A method, system and equipment for multi-parameter collaborative prediction of wind turbine clusters

Abstract

This invention discloses a method, system, and device for multi-parameter collaborative prediction of wind turbine clusters. At the data-driven level, physical quantities are jointly modeled through multi-channel temporal coding and temporal and cross-channel attention mechanisms to extract spatiotemporal dependency features among multiple parameters. At the digital twin layer, a dynamic mapping of multiple parameters between physical and virtual wind farms is constructed, and physical assimilation constraints are introduced to improve the physical reliability of the prediction results. At the fusion representation layer, transfer learning is used to align the feature distributions of different wind farms, enhancing the model's cross-site generalization ability. The method can achieve high-precision, physically consistent, and interpretable prediction of multiple parameters of wind turbine clusters under complex operating conditions and cross-site scenarios, providing reliable data-driven support for intelligent operation and maintenance and cluster scheduling of wind turbine clusters under high-proportion wind power access conditions, and realizing unified modeling and high-precision collaborative prediction of the multi-parameter operating status of wind turbine clusters.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance of wind power, and relates to a method, system and equipment for multi-parameter collaborative prediction of wind turbine groups. Background Technology

[0002] As the global energy structure accelerates its transformation towards low-carbon, clean, and high-proportion renewable energy, the installed capacity and power generation share of wind power continue to climb. However, with a high proportion of wind power connected to the grid, the output power and key operating parameters of wind turbines are significantly affected by atmospheric boundary layer turbulence, complex terrain, and nonlinear dynamics of the units, exhibiting significant randomness, intermittency, and strong coupling characteristics. This not only increases the pressure on grid-side peak shaving and frequency stability but also makes the assessment of unit operating status and fault early warning more complex. Therefore, power system operation and dispatch no longer focus solely on short-term predictions of the active power of a single wind turbine. Instead, there is an increasing need for joint modeling and high-precision collaborative prediction of key physical quantities such as wind speed, rotor speed, pitch angle, active power, main bearing temperature, generator temperature, and vibration intensity at the wind turbine cluster scale. This is to support the safe and economical operation of wind turbine clusters, optimized cluster dispatch, and intelligent operation and maintenance.

[0003] Existing wind power forecasting and operational status modeling methods can be categorized into three types: mechanism-based physical models, statistical time series models, and data-driven machine learning and deep learning models. Among these, deep learning models, with their automatic feature extraction and strong nonlinear representation capabilities, are widely used in short-term and ultra-short-term wind power forecasting. On the other hand, with the widespread deployment of Supervisory Control and Data Acquisition (SCADA) systems in wind farms, a large amount of multi-source operational data, including information on wind speed, power, rotational speed, temperature, and vibration, has been recorded over a long period, providing an important foundation for data-driven condition assessment and fault diagnosis. SCADA-based multivariate spatiotemporal feature modeling is gradually becoming an important direction for wind turbine condition monitoring. This paper systematically summarizes the research progress of SCADA-driven wind turbine condition monitoring and fault diagnosis from the perspectives of data quality control, adaptability to complex operating conditions, multi-source data fusion, and interpretability. It also points out that there are still shortcomings in deep fusion of multi-source information, physical consistency constraints, and engineering feasibility. These studies indicate that balancing the physical rationality of models with the value of operation and maintenance decisions while fully leveraging the advantages of SCADA data remains a key challenge in the field of intelligent operation and maintenance of wind power. Complementing data-driven methods, digital twin (DT) technology, by constructing virtual entities highly consistent with physical entities at the structural, behavioral, and rule levels, and achieving real-time bidirectional mapping between physical and digital spaces, is one of the important supporting technologies for intelligent operation and maintenance of complex energy equipment. In the wind power field, digital twins have been used in applications such as floating wind turbine dynamics simulation and state estimation, gearbox multi-objective optimization design, online monitoring of bearing fatigue damage, and online intelligent evaluation of gearboxes based on Vibration-Based Damage Monitoring Digital Twin (VBDM-DT). However, existing research on wind power digital twins largely focuses on the single-machine or single-component level.

[0004] In summary, the following issues still exist in the multi-parameter collaborative prediction and digital twin construction of wind turbine clusters:

[0005] 1. Insufficient unified characterization of multiple parameters at the cluster scale. Existing work is mostly focused on single wind turbines or single components. Although it has achieved good results in power prediction or local component health assessment, it lacks a unified state characterization framework that can simultaneously depict the coupling relationship of "multiple wind turbines-multiple components-multiple physical quantities" at the cluster scale. This makes it difficult to support the high-dimensional prediction and assessment needs for cluster-based optimized scheduling and collaborative operation and maintenance.

[0006] 2. Limited generalization capability of digital twins across wind farms and turbine models. Most wind power digital twin models are built for specific wind farms or turbine models, without considering the data distribution offset caused by differences in wind resources, terrain, and operation strategies between different wind farms.

[0007] 3. Insufficient physical constraints and engineering interpretability of data-driven models. Many deep learning models, while pursuing prediction accuracy, lack explicit modeling and constraints on the inherent physical relationships of wind speed, power, rotational speed, pitch angle, temperature, load, vibration, torque, etc. The prediction results often remain at the level of numerical approximation, which makes it difficult to fully reflect the constraint role of wind turbine operating mechanism and safety boundary in the multi-parameter prediction process. Under extreme operating conditions, this can easily lead to the risk of misjudgment or omission.

