A wind power prediction method based on meteorological and electrical multi-modal fusion
By employing the Mamba dual-branch coupled multi-scale Kronecker tensor interactive fusion method and the multi-expert knowledge reuse architecture of the diffusion model, the problems of difficult multimodal data fusion and low cross-domain migration efficiency in wind power prediction are solved, achieving high-precision, real-time wind power prediction and adaptive capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAIYIN INSTITUTE OF TECHNOLOGY
- Filing Date
- 2026-05-06
- Publication Date
- 2026-06-05
Smart Images

Figure CN122159204A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of new energy power generation prediction technology, and in particular relates to a wind power prediction method based on the fusion of meteorological and electrical multi-modal data. Background Technology
[0002] Wind energy, as a clean and renewable energy source, is a core force in building new power systems. However, its power output is significantly affected by meteorological conditions, exhibiting marked intermittency, volatility, and randomness, posing substantial challenges to grid dispatch, stable operation, and power consumption. High-precision wind power forecasting has become a crucial supporting technology. Current wind power forecasting technologies mainly include physical modeling, statistical methods, and deep learning methods. Physical modeling methods have clear physical meaning but low spatiotemporal resolution, making it difficult to capture microscale meteorological changes; statistical methods offer stable short-term predictions but cannot handle the strong nonlinear characteristics of the data; while deep learning methods improve accuracy, they also have significant bottlenecks. Existing technologies suffer from three core defects: First, meteorological and electrical multimodal data are mostly shallowly spliced together, failing to establish deep coupling relationships and resulting in insufficient data utilization; second, cross-domain generalization ability is poor, with differences in data distribution across different wind farms leading to decreased model transfer performance, and newly built wind farms struggling to achieve accurate forecasts due to data scarcity; third, adaptability to dynamic operating conditions such as extreme weather and power ramp-up is weak, with fixed model structure parameters unable to cope with time-varying complex scenarios. Existing technologies cannot meet the requirements of new power systems for "high accuracy, rapid adaptation, and wide applicability" in wind power forecasting. Therefore, there is an urgent need for a wind power forecasting system and method based on the fusion of meteorological and electrical multimodal data. Summary of the Invention
[0003] Objective: This invention addresses the challenges of multi-source heterogeneous data fusion, frequent concept drift caused by dynamic marine environments, scarce extreme event samples, and low efficiency of cross-domain transfer in offshore wind power forecasting. It provides a wind power forecasting method based on the fusion of meteorological and electrical multimodal data. Existing technologies suffer from the following drawbacks: single-modal data cannot comprehensively characterize the dynamic changes in power, limiting prediction accuracy; concept drift detection is lagging and lacks sufficient type recognition capabilities, preventing the model from adapting to distribution changes in a timely manner; the scarcity of extreme samples leads to poor prediction reliability in extreme scenarios; and cross-site model transfer is difficult, resulting in low knowledge reuse efficiency. By integrating multimodal features, real-time concept drift perception, incremental generation of extreme samples, and rapid cross-domain knowledge reuse, this invention comprehensively improves the accuracy, real-time performance, and generalization ability of offshore wind power forecasting, providing adaptive and transferable basic data support for intelligent operation and maintenance and grid dispatching of multi-site offshore wind power.
[0004] The method includes the following steps:
[0005] Step 1: Collect meteorological data and operational signal data from offshore wind farms to construct a multimodal raw data stream;
[0006] Step 2: Perform heterogeneous feature alignment on meteorological and operational signal data, eliminate differences through time synchronization and feature space mapping, and form a time-domain aligned multimodal feature sequence;
[0007] Step 3: Construct a Mamba dual-branch coupled multi-scale Kronecker tensor interaction fusion method; use the Mamba model to capture long-term temporal dependencies, realize cross-modal interaction through dual-branch coupling, use multi-scale tensor interaction to characterize nonlinear synergistic effects, and realize adaptive step-size discretization through extreme event sensitivity gating to output joint feature representation.
[0008] Step 4: Based on joint feature representation, a feature space offset sensing mechanism is used to detect joint distribution changes in real time, and high-dimensional statistical indicators are used to quantify drift intensity. The concept of drift type is identified based on the offset speed and pattern recognition.
[0009] Step 5: In high-risk scenarios, extreme samples are generated using Copula tail dependency modeling and time block consistency reconstruction to achieve incremental training and distribution adaptation. The generated extreme samples are then fed back to Step 2 for dynamic optimization of feature alignment parameters.
[0010] Step 6 introduces a multi-expert knowledge reuse architecture based on a diffusion model. In the meta-knowledge space, a dedicated expert model adapted to the new task is generated on demand. Expert predictions are integrated through a distributed consensus mechanism to output the final prediction results. The prediction results are compared with the real data, and the prediction bias is used to drive the closed-loop update of the model parameters. The updated knowledge is fed back to Step 3 to optimize the Mamba model parameters.
[0011] Step 1 includes: meteorological data including wind speed at time t. ,wind direction High waves At the first sampling frequency Data acquisition; operational signal data includes the shutdown signal at time t. Inertia response signal Primary frequency modulation signal At the second sampling frequency Collection, and ;No. The timestamp of each data item is This forms the raw data stream. :
[0012] ,
[0013] in, The first The meteorological data vector in the data and the first The running signal data vector in the data.
[0014] Step 2 includes:
[0015] Step 2.1, Time Synchronization;
[0016] Based on the timestamps of meteorological data, cubic spline interpolation is used to resample the operational signal data and transform the operational signal data onto the time axis of meteorological data. At the same time, a mutation point detection algorithm is used to identify the jump moments in the operational signal and align them with the corresponding events in the meteorological data.
