Wind power generation prediction method and system, electronic device and storage medium

Abstract

This invention relates to the field of wind power generation technology and discloses a wind power generation prediction method, system, electronic device, and storage medium. The method includes: acquiring a high-dimensional feature representation, which is obtained by mapping multi-source heterogeneous spatiotemporal sequence data to a decoupled latent space through a feature variational autoencoder; based on the high-dimensional feature representation, performing physical simulation and data-driven prediction in parallel to generate a power prediction sequence; constructing a digital twin environment for wind farm operation based on real-time wind farm operation data and the power prediction sequence; and in the digital twin environment for wind farm operation, solving the conflicts between multiple optimization objectives through multi-round iterative game to obtain a power generation plan and bidding strategy. This invention achieves autonomous collaborative optimization of the power generation plan and bidding strategy, improving the accuracy, continuity, and collaborative optimization capability of prediction.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, specifically to wind power generation prediction methods, systems, electronic devices, and storage media. Background Technology

[0002] Wind power forecasting is an important means of ensuring the safe and stable operation of the power grid. Currently, multi-source data fusion technology has been applied in wind power forecasting.

[0003] However, existing technologies still have the following shortcomings: First, the fusion of multi-source data usually remains at the level of simple spatiotemporal alignment and feature stitching. For example, wind speed and direction predicted by NWP (Numerical Weather Prediction) and power data collected by SCADA (Supervisory Control and Data Acquisition) are aligned on the time axis and directly used as input for machine learning models. This fails to fully explore the deep physical correlations and statistical coupling relationships between different data sources, resulting in insufficient information utilization. Second, when a key data source fails or communication is interrupted, the performance of the prediction system drops sharply, making it difficult to guarantee the continuity of predictions and affecting the reliability of power grid dispatch. Third, existing prediction systems only output power prediction values; subsequent trading decisions and dispatch arrangements still need to be completed manually, making it difficult to achieve coordinated optimization of multiple objectives such as power generation revenue, power grid safety, and equipment health. Summary of the Invention

[0004] This invention provides a wind power generation prediction method, system, electronic device, and storage medium to solve the problems in the prior art where traditional feature engineering methods are difficult to deeply integrate heterogeneous data, the prediction reliability drops sharply during faults, and traditional prediction methods only output power values, which still require manual decision-making for trading and scheduling, and cannot achieve global optimization.

[0005] In a first aspect, the present invention provides a wind power generation prediction method, the method comprising: High-dimensional feature representations are obtained by mapping multi-source heterogeneous spatiotemporal sequence data to a decoupled latent space through a feature variational autoencoder. Based on high-dimensional feature representation, physical simulation and data-driven prediction are executed in parallel to generate power prediction sequences; Based on real-time operation data and power prediction sequences of wind farms, a digital twin environment for wind farm operation is constructed. In the digital twin environment of wind farm operation, the conflict between multiple optimization objectives is solved through multiple rounds of iterative game to obtain the power generation plan and bidding strategy.

[0006] This invention provides a wind power generation prediction method that maps multi-source heterogeneous data to a decoupled latent space through a feature variational autoencoder, enabling each dimension of the latent variables to independently correspond to different physical factors. This achieves deep fusion at the feature level, improving the robustness and interpretability of feature representation. Based on this, it executes and fuses physical simulation and data-driven prediction in parallel, ensuring that the prediction results possess both statistical regularity and physical consistency. By constructing a digital twin environment for wind farm operation and performing multi-round iterative game theory within it to solve multi-objective conflicts, it achieves autonomous and collaborative optimization of power generation plans and bidding strategies. This invention solves the technical problems of insufficient multi-source data fusion, easy prediction interruption when data sources are missing, and difficulty in directly using prediction results for trading and scheduling decisions in existing technologies, thus improving the accuracy, continuity, and collaborative optimization capabilities of prediction.

[0007] In one optional implementation, obtaining the high-dimensional feature representation further includes: When the data source is missing, the corresponding dimensions are empty in the decoupled latent space and the complete high-dimensional feature representation is reconstructed by decoding. In the decoupled latent space, different dimensions correspond to at least one physical factor among wind speed, turbulence intensity, unit mechanical state, wake effect intensity and market volatility, and each dimension is associated with the corresponding data source.

[0008] In the above technical solution, by assigning different dimensions in the decoupled latent space to at least one physical factor among wind speed, turbulence intensity, unit mechanical state, wake effect intensity, and market volatility, and associating them with the corresponding data source, the physical interpretability of the feature dimensions is achieved. When the data source is missing, by emptying the corresponding dimensions in the decoupled latent space and decoding and reconstructing the complete high-dimensional feature representation, the system can still generate complete feature vectors even when the critical data source is interrupted, thereby ensuring the continuity of the prediction process. This solution solves the technical problems of the sharp decline in prediction performance and the difficulty in ensuring prediction continuity when the data source is missing in the existing technology, and improves the fault tolerance and reliability of the system.

[0009] In one optional implementation, based on high-dimensional feature representation, physical simulation and data-driven prediction are performed in parallel to generate a power prediction sequence, including: Based on high-dimensional feature representation, physical simulation is performed using a physical simulation model to generate physical priors; the physical priors include the initial wind speed and power distribution of each wind turbine in the wind farm in the future time period. Based on high-dimensional feature representation, data-driven prediction is performed using a data-driven model to generate statistical predictions. The power prediction sequence is generated by adaptively weighting and fusing physical priors and statistical data predictions through a physics-guided deep learning model.

[0010] In the above technical solution, physical simulation and data-driven prediction are executed in parallel based on the same high-dimensional feature representation to generate physical priors and statistical predictions, respectively. These two priors are then adaptively weighted and fused using a physics-guided deep learning model. This results in a power prediction sequence that integrates the mechanistic constraints provided by physical simulation and utilizes the statistical patterns mined from historical data by the data-driven model, achieving a balance between physical consistency and prediction accuracy. This solution addresses the technical problems of insufficient generalization ability under extreme conditions and potential deviations from physical laws in existing technologies that rely solely on data-driven models. It improves the physical rationality, temporal consistency, and adaptability to unknown conditions.

[0011] In one optional implementation, the real-time operating data of the wind farm includes: real-time wind measurement data, SCADA real-time monitoring data, and real-time grid connection point power. Based on real-time wind farm operation data and power prediction sequences, a digital twin environment for wind farm operation is constructed, including: Based on real-time wind measurement data and power prediction sequences, a three-dimensional dynamic distribution of the flow field in a wind farm is constructed. Based on SCADA real-time monitoring data, a model of the operating status and health of wind turbine units in a wind farm is constructed. Based on real-time grid connection point power and power prediction sequences, a dynamic model of wind farm power output and grid interaction is constructed. By coupling the three-dimensional flow field dynamic distribution, the operating status and health model, and the dynamic model of power output and grid interaction, a digital twin environment for wind farm operation is generated.

[0012] The above technical solution combines real-time wind measurement data, SCADA real-time monitoring data, and real-time grid-connected power with power prediction sequences to construct a three-dimensional dynamic distribution model of the wind farm's flow field, a model of the wind turbine's operating status and health, and a dynamic model of power output and grid interaction. These models are then coupled to generate a unified digital twin environment for wind farm operation, achieving multi-dimensional dynamic mapping of the wind farm's physical state, equipment health, and grid interaction. This solution addresses the technical problem of lacking a high-fidelity simulation environment for strategy deduction in existing technologies, providing a simulateable and iterative virtual deduction platform for subsequent multi-objective collaborative optimization, and improving the accuracy and reliability of decision-making deduction.