[0008] Therefore, a new technical solution is urgently needed to effectively solve the above problems, achieve high precision, physical consistency and interpretable prediction of multiple parameters of wind turbine clusters under complex operating conditions and cross-site scenarios, and provide reliable data-driven support for intelligent operation and maintenance and cluster scheduling of wind turbine clusters under high-proportion wind power access conditions. Summary of the Invention

[0009] This invention aims to provide a method, system, and device for collaborative prediction of multiple parameters of wind turbine clusters, enabling high-precision, physically consistent, and interpretable prediction of multiple parameters of wind turbine clusters under complex operating conditions and cross-site scenarios, and providing reliable data-driven support for intelligent operation and maintenance and cluster scheduling of wind turbine clusters under high-proportion wind power access conditions.

[0010] To achieve the above objectives, the technical solution adopted by this invention is: a multi-parameter collaborative prediction method for wind turbine groups, comprising:

[0011] S1. Collect historical operating sequence data of wind turbines in a reference wind farm, including physical quantities and corresponding feature vectors within the total number of time steps; use a sliding window mechanism to extract continuous time segments from the historical operating sequence data; each time segment includes the previous... The input time window and the subsequent time step The predicted horizon is calculated at each time step; the feature vectors within the input time window are combined to form the input sequence, and the physical quantities within the predicted horizon are combined to form the output label; the input sequence and the output label constitute a time segment sample; all time segment samples within the total number of time steps are combined to form a reference dataset.

[0012] S2. Based on the target wind farm, repeat S1 to obtain the target dataset;

[0013] S3. Construct a deep learning model, including a data-driven module and a physical assimilation constraint module; the data-driven module receives the input sequence of the reference dataset, extracts the reference fusion representation, and outputs the conditional probability distribution and predicted physical quantity values ​​within the predicted horizon; based on the output labels of the reference dataset and the conditional probability distribution within the predicted horizon, the deep learning model is pre-trained using a preset objective function; the physical assimilation constraint module applies physical assimilation constraints to the predicted physical quantity values ​​to obtain the pre-trained model; the network parameters of the pre-trained model include data-side parameters and physical-side parameters;

[0014] S4. Input the target dataset into the pre-trained model to obtain the target fusion representation; perform domain alignment between the reference fusion representation and the target fusion representation; based on the target dataset, fine-tune the pre-trained model through a layer-by-layer unfreezing strategy to minimize the constructed joint loss function and obtain a collaborative prediction model; wherein, the joint loss function is the sum of the preset objective function loss, the physical assimilation constraint loss and the domain alignment loss;

[0015] Preferably, it includes: S5, collecting real-time operating sequence data of wind turbines in the target wind farm, inputting the real-time operating sequence data into the collaborative prediction model, and obtaining the predicted physical quantity.

[0016] The solution provided by this invention performs joint modeling of key physical quantities at the data-driven level, extracting spatiotemporal dependency features among multiple parameters; at the digital twin layer, it constructs a dynamic mapping of multiple parameters between physical and virtual wind farms, introducing physical assimilation constraints for observation consistency, dynamic consistency, and prediction consistency to improve the physical credibility of prediction results; at the fusion representation layer, it utilizes transfer learning to align the feature distributions of different wind farms, enhancing the model's cross-site generalization capability; the method can achieve high-precision, physically consistent, and interpretable prediction of multiple parameters of wind turbine clusters under complex operating conditions and cross-site scenarios, providing reliable data-driven support for intelligent operation and maintenance and cluster scheduling of wind turbine clusters under high-proportion wind power access conditions, and realizing unified modeling and high-precision collaborative prediction of the multi-parameter operating status of wind turbine clusters.

[0017] According to embodiments of the present invention, the present invention can be further optimized, and the optimized technical solution is as follows:

[0018] In one preferred embodiment, in step S1, the physical quantities include wind turbine operating parameters and / or meteorological information; the wind turbine operating parameters include rotor speed, pitch angle, active power, main bearing temperature, generator temperature, and vibration intensity; the meteorological information includes wind speed, air temperature, humidity, and wind direction. The seven physical quantities selected in the above scheme are highly correlated with intelligent operation and maintenance and cluster scheduling of wind turbine groups.

[0019] In one preferred embodiment, in step S3, the data-driven module is a multi-dimensional adaptive long short-term memory network; the multi-dimensional adaptive long short-term memory network extracts fusion representations through multi-channel long short-term memory network encoding, temporal attention, and cross-channel attention mechanisms; and outputs the conditional probability distribution and predicted physical quantity values ​​within the predicted horizon through heteroscedastic probability decoding; wherein, the multi-channel long short-term memory network includes multiple long short-term memory networks configured in parallel for processing the feature vectors; the fusion representation is the dependency relationship of the physical quantity at different time scales and the nonlinear coupling characteristics between the physical quantities; the preset objective function includes a probabilistic objective function and a point estimation objective function.

[0020] The above scheme constructs a parallel coding structure oriented towards physical quantities, introduces time attention and cross-channel attention mechanisms, and adaptively identifies when and which channel is more important under a unified framework. It achieves multi-step prediction and uncertainty characterization through heteroscedastic probability decoding, which significantly improves the accuracy and robustness of multi-parameter collaborative prediction while enhancing the interpretability of the model for key signals and key moments.

