output of wind energy data at the prediction time, which is beneficial to improving the practicality and prediction response capability of this method in actual deployment.

[0039] In one specific embodiment, the online prediction process of the target wind energy prediction model is performed according to the following steps: First, real-time wind energy data for I consecutive time steps before the time to be predicted is obtained, where I is the number of nodes in the model input layer. The same preprocessing method as in the training phase is used to correct outliers in the real-time data to obtain standardized prediction input wind energy data. The predicted wind energy data is input into the input layer of the target wind energy prediction model to obtain the state of the input layer nodes. The superscript 0 represents the input layer, i.e., the 0th layer; Based on the weight matrix ωij(1) between the input layer and the first hidden layer, calculate the node state of the first hidden layer. The calculation formula is as follows: ;in, These are the bias parameters for the first hidden layer; If the target wind energy prediction model contains w hidden layers, then following the above method, based on the node states of the previous hidden layer and the weight parameters of the adjacent layers, the node states of the next hidden layer are calculated layer by layer until the node states of the last hidden layer, i.e., the w-th layer, are obtained. ; Let J be the number of nodes in the last hidden layer. last The corresponding output weight parameter is ω. j,out The output bias parameter is b. out Then the wind energy prediction value corresponding to the time to be predicted The output mapping formula is: ; Among them, the output weight parameters and output bias parameters are output layer parameters determined through supervised fine-tuning after the model training is completed, which can further improve the prediction accuracy of the model; the final output wind energy prediction value can be directly used in practical application scenarios such as power scheduling of wind farms, grid connection planning, and unit operation and maintenance planning.

[0040] It should be understood that, although Figure 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0041] In one embodiment, such as Figure 3 As shown, a deep learning-based wind energy prediction system is provided, including: a raw data processing module, an initialization parameter evaluation module, an initialization parameter determination module, a prediction model training module, and a prediction model determination module, wherein: The raw data processing module is used to acquire raw wind energy data and preprocess it to generate standard wind energy data, which is then divided into a first wind energy dataset and a second wind energy dataset. The initialization parameter evaluation module is used to generate multiple sets of model initialization parameters for initializing a preset deep learning prediction model, and to determine the parameter evaluation values ​​of the multiple sets of model initialization parameters based on the first wind energy dataset and the preset parameter evaluation function. The initialization parameter determination module is used to iteratively update multiple sets of model initialization parameters based on parameter evaluation values ​​until a preset number of iterations is reached. The model initialization parameter corresponding to the optimal parameter evaluation value among the updated multiple sets of model initialization parameters is determined as the target model initialization parameter. The prediction model training module is used to initialize the deep learning prediction model based on the initialization parameters of the target model, and to train the initialized deep learning prediction model based on the first wind energy dataset to obtain the trained deep learning prediction model. The prediction model determination module is used to determine whether the output accuracy of the trained deep learning prediction model meets expectations based on the second wind energy dataset. If so, the trained deep learning prediction model is determined as the target wind energy prediction model.

[0042] Specific limitations regarding the deep learning-based wind energy forecasting system can be found in the limitations of the deep learning-based wind energy forecasting method described above, and will not be repeated here. The modules in the aforementioned deep learning-based wind energy forecasting system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0043] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores wind energy forecasting data based on deep learning. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a deep learning-based wind energy forecasting method.

[0044] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0045] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned deep learning-based wind energy prediction method.

[0046] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the aforementioned deep learning-based wind energy prediction method.

[0047] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0048] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0049] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A wind energy prediction method based on deep learning, characterized in that, include: Raw wind energy data is acquired and preprocessed to generate standard wind energy data, which is then divided into a first wind energy dataset and a second wind energy dataset. Generate multiple sets of model initialization parameters for initializing a preset deep learning prediction model, and determine the parameter evaluation values ​​of the multiple sets of model initialization parameters based on the first wind energy dataset and the preset parameter evaluation function. Based on the parameter evaluation value, the multiple sets of model initialization parameters are iteratively updated until a preset number of iterations is reached. The model initialization parameter corresponding to the optimal parameter evaluation value among the multiple sets of updated model initialization parameters is determined as the target model initialization parameter. The deep learning prediction model is initialized based on the target model initialization parameters, and the initialized deep learning prediction model is trained based on the first wind energy dataset to obtain the trained deep learning prediction model. Based on the second wind energy dataset, determine whether the output accuracy of the trained deep learning prediction model meets expectations; if so, determine the trained deep learning prediction model as the target wind energy prediction model.

