Federal split learning-based offshore wind farm power generation power short-term prediction method and system

Abstract

The application provides a kind of offshore wind farm power generation power short-term prediction method and system based on federal segmentation learning, through the collaborative work of distributed client, edge server and cloud, realize the efficient processing and prediction of offshore wind farm power generation power data. Through the most efficient model segmentation, the machine learning model of the cloud is divided into two parts: feature extractor and regressor. Send the model information of the offshore wind farm to the cloud and apply to join the cooperative learning task. The cloud proposes a two-stage cooperative learning task, groups the client nodes according to their computing power, and uses different complexity models for federal learning. After completing the joint grouping learning, the local model is trained. After training, the model parameters are uploaded to the edge, and after two-stage verification, they are used for global model aggregation update. Finally, the cloud obtains the feature extractor and complex model of all groups for offshore wind farm power generation power short-term prediction.

Description

Technical Field

[0001] This invention belongs to the field of offshore wind power generation prediction technology, specifically a method and system for short-term prediction of offshore wind farm power generation based on federated segmentation learning. Background Technology

[0002] Offshore wind power is a renewable energy technology that uses wind turbines installed on the sea surface to generate electricity from wind power, and it is developing rapidly worldwide. This energy technology captures the power of sea breezes by installing various wind turbines on the ocean surface and converts it into clean electrical energy output. Compared with traditional onshore wind power, it is gradually becoming a key part of the future sustainable energy structure due to its advantages such as faster wind speeds, more stable wind directions, and the open sea area, which does not occupy precious land resources.

[0003] The high efficiency of offshore wind power lies in its high load factor. In areas rich in wind resources, generator sets can operate for longer periods, and the continuity and stability of average wind speed mean that the reliability and efficiency of power generation are significantly improved compared to onshore wind power. Therefore, offshore wind farms can provide more continuous and reliable power.

[0004] Currently, offshore wind farm power generation forecasting faces numerous challenges, including: limited data sources; traditional forecasting methods typically rely solely on historical data, neglecting to consider turbine parameters, geographical conditions, meteorological information, and wake effects, leading to low forecast accuracy; insufficient data privacy protection; and poor model generalization ability; traditional methods are often only applicable to specific time periods and regions, lacking adaptability to different areas and times. Summary of the Invention

[0005] To address the shortcomings and deficiencies of existing technologies, this invention considers the significant advantages of using federated segmentation learning for short-term prediction of offshore wind farm power generation. Furthermore, a distributed learning architecture—an edge-cloud model—can be considered for power generation prediction.

[0006] This process involves a large amount of wind turbine privacy. How to optimize the segmentation and aggregation models while protecting data privacy, improve the accuracy of short-term power prediction for offshore wind farms, and minimize training time is the main technical problem that this invention aims to solve.

[0007] In view of this, the present invention provides a method and system for short-term prediction of offshore wind farm power generation based on federated segmentation learning, combining federated learning and segmentation learning. Using segmentation learning, the dataset of the offshore wind farm model is segmented and integrated to construct a prediction model for offshore wind farm power generation. Through federated learning, information sharing and model fusion among wind turbine systems can be achieved, improving system efficiency and reliability. Simultaneously, federated learning can protect the data privacy of each wind turbine system, ensuring data security and privacy protection.

[0008] This solution aims to achieve efficient processing and prediction of offshore wind farm power generation data through the collaborative work of distributed clients, edge servers, and the cloud. By employing optimal model segmentation, the cloud-based machine learning model is divided into a feature extractor and a regressor. The offshore wind farm's model information is sent to the cloud to request participation in the collaborative learning task. The cloud proposes a two-stage collaborative learning task, inviting devices collecting local data from the offshore wind farm as clients. Based on the computing power of the client nodes, the cloud groups them and uses models of varying complexity for federated learning. After joint group learning, the client retrieves its model from the edge server for local model training. The trained model parameters are uploaded to the edge and, after two stages of validation, are used for global model aggregation and updates. Subsequently, the client downloads the feature extractor, extracts features from the local data, and uploads them to the edge. The cloud trains a complex model on a high-computing platform. Finally, the cloud obtains all the grouped feature extractors and complex models for short-term prediction of offshore wind farm power generation. Therefore, this framework helps protect data privacy, while updates are performed through semi-asynchronous model aggregation, which can solve the challenges of large-scale heterogeneity and reduce training time and communication overhead, enabling accurate power generation prediction for offshore wind farms.

[0009] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0010] A short-term prediction method for offshore wind farm power generation based on federated segmentation learning includes the following steps:

[0011] Step S1: Divide the cloud-based machine learning model into two parts, a feature extractor and a regressor, using the most efficient model segmentation method.

[0012] Step S2: Send the model information of the offshore wind farm to the cloud and apply to join the collaborative learning task;

[0013] Step S3: The cloud proposes a two-stage collaborative learning task, with the device that collects local data from the offshore wind farm joining the task as a client: The cloud groups the client nodes according to their computing power, and different groups use models of different complexities for federated learning.

[0014] Step S4: After completing the joint group learning phase, each client retrieves the client models of all groups from the edge server to complete the training of its local model. The model used for local training is divided into two parts: the client model and the edge server model. During the forward and backward propagation processes, the client and the edge server perform computation and parameter update operations respectively.