[0017] Step 2.2, Feature space mapping;
[0018] A dual-branch feature embedding network is constructed to extract features from meteorological data and operational signal data separately, mapping the meteorological data and operational signal data to a unified feature space, thus obtaining a meteorological data encoder. With running data encoder Output feature vectors of meteorological data with uniform dimensions. eigenvectors of running signal data A contrastive learning mechanism is introduced to make cross-modal features at the same time point closer to each other in space, while features at different times point further apart.
[0019] Step 2.3: Construct multimodal feature sequences;
[0020] The aligned meteorological features and operational features are then stitched together to form the data for each time step. The comprehensive feature vector:
[0021] ,
[0022] in, , The first The meteorological feature vector and the operational signal feature vector aligned at each time step This is a vector concatenation operation;
[0023] Obtain the temporally aligned multimodal feature sequence X:
[0024] ,
[0025] in, To indicate the first The comprehensive feature vector at each time step, i.e., the feature representation of the last time step in the time series. This represents the total length of the time series.
[0026] Step 3 includes:
[0027] Step 3.1: Meteorological-Operational Two-Branch Coupled State-Space Modeling;
[0028] A dual-branch coupling structure is constructed, which performs independent state evolution for meteorological and operational characteristics respectively, and achieves cross-modal interaction through coupling terms. Momentary state vector With running state vector The coupled evolution equation is:
[0029] ,
[0030] in, To represent the differential, , These are the state transition matrices for the meteorological branch and the operational branch, respectively. , The cross-modal coupling matrix represents the influence of operational signals on meteorological dynamics and the driving effect of meteorological conditions on operational response, respectively. , These are the input mapping matrices for the meteorological branch and the operational branch, respectively.
[0031] Step 3.2, Multi-scale cooperative state decomposition and inter-scale coupling mechanism;
[0032] To investigate the coordinated variation patterns of marine meteorological elements and wind power output at different time scales, meteorological state vectors are used. Decomposed into Subvectors at each time scale: , where the subvector at the i-th time scale Corresponding time scale The coupling between scales is achieved through tensor products:
[0033] ,
[0034] in, For the first The state transition matrix at the scale, It is a third-order coupled tensor. For the first Sub-state vectors at different scales For the Kronecker product operation, For the first The input mapping matrix at the scale, For the meteorological feature vector at the th Scale components;
[0035] Step 3.3, Adaptive discretization of sensitivity to extreme weather events;
[0036] Define an indicator function for extreme weather events:
[0037] ,
[0038] in, For a moment Intensity index of extreme weather events The L2 norm of the time gradient of meteorological characteristics. For indicator functions, Determine the threshold for extreme events;
[0039] Adaptive Discretization Step Size Represented as:
[0040] ,
[0041] in, As the reference discretization step size, The attenuation coefficient is... The modulation intensity coefficient, The hyperbolic tangent activation function is used. A multilayer perceptron network designed for extreme events.
[0042] The discretized state update equation is:
[0043] ,
[0044] in, For the first The system state vector at each time step This is the state vector from the previous time step. For the first The adaptive discretization step size of the step. This is the state transition matrix in a continuous-time state-space model. For continuous-time input mapping matrix, For the first The input feature vector at each time step;
[0045] Final output joint feature representation ,in, For the first A joint feature representation vector at each time step.
[0046] Step 4 includes:
[0047] Step 4.1, Joint distribution change detection;
[0048] A sliding window mechanism is used to calculate the maximum mean difference between the current window and the historical baseline window as the drift intensity index. An adaptive threshold is set, and when the drift intensity exceeds the threshold, a concept drift event is determined to have occurred.
[0049] Step 4.2, Drift type identification;
[0050] Based on the offset speed and offset pattern, concept drift is divided into three categories:
[0051] Sudden drift: Drift intensity increases dramatically in a short period of time, indicating extreme weather or sudden malfunction;
[0052] Gradual drift: The drift intensity increases slowly and continuously, indicating seasonal changes or equipment aging;
[0053] Repeated drift: The drift intensity exhibits periodic fluctuations, characterizing periodic changes such as tides and day / night cycles.
[0054] Step 5 includes:
[0055] Step 5.1, Tail dependency modeling based on Copula;
[0056] A multivariate Copula function is constructed to describe the tail dependency structure between meteorological variables and power output, and the joint probability distribution under extreme events is quantified. The Copula parameters are learned from historical extreme samples through maximum likelihood estimation.
[0057] Step 5.2, Time Block Consistency Reconstruction;
[0058] The original time series is divided into fixed-length time blocks. Conditional generative adversarial network (CGAN) or variational autoencoder (VAE) is used to generate new sample blocks with consistent temporal correlation with real extreme samples under Copula constraints. The consistency of time blocks is constrained by the autocorrelation function and cross-correlation function within the block.
[0059] Step 5.3, Incremental Training and Distribution Adaptation;
[0060] The generated extreme samples are incrementally added to the training set in a streaming manner. Empirical replay or gradient constraint methods are used to avoid catastrophic forgetting. At the same time, the difference between the generated sample distribution and the real extreme distribution is minimized by the maximum mean difference (MMD) or Wasserstein distance to achieve distribution adaptation. The generated extreme samples are fed back to step 2 and input together with the original multimodal feature sequence into the Mamba dual-branch coupled multi-scale Kronecker tensor interaction fusion method for incremental training.