[0013] In one optional implementation, the conflict between multiple optimization objectives is resolved through multi-round iterative game theory to obtain the power generation plan and bidding strategy, including: In the digital twin environment of wind farm operation, initialize the decision-making schemes of each virtual agent; In each iteration, with the decision schemes of other virtual agents remaining unchanged except for the current virtual agent, the optimal response scheme that maximizes the utility function of the current virtual agent is solved, and the decision scheme of the current virtual agent is updated to the optimal response scheme. The process is repeated iteratively until the change in the decision schemes of all virtual agents is less than a preset threshold, thus obtaining the optimal power generation plan and bidding strategy. The power generation plan and bidding strategy include: power prediction curves and trading strategy suggestions.

[0014] In the above technical solution, the decision-making schemes of each virtual agent are initialized in the digital twin environment of the wind farm operation. In each iteration, the optimal response scheme of each agent is solved sequentially under the condition that the strategies of the other agents remain unchanged, until the changes in the decision schemes of all agents converge to a preset threshold. This achieves the autonomous and collaborative solution of conflicts between multiple optimization objectives, resulting in the optimal power generation plan and bidding strategy. This solution solves the technical problems of existing technologies, such as only outputting power prediction values, the need for manual completion of subsequent trading decisions and scheduling arrangements, and the difficulty in balancing power generation revenue and grid security. It realizes automated decision-making for multi-objective collaborative optimization and improves the overall efficiency of wind farm operation.

[0015] In one alternative implementation, the method further includes: Deploy a global model in the cloud to aggregate model updates from multiple wind farms; Deploy local prediction processes at each wind farm and train them using local private data; The latent variable dimensions related to physical laws in the feature variational autoencoder are participated in the federated aggregation, while the latent variable dimensions related to wind farm station attributes are kept and updated locally. The computation graph of the physical loss term in the physics-guided deep learning model is synchronized with the local machine as public knowledge in the cloud, and the network parameters of the data-driven model are uploaded to the cloud for secure aggregation after being protected by differential privacy technology.

[0016] The aforementioned technical solution achieves a balance between collaborative evolution of models across wind farms and data privacy protection by deploying a global model in the cloud to aggregate model updates from multiple wind farms, deploying a prediction process locally for training using private data, and incorporating latent variables related to physical laws in the feature variational autoencoder into federated aggregation, retaining dimensions related to wind farm attributes for local updates, synchronizing the physical loss term computation graph as public knowledge, and uploading and aggregating data-driven model parameters after differential privacy protection. This solution addresses the technical challenges of existing technologies where models struggle to share general knowledge while protecting wind farm private data, resulting in low model update efficiency, and improves the generalization ability and data security of model training.

[0017] In one alternative implementation, the method further includes: Based on the digital twin environment of wind farm operation, the load spectrum of wind turbine components is simulated, and an equipment risk warning report is generated based on the component load spectrum; The power generation plan, bidding strategy, and equipment risk warning report are distributed to the wind farm's energy management system, power trading terminal, and predictive maintenance platform.

[0018] The above technical solution, by simulating the load spectrum of wind turbine components and generating equipment risk warning reports within a digital twin environment for wind farm operation, and simultaneously distributing power generation plans, bidding strategies, and equipment risk warning reports to the energy management system, power trading terminal, and predictive maintenance platform, achieves a complete closed loop from power prediction to multi-objective optimization decision-making, and then to equipment risk warning and instruction issuance. This solution solves the technical problems of existing technologies where prediction results only output power values, subsequent trading decisions and equipment maintenance still require manual completion, and multi-stage collaborative linkage is difficult to achieve, thus improving the automation level and overall collaborative efficiency of wind farm operation and management.

[0019] In a second aspect, the present invention provides a wind power generation prediction system, applied to the wind power generation prediction method of the first aspect or any corresponding embodiment thereof, the system comprising: The multimodal fusion and representation module is used to acquire multi-source heterogeneous spatiotemporal sequence data and map the multi-source heterogeneous spatiotemporal sequence data to a decoupled latent space through a feature variational autoencoder to obtain high-dimensional feature representations. The multi-fidelity collaborative prediction engine, connected to the multi-modal fusion and characterization module, is used to generate power prediction sequences by performing physical simulation and data-driven prediction in parallel based on high-dimensional feature characterization. A digital twin, connected to a multi-fidelity collaborative prediction engine, is used to construct a digital twin environment for wind farm operation based on real-time wind farm operation data and the power prediction sequence. The dynamic collaborative optimizer, connected to the digital twin, is used in the digital twin environment of wind farm operation to solve the conflicts between multiple optimization objectives through multi-round iterative game to obtain the power generation plan and bidding strategy.

[0020] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the wind power generation prediction method of the first aspect or any corresponding embodiment described above.

[0021] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the wind power generation prediction method of the first aspect or any corresponding embodiment thereof.

[0022] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to execute the wind power generation prediction method of the first aspect or any corresponding embodiment described above. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of an application scenario according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of the mode fusion and characterization module in the wind power generation prediction system according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the architecture of a multi-fidelity collaborative prediction engine in a wind power generation prediction system according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the architecture of the dynamic collaborative optimizer in the wind power generation prediction system according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the first process of the wind power generation prediction method according to an embodiment of the present invention; Figure 6 This is a schematic diagram of a second process for the wind power generation prediction method according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0027] As an optional application scenario of this invention, such as Figure 1 The diagram shown is an architecture diagram of the wind power generation prediction system provided by the present invention. The system includes: a multimodal fusion and characterization module 110, a multi-fidelity collaborative prediction engine 120, a digital twin 130, a dynamic collaborative optimizer 140, and an output and application execution module 150, wherein: The multimodal fusion and characterization module 110 is used to receive and fuse multi-source heterogeneous spatiotemporal sequence data from meteorological numerical forecasts, wind farm sensor networks, computational fluid dynamics simulations, and power market systems, and generate unified and robust high-dimensional feature representations through a feature decoupling and reconstruction mechanism.

[0028] The multi-fidelity collaborative prediction engine 120 is connected to the multi-modal fusion and characterization module 110. The multi-fidelity collaborative prediction engine integrates a physics-guided deep neural network to couple the mechanistic constraints provided by rapid physical simulation with the learning capabilities of the data-driven model, generating a high-precision, physically consistent power prediction sequence on the timescale of minutes to days in the future.

[0029] The digital twin 130 is connected to the multi-fidelity collaborative prediction engine 120. The digital twin is a digital twin for wind farm operation. The digital twin is used to construct and continuously update a dynamic virtual mapping of wind farm flow field, equipment status and grid connection point power based on real-time data and prediction sequences, and serves as a high-fidelity simulation environment for strategy deduction, i.e., the wind farm operation digital twin environment.

[0030] The dynamic co-optimizer 140 is connected to the digital twin 130. The dynamic co-optimizer has a built-in multi-agent decision-making framework based on non-cooperative game theory, which is used to simulate and coordinate the conflict of multiple objectives such as power generation revenue, grid dispatch compliance and equipment health life in the digital twin environment, and output Pareto optimal power generation plan and market bidding strategy. The output and application execution module 150 is connected to the dynamic collaborative optimizer 140. The output and application execution module is used to send the optimized power curve, market trading strategy and equipment risk warning instructions to the wind farm's energy management system, power trading terminal and predictive maintenance platform, respectively.