[0015] Step S5: The client uploads model parameters to the edge, and after transaction verification, the parameters are used for global model updates.

[0016] Step S6: After the corresponding group completes federated learning, the client downloads the feature extractor for that group from the edge. After downloading, it uses this feature extractor to extract features from the local data and utilizes multiple feature extractors to extract high-level features from various measurement data. These high-level features from different data sources are fused into a unified feature set with richer information. Finally, Laplacian noise is added to the fused feature set before it is uploaded to the edge.

[0017] Step S7: The cloud obtains the fusion features of local data from the edge and trains the final complex model on a cloud computing platform with higher computing power;

[0018] Step S8: Obtain feature extractors and high-precision complex models for all groups in the cloud: For actual samples, use feature extractors to extract features in the cloud and then input them into complex models to obtain prediction results for offshore wind farm power generation.

[0019] Furthermore, in step S1, the optimal model segmentation objective is to find an optimal segmentation ratio, while simultaneously modeling the memory usage and training time of the device collecting local data, as follows:

[0020]

[0021] stM(α)≤M sm

[0022]

[0023]

[0024] In the formula, α represents the optimal split ratio, T(α) represents the modeling of training time, and M(α) represents the modeling of memory usage. sm This represents the maximum memory constraint of the device, where L is the number of layers in the model, B is the batch size, and |w i | is the weight vector of the feature processor in the i-th layer, |a i| is the activation vector of the i-th layer, s represents the number of samples in the dataset, D represents the dataset, R represents the communication bandwidth, β represents the proportion of the training dataset, n represents the number of iterations, and P sm P represents the processing capacity of the equipment. es This indicates the processing capacity of the edge servers, where K represents the number of edge servers and st means "constrained by".

[0025] Further, in step S2, the model information of the offshore wind farm includes an offshore wind turbine model and an offshore wind farm model, wherein the power generation of a single wind turbine when the average wind speed is lower than the rated wind speed is expressed as:

[0026] P(U)=αU 3 +β

[0027] In the formula, U represents the average wind speed, and the coefficients α and β are expressed as:

[0028]

[0029] In the formula, P G U represents the rated generating capacity. in This represents the cut-in wind speed, where the rated wind speed U r Represented as:

[0030]

[0031] In the formula, ρ represents air density, D represents rotor diameter, and C P This indicates that the fan is operating at its optimal state, with a value of 0.48.

[0032] Based on the above formula, the power generation model of a single offshore wind turbine is expressed as follows:

[0033]

[0034] In the formula, U out To determine the wind speed, a value of 25 m / s is assumed in subsequent model training. in The value is 3 m / s:

[0035] An offshore wind farm consists of multiple offshore wind turbine models. In the offshore wind farm model, the input to the wind farm topology model is the total number of offshore wind turbines, N. T Given the area A of the offshore wind farm, the relationship between the area of ​​the offshore wind farm and the number of offshore wind turbines can be expressed as:

[0036]

[0037] In the formula, L T The distance between the wind turbines is expressed as [distance], and their floor area is expressed as [floor area]. S represents the layout density of the wind turbines.

[0038] Furthermore, by estimating the average wind speed in the windless environment and the wind speed inside the offshore wind farm, input is provided for training the federated segmentation learning model. The annual environmental wind speed statistics are quantified using a two-parameter Weibull distribution: the average wind speed probability density is expressed as:

[0039]

[0040] In the formula, x represents a random variable, k represents the Weibull shape parameter, and λ represents the Weibull scale parameter:

[0041] To determine the average power generation within an offshore wind farm, the Weibull scaling parameter λ is adjusted, and the power generation P of the offshore wind farm is... Fram Represented as:

[0042]

[0043]

[0044] In the formula, P WF,y Let Γ represent the power generation of a single wind turbine within an offshore wind farm, where Γ is an incomplete gamma function, and ε1 and ε2 are represented as:

[0045]

[0046]

[0047]

[0048] In the formula, G represents the geostrophic wind speed, f represents the Coriolis parameter, z0 and l0 represent the corresponding roughness lengths, κ represents a constant with a value of 0.41, and C T The rated value is 0.75, and h represents the wind speed distribution at the height of the hub.

[0049] Furthermore, in step S3, distributed optimization is employed, including the use of an auxiliary network, parallel processing, and knowledge distillation: wherein the auxiliary network W α Its purpose is to introduce the loss function W. e For parallel data updates between different loss functions, it is represented as:

[0050]

[0051]

[0052] In the formula, W r W p Used to minimize the loss function, L(w)s ), L(w c () represents the loss function, and D represents the database;

[0053] In step S3, client nodes are grouped according to their computing power. Clients, acting as devices collecting local data, are grouped using training time as a metric. Devices with similar training times are assigned to the same edge server. The edge server synchronously aggregates the models of devices within the cluster as follows:

[0054]

[0055] In the formula, Let k represent the model of the i-th device at time t+1. i Indicates the number of devices in the cluster. This represents the model of the k-th device at time t+1.