2. The wind energy prediction method based on deep learning according to claim 1, characterized in that, The process of acquiring raw wind energy data and preprocessing it to generate standard wind energy data includes: Acquire raw wind energy data and arrange it in chronological order of sampling time to generate raw wind energy time series data; In the original wind energy time series data, a window center sampling point is determined, and a moving data window centered on the wind energy data corresponding to the window center sampling point is determined according to a preset half-window length; Determine the median value of the moving data window, and determine the deviation between the wind energy data corresponding to the center sampling point of the window and the median value; Based on the absolute deviation of the wind energy data within the mobile data window relative to the median value, a deviation scale estimate is determined, and an anomaly detection threshold is determined based on a preset threshold coefficient and the deviation scale estimate. When the deviation value is greater than the anomaly determination threshold, the wind energy data corresponding to the sampling point at the center of the window is determined as an anomaly, and the anomaly value is replaced by the median value. When the deviation value is less than or equal to the anomaly determination threshold, the wind energy data corresponding to the sampling point at the center of the window is retained; The moving data window is moved along the time direction of the original wind energy time series data, and the outlier detection and data replacement processing are repeated until the preprocessing of the original wind energy time series data is completed, and the standard wind energy data is obtained.

3. The wind energy prediction method based on deep learning according to claim 1, characterized in that, The standard wind energy data is divided into a first wind energy dataset and a second wind energy dataset, including: Extract the data quality feature parameters and time distribution feature parameters corresponding to the standard wind energy data. The data quality feature parameters include at least one or more of the following: abnormal data ratio parameter and missing data ratio parameter. The time distribution feature parameters include at least a preset time period coverage parameter. The first sample size is determined based on the total sample size of the standard wind energy data, the data quality characteristic parameters, and the time distribution characteristic parameters; Wind energy data that meets the first sample size is selected from the standard wind energy data to form the first wind energy dataset, and the remaining standard wind energy data other than the first wind energy dataset are determined as the second wind energy dataset.

4. The wind energy prediction method based on deep learning according to claim 1, characterized in that, The process of generating multiple sets of model initialization parameters for initializing a preset deep learning prediction model, and determining parameter evaluation values ​​for each set of model initialization parameters based on the first wind energy dataset and the preset parameter evaluation function, includes: Determine multiple model initialization parameters to be optimized in the deep learning prediction model, as well as the parameter value ranges corresponding to the multiple model initialization parameters; Select parameter values ​​from the range of values ​​corresponding to multiple model initialization parameters, and combine them according to the correspondence between the multiple model initialization parameters to generate candidate parameter groups; Multiple sets of candidate parameter groups are determined as multiple sets of model initialization parameters, and the multiple sets of model initialization parameters are loaded into the deep learning prediction model respectively to perform pseudo-initialization of the deep learning prediction model and generate multiple pseudo-prediction models. At least one moment in the first wind energy dataset is selected as the time to be predicted. The wind energy data in the first wind energy dataset that are before the time to be predicted are input into multiple prediction models to obtain multiple wind energy prediction values ​​at the time to be predicted. The prediction deviation of the proposed prediction model is determined based on the difference between the predicted wind energy value at the proposed prediction time and the actual wind energy value in the first wind energy dataset at the proposed prediction time. Based on a preset parameter evaluation function and multiple prediction biases, parameter evaluation values ​​of multiple sets of model initialization parameters corresponding to multiple proposed prediction models are determined.