[0031] like Figure 2 As shown, the multimodal fusion and representation module specifically includes: The spatiotemporal alignment and data quality control unit is used to standardize timestamp alignment and spatial interpolate data from different sources, with different sampling frequencies and different delays to a unified grid. It also uses a hybrid algorithm based on sliding window and isolated forest to detect and repair abnormal data points.

[0032] The core unit of the feature variational autoencoder maps multi-source data that has undergone quality control to a low-dimensional latent space. By introducing a decoupling regularization term, it forces different dimensions in the latent space to independently correspond to underlying physical or business factors such as wind speed, turbulence intensity, unit mechanical state, wake effect intensity, and market volatility.

[0033] The latent variable manipulation and feature completion unit is connected to the core unit of the feature variational autoencoder. When a specific data source signal is missing, it can manually or automatically empty the corresponding latent variable dimension and use the decoder network to reconstruct a complete fused feature vector based on the remaining intact latent variable dimensions, so as to achieve uninterrupted prediction process.

[0034] like Figure 3 As shown, the multi-fidelity collaborative prediction engine includes: The low-fidelity physical simulation unit uses a parameterized Jensen wake model or a flow field super-resolution model based on a deep neural network to quickly simulate the future flow field of the entire wind farm in seconds and output the initial power distribution.

[0035] The high-fidelity data-driven prediction unit takes the fused features output by the multimodal fusion and representation module as the main input and adopts a hybrid architecture of a bidirectional long short-term memory network with an attention mechanism and a spatiotemporal graph convolutional network to learn complex nonlinear mapping relationships.

[0036] The physics-guided deep operator network fusion unit takes the outputs of the low-fidelity physics simulation unit and the high-fidelity data-driven prediction unit as dual-channel inputs and performs adaptive weighting through a fusion gating network. The training loss function of this fusion network explicitly incorporates a physical consistency constraint term based on the fluid dynamics equation.

[0037] like Figure 4 As shown, the dynamic collaborative optimizer includes a multi-agent decision-making framework, which comprises the following three virtual agents: The utility function of the power generation revenue agent is the total revenue calculated based on the spot market price curve and ancillary service prices within the forecast period, minus the assessment penalty caused by forecast deviation. The power grid security intelligent agent has a utility function that is the negative of the cumulative amount of tracking deviation of dispatch instructions, and also takes into account its potential support capability for power grid frequency and voltage stability. The device health intelligent agent has a utility function that is the negative value of the cumulative fatigue load of key components based on the rainflow counting method. The load is estimated in real time through a high-fidelity structural dynamics model in the digital twin.

[0038] It should be noted that the wind power prediction system adopts a hierarchical federated learning architecture to achieve continuous model evolution and privacy protection, specifically as follows: Deploy a global model on a cloud server to aggregate model updates from multiple wind farms; Edge computing nodes are deployed locally at each wind farm to run a complete local prediction system and train it using local private data; In the feature variational autoencoder, the decoupled latent variables whose dimensions are related to general physical laws participate in federated aggregation; while the dimensions that are highly related to the station-specific attributes do not participate in aggregation and are only updated locally. The computation graph of the physical loss term in the physics-guided deep operator network is synchronized as public knowledge in the cloud and at the edge, while the network parameters of the data-driven part are uploaded to the cloud for secure aggregation after being protected by homomorphic encryption or differential privacy technology.

[0039] The output and application execution module 150 specifically outputs the following three types of results: High-precision power prediction curves, with time resolution supporting minute and 15-minute levels; Market trading strategy recommendations include the order volume and price curves for the next day's time-of-use electricity spot market, as well as capacity ordering recommendations for ancillary service markets such as frequency regulation and reserve. The equipment risk warning report, based on the component load spectrum derived from digital twins and a dynamic Bayesian network fault inference model, generates early failure probabilities and preventative maintenance time window suggestions. According to an embodiment of the present invention, a wind power generation prediction method embodiment is provided. It should be noted that the steps shown in the flowcharts in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0040] This embodiment provides a wind power generation prediction method, which can be used in the aforementioned wind power generation prediction system. Figure 5 This is a flowchart of a wind power generation prediction method according to an embodiment of the present invention, such as... Figure 5 As shown, the process includes the following steps: Step S501: Obtain high-dimensional feature representation. The high-dimensional feature representation is obtained by mapping multi-source heterogeneous spatiotemporal sequence data to a decoupled latent space through a feature variational autoencoder.

[0041] High-dimensional feature representation refers to the unified, compact, and information-rich high-dimensional feature vector generated after feature extraction and fusion of multi-source heterogeneous spatiotemporal sequence data through feature variational autoencoders.

[0042] Specifically, the multimodal fusion and representation module receives multi-source heterogeneous spatiotemporal sequence data from numerical weather prediction, SCADA, lidar, CFD simulation, and power market systems. After data preprocessing by the spatiotemporal alignment and data quality control unit, the feature variational autoencoder core unit maps the data to a decoupled latent space. This latent space, through decoupling regularization, ensures that each dimension independently corresponds to physical or operational factors such as wind speed, turbulence intensity, unit mechanical state, wake effect intensity, and market volatility, thereby generating a unified high-dimensional feature representation. Simultaneously, the latent variable operation and feature completion unit sets the corresponding dimension to empty and reconstructs the complete feature vector through the decoder when a specific data source is missing, ensuring the continuity of the prediction process. This process achieves deep fusion and robust representation of multi-source heterogeneous data, providing physically interpretable and fault-tolerant high-dimensional feature inputs for downstream prediction tasks.

[0043] The wind power prediction method provided in this embodiment maps multi-source heterogeneous data to a decoupled latent space through a feature variational autoencoder. A decoupling loss term is introduced during training, so that each dimension of the latent variable independently corresponds to physical or operational factors such as wind speed, turbulence, and equipment status. This deep fusion at the feature level, compared with traditional simple data splicing, can more fully explore the inherent correlation between data, improve the robustness and interpretability of feature representation, and furthermore, when a specific data source is missing, the system can empty the corresponding dimension in the latent space and reconstruct the complete feature vector through the decoder, thereby ensuring that the prediction process is not interrupted and significantly improving the fault tolerance and reliability of the system.

[0044] Step S502: Based on high-dimensional feature representation, physical simulation and data-driven prediction are executed in parallel to generate a power prediction sequence.

[0045] Physical simulation refers to the process of simulating and calculating the future flow field and power generation of a wind farm using physical models based on the principles of fluid mechanics and aerodynamics.

[0046] Data-driven forecasting refers to the process of predicting wind power output by using machine learning or deep learning models to learn statistical patterns from historical data.

[0047] Specifically, the multi-fidelity collaborative prediction engine receives high-dimensional feature representations output by the multi-modal fusion and representation module, and executes two prediction paths in parallel: First, the low-fidelity physical simulation unit uses a parameterized wake model to quickly simulate the future flow field of the wind farm based on physical factors such as wind speed, wind direction, and turbulence from the high-fidelity feature representations, generating physical priors that include the initial wind speed and power distribution at each turbine location; second, the high-fidelity data-driven prediction unit uses all the fused features of the high-fidelity feature representations as input, employing a hybrid architecture of a bidirectional long short-term memory network with an attention mechanism and a spatiotemporal graph convolutional network to mine complex nonlinear statistical patterns in historical data and generate statistical data predictions. Subsequently, the physics-guided deep operator network fusion unit uses the physical priors and statistical data predictions as dual-channel inputs, performs adaptive weighted fusion through a fusion gating network, and explicitly adds physical residual constraint terms based on fluid dynamics equations to the training loss function, ensuring that the fusion result conforms to both statistical data patterns and physical consistency, ultimately generating a power prediction sequence for the next few minutes to several days. This process achieves a balance between prediction accuracy and physical rationality through the collaboration and fusion of physical simulation and data-driven prediction.