[0056] Furthermore, in step S4, the operations of calculation and parameter update performed by the client and the edge server during the forward and backward propagation processes are specifically as follows:

[0057] A1: The device uses local data to complete the forward propagation of the wind turbine model: after obtaining the output features of the wind turbine, Laplace noise is added to it and uploaded to the edge server.

[0058] A2: The edge server uses the received features to complete the forward propagation process of the edge model and calculate the loss.

[0059] A3: The edge server performs backpropagation and updates the parameters on the edge model: the gradient from the edge is sent to the wind turbine equipment.

[0060] A4: The wind turbine equipment receives the gradient returned by the edge server and completes the backpropagation and parameter update of the wind turbine model.

[0061] Further, in step S5, the transaction verification specifically includes:

[0062] Initial global models of varying complexity are generated in the cloud and uploaded to the edge nodes participating in federated segmentation learning. The client obtains model parameters from the edge nodes for local model training. The model parameters, verified by witnesses in two stages, are used for aggregation of global model parameters. The updated global model parameters are written to new edge nodes and broadcast to other witnesses using the Gossip protocol. After other witnesses verify the validity of the nodes, the new edge nodes are added to the edge nodes. At the end of the grouped federated learning, each client obtains the corresponding feature extractor from the edge nodes to extract features from the local data, concatenates them, and uploads them to the edge nodes.

[0063] In step S7, the model parameter updates between the edge and the cloud are performed using asynchronous aggregation, as shown in the following model:

[0064]

[0065] In the formula, w t+1 This represents the model after aggregation at time (t+1), τ i Denotes the aggregation factor of the i-th cluster. Let represent the aggregation model of the i-th cluster at time t+1'.

[0066] Correspondingly, a short-term prediction system for offshore wind farm power generation based on federated segmentation learning is also provided, including distributed clients, edge servers and the cloud, adopting an end-edge-cloud framework;

[0067] The system collects local data including information from offshore wind farm models and historical measurements of offshore wind power generation, wind turbine parameters, geographical conditions, meteorological information, model training time, equipment memory usage, and communication overhead to initially predict power generation and its corresponding time periods. Local data is used to complete the forward propagation of the client model. After obtaining the output features from the edge, Laplace noise is added and uploaded to the edge server. The edge server uses the received features to complete the forward propagation process of the edge model and calculates the loss. The edge server performs backpropagation and updates the parameters on the edge model. The gradient from the edge is sent to the wind turbine equipment. The wind turbine equipment receives the gradient returned by the edge server, completing the backpropagation and parameter update of the offshore wind turbine model. The cloud obtains the stitched features of the dataset from the edge and trains the final complex model on a high-performance cloud computing platform. The cloud obtains feature extractors for all groups and a high-precision complex model for predicting offshore wind farm power generation.

[0068] Furthermore, its operating mechanism includes the following steps:

[0069] Step 1: Collect local data, including input data for offshore wind farm models, historical measurements of offshore wind power generation, wind turbine parameters, geographical conditions, meteorological information, model training time, equipment memory usage, and communication overhead.

[0070] Step 2: Send the local data collected in Step 1 to the cloud and apply to join the collaborative learning task; set the number of federated learning groups T, initialize T model parameters of different complexities in the cloud, and upload them to the edge.

[0071] Each group is processed in parallel, for a total of T groups; and in the global iteration, each client C in the T groups is processed simultaneously; model parameters are downloaded from the edge. It is represented as the parameter of the c-th wind turbine in the (i-1)-th round of the t-th group, which is updated to P through the customer update algorithm. i t,c The algorithm is expressed as follows:

[0072] B1: Input the initial model parameters, number of local iterations, number of batches per local iteration, and learning efficiency of local training;

[0073] B2: Divide the initialized model parameters into client-side models and edge-side models;

[0074] B3: Run each local iteration and its batch on each client and its corresponding edge server. The process involves the device using the local dataset to complete the forward propagation of the wind turbine model parameters, obtaining the output features of the wind turbine, adding Laplace noise to them, and uploading them to the edge server. Forward and backward propagation are performed on the edge server, and its model parameters are updated based on the learning efficiency of local training and the calculated loss. Subsequently, the gradients from the edge server are sent to the client. The wind turbine device receives the gradients returned by the edge server, completes the backward propagation of the wind turbine model, and updates the parameters based on the learning efficiency of local training and the calculated loss.

[0075] B4: Repeat steps B1-B3 until all client-side model parameters and edge-side model parameters are updated. Then, integrate the updated parameters to form a new local training model and upload it to the edge.

[0076] Step 3: Validate each parameter. Only parameters from clients that pass Multi-KRUM validation are allowed for global model updates. Client validation is determined by the discreteness, expressed as:

[0077]

[0078] In the formula, This is expressed as dispersion. This represents the number of model parameters that are closest to the validator during validation;

[0079] Step 4: Download from the edge and their first few layers As a feature extractor, it utilizes multiple feature extractors to extract high-level features from various measurement data. 1 x 2 x 3 , ..., x i High-level features from different measurements are fused into a unified feature x with richer information. s Finally, Laplace noise was added to the fused features to obtain... Upload to the edge;

[0080] Repeat step 4 on all wind turbine units until all Laplace fusion features are added and uploaded to the edge.