[0061] Step 6 includes:
[0062] Step 6.1, Riemannian manifold modeling of the meta-knowledge space;
[0063] The parameters of the expert models trained from each source domain task are considered as points on a high-dimensional manifold, and a Riemannian metric is defined on the manifold. Introducing exponential mapping With logarithmic mapping This enables bidirectional transformations between manifolds and tangent spaces.
[0064] Step 6.2, manifold generalization of the conditional diffusion model;
[0065] To generate new expert parameters on the manifold, a conditional backdiffusion process is defined:
[0066] ,
[0067] in, For the first Step diffusion state and condition vector Next, generate the state of the previous time step. The probability density function, For Riemann index mapping, It follows a Gaussian distribution (normal distribution). For the Riemannian logarithmic mapping, The mean value predicted by the neural network. To preset the noise level, It is the identity matrix;
[0068] Step 6.3, noise prediction of physical information embedding;
[0069] Embed physical constraints in the noise prediction network and define the following noise prediction function:
[0070] ,
[0071] in, The output of the noise prediction network (the predicted noise). For noise prediction terms that are purely data-driven, These are the weighting coefficients for physical constraints. For model parameters gradient operator, physical operator Deviation between the quantization model output and the theoretical power curve and inertial response characteristics:
[0072] ,
[0073] in, For wind speed, For the indicator function (indicator function), when the wind speed Cutting in wind speed With rated wind speed The value is 1 when it is between 1 and 0; otherwise, it is 0. To represent partial derivatives, To predict power for the model, This is the theoretical power curve. The inertial response power predicted by the model. This is the theoretical inertial response value. These are the weighting coefficients.
[0074] Step 6.4, Distributed consensus guided by manifold geometry;
[0075] generate After the first expert model, a distributed consensus mechanism is used for collaborative prediction, defining the first... One expert on the sample Fusion weights :
[0076] ,
[0077] in, The first The parameter vector of the first expert model and the first... The parameter vector of an expert model The geodesic distance on the manifold. For the first The confidence level of each expert's prediction for the current sample. To assess the fit between the sample and the expert model, Temperature coefficient;
[0078] The final prediction is:
[0079] ;
[0080] in, For the sample The final power forecast value (including meteorological characteristics and operational signals), For the first An expert model for samples Individual power prediction values;
[0081] Step 6.5, closed-loop evolution mechanism.
[0082] Step 6.5 includes: after new task data has accumulated, fine-tuning the generated expert and updating the parameters. Add to Meta-Knowledge Space The Riemann metric is updated using gradient flow:
[0083] ,
[0084] in, For the Riemannian metric defined on the Riemannian manifold M, manifold parameters gradient operator, For the meta-knowledge space Expert model parameters Seeking expectations, The transition probability density induced by the conditional diffusion model. The reference distribution (such as historical parameter distribution or prior distribution); the updated parameters The data is then fed back to step 3 to optimize the Mamba model parameters, ultimately outputting an adaptive, transferable, and high-precision multi-site offshore wind power prediction.
[0085] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.
[0086] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.
[0087] Beneficial Effects: This invention addresses key challenges in offshore wind power prediction, including difficulties in fusing multi-source heterogeneous data, frequent concept drift caused by dynamic marine environments, scarcity of extreme event samples, and low efficiency in cross-domain migration across multiple sites. Through collaborative innovation in multimodal feature fusion, real-time perception of concept drift, incremental generation of extreme samples, and rapid reuse of cross-domain knowledge, it achieves a unified improvement in prediction accuracy, real-time performance, and generalization ability. Specific effects are as follows:
[0088] This invention innovatively designs a multimodal feature fusion mechanism based on the Mamba state-space model. Through a bi-branch coupled state-space equation for meteorological and operational data, multi-scale collaborative state decomposition, and inter-scale coupling mechanisms, it explicitly couples and models meteorological data with operational signal data. The linear complexity of the Mamba model makes it significantly more efficient than the Transformer when processing long-term time-series data. Simultaneously, the selective state-space mechanism enables adaptive modeling of the time-varying coupling relationship between meteorological and power, greatly improving the accuracy of power prediction.
[0089] This invention constructs a multi-expert knowledge reuse architecture based on a diffusion model. It models the expert parameter space as a Riemannian manifold and uses conditional diffusion to generate customized expert models for new tasks from the meta-knowledge space on demand. A distributed consensus mechanism is then used to collaboratively integrate multiple generating experts. This architecture overcomes the technical bottlenecks of traditional multi-expert systems, such as a fixed number of experts and high adaptation costs for new tasks. New sites or tasks require only a small number of samples to obtain a suitable model, significantly reducing training costs. Furthermore, through manifold evolution dynamics, a closed-loop continuous learning mechanism of generation-use-feedback-regeneration is formed, enabling the system to have lifelong learning capabilities. Attached Figure Description
[0090] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.
[0091] Figure 1 This is a structural diagram of the multimodal feature fusion mechanism based on the Mamba state space model in this invention.
[0092] Figure 2 This is a schematic diagram of the multi-expert knowledge reuse architecture based on the diffusion model in this invention.