[0048] The wind power generation prediction method provided in this embodiment combines the mechanism priors provided by rapid physical simulation with the high-precision learning capabilities of data-driven models through a multi-fidelity collaborative prediction engine. This engine explicitly introduces physical consistency constraints based on fluid dynamics equations during training and achieves backpropagation of physical residuals through a differentiable physical computation layer, so that the final prediction results not only conform to the statistical laws of data but also satisfy the basic physical laws. This effectively alleviates the problem of insufficient generalization ability of pure data-driven models under extreme weather or historically unseen conditions, and improves the physical rationality and temporal consistency of the prediction.

[0049] Step S503: Based on the real-time operation data and power prediction sequence of the wind farm, construct a digital twin environment for wind farm operation.

[0050] Among them, the wind farm operation digital twin environment refers to a high-fidelity virtual simulation environment constructed by coupling multi-dimensional dynamic models based on real-time wind farm operation data and power prediction sequences. It can map the flow field distribution, equipment status and grid interaction behavior of the physical wind farm in real time, and serve as a simulation platform for strategy deduction and optimization decision-making.

[0051] Specifically, the digital twin receives the power prediction sequence output by the multi-fidelity collaborative prediction engine and uses it together with the real-time wind farm operation data (including real-time wind measurement data, SCADA real-time monitoring data, and real-time grid-connected power) provided by the multi-modal fusion and characterization module as input. Through multi-dimensional modeling and coupling, a unified digital twin environment for wind farm operation is generated: First, based on the wind speed and direction predictions in the real-time wind measurement data and power prediction sequence, a three-dimensional dynamic distribution of the wind farm flow field is constructed; second, based on the SCADA real-time monitoring data, an operation status and health model of key components of the wind turbine is constructed; third, based on the real-time grid-connected power and power prediction sequence, a dynamic model of wind farm power output and grid interaction is constructed; finally, the above three models are coupled to form a unified digital twin environment that can dynamically map the physical wind farm flow field, equipment status, and grid interaction behavior, and is continuously updated synchronously based on newly acquired real-time data and updated prediction sequences to maintain high-fidelity consistency with the physical wind farm.

[0052] This digital twin environment serves as a high-fidelity simulation platform for subsequent multi-objective collaborative optimization, providing an iterative and risk-free virtual simulation sandbox for strategy deduction.

[0053] Step S504: In the digital twin environment of wind farm operation, the conflict between multiple optimization objectives is solved through multi-round iterative game to obtain the power generation plan and bidding strategy.

[0054] Among them, multiple optimization objectives refer to three core objectives that need to be considered simultaneously but are in conflict during the operation of a wind farm: power generation revenue, grid security, and equipment health.

[0055] Among them, the core demand for power generation revenue is to maximize the economic income of wind farms in the electricity market. The source of conflict in power generation revenue is that pursuing high revenue may require full power generation or high price bidding, but may deviate from dispatch instructions or accelerate equipment wear and tear.

[0056] The core objective of power grid security is to maximize compliance with power grid dispatch instructions and ensure stable power grid operation. The source of conflict in power grid security is that strictly following dispatch instructions may limit power generation and affect revenue.

[0057] The core objective of equipment health is to maximize the remaining lifespan of critical components and reduce the risk of failure. The source of conflict in equipment health is that limiting power generation to protect equipment may sacrifice power generation revenue.

[0058] Specifically, the dynamic collaborative optimizer is based on a digital twin environment for wind farm operation constructed from digital twins. It incorporates a multi-agent decision-making framework based on non-cooperative game theory. This framework includes virtual agents representing three optimization objectives: power generation revenue, grid security, and equipment health. Each agent has a quantified utility function. The dynamic collaborative optimizer first initializes decision schemes for each agent, then enters a multi-round iterative game: in each iteration, each agent acts sequentially. Assuming the decision schemes of other agents remain unchanged, it uses the digital twin environment for simulation to find the optimal response scheme that maximizes its own utility function, and updates its own decision scheme to this optimal response scheme. This iteration is repeated until the change in the decision schemes of all agents is less than a preset threshold. At this point, the strategy combination converges to Nash equilibrium, outputting a Pareto-optimal power generation plan and bidding strategy.

[0059] This process autonomously coordinates the conflicts between power generation revenue, grid security, and equipment health in a virtual environment through a multi-agent game mechanism, achieving multi-objective collaborative optimization decision-making. The output power generation plan and bidding strategy can be directly used to guide wind farm operation and market transactions.

[0060] The wind power generation prediction method provided in this embodiment constructs a digital twin of wind farm operation and deploys a multi-agent dynamic collaborative optimizer based on non-cooperative game theory. This optimizer models multiple conflicting objectives such as power generation revenue, grid dispatch compliance, and equipment health life as utility functions of different agents. Through multi-round game simulation in the digital twin environment, it autonomously finds the Pareto optimal power generation plan and market strategy, realizing a leap from simple power prediction to an integrated closed loop of "prediction-optimization-decision". This significantly improves the economy, safety, and equipment reliability of wind farm operation and reduces reliance on human decision-making and subjective bias.

[0061] This embodiment provides a wind power generation prediction method, which can be used for the aforementioned wind power generation prediction. Figure 6 This is a flowchart of a wind power generation prediction method according to an embodiment of the present invention, such as... Figure 6 As shown, the process includes the following steps: Step S601: Obtain high-dimensional feature representation. The high-dimensional feature representation is obtained by mapping multi-source heterogeneous spatiotemporal sequence data to a decoupled latent space through a feature variational autoencoder.

[0062] The training objective function of the feature variational autoencoder is defined as a weighted sum of the following three terms: (1); in, The mean square error between the input data and the reconstructed data; The Kullback-Leibler divergence between the latent space variable distribution and the standard normal prior distribution is given by the hyperparameter. Control its intensity; To decouple the loss term, a method based on total correlation or group distance is used to minimize the statistical dependencies between the dimensions of the latent space variables, determined by hyperparameters. Control its intensity; Decoupling loss term The following formula is used for calculation: (2); or: (3); in, , The i and j-th dimensions of the latent variables are represented, and corr represents the correlation coefficient calculated on the batch data. Denotes KL divergence, For the joint empirical distribution of latent variables, The product of the edge distributions of each dimension, where D is the total dimension of the latent space.

[0063] Specifically, this step is implemented through a multimodal fusion and representation module, whose multimodal data fusion method includes the following steps: Step 1: Receive multi-source heterogeneous spatiotemporal sequence data from at least numerical weather prediction, wind farm monitoring systems, lidar wind measurement devices, computational fluid dynamics simulation software, and electricity market information platforms; Numerical weather forecast data includes data from the European Centre for Medium-Range Weather Forecasts (ECMWF) and the Global Forecast System (GFS), acquiring gridded meteorological data with a temporal resolution of 1 hour and a spatial resolution of 0.1°×0.1° for the next 72 hours. The data includes 20 meteorological elements such as wind speed (10m, 50m, 100m altitude), wind direction, temperature, air pressure, and humidity.