[0081] Step 5: Download all from the cloud. The final high-precision complex model W is trained on a cloud computing platform with higher computing power, i.e., the feature extractors for all groups are obtained in the cloud. And a high-precision complex model W; for actual samples, the cloud uses these feature extractors to extract features, and then inputs them into the complex model to obtain high-precision prediction results of offshore wind farm power generation.

[0082] Compared to existing technologies, this invention and its preferred embodiment combine federated learning and segmentation learning. Using segmentation learning, the dataset of an offshore wind farm model is segmented and integrated to construct a prediction model for offshore wind farm power generation. Through federated learning, information sharing and model fusion among wind turbine systems can be achieved, improving system efficiency and reliability. Simultaneously, federated learning can protect the data privacy of each wind turbine system, ensuring data security and privacy protection. Attached Figure Description

[0083] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0084] Figure 1 This is a schematic diagram of the offshore wind farm power generation prediction process based on federated segmentation learning, as described in an embodiment of the present invention.

[0085] Figure 2 This is a flowchart illustrating the system workflow of an embodiment of the present invention. Detailed Implementation

[0086] To make the features and advantages of this patent more apparent and understandable, specific embodiments are provided below for detailed explanation:

[0087] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0088] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0089] This invention first provides a short-term prediction method for offshore wind farm power generation based on federated segmentation learning, such as... Figure 1 As shown, the specific steps include S1 to S8:

[0090] S1: The cloud-based machine learning model is divided into two parts, a feature extractor and a regressor, through the most efficient model segmentation.

[0091] S2: Send the model information of the offshore wind farm to the cloud and apply to join the collaborative learning task;

[0092] S3: The cloud proposes a two-stage collaborative learning task, incorporating local data collected from offshore wind farms as clients. The cloud groupes these client nodes based on their computing power, with different groups using models of varying complexity for federated learning.

[0093] S4: After completing the joint group learning phase, each client retrieves the client models of all groups from the edge server to complete the training of its local model. The model used for local training is divided into two parts: the client model and the edge server model. During the forward and backward propagation processes, the client and the edge server each perform four phases of operations;

[0094] S5: The client uploads model parameters to the edge, and the parameters can only be used for global model updates after passing two-stage verification.

[0095] S6: After the corresponding group completes federated learning, the client downloads the feature extractor for that group from the edge. This feature extractor is then used to extract features from the local data, and multiple feature extractors are used to extract high-level features from various measurement data. These high-level features from different data sources are fused into a unified feature set with richer information. Finally, Laplacian noise is added to the fused feature set before it is uploaded to the edge.

[0096] S7: The cloud acquires the fusion features of local data from the edge and trains the final complex model on a cloud computing platform with higher computing power.

[0097] S8: The cloud obtains feature extractors for all groups and high-precision complex models. For test samples in real-world applications, the cloud uses these feature extractors to extract features, which are then input into the complex model to obtain prediction results for the power generation of offshore wind farms.

[0098] As a preferred embodiment, in step S1, the goal is to find an optimal segmentation ratio, while simultaneously modeling the memory usage and training time of the device collecting local data, which can be expressed as:

[0099]

[0100] stM(α)≤M sm

[0101]

[0102]

[0103] In the formula, α represents the optimal split ratio, T(α) represents the modeling of training time, and M(α) represents the modeling of memory usage. sm This represents the maximum memory constraint of the device, where L is the number of layers in the model, B is the batch size, and |w i | is the weight vector of the i-th layer, |a i | is the activation vector of the i-th layer, s represents the number of samples in the dataset, D represents the dataset, R represents the communication bandwidth, β represents the proportion of the training dataset, n represents the number of iterations, and P sm P represents the processing capacity of the equipment. es This represents the processing capacity of the edge servers, and K represents the number of edge servers.

[0104] As a preferred embodiment, in step S2, the model information of the offshore wind farm includes an offshore wind turbine model and an offshore wind farm model, wherein the power generation of a single wind turbine when the average wind speed is lower than the rated wind speed can be expressed as:

[0105] P(U)=αU 3 +β

[0106] In the formula, U represents the average wind speed, and the coefficients α and β can be expressed as:

[0107]

[0108] In the formula, P G U represents the rated generating capacity. in This represents the cut-in wind speed, where the rated wind speed U r It can be represented as:

[0109]

[0110] In the formula, ρ represents air density, D represents rotor diameter, and C P This indicates that the fan is operating at its optimal state, and its value is 0.48.

[0111] Based on the above formula, the power generation model of a single offshore wind turbine is expressed as follows:

[0112]

[0113] In the formula, U out To determine the wind speed, a value of 25 m / s is assumed in subsequent model training. in The value is 3 m / s.

[0114] An offshore wind farm can be composed of multiple offshore wind turbine models. To facilitate model training, this invention simplifies the offshore wind farm model, making the input of the wind farm topology model the total number of offshore wind turbines N. T Given the area A of the offshore wind farm, the relationship between the area of ​​the offshore wind farm and the number of offshore wind turbines can be expressed as:

[0115]

[0116] In the formula, L T The distance between wind turbines can be expressed as the area they occupy.

[0117] As a preferred embodiment, in order to more accurately predict the power generation of offshore wind farms and further provide input for training the federated segmentation learning model, it is necessary to estimate the average wind speed statistics of sites without wind turbines (i.e., environmental wind speed statistics) and the average wind speed statistics inside the offshore wind farm.