[0093] Figure 3 This is a schematic diagram of the overall system flow provided by the present invention. Detailed Implementation
[0094] This invention provides a wind power prediction method based on the fusion of meteorological and electrical multimodal data, such as... Figure 1 , Figure 2 and Figure 3 As shown, it includes:
[0095] Step 1: Collect meteorological data and operational signal data from offshore wind farms to construct a multimodal raw data stream;
[0096] Step 2: Perform heterogeneous feature alignment on meteorological and operational signal data, eliminate differences through time synchronization and feature space mapping, and form a time-domain aligned multimodal feature sequence;
[0097] Step 3: Construct a Mamba dual-branch coupled multi-scale Kronecker tensor interaction fusion method; use the Mamba model to capture long-term temporal dependencies, realize cross-modal interaction through dual-branch coupling, use multi-scale tensor interaction to characterize nonlinear synergistic effects, and realize adaptive step-size discretization through extreme event sensitivity gating to output joint feature representation.
[0098] Step 4: Based on joint feature representation, a feature space offset sensing mechanism is used to detect joint distribution changes in real time, and high-dimensional statistical indicators are used to quantify drift intensity. The concept of drift type is identified based on the offset speed and pattern recognition.
[0099] Step 5: In high-risk scenarios, extreme samples are generated using Copula tail dependency modeling and time block consistency reconstruction to achieve incremental training and distribution adaptation. The generated extreme samples are then fed back to Step 2 for dynamic optimization of feature alignment parameters.
[0100] Step 6 introduces a multi-expert knowledge reuse architecture based on a diffusion model. In the meta-knowledge space, a dedicated expert model adapted to the new task is generated on demand. Multiple expert predictions are integrated through a distributed consensus mechanism to output the final prediction result. The prediction result is compared with the real data, and the prediction bias is used to drive the closed-loop update of the model parameters. The updated knowledge is fed back to Step 3 to optimize the parameters of the Mamba fusion model.
[0101] Step 1 includes the following steps:
[0102] Collect meteorological and operational signal data from offshore wind farms to construct a multimodal raw data stream.
[0103] Meteorological data includes wind speed ,wind direction High waves At the first sampling frequency Data acquisition; operational signal data, including shutdown signals. Inertia response signal Primary frequency modulation signal At the second sampling frequency Collection, and ;
[0104] Each data entry is accompanied by a high-precision timestamp. This forms the original data stream:
[0105] ,
[0106] in, The first The meteorological data vector in the data and the first The running signal data vector in the data.
[0107] Step 2 includes the following steps:
[0108] Step 2.1, Time Synchronization;
[0109] Using the timestamps of meteorological data as a benchmark, cubic spline interpolation is employed to resample the operational signal data and transform it onto the time axis of the meteorological data. Simultaneously, a mutation point detection algorithm is used to identify abrupt changes in the operational signal and precisely align them with corresponding events in the meteorological data, ensuring the temporal consistency of key dynamic processes.
[0110] Step 2.2, Feature space mapping;
[0111] A dual-branch feature embedding network is constructed to extract features from meteorological data and operational signal data separately, mapping them to a unified feature space to obtain a meteorological data encoder. With running data encoder Output feature vectors of meteorological data with uniform dimensions. eigenvectors of running signal data A contrastive learning mechanism is introduced to make cross-modal features at the same time move closer to each other in space, while features at different times move further apart, thereby enhancing the semantic consistency between modalities and eliminating differences in physical attributes and dimensions.
[0112] Step 2.3, Construction of multimodal feature sequences;
[0113] The aligned meteorological features and operational features are concatenated to form a comprehensive feature vector for each time step: ,
[0114] Obtain temporally aligned multimodal feature sequences This provides standardized input for the multimodal feature fusion in step 3.
[0115] Step 3 includes the following steps:
[0116] Step 3.1, meteorological-operational dual-branch coupled state-space modeling;
[0117] A bi-branch coupling structure is constructed to independently perform state evolution for meteorological and operational characteristics, and cross-modal interaction is achieved through coupling terms. Definition Momentary state vector With running state vector Its coupling evolution equation is:
[0118] ,
[0119] in, , These are the state transition matrices for the meteorological branch and the operational branch, respectively. , The cross-modal coupling matrix represents the influence of operational signals on meteorological dynamics and the driving effect of meteorological conditions on operational response, respectively. , These are the input mapping matrices for the meteorological branch and the operational branch, respectively. This dual-branch structure enables explicit coupled modeling of meteorological elements and operational signals.
[0120] Step 3.2, Multi-scale cooperative state decomposition and inter-scale coupling mechanism;
[0121] To investigate the coordinated changes in marine meteorological elements and wind power output across different time scales, the state vector is decomposed into multiple time-scale subspaces, and the nonlinear interactions between different scales are quantified using an inter-scale coupling matrix. (Meteorological state vector) Decomposed into Subvectors at each time scale: , where the subvector at the i-th time scale Corresponding time scale The coupling between scales is achieved through tensor products:
[0122] ,
[0123] in, For the first The state transition matrix at the scale, It is a third-order coupled tensor. For the Kronecker product operation, For the first The input mapping matrix at different scales. This equation captures the synergistic amplification effect between meteorological elements at different scales through second-order interaction terms, thus achieving explicit modeling of multi-scale nonlinear synergistic laws.