[0021] In one preferred embodiment, the expression for the multi-channel long short-term memory network encoding is:

[0022] ;

[0023] in, The number of input channels for the multi-channel long short-term memory network. The input channel is numbered. For a historic moment The The hidden states of channel long short-term memory networks. , For the current moment, The length of the input time window. For a historic moment The The memory state of the channel long short-term memory network. For a historic moment The Hidden states of channel long short-term memory networks, For a historic moment The The memory state of the channel long short-term memory network. For the first Channel Long Short-Term Memory Network, For the first Channel Long Short-Term Memory Networks in Historical Moments The input feature vector, For the first Feature dimensions of channel long short-term memory networks;

[0024] The expression for the time attention is:

[0025] ;

[0026] ;

[0027] ;

[0028] in, For the first The intermediate vector of the channel long short-term memory network after time-attention aggregation. For the first Channel Long Short-Term Memory Network at time Historical moments Time attention weights The scoring function for time attention weights. , , These are all parameters for a time-attention scoring network. For time step index, For attention query vectors;

[0029] The expression for cross-channel attention is:

[0030] ;

[0031] ;

[0032] ;

[0033] in, This is the latent space representation vector after cross-channel attention fusion, i.e., the fused representation. For the first Channel Long Short-Term Memory Network at the current moment Cross-channel attention weights The scoring function for cross-channel attention weights. , , All are parameters for cross-channel attention scoring networks;

[0034] The expression for the heteroscedasticity probability decoding is:

[0035] ;

[0036] ;

[0037] ;

[0038] in, Let be the conditional probability density function of the multidimensional adaptive long short-term memory network under the Gaussian assumption. These are the trainable parameters of the multi-dimensional adaptive long short-term memory network. For the first A channel long short-term memory network in The input sequence within the window, For the current moment Starting from the first The output label of the step, For a moment No. The predicted mean of each output dimension For a moment No. The prediction variance of each output dimension The vector dimension of the predicted physical quantity value. , Output dimension index, , All are the first Weights for linear decoding , All are the first The bias of the linear decoding step. For the multi-dimensional adaptive long short-term memory network in the first... The predicted values ​​of physical quantities output by the step;

[0039] The point estimation objective function is a weighted mean square error function; the probabilistic objective function is a negative log-likelihood function, and the corresponding expression for the negative log-likelihood loss is:

[0040] ;

[0041] in, Let be the normalization constant term of the Gaussian negative log-likelihood. For a moment The output label of the first One output dimension;

[0042] The multidimensional adaptive long short-term memory network also incorporates a simplex constraint, expressed as follows:

[0043] .

[0044] In the above scheme, negative log-likelihood is used for theoretical derivation to maintain statistical interpretability, and weighted mean square error function is used for parameter optimization. The two are connected through a unified weight coefficient or learning rate, so that the deep learning model can fully learn the temporal dependence and coupling mode between various physical quantities.

[0045] In one preferred embodiment, step S3, the physical assimilation constraint includes observational consistency constraint, dynamic consistency constraint, and prediction consistency constraint, the specific expressions of which are as follows:

[0046] ;

[0047] in, This refers to the physical assimilation constraint. The total number of timing steps. For the current moment, For the current moment physical quantity, For the wind turbine at the current moment The hidden state, To map hidden states to observation functions of physical quantities, To observe the inverse of the noise covariance matrix, For the state update function that includes physical priors and learnable residuals, For the wind turbine unit at the previous moment The hidden state, For the wind turbine unit at the previous moment Exogenous input, The inverse of the process noise covariance matrix. For observation function and state update function The parameter is η, which is the prediction consistency weight coefficient. For the predicted horizon length, For the prediction step index, For the multi-dimensional adaptive long short-term memory network in the first... The predicted values ​​of physical quantities output by the step. The predicted hidden state is obtained by the physical assimilation constraint module through attention rolling extrapolation on the prediction horizon.

[0048] The above scheme integrates physical priors and learnable residuals into the state equation at the digital twin level. Through constraints such as observation consistency, dynamic consistency, and prediction consistency, it tightly couples data-driven prediction with physical model attention-based rolling extrapolation, enabling the deep learning model to maintain physical feasibility and numerical stability under complex operating conditions and extreme events. This provides a reliable digital twin carrier for fleet operation status assessment, early fault identification, and operation and maintenance strategy optimization.

[0049] In one preferred embodiment, the specific formula for the joint loss function in step S4 is as follows:

[0050] ;

[0051] in, Let the joint loss function be... These are the trainable parameters of the multi-dimensional adaptive long short-term memory network. The number of input channels for the multi-channel long short-term memory network. The input channel is numbered. The length of the input time window. For the first Each channel is in The input sequence within the window, For the current moment Starting from the first The output label of the step, Let be the conditional probability density function of the multidimensional adaptive long short-term memory network under the Gaussian assumption. For domain alignment distance, For the reference fusion representation set, For the target fusion representation set, These are the weighting coefficients for the physical assimilation loss. is the weighting coefficient for the domain alignment loss.

[0052] In one preferred embodiment, the domain alignment distance is obtained by a weighted combination of the higher-order distribution difference and the second-order statistical difference between the reference fusion representation and the target fusion representation, as calculated by the following formula:

[0053] ;

[0054] ;

[0055] ;

[0056] in, The distance between the reference fusion representation and the target fusion representation is the maximum mean difference. For kernel function, The number of samples in the reference dataset. The number of samples in the target dataset. For the reference dataset of the first The sample number, The first reference dataset The sample number, For the target dataset of the first The sample number, The first of the target dataset The sample number, The second-order statistical difference distance is defined as the distance between the covariance matrix of the reference fusion representation and the covariance matrix of the target fusion representation. Let be the covariance matrix of the reference fusion representation. Let be the covariance matrix representing the fusion of the target.