[0118] Annual ambient wind speed statistics (typically based on 10 or 30-minute intervals) are quantified using a two-parameter Weibull distribution. The average wind speed probability density can be expressed as:

[0119]

[0120] In the formula, x represents a random variable, k represents the Weibull shape parameter, and λ represents the Weibull scale parameter.

[0121] The wind speed statistics within an offshore wind farm differ from the ambient wind speed statistics due to the wake effect of nearby wind turbines. To determine the average power generation within an offshore wind farm, the Weibull scale parameter λ needs to be adjusted; therefore, the power generation P of the offshore wind farm... Fram It can be represented as:

[0122]

[0123]

[0124] In the formula, P WF,y Let Γ represent the power generation of a single wind turbine within an offshore wind farm, where Γ represents an incomplete gamma function, and ε1 and ε2 can be expressed as:

[0125]

[0126]

[0127]

[0128] In the formula, G represents the geostrophic wind speed, f represents the Coriolis parameter, z0 and l0 represent the corresponding roughness lengths, κ represents a constant with a value of 0.41, and C T The rated value is 0.75, where h represents the wind speed distribution at the height of the wheel hub.

[0129] Based on the above design, it can be seen that the local data required in this invention includes the input data required by the above formulas and historical measurements of offshore wind power generation. From a physical perspective, the amount of offshore wind power generation reflects the magnitude of aerodynamic energy, which is directly or indirectly related to various factors. Therefore, it also includes information such as wind turbine parameters, geographical conditions, and meteorological information, along with model training time, equipment memory usage, and communication overhead to form local data. Simultaneously, it is ensured that the data has been preprocessed and standardized.

[0130] As a preferred embodiment, the distributed optimization in step S3 utilizes an auxiliary network, parallel processing, and knowledge distillation. Specifically, an auxiliary network W is added. α The purpose is to introduce a new loss function W. e This facilitates parallel data updates between different loss functions and promotes the exchange of useful information between models. It can be expressed as:

[0131]

[0132]

[0133] In the formula, W r W p Used to minimize the loss function, L(w) s ), L(w c ) represents the loss function, and D represents the database.

[0134] In a preferred embodiment, in step S3, client nodes are grouped according to their computing power. Since clients are devices that collect local data, and computing power cannot be directly measured, training time is chosen as the metric. In this case, training time can be considered an alternative measure of computing power, as it is typically associated with the required computing resources and algorithm complexity. Devices with similar training times are assigned to the same edge server, and the model for edge servers synchronously aggregating devices within a cluster is as follows:

[0135]

[0136] In the formula, Let k represent the model of the i-th device at time (t+1). i Indicates the number of devices in the cluster. The model represents the k-th device at time (t+1).

[0137] As a preferred embodiment, the four stages involved in step S4 specifically include:

[0138] A1: The device uses local data to complete the forward propagation of the wind turbine model. After obtaining the output characteristics of the wind turbine, Laplace noise is added to it and uploaded to the edge server.

[0139] A2: The edge server uses the received features to complete the forward propagation process of the edge model and calculate the loss.

[0140] A3: The edge server performs backpropagation and updates the parameters on the edge model. The gradients from the edge are then sent to the wind turbine equipment.

[0141] A4: The wind turbine equipment receives the gradient returned by the edge server and completes the backpropagation and parameter update of the wind turbine model.

[0142] As a preferred embodiment, in step S5, the specific transaction verification includes:

[0143] Initial global models of varying complexity are generated in the cloud and uploaded to the edge nodes participating in the federated segmentation learning. Clients use the model parameters obtained from the edge nodes to train their local models. Model parameters validated in a two-stage process by witnesses are used to aggregate the global model parameters. The updated global model parameters are written to new edge nodes and broadcast to other witnesses using the Gossip protocol. After other witnesses verify the node's validity, the new edge node is added to the edge node. At the end of the grouped federated learning, each client obtains the corresponding feature extractor from the edge node to extract features from its local data, concatenates the data, and uploads it to the edge node.

[0144] As a preferred embodiment, in step S7, the model parameter update between the edge and the cloud adopts an asynchronous aggregation method, as shown in the following model:

[0145]

[0146] In the formula, w t+1 This represents the model after aggregation at time (t+1), τ i Denotes the aggregation factor of the i-th cluster. This represents the aggregation model of the i-th cluster at time t+1.