[0124] Step 3.3, Adaptive discretization of sensitivity to extreme weather events;
[0125] An extreme event sensitivity gating mechanism is introduced to adaptively adjust the discretization step size and state update intensity according to the degree of abrupt change in meteorological characteristics; an extreme meteorological event indicator function is defined:
[0126] ,
[0127] in, For a moment Intensity index of extreme weather events The L2 norm of the time gradient of meteorological characteristics. For indicator functions, To determine the threshold for extreme events, the adaptive discretization step size is expressed as:
[0128] ,
[0129] in, As the reference discretization step size, The attenuation coefficient is... The modulation intensity coefficient, The hyperbolic tangent activation function is used. A multilayer perceptron network designed for extreme events; the discretized state update equation is:
[0130] ,
[0131] This mechanism adaptively reduces the discretization step size during extreme weather events, enhancing the temporal resolution of abrupt changes while maintaining high computational efficiency during stable periods. Through this mechanism, a unified modeling of the meteorological-operational coupling relationship, multi-scale synergistic patterns, and sensitivity to extreme events is achieved, ultimately outputting a joint feature representation. .
[0132] Step 4 includes the following steps:
[0133] Step 4.1, Joint distribution change detection;
[0134] A sliding window mechanism is used to calculate the maximum mean difference (MMD) between the current window and the historical baseline window as an indicator of drift intensity. An adaptive threshold is set, and a concept drift event is determined to have occurred when the drift intensity exceeds the threshold.
[0135] Step 4.2, Drift type identification;
[0136] Based on the offset speed and offset pattern, concept drift is divided into three categories:
[0137] Sudden drift: Drift intensity increases dramatically in a short period of time, indicating extreme weather or sudden malfunction;
[0138] Gradual drift: The drift intensity increases slowly and continuously, indicating seasonal changes or equipment aging;
[0139] Repeated drift: The drift intensity exhibits periodic fluctuations, characterizing periodic changes such as tides and day / night cycles.
[0140] Through the above mechanism, we can accurately perceive the evolution of data and provide a triggering basis for the generation of subsequent extreme samples.
[0141] Step 5 includes the following steps:
[0142] Step 5.1, Tail dependency modeling based on Copula;
[0143] A multivariate Copula function is constructed to describe the tail dependency structure between meteorological variables and power output, and to quantify the joint probability distribution under extreme events. The Copula parameters are learned from historical extreme samples through maximum likelihood estimation.
[0144] Step 5.2, Time Block Consistency Reconstruction;
[0145] The original time series is divided into fixed-length time blocks. A new sample block with consistent temporal correlation with the real extreme samples is generated under Copula constraints using a conditional generative adversarial network (CGAN) or variational autoencoder (VAE). The consistency of the time blocks is constrained by the autocorrelation function and cross-correlation function within the block.
[0146] Step 5.3, Incremental Training and Distribution Adaptation;
[0147] The generated extreme samples are incrementally added to the training set in a streaming manner. Empirical replay or gradient constraint methods are used to avoid catastrophic forgetting. At the same time, the difference between the generated sample distribution and the true extreme distribution is minimized by the maximum mean difference (MMD) or Wasserstein distance to achieve distribution adaptation. The generated extreme samples are fed back to step 2 and input together with the original multimodal feature sequence into the fusion method for incremental training.
[0148] Step 6 includes the following steps:
[0149] Step 6.1, Riemannian manifold modeling of the meta-knowledge space;
[0150] The expert model parameters obtained from training each source domain task Consider it as a high-dimensional manifold Points on a manifold, the Riemannian metric is defined. To preserve the geometry of the parameter space, an exponential mapping is introduced. With logarithmic mapping Realizing manifolds and tangent spaces Bidirectional transformation between them.
[0151] Step 6.2, manifold generalization of the conditional diffusion model;
[0152] To generate new expert parameters on the manifold, a conditional backdiffusion process is defined:
[0153] ,
[0154] in, For the first Step diffusion state and condition vector Next, generate the state of the previous time step. The probability density function, For Riemann index mapping, It follows a Gaussian distribution (normal distribution). For the Riemannian logarithmic mapping, The mean value predicted by the neural network. To preset the noise level, It is the identity matrix;
[0155] This formula extends traditional Euclidean space diffusion to Riemannian manifolds, ensuring that the generation process proceeds along geodesics and maintains the geometric consistency of parameter distribution.
[0156] Step 6.3, noise prediction of physical information embedding;
[0157] To ensure that the generated expert model conforms to the physical laws of wind power generation, physical constraints are embedded in the noise prediction network, and the following noise prediction function is defined:
[0158] ,
[0159] Among them, physical operators Deviation between the quantization model output and the theoretical power curve and inertial response characteristics:
[0160] ,
[0161] In the formula, To predict power for the model, This is the theoretical power curve. The inertial response power predicted by the model. This is the theoretical inertial response value. These are the weighting coefficients. The noise prediction function backpropagates physical biases to the diffusion process through the gradient term, ensuring that the generated expert parameters automatically satisfy the physical constraints.
[0162] Step 6.4, Distributed consensus guided by manifold geometry;
[0163] generate After the first expert model, a distributed consensus mechanism is used for collaborative prediction. Define the first... One expert on the sample Fusion weights:
[0164] ,
[0165] in, The geodesic distance on the manifold. For the first The confidence level of each expert's prediction for the current sample. The fit between the sample and the expert model (which can be calculated from the negative log-likelihood predicted by the model). The temperature coefficient is used. This formula incorporates the geometric relationships between experts into the weighting calculation, ensuring that the fusion result reflects both individual confidence levels and avoids the excessive influence of isolated experts.