[0057] The aforementioned scheme simultaneously measures and compresses the higher-order distribution differences and second-order statistical differences between the reference and target wind farms at the fusion representation layer, incorporating them into the domain alignment distance through a weighted combination. This scheme achieves cross-site generalization, significantly reducing domain offset and stabilizing transfer performance even with limited target wind farm samples; it achieves complementary alignment, adapting higher-order distribution differences to nonlinear differences, and using second-order statistical differences for low-cost robustness, with weighting both to improve alignment quality; it achieves controllable transfer, with the alignment layer and weights explicitly set, facilitating a trade-off between bias and variance under different domain differences and data volume scenarios. The deep learning model, through a two-stage learning optimization process of pre-training in the reference wind farm and fine-tuning in the target wind farm, effectively alleviates the distribution offset and sample imbalance problems between different wind farms, enabling the multi-parameter co-evolution laws learned in data-rich wind farms to be transferred to new wind farms with scarce data, thereby improving the generalization ability and sample efficiency of the deep learning model in practical engineering deployments.

[0058] In one preferred embodiment, the optimization of the collaborative prediction model in step S4 is specifically implemented using the following formula:

[0059] ;

[0060] ;

[0061] ;

[0062] in, For the data-side parameters, For the physical side parameters, Updated during optimization process The learning rate Updated during optimization process The learning rate The joint loss function is relative to The estimated gradient, For the physical assimilation constraint relative to The estimated gradient, The input sequence is standardized. For the standardized output labels, For the first Channel Long Short-Term Memory Networks in Historical Moments The input feature vector, The input sequence of the reference dataset is in the th... The mean of the channel, The input sequence of the reference dataset is in the th... Standard deviation of the channel The output labels for the reference dataset are in the 1st... The mean of the channel, The output labels for the reference dataset are in the 1st... Standard deviation of the channel These are the weighting coefficients for different prediction step sizes. .

[0063] The above scheme alleviates gradient conflicts between the data side and the physical side by updating in blocks, thereby improving convergence. By calculating the mean and standard deviation required for standardization on the training samples, the real-time data received during the evaluation phase and future deployment are uniformly standardized, ensuring consistent data preprocessing standards in the model training, evaluation, and deployment phases. This avoids data leakage problems caused by using global statistical information and guarantees generalization performance and stability in practical applications. Step-size weighting allows the collaborative prediction model to be optimized for short-term scheduling or medium-term rolling plans. Unified optimization and regularization strategies ensure comparability and reproducibility between different data batches and sites.

[0064] Based on the same concept, the present invention also provides a multi-parameter collaborative prediction system for wind turbine clusters, comprising:

[0065] The reference dataset module is used to collect historical operating sequence data of wind turbines in a reference wind farm, including physical quantities and corresponding feature vectors within the total number of time steps. A sliding window mechanism is used to extract continuous time segments from the historical operating sequence data; each time segment includes the preceding... The input time window and the subsequent time step The predicted horizon is calculated at each time step; the feature vectors within the input time window are combined to form the input sequence, and the physical quantities within the predicted horizon are combined to form the output label; the input sequence and the output label constitute a time segment sample; all time segment samples within the total number of time steps are combined to form a reference dataset.

[0066] The target dataset module is used to obtain the target dataset by repeating S1 based on the target wind farm.

[0067] A pre-trained model module is used to construct a deep learning model, including a data-driven module and a physical assimilation constraint module. The data-driven module receives the input sequence of the reference dataset, extracts the reference fusion representation, and outputs the conditional probability distribution and predicted physical quantity values ​​within the predicted horizon. Based on the output labels of the reference dataset and the conditional probability distribution within the predicted horizon, the deep learning model is pre-trained using a preset objective function. The physical assimilation constraint module applies physical assimilation constraints to the predicted physical quantity values ​​to obtain the pre-trained model. The network parameters of the pre-trained model include data-side parameters and physical-side parameters.

[0068] A collaborative prediction model module is used to input the target dataset into the pre-trained model to obtain a target fusion representation; the reference fusion representation and the target fusion representation are aligned in domains; based on the target dataset, the pre-trained model is fine-tuned through a layer-by-layer unfreezing strategy to minimize the constructed joint loss function, thereby obtaining a collaborative prediction model; wherein, the joint loss function includes a preset objective function loss, a physical assimilation constraint loss, and a domain alignment loss;

[0069] The real-time prediction module is used to collect real-time operating sequence data of wind turbines in the target wind farm, and input the real-time operating sequence data into the collaborative prediction model to obtain the predicted physical quantities.

[0070] Based on the same concept, the present invention also provides an electronic device, including a memory, a processor, and a computer program / instructions stored in the memory, wherein the processor executes the computer program / instructions to implement the multi-parameter collaborative prediction method for wind turbine groups as described above.

[0071] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a method, system, and device for multi-parameter collaborative prediction of wind turbine clusters; at the data-driven level, key physical quantities are jointly modeled through multi-channel time-series coding and time attention and cross-channel attention mechanisms to extract spatiotemporal dependency features between multiple parameters; at the digital twin layer, a multi-parameter dynamic mapping between physical and virtual wind fields is constructed, and physical assimilation constraints of observation consistency, dynamic consistency, and prediction consistency are introduced to improve the physical credibility of prediction results; at the fusion representation layer, transfer learning is used to align the feature distributions of different wind farms to enhance the model's cross-site generalization ability; the method can achieve high-precision, physically consistent, and interpretable prediction of multiple parameters of wind turbine clusters under complex operating conditions and cross-site scenarios, providing reliable data-driven support for intelligent operation and maintenance and cluster scheduling of wind turbine clusters under high-proportion wind power access conditions, and realizing unified modeling and high-precision collaborative prediction of the multi-parameter operating status of wind turbine clusters. Attached Figure Description

[0072] Figure 1This is a flowchart of a multi-parameter collaborative prediction method for wind turbine groups according to an embodiment of the present invention;