[0147] Based on the above method design, the embodiments of the present invention further provide the following construction scheme for the corresponding system:

[0148] To address the challenges of predicting offshore wind farm power generation due to uncertainties in wind speed and direction, complex marine meteorological conditions, difficulties in data acquisition, and inaccurate meteorological models, this invention collects local data that includes not only information from offshore wind farm models and historical measurements of offshore wind power generation, but also information such as wind turbine parameters, geographical conditions, meteorological information, model training time, device memory usage, and communication overhead. This allows for a preliminary prediction of power generation and its corresponding time periods. The device uses local data to complete the forward propagation of the client model. After obtaining the output features at the edge, Laplace noise perturbation is added and uploaded to the edge server. The edge server uses the received features to complete the forward propagation process of the edge model and calculates the loss. The edge server performs backpropagation and updates the parameters on the edge model. The gradient from the edge is sent to the wind turbine equipment. The wind turbine equipment receives the gradient returned by the edge server, completes the backpropagation of the offshore wind turbine model, and updates the parameters. The cloud obtains the stitched features of the dataset from the edge and trains the final complex model on a high-performance cloud computing platform. The cloud-based system obtains feature extractors and high-precision complex models for predicting the power generation of offshore wind farms. During the prediction phase, the cloud uses feature extractors to extract features from test samples in real-world applications, then inputs these features into the complex model to obtain prediction results. These predictions help optimize power system operation, improve wind farm efficiency, reduce operation and maintenance costs, and promote energy market transactions. This process allows for model training without sharing raw data, effectively protecting the privacy of wind turbine equipment. Offshore wind farms are typically distributed across different locations on the sea surface; segmentation learning allows for local model training on each node, fully utilizing distributed computing resources and improving computational efficiency. This invention updates the data using parameters from the next time step. Simultaneously, it is necessary to observe at each time step whether the changes in model parameters converge or are very subtle. If so, the iteration ends; otherwise, the above steps are repeated to iteratively optimize the model to adapt to the prediction of offshore wind farm power generation.

[0149] Specifically, such as Figure 2 As shown, please refer to the following steps:

[0150] Step 1: Create an edge-cloud framework, with corresponding servers for each stage to train models of varying complexity.

[0151] Step 1-1: Collect local data, including input data for offshore wind farm models, historical measurements of offshore wind power generation, wind turbine parameters, geographical conditions, meteorological information, model training time, equipment memory usage, and communication overhead.

[0152] Step 2: Send local data to the cloud and apply to join the collaborative learning task. Let there be T federated learning groups. The cloud initializes T model parameters of different complexities and uploads them to the edge.

[0153] Step 2-1: Process each group in parallel, for a total of T groups. In the global iteration, process each client C in the T groups simultaneously. Download model parameters from the edge. It is represented as the parameter of the c-th wind turbine in the (i-1)-th round of the t-th group, which is updated to P through the customer update algorithm. i t,c The algorithm is expressed as follows:

[0154] B1: Input initial model parameters, number of local iterations, number of batches per local iteration, and learning efficiency of local training.

[0155] B2: Divide the initialized model parameters into client-side models and edge-side models.

[0156] B3: Run each local iteration and its batch on each client and its corresponding edge server. The process involves the device using the local dataset to perform forward propagation of the wind turbine model parameters, obtaining the output features of the wind turbine, adding Laplace noise, and uploading it to the edge server. Forward and backward propagation are performed on the edge server, and its model parameters are updated based on the learning efficiency of local training and the calculated loss. The gradients from the edge server are then sent to the client. The wind turbine device receives the gradients returned from the edge server, completes the backward propagation of the wind turbine model, and updates the parameters based on the learning efficiency of local training and the calculated loss.

[0157] B4: Repeat the above steps until all client-side model parameters and edge-side model parameters are updated. Then, integrate the updated parameters to form a new local training model and upload it to the edge.

[0158] Step 3: Validate each parameter. Only parameters from clients that pass Multi-KRUM validation can be used for global model updates. The algorithm determines whether a client passes validation based on its discreteness, which can be represented as:

[0159]

[0160] In the formula, This is expressed as dispersion. This represents the number of model parameters that are closest to the validator during validation.

[0161] Step 4: Download from the edge and their first few layers As a feature extractor, it utilizes multiple feature extractors to extract high-level features from various measurement data. 1 x 2 x 3 , ..., x i High-level features from different measurements are fused into a unified feature x with richer information. s Finally, Laplace noise was added to the fused features to obtain... Upload to the edge.

[0162] Step 4-1: Repeat the above operation for all wind turbine equipment until all Laplace fusion features are added and uploaded to the edge.

[0163] Step 5: Download all from the cloud. The final high-precision complex model W is trained on a cloud computing platform with higher computing power;

[0164] Step 5-1: Obtain the feature extractor for all groups in the cloud. And a high-precision, complex model W. For test samples in practical applications, the cloud uses these feature extractors to extract features, which are then input into the complex model to obtain high-precision prediction results of offshore wind farm power generation.

[0165] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0168] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0169] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0170] This patent is not limited to the above-described preferred embodiments. Anyone can derive other forms of short-term prediction methods and systems for offshore wind farm power generation based on federated segmentation learning under the guidance of this patent. All equivalent changes and modifications made within the scope of this patent application shall fall within the scope of this patent.