[0166] The final prediction is:
[0167] ;
[0168] in, For the sample The final power forecast value (including meteorological characteristics and operational signals), For the first An expert model for samples Individual power prediction values;
[0169] Step 6.5, closed-loop evolution mechanism;
[0170] Once new task data has accumulated, the generative expert is fine-tuned, and the updated parameters are applied. Incorporate a meta-knowledge space. To maintain the consistency of the manifold structure, gradient flow is used to update the Riemann metric:
[0171] ,
[0172] in, For the Riemannian metric defined on the Riemannian manifold M, manifold parameters gradient operator, For the meta-knowledge space Expert model parameters Seeking expectations, The transition probability density induced by the conditional diffusion model. The reference distribution (such as historical parameter distribution or prior distribution); the updated parameters The data is then fed back to step 3 to optimize the Mamba model parameters, ultimately outputting an adaptive, transferable, and high-precision multi-site offshore wind power prediction.
[0173] This embodiment is applied to an offshore wind farm with an installed capacity of 400MW, consisting of 50 wind turbine generators, each with a capacity of 8MW. The specific implementation is as follows.
[0174] I. Data Collection Stage
[0175] Multi-source data was collected simultaneously through weather stations, lidar wind measurement systems, and SCADA systems: meteorological data included wind speed. ,wind direction High waves The sampling frequency is 0.2Hz; the operating signal data includes the shutdown signal. Inertia response signal Primary frequency modulation signal The sampling frequency is 20Hz. After being aligned with millisecond-level timestamps, each data point is transmitted to the edge computing node via dual fiber optic links to form the original dataset (including wind speed, wind direction, wave height, shutdown, inertial response, primary frequency modulation, and actual power tags).
[0176] II. Data Preprocessing and Feature Alignment
[0177] The edge computing node automatically executes steps 2.1–2.3: Based on the meteorological data timestamps, cubic spline interpolation is used to resample the operational signal to the meteorological time axis, while the CUSUM mutation detection algorithm is used to align abrupt events. A dual-branch feature embedding network is constructed (the meteorological encoder and the operational encoder each contain 3 fully connected layers, with a 64-dimensional output). A contrastive learning loss is used to bring cross-modal features at the same time point closer together and those at different times further apart. The aligned meteorological features and operational features are concatenated to obtain the comprehensive feature vector for each time step, constructing the input feature matrix. ,in The goal is 17,280 steps per day (one step every 5 seconds).
[0178] III. Mamba Two-Branch Multi-Scale Fusion Process
[0179] First Channel (Meteorological Branch): Based on current meteorological characteristics Calculate state evolution, extract meteorological dynamics, and output meteorological state. .
[0180] Second channel (running branch): This channel will display the running signal characteristics. Input the running branch, and output the running status after interaction via the coupling matrix. .
[0181] Multi-scale decomposition: The meteorological state is decomposed into four time scales (10 seconds, 5 minutes, 1 hour, and 24 hours). Nonlinear interaction between scales is achieved through third-order coupling tensors and Kronecker product to capture the synergistic amplification effect of gusts and tides.
[0182] Adaptive Discretization: Defining Extreme Event Indicator Functions Adaptive step size It employs exact zero-order preserved discretization to update the state and outputs a joint feature representation. .
[0183] IV. Concept Drift Detection
[0184] Drift detection: A sliding window (window length 1 hour) is used to calculate the maximum mean difference (MMD) between the current window and the historical baseline window in real time as the drift intensity. Set adaptive threshold .like This is then classified as a drift event.
[0185] Pattern recognition: based on the rate of change of drift intensity :
[0186] like and If the sudden increase is greater than twice the historical average, switch to extreme event response mode and activate sample generation in step 5;
[0187] like And it continues to rise, switching to progressive adaptation mode, triggering a slow update of model parameters;
[0188] If the drift is periodic (the Fourier principal frequency is close to the tidal period), switch to periodic follow mode and only update the periodic term.
[0189] V. Extreme Sample Generation and Incremental Training
[0190] Tail dependency modeling: Clayton Copula is used to model the lower tail dependency (parameters) of extreme low-power events. GumbelCopula models the upper tail dependency (parameters) of extreme high-power events. ).
[0191] Time block reconstruction: The historical sequence is divided into blocks of 100 steps each. The Conditional Generative Adversarial Network (CGAN) is trained to generate new sample blocks with consistent autocorrelation and cross-correlation with real extreme samples under Copula constraints (autocorrelation coefficient matching degree > 0.85).
[0192] Incremental training: Generated samples are added to the training set in a streaming manner, using an empirical replay buffer (capacity...). To prevent forgetting, the difference between the generated distribution and the true distribution is minimized using the MMD distance. The generated extreme samples are fed back to step 2 to optimize the feature embedding network parameters, forming the first feedback loop.
[0193] VI. Diffusion of Multi-Expert Knowledge Reuse and Predictive Output
[0194] Riemannian manifold modeling: The parameters of the expert model obtained from training three wind farms in the source domain are modeled as Riemannian manifolds, and the metric and exponential / logarithmic mapping are defined.
[0195] Conditional diffusion generation: For a new task (test wind farm), its meteorological statistical characteristics, operational signal characteristics, and drift intensity are extracted as conditions, and five dedicated expert models are generated through a back-diffusion process. The noise prediction network embeds physical constraints (theoretical power curve gradient deviation, inertial response deviation), and weighting coefficients. .
[0196] Distributed consensus: Calculate the fusion weights of each expert on the sample (considering geodesic distance, prediction confidence, sample fit, and temperature coefficient). The final power prediction value is obtained by weighted integration. .
[0197] Closed-loop evolution: The predicted results are compared with the actual power weekly. The prediction bias is used to drive the fine-tuning of the diffusion model parameters. The updated knowledge is fed back to step 3 to optimize the Mamba model parameters, forming a second feedback loop. The dual closed-loop collaboration enables continuous learning.