[0073] Figure 2 This is a diagram of a multi-dimensional adaptive long short-term memory network framework according to an embodiment of the present invention;

[0074] Figure 3 This is a multi-parameter digital twin architecture diagram according to an embodiment of the present invention;

[0075] Figure 4 This is a diagram illustrating the transfer learning mechanism of one embodiment of the present invention;

[0076] Figure 5 This is a loss convergence curve during the pre-training and fine-tuning stages of another embodiment of the present invention. Figure 5 (a) represents the MSE loss during the pre-training phase. Figure 5 (b) represents the MSE loss during the fine-tuning phase;

[0077] Figure 6 This is a comparison chart of measured data and MDA-LSTM multi-physical quantity time series predictions according to another embodiment of the present invention. Figure 6 (a) demonstrates the dynamic response characteristics of the rotor speed under wind speed drive. Figure 6 (b) demonstrates the tracking performance for pitch angle. Figure 6 (c) illustrates the shape of the active power nonlinear segment. Figure 6 (d) illustrates the gradual dynamic process of bearing temperature change. Figure 6 (e) illustrates the correlation between generator temperature and power level. Figure 6 (f) demonstrates that the wind speed accurately reproduces the slow undulations and gust outlines of the background wind. Figure 6 (g) demonstrates the reproduction of vibration intensity on multi-peak fluctuations and transient enhancement;

[0078] Figure 7 This is a comparison chart of the performance indicators of a collaborative prediction method based on MDA-LSTM and MP-DT combined with transfer learning, according to another embodiment of the present invention, and the baseline method. Figure 7 (a) is a comparison chart of parameter-by-parameter indicators. Figure 7 (b) is a comparison chart of average indicators;

[0079] Figure 8 This is a comparison chart of the performance indicators of a collaborative prediction method based on MDA-LSTM and MP-DT combined with transfer learning, according to another embodiment of the present invention, and the baseline method. Figure 8 (a) is a comparison chart of overall indicators. Figure 8 (b) is a comparison chart of parameter-by-parameter indicators;

[0080] Figure 9 This is a cross-condition time-domain fitting and generalization analysis diagram of another embodiment of the present invention. Figure 9 (a) is a comparison chart of predictions and actual measurements for the entire sample without different working conditions. Figure 9 (b) is a comparison chart of predicted and measured values ​​under low wind speed conditions. Figure 9 (c) is a comparison chart of predicted and measured values ​​under near-rated operating conditions. Figure 9 (d) is a comparison chart of predicted and measured high wind speed conditions;

[0081] Figure 10 This is a consistency diagram of the predicted-measured probability distribution according to another embodiment of the present invention. Figure 10 (a) is a comparison diagram of the probability distribution of rotor speed. Figure 10 (b) is a comparison diagram of the probability distribution of propeller pitch angle. Figure 10 (c) is a comparison diagram of the probability distribution of active power. Figure 10 (d) is a comparison diagram of the probability distribution of bearing temperature. Figure 10 (e) is a comparison diagram of the probability distribution of generator temperature. Figure 10 (f) is a comparison diagram of the probability distribution of wind speed. Figure 10 (g) is a comparison diagram of the probability distribution of vibration intensity;

[0082] Figure 11 This is a violin plot of the residual distribution and a CDF plot of the cumulative distribution function of the absolute error of MDA-LSTM, according to another embodiment of the present invention. Figure 11 (a) is a violin plot of the residual distribution of the physical quantities. Figure 11 (b) is a graph of the cumulative distribution function of the model's absolute error;

[0083] Figure 12 This is a comparison chart of the importance of various physical quantity characteristics according to another embodiment of the present invention;

[0084] Figure 13 This is a correlation thermogram of various physical quantities according to another embodiment of the present invention;

[0085] Figure 14 This is a partial contribution diagram of each physical quantity according to another embodiment of the present invention;

[0086] Figure 15 This is a parameter-by-parameter performance comparison chart of the ablation sequence according to another embodiment of the present invention;

[0087] Figure 16 This is a comparison chart of the overall performance of the ablation sequence according to another embodiment of the present invention;

[0088] Figure 17 This is a comparison chart of the overall performance of another embodiment of the present invention with the baseline method;

[0089] Figure 18 This is a parameter-by-parameter performance thermogram of the ablation sequence according to another embodiment of the present invention. Detailed Implementation

[0090] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0091] Example 1

[0092] This embodiment 1 addresses the collaborative prediction needs of wind turbine clusters in multi-parameter, cross-site scenarios. It proposes a collaborative prediction method based on a multi-parameter digital twin (MP-DT) of the wind turbine cluster constructed using a multi-dimensional adaptive long short-term memory network (MDA-LSTM) and a transfer learning mechanism. The flowchart of the method is shown below. Figure 1 As shown, the specific steps are as follows:

[0093] S1. Collect historical operating sequence data of wind turbines in a reference wind farm, including physical quantities and corresponding feature vectors within the total number of time steps; use a sliding window mechanism to extract continuous time segments from the historical operating sequence data; each time segment includes the previous... The input time window and the subsequent time step The predicted horizon is calculated at each time step; the feature vectors within the input time window are combined to form the input sequence, and the physical quantities within the predicted horizon are combined to form the output label; the input sequence and the output label constitute a time segment sample; all time segment samples within the total number of time steps are combined to form a reference dataset.