Claims

1. A short-term prediction method for offshore wind farm power generation based on federated segmentation learning, characterized in that: Includes the following steps: Step S1: Divide the cloud-based machine learning model into two parts, a feature extractor and a regressor, using the most efficient model segmentation method. Step S2: Send the model information of the offshore wind farm to the cloud and apply to join the collaborative learning task; Step S3: The cloud proposes a two-stage collaborative learning task, with the device that collects local data from the offshore wind farm joining the task as a client: The cloud groups the client nodes according to their computing power, and different groups use models of different complexities for federated learning. Step S4: After completing the joint group learning phase, each client retrieves the client models of all groups from the edge server to complete the training of its local model. The model used for local training is divided into two parts: the client model and the edge server model. During the forward and backward propagation processes, the client and the edge server perform computation and parameter update operations respectively. Step S5: The client uploads model parameters to the edge, and after transaction verification, the parameters are used for global model updates. Step S6: After the corresponding group completes federated learning, the client downloads the feature extractor for that group from the edge. After downloading, it uses this feature extractor to extract features from the local data and utilizes multiple feature extractors to extract high-level features from various measurement data. These high-level features from different data sources are fused into a unified feature set with richer information. Finally, Laplacian noise is added to the fused feature set before it is uploaded to the edge. Step S7: The cloud obtains the fusion features of local data from the edge and trains the final complex model on a cloud computing platform with higher computing power; Step S8: Obtain feature extractors and high-precision complex models for all groups in the cloud: For actual samples, use feature extractors to extract features in the cloud and then input them into complex models to obtain prediction results for offshore wind farm power generation. In step S2, the model information of the offshore wind farm includes an offshore wind turbine model and an offshore wind farm model, wherein the power generation of a single wind turbine when the average wind speed is lower than the rated wind speed is expressed as follows: In the formula, This represents the average wind speed, where the coefficient is... , Represented as: In the formula, Indicates the rated power generation capacity. This indicates the cut-in wind speed, where the rated wind speed is... Represented as: In the formula, Indicates air density, Indicates the rotor diameter. This indicates that the fan is operating at its optimal state, with a value of 0.

48. Based on the above formula, the power generation model of a single offshore wind turbine is expressed as follows: In the formula, To determine the wind speed, a value of 25 m / s was assumed during subsequent model training. The value is 3 m / s: An offshore wind farm consists of multiple offshore wind turbine models. In the offshore wind farm model, the input to the wind farm topology model is the total number of offshore wind turbines. and offshore wind farm area The relationship between the area of ​​an offshore wind farm and the number of offshore wind turbines can be expressed as follows: In the formula, The distance between the wind turbines is expressed as [distance], and their floor area is expressed as [floor area]. S represents the layout density of the wind turbines.

2. The method for short-term prediction of offshore wind farm power generation based on federated segmentation learning according to claim 1, characterized in that: In step S1, the optimal model segmentation objective is to find an optimal segmentation ratio, while simultaneously modeling the memory usage and training time of the device collecting local data, as follows: In the formula This represents the optimal split ratio. This represents modeling of the training time. Modeling of memory usage This indicates the maximum memory constraint of the device. It is the number of layers in the model. It is the batch size. It is the weight vector of the i-th layer. It is the activation vector of the i-th layer. This represents the number of samples in the dataset. Represents a dataset, Indicates communication bandwidth. This indicates the proportion of the training dataset. Indicates the number of iterations. Indicates the equipment's processing capacity. This indicates the processing capacity of the edge server. This indicates the number of edge servers.

3. The short-term prediction method for offshore wind farm power generation based on federated segmentation learning according to claim 1, characterized in that: By estimating the average wind speed in the windless environment and the wind speed inside the offshore wind farm, we provide input for training the federated segmentation learning model. The annual environmental wind speed statistics are quantified using a two-parameter Weibull distribution: the average wind speed probability density is expressed as: In the formula, Represented as a random variable, Represented as Weibull shape parameters, Represented as Weibull scaling parameters: To determine the average power generation within an offshore wind farm, the Weibull scaling parameters were adjusted. The power generation of offshore wind farms Represented as: In the formula, This is expressed as the power generation expression of a single wind turbine located within an offshore wind farm. Represented as an incomplete gamma function, where , Represented as: ， In the formula, Expressed as geostrophic wind speed, Represented as Coriolis parameters, , This is represented by the corresponding roughness length. It is represented as a constant with a value of 0.

41. The rated value is 0.

75. This represents the wind speed distribution at the height of the wheel hub.

4. The short-term prediction method for offshore wind farm power generation based on federated segmentation learning according to claim 1, characterized in that: In step S3, distributed optimization is employed, including the use of an auxiliary network, parallel processing, and knowledge distillation: wherein, the auxiliary network Its purpose is to introduce a loss function. For parallel data updates between different loss functions, it is represented as: In the formula, , Used to minimize the loss function, , Represents the loss function. Represents a database; In step S3, client nodes are grouped according to their computing power. Clients, acting as devices collecting local data, are grouped using training time as a metric. Devices with similar training times are assigned to the same edge server. The edge server synchronously aggregates the models of devices within the cluster as follows: In the formula, This represents the model of the i-th device at time t+1. Indicates the number of devices in the cluster. This represents the model of the k-th device at time t+1.

5. The short-term prediction method for offshore wind farm power generation based on federated segmentation learning according to claim 1, characterized in that: In step S4, the operations of calculation and parameter update performed by the client and edge server during the forward and backward propagation processes are specifically as follows: A1: The device uses local data to complete the forward propagation of the wind turbine model: after obtaining the output features of the wind turbine, Laplace noise is added to it and uploaded to the edge server. A2: The edge server uses the received features to complete the forward propagation process of the edge model and calculate the loss. A3: The edge server performs backpropagation and updates the parameters on the edge model: the gradient from the edge is sent to the wind turbine equipment. A4: The wind turbine equipment receives the gradient returned by the edge server and completes the backpropagation and parameter update of the wind turbine model.