[0198] VII. Verification and Iteration
[0199] During operation, the predicted results are compared with the actual power in real time, and the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination are calculated. When the prediction performance drops by more than 10% for three consecutive weeks, the online fine-tuning process is automatically triggered, retraining the model using data from the most recent three months (learning rate...). (Iterate 200 times) to ensure that the prediction performance meets the grid dispatch requirements.
[0200] To comprehensively evaluate the predictive performance of the method of this invention, on the same offshore wind farm dataset (including normal periods and two typhoon events), the method of this invention was compared with four existing methods (persistent prediction method, pure LSTM network, standard Transformer, and standard Mamba) in terms of RMSE, MAE, and R. 2 The comparison is made based on several core metrics.
[0201] The main performance comparisons are shown in Table 1 below.
[0202] Table 1
[0203]
[0204] This invention provides a wind power prediction method based on the fusion of meteorological and electrical multimodal data. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.
Claims
1. A wind power prediction method based on the fusion of meteorological and electrical multimodal data, characterized in that, Includes the following steps: Step 1: Collect meteorological data and operational signal data from offshore wind farms to construct a multimodal raw data stream; Step 2: Perform heterogeneous feature alignment on meteorological and operational signal data, eliminate differences through time synchronization and feature space mapping, and form a time-domain aligned multimodal feature sequence; Step 3: Construct a Mamba dual-branch coupled multi-scale Kronecker tensor interaction fusion method; use the Mamba model to capture long-term temporal dependencies, realize cross-modal interaction through dual-branch coupling, use multi-scale tensor interaction to characterize nonlinear synergistic effects, and realize adaptive step-size discretization through extreme event sensitivity gating to output joint feature representation. Step 4: Based on joint feature representation, a feature space offset sensing mechanism is used to detect joint distribution changes in real time, and high-dimensional statistical indicators are used to quantify drift intensity. The concept of drift type is identified based on the offset speed and pattern recognition. Step 5: In high-risk scenarios, extreme samples are generated using Copula tail dependency modeling and time block consistency reconstruction to achieve incremental training and distribution adaptation. The generated extreme samples are then fed back to Step 2 for dynamic optimization of feature alignment parameters. Step 6 introduces a multi-expert knowledge reuse architecture based on a diffusion model. In the meta-knowledge space, a dedicated expert model adapted to the new task is generated on demand. Expert predictions are integrated through a distributed consensus mechanism to output the final prediction results. The prediction results are compared with the real data, and the prediction bias is used to drive the closed-loop update of the model parameters. The updated knowledge is fed back to Step 3 to optimize the Mamba model parameters.
2. The method according to claim 1, characterized in that, Step 1 includes: meteorological data including wind speed at time t. ,wind direction High waves At the first sampling frequency Data acquisition; operational signal data includes the shutdown signal at time t. Inertia response signal Primary frequency modulation signal At the second sampling frequency Collection, and ;No. The timestamp of each data item is This forms the raw data stream. : , in, The first The meteorological data vector in the data and the first The running signal data vector in the data.
3. The method according to claim 2, characterized in that, Step 2 includes: Step 2.1, Time Synchronization; Based on the timestamps of meteorological data, cubic spline interpolation is used to resample the operational signal data and transform the operational signal data onto the time axis of meteorological data. At the same time, abrupt change detection algorithm is used to identify the jump moments in the operational signal and align them with the corresponding events in the meteorological data. Step 2.2, Feature space mapping; A dual-branch feature embedding network is constructed to extract features from meteorological data and operational signal data separately, mapping the meteorological data and operational signal data to a unified feature space, thus obtaining a meteorological data encoder. With running data encoder Output feature vectors of meteorological data with uniform dimensions. eigenvectors of running signal data A contrastive learning mechanism is introduced to make cross-modal features at the same time point closer to each other in space, while features at different times point further apart. Step 2.3: Construct multimodal feature sequences; The aligned meteorological features and operational features are then stitched together to form the data for each time step. The comprehensive feature vector: , in, , The first The meteorological feature vector and the operational signal feature vector aligned at each time step This is a vector concatenation operation; Obtain the temporally aligned multimodal feature sequence X: , in, To indicate the first The comprehensive feature vector at each time step, This represents the total length of the time series.
4. The method according to claim 3, characterized in that, Step 3 includes: Step 3.1: Meteorological-Operational Two-Branch Coupled State-Space Modeling; A dual-branch coupling structure is constructed, which performs independent state evolution for meteorological and operational characteristics respectively, and achieves cross-modal interaction through coupling terms. Momentary state vector With running state vector The coupled evolution equation is: , in, To represent the differential, , These are the state transition matrices for the meteorological branch and the operational branch, respectively. , The cross-modal coupling matrix represents the influence of operational signals on meteorological dynamics and the driving effect of meteorological conditions on operational response, respectively. , These are the input mapping matrices for the meteorological branch and the operational branch, respectively. Step 3.2, Multi-scale cooperative state decomposition and inter-scale coupling mechanism; To investigate the coordinated changes in marine meteorological elements and wind power output at different time scales, meteorological state vectors are used. Decomposed into Subvectors at each time scale: , where the subvector at the i-th time scale Corresponding time scale The coupling between scales is achieved through tensor products: , in, For the first The state transition matrix at the scale, It is a third-order coupled tensor. For the first Sub-state vectors at different scales For the Kronecker product operation, For the first The input mapping matrix at the scale, For the meteorological feature vector at the th Scale components; Step 3.3, Adaptive discretization of sensitivity to extreme weather events; Define an indicator function for extreme weather events: , in, For a moment Intensity index of extreme weather events The L2 norm of the time gradient of meteorological characteristics. For indicator functions, Thresholds are used to determine extreme events; Adaptive Discretization Step Size Represented as: , in, As the reference discretization step size, The attenuation coefficient is... The modulation intensity coefficient, The hyperbolic tangent activation function is used. Multilayer perceptron networks designed for extreme events; The discretized state update equation is: , in, For the first The system state vector at each time step This is the state vector from the previous time step. For the first The adaptive discretization step size of the step. This is the state transition matrix in a continuous-time state-space model. For continuous-time input mapping matrix, For the first The input feature vector at each time step; Final output joint feature representation ,in, For the first A joint feature representation vector at each time step.