[0094] The physical quantities include wind turbine operating parameters and / or meteorological information; the wind turbine operating parameters include rotor speed, pitch angle, active power, main bearing temperature, generator temperature, and vibration intensity; the meteorological information includes wind speed, air temperature, humidity, and wind direction. In this embodiment 1, based on wind farm SCADA data and meteorological monitoring data, multiple types of meteorological information (such as wind speed, air temperature, humidity, and wind direction) are integrated with turbine operating parameters (such as rotor speed, pitch angle, active power, main bearing temperature, generator temperature, and vibration intensity). It is understood that the selection of physical quantities is adjusted according to actual application.

[0095] S2. Based on the target wind farm, repeat S1 to obtain the target dataset;

[0096] S3. Construct a deep learning model, including a data-driven module and a physical assimilation constraint module; the data-driven module receives the input sequence of the reference dataset, extracts the reference fusion representation, and outputs the conditional probability distribution and predicted physical quantity values ​​within the predicted horizon; based on the output labels of the reference dataset and the conditional probability distribution within the predicted horizon, the deep learning model is pre-trained using a preset objective function; the physical assimilation constraint module applies physical assimilation constraints to the predicted physical quantity values ​​to obtain the pre-trained model; the network parameters of the pre-trained model include data-side parameters and physical-side parameters;

[0097] The processed data from step S1 is input into parallel sub-encoders corresponding to each physical quantity. Through temporal attention and cross-channel attention, cross-scale and cross-physical quantity information adaptive fusion is achieved, resulting in multi-source feature representations with clear physical orientation. Based on this, the multi-source feature representations are tightly coupled with the virtual wind field model in MP-DT. Physical assimilation constraints enable closed-loop collaboration between data-side prediction and physical-side attention rolling extrapolation, thereby constructing a multi-parameter digital twin of the wind turbine cluster with predictive and diagnostic capabilities.

[0098] This invention addresses the limitations of traditional statistical models, machine learning models, and single-unit LSTM in multi-parameter coupled modeling and cross-unit collaborative prediction by providing MDA-LSTM to meet the multi-parameter collaborative prediction needs of wind turbine clusters. MDA-LSTM uses "multi-channel encoding - dual attention - heteroscedasticity decoding" as its core concept. It simultaneously receives multiple physical quantities such as wind speed, rotor speed, pitch angle, active power, main bearing temperature, generator temperature, and vibration intensity. It explicitly distinguishes the differences in dimensions and time scales of different physical quantities, characterizes key moments through time attention, and characterizes key parameters through cross-channel attention, thereby more fully capturing the nonlinear coupling relationships between physical quantities and outputting high-quality multi-parameter prediction results suitable for cluster-level collaborative scheduling.

[0099] like Figure 2The diagram illustrates the overall framework of MDA-LSTM, visually demonstrating the relationships between key units such as multi-channel encoding, dual attention, and heteroscedastic decoding. Specifically, on the data-driven side, each physical quantity is modeled by an independent LSTM sub-network to preserve different dimensions, time scales, and dynamic features, avoiding dimensional contamination and information weakening caused by simple splicing. Based on this, the temporal attention mechanism adaptively focuses on key moments that contribute more to subsequent predictions within a given input window, suppressing redundant and high-noise sampling points. Cross-channel attention dynamically weights each physical quantity in the form of probability weights, obtaining a visualized and diagnosable distribution of variable importance, thus characterizing when and which channel is more important. At the output end, MDA-LSTM employs heteroscedastic probabilistic decoding, simultaneously providing the conditional mean and diagonal variance of multi-step predictions. It is trained using negative log-likelihood (NLL), and numerical stability and statistical interpretability are ensured through simplex constraints and lightweight regularization terms in temporal and cross-channel attention. Furthermore, the model introduces weighting coefficients for different prediction step sizes in the loss function, allowing for flexible emphasis on short-term scheduling accuracy or medium-term rolling plan accuracy based on actual operational preferences. The above design collectively enhances the model's ability to simultaneously characterize drastic transient signals and slowly changing variables, effectively suppresses performance degradation caused by strong multivariate coupling and error propagation, and provides high-quality data-side representation for subsequent physical assimilation and cross-site alignment.

[0100] Compared to traditional LSTM models, which typically rely on single time-series data, such as historical wind speed or power data, and neglect other meteorological factors affecting wind turbine power output, such as temperature, humidity, and wind direction, thus limiting the model's prediction accuracy, the MDA-LSTM provided in this embodiment of the invention introduces multi-dimensional turbine operating parameters and considers multiple meteorological factors simultaneously, enhancing the model's adaptability to the complex changes in wind turbine clusters. MDA-LSTM processes different types of data through multiple input channels, enabling the model to comprehensively understand the impact of each input parameter on the wind turbine, thereby improving the accuracy and stability of predictions.

[0101] To clearly describe various variables, parameters, hidden states, loss functions, and constraints, and to facilitate logical connection and comparative analysis, the embodiments of this invention use standardized conventions for the symbols involved in the method, as shown in Table 1, which lists the main general symbols and their meanings. Table 2 shows the symbols and meanings related to MDA-LSTM, facilitating a rigorous description of the structure and mathematical form of MDA-LSTM.

[0102] The formula for a traditional LSTM is as follows:

[0103] ;

[0104] The multi-step decoding expression is:

[0105] ;

[0106] The weighted multistep loss expression is:

[0107] ;

[0108] Traditional LSTM can model long dependencies, but single-stream input and static fusion cannot distinguish the differentiated contributions of multiple physical quantities; early splicing amplifies scale heterogeneity and correlation differences; and feature importance cannot be adaptively adjusted when operating conditions change. Therefore, it suffers from insufficient performance and interpretability in wind turbine group collaborative prediction and scheduling scenarios. The MDA-LSTM provided in this embodiment of the invention has the following construction formula:

[0109] (1);

[0110] Formula (1) satisfies the following sub-formulas and constraints:

[0111] A1. The channel-level encoding expression is as follows:

[0112] (2);

[0113] Channel-level encoding allows each physical quantity to be modeled independently in its own LSTM subspace to preserve scale and dynamic differences, thus avoiding dimensional contamination caused by early splicing from the source.