[0064] The wind farm monitoring system collects real-time operating data of the wind turbines through the SCADA system, including: electrical parameters: active power, reactive power, voltage, current, power factor, with a sampling frequency of 1 second; mechanical parameters: pitch angle, speed, yaw angle, bearing temperature, gearbox temperature, generator temperature, with a sampling frequency of 10 seconds; and vibration parameters: tower vibration acceleration, nacelle vibration acceleration, with a sampling frequency of 100Hz.

[0065] LiDAR wind measurement device: Deployed at key locations in the wind farm to acquire: 3D wind speed profile: vertical from 10m to 200m, with a measurement point every 10m; turbulence intensity: calculated based on a 10-minute sliding window; wind direction change: based on 5-minute time series analysis data update frequency: 1 minute.

[0066] Computational fluid dynamics simulation software: High-fidelity flow field simulation based on the OpenFOAM platform. Inputs include: Topographic data: Digital Elevation Model (DEM), resolution 5m; Roughness data: Classified by land use type; Boundary conditions: Time-varying boundary conditions based on NWP data; Output results: Three-dimensional flow field of the entire wind farm, including wind speed, turbulence intensity, and pressure field, with a spatial resolution of 10m×10m×5m and a time resolution of 5 minutes.

[0067] Electricity Market Information Platform: Accesses the regional electricity trading center's API (Application Programming Interface) to obtain: Day-ahead market: Electricity price forecast curves for 96 time nodes (every 15 minutes) in the next 24 hours; Real-time market: Rolling electricity price forecasts for the current hour and the next hour; Ancillary services market: Frequency regulation capacity demand, reserve capacity demand, and ramp rate requirements; Grid dispatch instructions: AGC (Automatic Generation Control) instructions, AVC (Automatic Voltage Control) instructions, and planned power curves.

[0068] Step 2: Perform spatiotemporal alignment, missing value imputation, and outlier smoothing on the multi-source data streams to form tensors on a unified spatiotemporal grid.

[0069] Time alignment is performed in the following way: 1. Unified Time Standard: Convert all data timestamps to Coordinated Universal Time (UTC) and correct for time zone deviations; 2. Sampling frequency standardization: For second-level data (SCADA vibration data): use moving average filtering to downsample to a 1-minute interval; for minute-level data (LiDAR, CFD simulation): maintain the original sampling rate; for hour-level data (Numerical Weather Prediction, NWP): use cubic spline interpolation to upsample to a 15-minute interval. 3. Delay Compensation: Based on data source communication delay analysis (SCADA delay <1s, LiDAR delay <3s, CFD simulation delay <5 minutes, NWP delay <30 minutes), an advanced prediction model is used to compensate for data delay; Spatial alignment is performed in the following way: 1. Mesh processing: Establish a three-dimensional spatial coordinate system for the wind farm, with the origin located at the center of the wind farm. Mesh resolution: horizontal direction: 50m×50m grid; vertical direction: 10m layer, from the ground to a height of 200m. 2. Spatial interpolation methods: NWP grid data → wind farm grid: bilinear interpolation combined with terrain correction is used; discrete point data (wind turbine locations) → wind farm grid: inverse distance weighting (IDW) is used, with a weighting exponent p=2; lidar profile data → 3D grid: kriging interpolation is used, taking anisotropy into account; Data quality control and anomaly handling are carried out in the following ways: 1. Anomaly detection algorithms: 3σ criterion based on sliding window: detects short-term burst anomalies; Isolation forest algorithm: detects outliers in high-dimensional data; Long Short-Term Memory (LSTM) network: predicts normal data sequences and detects anomalies that deviate from the predicted values. 2. Anomaly Repair Strategy: For isolated outliers: use linear interpolation of valid data before and after to repair; for continuous outlier segments (<10 sampling points): use LSTM prediction sequences for repair; for long-term missing data (>30 minutes): start Generative Adversarial Network (GAN) for data synthesis, and use conditional GAN ​​to generate synthetic data that meets physical constraints; Data quality index calculation: define Data Availability Rate (DAR) = (number of valid data points / total number of data points) × 100%, and trigger an alert when DAR < 95%.

[0070] Step 3: Input the processed tensor into a pre-trained feature variational autoencoder, which maps the data to a decoupled latent space, where the different dimensions of the latent vectors are constrained to be independently related to a specific physical or business concept. The training and optimization of the feature variational autoencoder specifically involves designing the latent space dimension and optimizing the decoupling loss term, as follows: Implicit Space Dimension Design: Implicit space variable dimensions can be flexibly set according to business needs. For example: Dimensions 1-3: wind speed, wind direction, turbulence intensity; Dimensions 4 to 6: wind turbine mechanical status (bearing temperature, gearbox vibration, generator efficiency); Dimensions 7 to 8: wake effect intensity, overall efficiency of wind farm; Dimensions 9 to 10: market electricity price volatility, dispatch command change trend.

[0071] Optimization of the decoupling loss term: In addition to the total correlation or the grouped KL (Kullback-Leibler Divergence, relative entropy) divergence, mutual information can be introduced to minimize the loss, further strengthening the independence between the dimensions of latent variables; The training strategy adopts a progressive training approach, first training the reconstruction and KL loss, and then gradually introducing the decoupling loss to avoid training instability.

[0072] It should be noted that when the data source is missing, the corresponding dimensions are empty in the decoupled latent space and the complete high-dimensional feature representation is reconstructed by decoding. In the decoupled latent space, different dimensions correspond to at least one physical factor among wind speed, turbulence intensity, unit mechanical state, wake effect intensity and market volatility, and each dimension is associated with the corresponding data source.

[0073] Specifically, in the online prediction stage, if some data sources are detected to be missing, the corresponding concept dimension is set to zero or prior distribution sampling is performed in the latent space, and then the complete high-dimensional feature representation is reconstructed by the decoder for use by the downstream prediction model.

[0074] As shown above, by mapping multi-source heterogeneous data to a decoupled latent space through feature variational autoencoders and introducing a decoupling loss term during training, each dimension of the latent variables independently corresponds to physical or operational factors such as wind speed, turbulence, and equipment status. This deep fusion at the feature level, compared to traditional simple data splicing, can more fully explore the inherent correlation between data, improve the robustness and interpretability of feature representation, and, in addition, when a specific data source is missing, the system can empty the corresponding dimension in the latent space and reconstruct the complete feature vector through the decoder, thereby ensuring that the prediction process is not interrupted and significantly improving the system's fault tolerance and reliability.

[0075] Step S602: Based on high-dimensional feature representation, physical simulation and data-driven prediction are executed in parallel to generate a power prediction sequence.

[0076] Specifically, step S602 includes: Step S6021: Based on high-dimensional feature representation, perform physical simulation using a physical simulation model to generate physical priors; the physical priors include the initial wind speed and power distribution of each wind turbine in the wind farm in the future time period.

[0077] Specifically, a parameterized Jensen wake model or a flow field super-resolution model based on a deep neural network is used to quickly simulate the future flow field of the entire wind farm in seconds and output the initial power distribution.

[0078] Step S6022: Based on high-dimensional feature representation, data-driven prediction is performed using a data-driven model to generate data statistical prediction.

[0079] Specifically, using a data-driven model, based on historical data and high-dimensional feature representations, a power prediction purely based on statistical data patterns is generated, i.e., data statistical prediction. In step S6023, the physical prior and data statistical prediction are adaptively weighted and fused through a physics-guided deep learning model to generate a power prediction sequence.