5. The method according to claim 4, characterized in that, Step 4 includes: Step 4.1, Joint distribution change detection; A sliding window mechanism is used to calculate the maximum mean difference between the current window and the historical baseline window as the drift intensity index. An adaptive threshold is set, and when the drift intensity exceeds the threshold, a concept drift event is determined to have occurred. Step 4.2, Drift type identification; Based on the offset speed and offset pattern, concept drift is divided into three categories: Sudden drift: Drift intensity increases dramatically in a short period of time, indicating extreme weather or sudden malfunction; Gradual drift: The drift intensity increases slowly and continuously, indicating seasonal changes or equipment aging; Repeated drift: The drift intensity exhibits periodic fluctuations, characterizing periodic changes such as tides and day / night cycles.
6. The method according to claim 5, characterized in that, Step 5 includes: Step 5.1, Tail dependency modeling based on Copula; A multivariate Copula function is constructed to describe the tail dependency structure between meteorological variables and power output, and the joint probability distribution under extreme events is quantified. The Copula parameters are learned from historical extreme samples through maximum likelihood estimation. Step 5.2, Time Block Consistency Reconstruction; The original time series is divided into fixed-length time blocks. Conditional generative adversarial network (CGAN) or variational autoencoder (VAE) is used to generate new sample blocks with consistent temporal correlation with real extreme samples under Copula constraints. The consistency of time blocks is constrained by the autocorrelation function and cross-correlation function within the block. Step 5.3, Incremental Training and Distribution Adaptation; The generated extreme samples are incrementally added to the training set in a streaming manner. Empirical replay or gradient constraint methods are used, and the difference between the generated sample distribution and the true extreme distribution is minimized by the maximum mean difference (MMD) or Wasserstein distance to achieve distribution adaptation. The generated extreme samples are fed back to step 2 and input together with the original multimodal feature sequence into the Mamba dual-branch coupled multi-scale Kronecker tensor interaction fusion method for incremental training.
7. The method according to claim 6, characterized in that, Step 6 includes: Step 6.1, Riemannian manifold modeling of the meta-knowledge space; The parameters of the expert models trained from each source domain task are considered as points on a high-dimensional manifold, and a Riemannian metric is defined on the manifold. Introducing exponential mapping With logarithmic mapping This enables bidirectional transformations between manifolds and tangent spaces. Step 6.2, manifold generalization of the conditional diffusion model; To generate new expert parameters on the manifold, a conditional backdiffusion process is defined: , in, For the first Step diffusion state and condition vector Next, generate the state of the previous time step. The probability density function, For Riemann index mapping, It follows a Gaussian distribution. For the Riemannian logarithmic mapping, The mean value predicted by the neural network. To preset the noise level, It is the identity matrix; Step 6.3, noise prediction of physical information embedding; Embed physical constraints in the noise prediction network and define the following noise prediction function: , in, The output of the noise prediction network, For noise prediction terms that are purely data-driven, These are the weighting coefficients for physical constraints. For model parameters gradient operator, physical operator Deviation between the quantization model output and the theoretical power curve and inertial response characteristics: , in, For wind speed, As an indicator function, when wind speed Cutting in wind speed With rated wind speed The value is 1 when it is between 1 and 0; otherwise, it is 0. To represent partial derivatives, To predict power for the model, This is the theoretical power curve. The inertial response power predicted by the model. This is the theoretical inertial response value. These are the weighting coefficients; Step 6.4, Distributed consensus guided by manifold geometry; generate After the first expert model, a distributed consensus mechanism is used for collaborative prediction, defining the first... One expert on the sample Fusion weights : , in, The first The parameter vector of the first expert model and the first... The parameter vector of an expert model The geodesic distance on the manifold. For the first The confidence level of each expert's prediction for the current sample. To assess the fit between the sample and the expert model, Temperature coefficient; The final prediction is: ; in, For the sample The final power prediction value, For the first An expert model for samples Individual power prediction values; Step 6.5, closed-loop evolution mechanism.
8. The method according to claim 7, characterized in that, Step 6.5 includes: after new task data has accumulated, fine-tuning the generated expert and updating the parameters. Add to Meta-Knowledge Space The Riemann metric is updated using gradient flow: , in, For the Riemannian metric defined on the Riemannian manifold M, manifold parameters gradient operator, For the meta-knowledge space Expert model parameters Seeking expectations, The transition probability density induced by the conditional diffusion model. For reference distribution; updated parameters The data is then fed back to step 3 to optimize the Mamba model parameters, ultimately outputting an adaptive, transferable, and highly accurate multi-site offshore wind power prediction.
9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.
10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.