5. The wind energy prediction method based on deep learning according to claim 4, characterized in that, The step of iteratively updating multiple sets of model initialization parameters based on the parameter evaluation values ​​until a preset number of iterations is reached, and determining the model initialization parameter corresponding to the optimal parameter evaluation value among the updated multiple sets of model initialization parameters as the target model initialization parameter, includes: For multiple sets of model initialization parameters, respectively set the corresponding parameter update step size, local search trigger threshold and parameter acceptance threshold; Based on the parameter update step size, the local search trigger threshold, the parameter acceptance threshold, and the parameter evaluation value, multiple sets of model initialization parameters are iteratively updated, wherein the iterative update includes at least the following steps: Based on the current parameter evaluation values ​​corresponding to the multiple sets of model initialization parameters, the model initialization parameter corresponding to the smallest current parameter evaluation value is determined as the current optimal model initialization parameter; For each set of model initialization parameters, a corresponding random update factor is generated. The model initialization parameters are updated based on the parameter difference between the model initialization parameters and the current optimal model initialization parameters, the parameter update step size corresponding to the model initialization parameters, and the random update factor, to obtain the updated model initialization parameters. A first random number is generated based on a preset randomization strategy; In response to the first random number being greater than the local search trigger threshold corresponding to the updated model initialization parameters, the updated model initialization parameters are locally adjusted within a preset perturbation range using the current optimal model initialization parameters as a perturbation reference, generating locally adjusted model initialization parameters, and the locally adjusted model initialization parameters are determined as candidate model initialization parameters. Determine the parameter evaluation values ​​corresponding to the updated model initialization parameters and the parameter evaluation values ​​corresponding to the candidate model initialization parameters, respectively. A second random number is generated based on the aforementioned randomization strategy; In response to the second random number being less than the parameter acceptance threshold corresponding to the updated model initialization parameter, and the parameter evaluation value corresponding to the candidate model initialization parameter being less than the parameter evaluation value corresponding to the updated model initialization parameter, the updated model initialization parameter is replaced with the candidate model initialization parameter. The parameter update step size, local search trigger threshold, and parameter acceptance threshold corresponding to the multiple sets of model initialization parameters are updated respectively, so as to execute the next iteration update, until the preset number of iterations is reached.

6. The wind energy prediction method based on deep learning according to claim 1, characterized in that, The initialization of the preset deep learning prediction model based on the target model initialization parameters includes: Construct the preset deep learning prediction model, which includes at least multiple input layer nodes and multiple hidden layer nodes; The first wind energy dataset is input into the input layer to determine the node status of multiple input layer nodes; The node states of the hidden layer nodes are determined based on the node states of the multiple input layer nodes and the initialization parameters of the target model. Update the node states of the input layer nodes based on the node states of the hidden layer nodes. Based on the updated node states of the multiple input layer nodes, the node states of the multiple hidden layer nodes are updated again. Based on the updated node states of the multiple hidden layer nodes, the initialization of the preset deep learning prediction model is completed.

7. The wind energy prediction method based on deep learning according to claim 6, characterized in that, The step of training the initialized deep learning prediction model based on the first wind energy dataset to generate the target wind energy prediction model includes: The preset deep learning prediction model is given a candidate number of hidden layers, and the initialized preset deep learning prediction model is trained in multiple rounds based on the candidate number of hidden layers. The training includes at least the following steps: at the candidate number of hidden layers, the node state of the previous hidden layer node is used as the input of the next hidden layer node to update the node state of the multiple hidden layer nodes of the preset deep learning prediction model layer by layer, so as to complete one round of training of the multiple hidden layer nodes at the candidate number of hidden layers; after completing one round of training, the weight parameters of the preset deep learning prediction model are updated based on the training samples selected from the first wind energy dataset. Repeat the layer-by-layer update and weight parameter update until the preset training termination condition is met, and obtain the prediction result data corresponding to the candidate hidden layer number; Based on the prediction result data corresponding to the candidate hidden layer number, the reconstruction error corresponding to the candidate hidden layer number is determined, and based on the reconstruction error, the number of hidden layers of the preset deep learning prediction model is increased to obtain the increased candidate hidden layer number. Based on the increased number of candidate hidden layers, the layer-by-layer update, weight parameter update, prediction result data determination, and reconstruction error determination are repeatedly performed on the initialized preset deep learning prediction model to obtain the reconstruction error corresponding to the increased number of candidate hidden layers. Compare the reconstruction error corresponding to the increased number of candidate hidden layers with the reconstruction error corresponding to the number of candidate hidden layers before the increase. If the reconstruction error corresponding to the increased number of candidate hidden layers is greater than the reconstruction error corresponding to the number of candidate hidden layers before the increase, then the increase of the number of hidden layers is stopped, and the number of candidate hidden layers before the increase is determined as the target number of hidden layers. If the reconstruction error corresponding to the increased number of candidate hidden layers is less than or equal to the reconstruction error corresponding to the number of candidate hidden layers before the increase, then the increased number of candidate hidden layers is determined as the new number of candidate hidden layers, and the steps of increasing the number of hidden layers and determining the corresponding reconstruction error are repeated until the target number of hidden layers is determined. Based on the training results corresponding to the target hidden layer number, the target wind energy prediction model is generated.