[0080] Specifically, the aforementioned physical priors and statistical data predictions are input into a physics-guided deep operator network (i.e., a physics-guided deep learning model). This network calculates the residuals between the prediction results and the fundamental physical laws through a differentiable physical computation layer. When training the network, the norm of the physical residuals and the data prediction errors are included in the loss function for backpropagation, enabling the network to learn to generate final prediction values ​​that conform to both statistical data laws and physical constraints.

[0081] A physics-guided deep operator network fusion unit is used for adaptive weighting through a fusion gating network. The training loss function of this fusion network explicitly incorporates a physical consistency constraint term (physical residual loss function) based on fluid dynamics equations, where: The training loss function consists of two parts: a common data loss (the error between predicted and actual power) and a unique physical loss (the norm of the aforementioned physical residuals). The strength of the physical constraints is controlled by an adjustable weight coefficient λ, and the total training loss function is... Defined as: (4); in, To predict power Compared to actual power The loss function between; The physical residual loss function; This represents the intermediate physical field provided by the physical simulation unit; and To balance the weight coefficients of the two losses, and in the early stages of training Larger values ​​are guided by strong physics, later on. The value is increased in pursuit of higher data accuracy; Physical residual loss function Specifically, it is defined as the residual norm based on the simplified Navier-Stokes equations or the laws of conservation of mass and momentum: (5); in, For wind speed vector field, For pressure field, air density, This refers to kinematic viscosity.

[0082] When training a physics-guided deep operator network, the backpropagation algorithm teaches the network not only how to reduce data errors but also how to reduce physical residuals. This means that the network parameter updates are directed towards ensuring that the prediction results simultaneously satisfy historical data patterns and fundamental physical laws. In the early stages of training, a larger λ value can be set to force the network to output results with strong physical consistency. As training progresses, the weight α of data loss is gradually increased to bring the prediction accuracy closer to the measured data, ultimately resulting in a model that achieves the best balance between data fitting and physical reasonableness. The output after processing by the physics-guided fusion network is the system's final wind power prediction sequence. This prediction utilizes the complex patterns mined from history by the data-driven model and is constrained by physical laws, thus improving prediction accuracy, temporal consistency, and generalization ability to unknown scenarios.

[0083] This embodiment combines the mechanistic priors provided by rapid physical simulation with the high-precision learning capabilities of data-driven models through a multi-fidelity collaborative prediction engine. The engine explicitly introduces physical consistency constraints based on fluid dynamics equations during training and achieves backpropagation of physical residuals through a differentiable physical computation layer. This ensures that the final prediction results conform to both statistical data patterns and fundamental physical laws, effectively alleviating the problem of insufficient generalization ability of pure data-driven models under extreme weather or historically unseen conditions, and improving the physical rationality and temporal consistency of the predictions.

[0084] Step S603: Based on the real-time operation data of the wind farm and the power prediction sequence, construct a digital twin environment for wind farm operation.

[0085] Specifically, the real-time operation data of the wind farm includes: real-time wind measurement data, SCADA real-time monitoring data, and real-time grid-connected power; step S603 above includes: Step S6031: Based on real-time wind measurement data and power prediction sequence, construct the three-dimensional dynamic distribution of the wind farm flow field.

[0086] Specifically, the system receives real-time wind measurement data collected by devices such as lidar, obtains three-dimensional wind speed profiles and turbulence intensity information at key locations in the wind farm, and combines the wind speed and direction prediction results for future periods in the power prediction sequence. Through spatial interpolation and flow field reconstruction algorithms, the discrete point wind measurement data is extended to the entire three-dimensional spatial grid of the wind farm, generating a three-dimensional dynamic distribution of the flow field that includes wind speed vectors, turbulence intensity, and wind shear characteristics at each height level. This distribution can reflect the continuous change law of wind resources in the spatiotemporal dimension, providing a basis for subsequent equipment load calculation and wake effect analysis.

[0087] Step S6032: Based on SCADA real-time monitoring data, construct a model of the operating status and health of wind turbine units in the wind farm.

[0088] Specifically, the SCADA system collects electrical parameters (active power, reactive power, voltage, current), mechanical parameters (pitch angle, speed, bearing temperature, gearbox temperature), and vibration parameters (tower vibration acceleration, nacelle vibration acceleration) of each wind turbine in real time. By combining the equipment mechanism model and the data-driven model, key features reflecting the operating status of the equipment are extracted, and a real-time operating status and health model of the key components of each unit (blades, main shaft, gearbox, generator) is constructed. This model can quantify the current health status of the equipment and predict its degradation trend, providing a basis for subsequent equipment risk warning and maintenance decisions.

[0089] Step S6033: Based on the real-time grid connection point power and power prediction sequence, construct a dynamic model of wind farm power output and grid interaction.

[0090] Specifically, the system receives real-time active and reactive power data from the wind farm's grid connection point. Simultaneously, it combines power output predictions for future periods in the power prediction sequence with a power system dynamic response model to simulate the grid connection characteristics of the wind farm under different output conditions. This model constructs a dynamic model of power output and grid interaction that includes power ramping capability, frequency response characteristics, voltage regulation capability, and dispatch command tracking capability. This model can reflect the two-way interaction between the wind farm as a power generation unit and the grid, providing a simulation basis for subsequent grid security target optimization.

[0091] Step S6034: Couple the three-dimensional flow field dynamic distribution, the operating status and health model, and the dynamic model of power output and grid interaction to generate a digital twin environment for wind farm operation.

[0092] Specifically, the three models mentioned above are coupled in a unified spatiotemporal coordinate system: the three-dimensional flow field dynamic distribution provides the wind load input of each unit's location to the operation status and health model, affecting equipment load calculation and health status inference; the operation status and health model feeds back the actual output capacity and operating constraints of the equipment to the power output and grid interaction dynamic model, correcting the executable power range; the power output and grid interaction dynamic model feeds back the impact of power adjustment on the wake to the flow field distribution model according to grid dispatch requirements. Through bidirectional data interaction and state synchronization among the three, the system generates a unified digital twin environment that can dynamically map the physical wind farm flow field, equipment status, and grid interaction behavior, and continuously updates synchronously based on newly acquired real-time data and updated prediction sequences, maintaining high-fidelity consistency with the physical wind farm.

[0093] Step S604: In the digital twin environment of wind farm operation, the conflict between multiple optimization objectives is solved through multi-round iterative game to obtain the power generation plan and bidding strategy.

[0094] Specifically, this step is implemented by a multi-agent decision-making framework in a dynamic collaborative optimizer, and step S604 includes: Step S6041: In the digital twin environment of the wind farm operation, initialize the decision-making schemes of each virtual agent.

[0095] Specifically, virtual intelligent agents include: The utility function of an intelligent agent that generates electricity revenue The total revenue calculated based on the spot market price curve and ancillary service prices within the forecast period is minus the performance penalty caused by forecast deviation. The utility function of a power grid security intelligent agent To track the negative value of the accumulated deviation for dispatch instructions, and to additionally consider the potential support capability for grid frequency and voltage stability; Device health intelligent agent, its utility function The negative value of the cumulative fatigue load of the key component is based on the rainflow counting method. The load is estimated in real time by a high-fidelity structural dynamics model in a digital twin.

[0096] In addition to the three virtual intelligent bodies of economy, safety, and equipment health, an environmental emission intelligent body (considering carbon emission constraints) and a grid flexibility intelligent body (participating in demand response) can be added. The economic intelligent body can introduce a risk aversion coefficient to optimize revenue fluctuations; the safety intelligent body can consider reserve capacity under grid failure scenarios; and the equipment health intelligent body can introduce a remaining life prediction model to dynamically adjust maintenance strategies.