6. The short-term prediction method for offshore wind farm power generation based on federated segmentation learning according to claim 1, characterized in that: In step S5, the transaction verification specifically includes: Initial global models of varying complexity are generated in the cloud and uploaded to the edge nodes participating in federated segmentation learning. The client obtains model parameters from the edge nodes for local model training. The model parameters, verified by witnesses in two stages, are used for aggregation of global model parameters. The updated global model parameters are written to new edge nodes and broadcast to other witnesses using the Gossip protocol. After other witnesses verify the validity of the nodes, the new edge nodes are added to the edge nodes. At the end of the grouped federated learning, each client obtains the corresponding feature extractor from the edge nodes to extract features from the local data, concatenates them, and uploads them to the edge nodes.

7. The short-term prediction method for offshore wind farm power generation based on federated segmentation learning according to claim 1, characterized in that: In step S7, the model parameter updates between the edge and the cloud are performed using asynchronous aggregation, as shown in the following model: In the formula, This represents the model after aggregation at time (t+1). Denotes the aggregation factor of the i-th cluster. Let represent the aggregation model of the i-th cluster at time t+1'.

8. A short-term prediction system for offshore wind farm power generation based on federated segmentation learning, used to implement the method as described in claim 1, characterized in that: It includes distributed clients, edge servers, and the cloud, adopting an end-edge-cloud framework; The local data it collects includes information on offshore wind farm models and historical measurements of offshore wind power generation, wind turbine parameters, geographical conditions, meteorological information, model training time, equipment memory usage, and communication overhead, in order to make preliminary predictions of power generation and its corresponding time periods. The forward propagation of the client model is completed using local data; after obtaining the output features of the edge end, Laplace noise perturbation is added to them and uploaded to the edge end server; The edge server uses the received features to complete the forward propagation process of the edge model and calculate the loss. The edge server performs backpropagation and updates the parameters on the edge model; the gradient from the edge is sent to the wind turbine equipment; the wind turbine equipment receives the gradient returned by the edge server and completes the backpropagation and parameter update of the offshore wind turbine model. The cloud acquires the spliced ​​features of the dataset from the edge and trains the final complex model on a high-performance cloud computing platform; the cloud obtains feature extractors for all groups and a high-precision complex model for predicting the power generation of offshore wind farms.

9. The short-term prediction system for offshore wind farm power generation based on federated segmentation learning according to claim 8, characterized in that: Its operating mechanism includes the following steps: Step 1: Collect local data, including input data for offshore wind farm models, historical measurements of offshore wind power generation, wind turbine parameters, geographical conditions, meteorological information, model training time, equipment memory usage, and communication overhead. Step 2: Send the local data collected in Step 1 to the cloud and apply to join the collaborative learning task; set the number of federated learning groups T, initialize T model parameters of different complexities in the cloud, and upload them to the edge. Each group is processed in parallel, for a total of T groups; and in the global iteration, each client C in the T groups is processed simultaneously; model parameters are downloaded from the edge. It is represented as the parameter of the c-th wind turbine in the (i-1)-th round of the t-th group, which is updated by the customer update algorithm. The algorithm is expressed as follows: B1: Input the initial model parameters, number of local iterations, number of batches per local iteration, and learning efficiency of local training; B2: Divide the initialized model parameters into client-side models and edge-side models; B3: Run each local iteration and its batch on each client and its corresponding edge server. The process involves the device using the local dataset to complete the forward propagation of the wind turbine model parameters, obtaining the output features of the wind turbine, adding Laplace noise to them, and uploading them to the edge server. Forward and backward propagation are performed on the edge server, and its model parameters are updated based on the learning efficiency of local training and the calculated loss. Subsequently, the gradients from the edge server are sent to the client. The wind turbine device receives the gradients returned by the edge server, completes the backward propagation of the wind turbine model, and updates the parameters based on the learning efficiency of local training and the calculated loss. B4: Repeat steps B1-B3 until all client-side model parameters and edge-side model parameters are updated. Then, integrate the updated parameters to form a new local training model and upload it to the edge. Step 3: Validate each parameter. Only parameters from clients that pass Multi-KRUM validation are allowed for global model updates. Client validation is determined by the discreteness, expressed as: In the formula, This is expressed as dispersion. This represents the number of model parameters that are closest to the validator during validation; Step 4: Download from the edge , , , ..., and their first few layers , , , ..., As a feature extractor, it utilizes multiple feature extractors to extract high-level features from various measurement data. , , , ..., High-level features from different measurements are fused into a unified feature with richer information. Finally, Laplace noise was added to the fused features to obtain... Upload to the edge; Repeat step 4 on all wind turbine units until all Laplace fusion features are added and uploaded to the edge. Step 5: Download all from the cloud. The final high-precision complex model is trained on a cloud computing platform with higher computing power. That is, obtaining feature extractors for all groups in the cloud. , , , ..., and high-precision complex models For actual samples, the cloud uses these feature extractors to extract features, which are then input into complex models to obtain high-precision prediction results of offshore wind farm power generation.

Images