[0114] A2. The expression for time attention is as follows:

[0115] (3);

[0116] (4);

[0117] (5);

[0118] By employing a time-attention mechanism, key moments are adaptively amplified and redundant points are suppressed within each channel, thereby achieving a compact... Improve noise immunity and interpretability.

[0119] A3. The cross-channel attention expression is as follows:

[0120] (6);

[0121] (7);

[0122] (8);

[0123] Cross-channel attention mechanism with probability weights The contribution of each physical quantity is dynamically aggregated, enabling a rapid response based on which quantity is most important under the current operating condition, and it can be directly used for visual diagnosis.

[0124] A4. The expression for decoding heteroscedasticity probability is as follows:

[0125] (9);

[0126] (10);

[0127] (11);

[0128] By simultaneously outputting point predictions and diagonal variances through heteroscedastic probability decoding, uncertainty is characterized, providing a statistical basis for interval assessment and risk constraints.

[0129]

[0130] A5. The expressions for negative log-likelihood and simplex constraints are as follows:

[0131] (12);

[0132] (13);

[0133] This invention employs a differentiable and stable negative log-likelihood model training method and simplex constraints to ensure the probabilistic interpretation and numerical controllability of attention weights. Negative log-likelihood is used for theoretical derivation to maintain statistical interpretability, and a weighted mean square error function is used for parameter optimization. These two methods are linked through a unified weighting coefficient or learning rate, allowing the model to fully learn the temporal dependencies and coupling patterns between various physical quantities. Those skilled in the art should understand that, beyond the negative log-likelihood, other functions based on probability distribution metrics can also be used; similarly, beyond the weighted mean square error, other functions based on prediction error metrics can also be used.

[0134] In summary, MDA-LSTM firstly through Each physical channel is encoded independently, then time attention is used to focus on key moments within a window, followed by cross-channel attention to adaptively fuse the contributions of different physical quantities, and finally, a heteroscedastic probability head is output simultaneously. The theoretical derivation of the step mean and variance uses negative log-likelihood to maintain statistical interpretability, while empirical training uses weighted mean square error. Both are implemented with a unified step size and weighting caliber. MDA-LSTM enables interpretable modeling of multi-source heterogeneous data. , It describes when and which channels are more important in a probabilistic form, which facilitates visualization and diagnosis; it achieves condition-adaptive operation, and cross-channel attention makes the feature importance under different wind conditions (low wind, near rated, high wind) change dynamically over time; it achieves uncertainty quantification, and heteroscedasticity output supports interval coverage and risk constraints; it achieves robustness and reproducibility, and negative log-likelihood and probabilistic constraints suppress attention collapse and numerical drift, ensuring consistent standards.

[0135]

[0136] Under the background of high-proportion wind power grid connection and complex operating conditions, prediction-scheduling models based on single parameters or low-dimensional features show significant limitations in abnormal condition identification, early fault warning, and group control optimization. This invention provides an MP-DT construction approach, treating the wind turbine cluster as a multi-source, multi-scale coupled system. Through digital twin technology, a real-time mapping is established between the physical and digital spaces, providing a high-precision, physically constrained virtual carrier for cluster operation status perception and scheduling optimization. Specifically, MP-DT is based on wind farm SCADA data and meteorological monitoring data, integrating multiple types of meteorological information (such as wind speed, temperature, humidity, and wind direction) with unit operating parameters (such as rotor speed, pitch angle, active power, main bearing temperature, generator temperature, and vibration intensity) to construct a virtual wind farm model in the state space that simultaneously reflects the multi-dimensional operating characteristics of the wind turbine cluster. By unifying the encoding and mapping of multi-parameter data, MP-DT can not only finely characterize the operating status at the individual unit and cluster levels, but also explicitly embed physical mechanisms and engineering constraints into the state equations and observation equations, enabling rolling predictions and reliable extrapolations of future performance. Compared with traditional methods that rely solely on single-channel information such as power or wind speed, MP-DT can process multi-dimensional data channels within a unified framework, fully considering the temporal correlation and physical coupling relationships between key operating parameters, thereby significantly improving the accuracy and reliability of wind turbine cluster prediction-scheduling integration. MP-DT achieves closed-loop feedback between the physical-side model and the data-side observations through a cycle of "state evolution - observation mapping - assimilation update": the hidden state of the cluster is rolled through state equations based on physical priors and learnable residuals; the hidden state is projected onto the multi-parameter output space through the observation model, and the virtual model is assimilated and corrected using real-time SCADA data, ensuring that the digital twin maintains high consistency with the real wind field throughout its entire lifecycle.

[0137] like Figure 3 As shown, the overall architecture of MP-DT is illustrated. Table 3 shows the relevant symbols and mathematical notation for Multi-Parameter Digital Twin (MP-DT), which facilitates the formal derivation and analysis of key modules such as the state equation, observation equation, and assimilation loss function.

[0138] MP-DT includes a data fusion module, a virtual model construction module, and a real-time monitoring and scheduling optimization module.