[0097] Each virtual agent initializes its decision-making scheme based on the current digital twin environment of the wind farm operation (e.g., the power generation revenue agent initializes a revenue maximization plan).

[0098] Step S6042: In each iteration, with the decision schemes of other virtual agents remaining unchanged, solve for the optimal response scheme that maximizes the utility function of the current virtual agent, and update the decision scheme of the current virtual agent to the optimal response scheme.

[0099] Specifically, in each iteration, each virtual agent acts in turn, simulating and deducing in the digital twin while assuming that the decision schemes of other agents remain unchanged, to find the optimal response scheme that maximizes its own utility function, and then updating the policy of each virtual agent to the optimal response scheme.

[0100] Step S6043, repeat the iteration until the change in the decision scheme of all virtual agents is less than the preset threshold, and obtain the optimal power generation plan and bidding strategy; the power generation plan and bidding strategy include: power prediction curve and trading strategy suggestion.

[0101] Specifically, repeat steps S6041 and S6042 until the policy changes of all virtual agents are less than the preset threshold or the maximum number of iterations is reached. At this point, the policy combination converges to the Nash equilibrium point, which is the final output of the system.

[0102] More specifically, (taking the example of a power generation revenue agent acting first): Step a1: Assuming the policies of the other two agents remain unchanged, the power generation revenue agent searches for the optimal power generation plan that maximizes its own utility function in a digital twin simulation environment (e.g., generating more power when electricity prices are high), and updates its own policy to the optimal plan. Step a2, (Grid Safety Agent Action): The grid safety agent observes the updated strategy of the power generation revenue agent, but assumes that the equipment health agent strategy remains unchanged. It simulates in the digital twin to find how to adjust the wind farm operating parameters (such as limiting the ramp rate) to maximize its own utility (i.e., to be closer to the dispatch instructions) under the current power generation plan, and then updates its own strategy. Step a3, (Equipment Health Agent Action): Based on the updated strategies from the previous two steps, the equipment health agent searches for and updates operational strategies that can reduce equipment fatigue damage in the simulation (such as appropriately reducing the load when there is high turbulence). Step a 4. In each round, each agent makes the most advantageous response based on the latest policies of other agents. The policy changes will become smaller and smaller. When the policy changes of all agents are less than a preset small threshold, the system considers that a Nash equilibrium has been reached. At the equilibrium point, any agent that unilaterally changes its own policy can no longer improve its own utility. This equilibrium policy combination is the best compromise solution for multiple conflicting objectives under the current conditions.

[0103] This embodiment requires further explanation of the multi-objective collaborative optimization method implemented using a dynamic collaborative optimizer, which specifically includes the following steps: Step b1: In the digital twin environment of wind farm operation, define multiple virtual intelligent agents that represent economic, safety and equipment health benefits respectively. This can be extended to environmental (carbon emissions), social (community impact) and other aspects. Create a corresponding virtual intelligent agent for each important goal and design a quantified utility function for it to mathematize its benefit claims. Step b2: Establish a quantified utility function for each agent and encode its preference for the power generation plan into adjustable policy parameters; Specifically, a high-fidelity digital twin of wind farm operation is used as a game simulation sandbox. Based on real-time data and predicted sequences, the digital twin can dynamically simulate the flow field changes, equipment response, and interaction with the power grid of the wind farm. Any strategy adjustment of each agent can be quickly simulated in this environment, and the impact on its own and other parties' utility functions can be calculated. Step b3: Design a multi-round game protocol. Each agent, based on its current digital twin state and the strategies of other agents, finds its optimal response strategy through simulation and updates it sequentially. Step b4: When the game reaches Nash equilibrium, the equilibrium strategy is decoded into a specific wind farm power setpoint curve, market application scheme and recommended operating limits. This strategy represents the best compromise solution for multi-objective conflicts under given conditions.

[0104] Once the game converges, the final equilibrium strategy is decoded into specific, executable instructions: Power generation plan curve: Issued to the wind farm energy management system to guide the power setting of each wind turbine; Market bidding strategy: This includes the electricity volume and price curves in the spot electricity market, as well as capacity application suggestions in the ancillary services market, which are sent to the electricity trading terminal. Equipment maintenance recommendations: Based on the deduction results of the equipment health intelligence agent in the game, early risk warnings and preventive maintenance time window suggestions for key components are generated and pushed to the maintenance platform.

[0105] Specifically, the load spectrum of wind turbine components is simulated based on the digital twin environment of wind farm operation, and an equipment risk warning report is generated based on the component load spectrum. The power generation plan, bidding strategy, and equipment risk warning report are then distributed to the wind farm's energy management system, power trading terminal, and predictive maintenance platform.

[0106] Step S605: Deploy a global model in the cloud to aggregate model updates from multiple wind farms; deploy a local prediction process in each wind farm and train it using local private data; participate in federated aggregation of latent variable dimensions related to physical laws in the feature variational autoencoder, and keep the latent variable dimensions related to wind farm attributes in local updates; synchronize the physical loss term computation graph in the physics-guided deep learning model as public knowledge in the cloud and locally, and upload the network parameters of the data-driven model to the cloud for secure aggregation after protecting them with differential privacy technology.

[0107] The wind power generation prediction method provided in this embodiment constructs a digital twin environment for wind farm operation and deploys a multi-agent dynamic collaborative optimizer based on non-cooperative game theory. This optimizer models multiple conflicting objectives such as power generation revenue, grid dispatch compliance, and equipment health lifespan as utility functions of different agents. Through multi-round game simulation in the digital twin environment, it autonomously finds the Pareto optimal power generation plan and market strategy, realizing a leap from simple power prediction to an integrated closed loop of "prediction-optimization-decision". This significantly improves the economy, safety, and equipment reliability of wind farm operation and reduces reliance on human decision-making and subjective bias.

[0108] As one or more specific application embodiments of the present invention, the wind power generation prediction method provided by the present invention will be further described in detail in conjunction with a wind power generation prediction system, specifically including: First, a multimodal fusion and representation module is used to receive multi-source heterogeneous data in real time from numerical weather prediction, wind farm SCADA systems, lidar, CFD simulation, and the power market information platform. After spatiotemporal alignment, quality control, and anomaly repair, the data is input into a pre-trained feature variational autoencoder. This encoder maps the data to a decoupled latent space, where different dimensions correspond to independent physical or operational factors. If a data source (such as lidar) signal is interrupted, the system zeros out or performs prior sampling on the corresponding dimension of that factor, and then reconstructs the complete fused feature vector through the decoder, thereby ensuring the robustness of subsequent processes.

[0109] Next, a multi-fidelity collaborative prediction engine receives the fused features and launches a low-fidelity physical simulation unit and a high-fidelity data-driven prediction unit in parallel. The physical simulation unit quickly generates physical priors for future flow fields and power distributions based on a parameterized wake model. The data-driven unit mines historical statistical patterns based on an attention mechanism and a spatiotemporal graph convolutional network. The outputs of both are fed into a physics-guided deep operator network fusion unit. This network optimizes data fitting errors and physical equation residuals during training through a differentiable physical computation layer, ultimately generating a wind power prediction sequence that is both accurate and physically consistent. This prediction sequence is synchronously input into the digital twin with real-time data, driving it to continuously update the high-fidelity dynamic virtual mapping of wind farm flow fields, equipment status, and grid connection point power. On this basis, the dynamic collaborative optimizer launches its built-in multi-agent game framework: three virtual agents—power generation revenue, grid security, and equipment health—each perform strategy deduction and iterative response based on their own utility functions in the digital twin environment until the strategy combination converges to a Nash equilibrium. This equilibrium point is the best compromise solution for multiple operational objectives under the current prediction and constraints.