8. The wind energy prediction method based on deep learning according to claim 7, characterized in that, The step of determining whether the output accuracy of the trained deep learning prediction model meets expectations based on the second wind energy dataset includes: Multiple times are selected from the second wind energy dataset as prediction times. Wind energy data in the second wind energy dataset that are before the multiple prediction times are input into the trained deep learning prediction model to obtain wind energy prediction values ​​corresponding to the multiple prediction times. The wind energy prediction values ​​corresponding to multiple prediction times are compared with the corresponding actual wind energy values ​​in the second wind energy dataset to determine the wind energy prediction difference for multiple prediction times. Based on the preset accuracy evaluation rules and the wind energy prediction difference at multiple prediction times, the prediction accuracy evaluation result of the trained deep learning prediction model is determined. The prediction accuracy evaluation result is compared with the expected accuracy evaluation result to determine whether the prediction accuracy of the trained deep learning prediction model meets expectations.

9. The wind energy prediction method based on deep learning according to claim 7, characterized in that, After determining the trained deep learning prediction model as the target wind energy prediction model, the method further includes: The real-time wind energy data before the time to be predicted is obtained and preprocessed to obtain the predicted input wind energy data; The predicted wind energy data is input into the input layer of the target wind energy prediction model to determine the node status of multiple input layer nodes; Based on the node states of multiple input layer nodes and the weight parameters between the input layer and the first hidden layer in the target wind energy prediction model, the node states of multiple hidden layer nodes in the first hidden layer are determined. If the target wind energy prediction model includes multiple hidden layers, then based on the node states of multiple hidden layer nodes in the previous hidden layer and the weight parameters between adjacent layers, the node states of multiple hidden layer nodes in the next hidden layer are determined layer by layer until the node states of multiple hidden layer nodes in the last hidden layer are obtained. The node states of multiple hidden layer nodes in the last hidden layer are output and mapped to obtain the wind energy prediction data corresponding to the time to be predicted.

10. A wind energy forecasting system based on deep learning, characterized in that, The system includes: The raw data processing module is used to acquire raw wind energy data and preprocess it to generate standard wind energy data, and divide the standard wind energy data into a first wind energy dataset and a second wind energy dataset. The initialization parameter evaluation module is used to generate multiple sets of model initialization parameters for initializing a preset deep learning prediction model, and to determine the parameter evaluation values ​​of the multiple sets of model initialization parameters based on the first wind energy dataset and the preset parameter evaluation function. The initialization parameter determination module is used to iteratively update multiple sets of model initialization parameters according to the parameter evaluation value until a preset number of iterations is reached, and determine the model initialization parameter corresponding to the optimal parameter evaluation value among the updated multiple sets of model initialization parameters as the target model initialization parameter. The prediction model training module is used to initialize the deep learning prediction model based on the target model initialization parameters, and train the initialized deep learning prediction model based on the first wind energy dataset to obtain the trained deep learning prediction model. The prediction model determination module is used to determine whether the output accuracy of the trained deep learning prediction model meets expectations based on the second wind energy dataset. If so, the trained deep learning prediction model is determined as the target wind energy prediction model.

Images