[0110] Finally, the output and application execution module decodes the equalization strategy into three types of executable instructions: a high-precision power prediction curve is sent to the power station energy management system to optimize internal power allocation; market trading strategy suggestions are sent to the power trading terminal to guide spot and ancillary service market applications; and equipment risk warning reports are pushed to the predictive maintenance platform to suggest maintenance time windows and measures.

[0111] The wind power prediction method provided in this embodiment maps multi-source heterogeneous data to a decoupled latent space through a feature variational autoencoder. A decoupling loss term is introduced during training, so that each dimension of the latent variable independently corresponds to wind speed, turbulence, and equipment state physical factors. This deep fusion at the feature level, compared with the traditional simple data splicing, can more fully explore the inherent correlation between data, improve the robustness and interpretability of feature representation. In addition, when a specific data source is missing, the system can empty the corresponding dimension in the latent space and reconstruct the complete feature vector through the decoder, thereby ensuring that the prediction process is not interrupted and significantly improving the fault tolerance and reliability of the system.

[0112] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0113] The following is a detailed reference. Figure 7 This diagram illustrates a suitable structural schematic for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 701, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 702 or a program loaded from memory 708 into random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the electronic device. The processor 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0114] Typically, the following devices can be connected to I / O interface 705: input devices 706 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 707 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 708 including, for example, magnetic tapes, hard disks, etc.; and communication devices 709. Communication device 709 allows electronic devices to exchange data via wireless or wired communication with other devices. Although Figure 7 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0115] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 709, or installed from a memory 708, or installed from a ROM 702. When the computer program is executed by the processor 701, it performs the functions defined in the wind power generation prediction method of the embodiments of the present invention.

[0116] Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0117] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the wind power generation prediction method shown in the above embodiments is implemented.

[0118] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0119] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for predicting wind power generation, characterized in that, The method includes: A high-dimensional feature representation is obtained by mapping multi-source heterogeneous spatiotemporal sequence data to a decoupled latent space through a feature variational autoencoder; Based on the high-dimensional feature representation, physical simulation and data-driven prediction are executed in parallel to generate a power prediction sequence; Based on the real-time operation data of the wind farm and the power prediction sequence, a digital twin environment for wind farm operation is constructed. In the digital twin environment of the wind farm operation, the conflict between multiple optimization objectives is solved through multiple rounds of iterative game to obtain the power generation plan and bidding strategy.

2. The method according to claim 1, characterized in that, The acquisition of high-dimensional feature representation also includes: When the data source is missing, the corresponding dimensions are empty in the decoupled latent space and the complete high-dimensional feature representation is reconstructed by decoding. In the decoupled latent space, different dimensions correspond to at least one physical factor among wind speed, turbulence intensity, unit mechanical state, wake effect intensity and market volatility, and each dimension is associated with the corresponding data source.

3. The method according to claim 1, characterized in that, The process of generating a power prediction sequence by performing parallel physical simulation and data-driven prediction based on the high-dimensional feature representation includes: Based on the high-dimensional feature representation, physical simulation is performed using a physical simulation model to generate physical priors; the physical priors include the initial wind speed and power distribution of each wind turbine in the wind farm in the future time period. Based on the high-dimensional feature representation, a data-driven model is used to perform data-driven prediction and generate data statistical prediction. The physical priors and the statistical data predictions are adaptively weighted and fused using a physics-guided deep learning model to generate a power prediction sequence.

4. The method according to claim 1, characterized in that, The real-time operation data of the wind farm includes: real-time wind measurement data, SCADA real-time monitoring data, and real-time grid connection point power. The construction of a digital twin environment for wind farm operation based on real-time wind farm operation data and the power prediction sequence includes: Based on the real-time wind measurement data and the power prediction sequence, a three-dimensional dynamic distribution of the flow field of the wind farm is constructed. Based on the SCADA real-time monitoring data, a model of the operating status and health of wind turbine units in a wind farm is constructed. Based on the real-time grid connection point power and the power prediction sequence, a dynamic model of wind farm power output and grid interaction is constructed. The three-dimensional flow field dynamic distribution, the operating status and health model, and the dynamic model of power output and grid interaction are coupled to generate a digital twin environment for wind farm operation.

5. The method according to claim 1, characterized in that, By resolving conflicts among multiple optimization objectives through multi-round iterative game theory, the power generation plan and bidding strategy are obtained, including: In the digital twin environment of the wind farm operation, the decision-making schemes of each virtual agent are initialized; In each iteration, with the decision schemes of other virtual agents remaining unchanged except for the current virtual agent, the optimal response scheme that maximizes the utility function of the current virtual agent is solved, and the decision scheme of the current virtual agent is updated to the optimal response scheme. The process is repeated iteratively until the change in the decision schemes of all virtual agents is less than a preset threshold, thus obtaining the optimal power generation plan and bidding strategy; the power generation plan and bidding strategy include: power prediction curve and trading strategy suggestions.

6. The method according to claim 3, characterized in that, The method further includes: Deploy a global model in the cloud to aggregate model updates from multiple wind farms; Deploy local prediction processes at each wind farm and train them using local private data; The latent variable dimensions related to physical laws in the feature variational autoencoder are participated in federated aggregation, while the latent variable dimensions related to wind farm station attributes are kept and updated locally. The computation graph of the physical loss term in the physics-guided deep learning model is synchronized with the local machine as public knowledge in the cloud, and the network parameters of the data-driven model are uploaded to the cloud for secure aggregation after being protected by differential privacy technology.

7. The method according to claim 1, characterized in that, The method further includes: The load spectrum of wind turbine components is simulated based on the digital twin environment of wind farm operation, and an equipment risk warning report is generated based on the component load spectrum; The power generation plan, bidding strategy, and equipment risk warning report are distributed to the wind farm's energy management system, power trading terminal, and predictive maintenance platform.

8. A wind power generation prediction system, characterized in that, The system, applied to the wind power generation prediction method as described in any one of claims 1 to 7, comprises: The multimodal fusion and representation module is used to acquire multi-source heterogeneous spatiotemporal sequence data and map the multi-source heterogeneous spatiotemporal sequence data to a decoupled latent space through a feature variational autoencoder to obtain high-dimensional feature representations. A multi-fidelity collaborative prediction engine is connected to the multi-modal fusion and characterization module. The multi-fidelity collaborative prediction engine is used to perform physical simulation and data-driven prediction in parallel based on the high-dimensional feature characterization to generate a power prediction sequence. A digital twin, connected to the multi-fidelity collaborative prediction engine, is used to construct a digital twin environment for wind farm operation based on real-time wind farm operation data and the power prediction sequence. A dynamic collaborative optimizer, connected to a digital twin, is used in the digital twin environment of the wind farm operation to solve the conflicts between multiple optimization objectives through multi-round iterative game theory, and obtain the power generation plan and bidding strategy.

9. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory stores computer instructions, and the processor executes the wind power generation prediction method according to any one of claims 1 to 7 by executing the computer instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the wind power generation prediction method according to any one of claims 1 to 7.

Images