[0139] The data fusion module integrates various operational data from wind turbines with real-time meteorological data to construct a complete wind turbine cluster dataset. The virtual model construction module builds a virtual model of the wind turbine cluster based on the multi-parameter dataset, and synchronously updates the virtual model with the actual state of the wind turbine cluster using digital twin technology. The real-time monitoring and scheduling optimization module transmits the current state of the wind farm to the digital twin for updates through a real-time feedback mechanism, adjusts scheduling strategies, and achieves real-time optimization. The specific construction formula for MP-DT is as follows:

[0140] B1. The state equation is expressed as follows:

[0141] (14);

[0142] By combining interpretable physical priors with learnable residuals through state equations, interpretability and fitting strength are balanced, and process noise is used to absorb unmodeled dynamics.

[0143] B2. The expression for the observation equation is as follows:

[0144] (15);

[0145] By mapping hidden states to observables through observational models, a data-state bridge is provided for physical assimilation.

[0146] B3. The expression for attention scrolling extrapolation is as follows:

[0147] (16);

[0148] Predicting the horizon The future projection of the physical side is generated and used for the data side to predict physical quantities for consistent comparison.

[0149]

[0150] B4. The assimilation target expression is as follows:

[0151] (17);

[0152] Formula (17) restricts model learning to the physical feasible region and suppresses out-of-distribution drift by three terms: observation consistency, dynamic consistency and prediction consistency.

[0153] In summary, MP-DT relies on physical priors. and learnable residuals This constitutes state dynamics, which are connected to measurable quantities through observation modules, and then the state attention is rolled out and extrapolated to the future. The next step involves physically assimilating the state and parameters through observational consistency, dynamic consistency, and prediction consistency. Consistency constraints limit model learning to the physically feasible region, reducing drift of the pure data model under out-of-distribution conditions; and bidirectional correction is achieved, with the prediction results on the data side... Projection with physical side Mutual reference enhances robustness to extreme and rare events; it enables explainability and adjustability. Explicitly characterizing the process and observational uncertainties facilitates engineering implementation and sensitivity analysis.

[0154] S4. Input the target dataset into the pre-trained model to obtain the target fusion representation; perform domain alignment between the reference fusion representation and the target fusion representation; optimize the pre-trained model based on the target dataset using a layer-by-layer unfreezing strategy to minimize the constructed joint loss function and obtain a collaborative prediction model; wherein, the joint loss function is the sum of the preset objective function loss, the physical assimilation constraint loss, and the domain alignment loss;

[0155] This invention explicitly introduces a transfer learning mechanism in the collaborative prediction stage to further improve the prediction accuracy and generalization ability of the digital twin model in cross-site and multi-condition scenarios. Specifically, the MP-DT model based on MDA-LSTM is combined with transfer learning and cross-domain alignment strategies to perform end-to-end multi-parameter collaborative prediction of wind turbine clusters. MDA-LSTM simultaneously receives multi-source time-series data through a multi-channel input structure, including wind speed, rotor speed, pitch angle, active power, main bearing temperature, generator temperature, and vibration intensity. It utilizes time attention and cross-channel attention to capture long-term dependencies at different time scales and nonlinear coupling characteristics between multiple physical quantities, providing a structured and physically meaningful fusion representation for subsequent transfer learning. By treating the source wind farm as the source domain and the target wind farm as the target domain, and introducing domain alignment indices such as Maximum Mean Difference (MMD) and Correlation Alignment (CORAL, which minimizes the second-order statistical difference between the source and target domains) into the fusion representation layer, the feature distributions and covariance structures of the source and target domains are jointly constrained. This reduces the domain shift caused by different wind resource conditions, terrain features, and operational strategies, significantly improving the model's adaptability and sample efficiency on the target wind farm in two stages: source farm pre-training and target farm fine-tuning. The synergistic mechanism of the MP-DT model based on MDA-LSTM and transfer learning not only maintains high-precision characterization of multi-parameter dynamics within a single wind farm, but also robustly transfers the learned cluster co-evolution laws to the relatively data-scarce target wind farm, providing a unified and reproducible modeling path for cross-site digital twin construction and group control scheduling.

[0156] Transfer learning is deeply coupled with MDA-LSTM and MP-DT in the same training framework through joint loss functions and optimization strategies: In the source wind farm stage, by minimizing data fitting terms such as weighted mean square error and negative log-likelihood, the MP-DT model based on MDA-LSTM fully learns the temporal dependencies and coupling patterns among multiple parameters, and anchors the data-driven representation within the physical feasible domain under physical assimilation constraints; In the target wind farm stage, while keeping the backbone network parameters relatively stable, domain alignment indicators such as maximum mean difference and correlation alignment are introduced, and a fine-tuning strategy of freezing first and then unfreezing layer by layer is adopted to enable the model to complete a smooth transfer from the source domain to the target domain under limited sample conditions, avoiding the forgetting of learned knowledge or fluctuations in prediction performance due to a large update at once. Furthermore, by setting reasonable weight coefficients for different loss terms and combining block updates and early stopping, a balance is achieved between data fitting, physical consistency, and cross-domain alignment during the training process. This ensures training convergence while further improving the robustness and engineering usability of the model under multiple sites and operating conditions, providing reliable multi-parameter prediction support for subsequent cluster scheduling optimization and intelligent operation and maintenance.

[0157] Table 4 shows the symbols and meanings related to transfer learning and domain alignment. Figure 4 The above is a schematic diagram of source-target domain knowledge transfer using a transfer learning mechanism.

[0158]

[0159] Transfer learning transfers data knowledge from the source wind farm to the target wind farm, enabling the MDA-LSTM-based MP-DT model to adapt to the environments and needs of different wind farms. By pre-training on historical data from multiple wind farms, the model can learn universal patterns of wind turbine clusters, improving cross-wind farm collaborative prediction capabilities and enhancing the scheduling and optimization capabilities of wind turbine clusters. This is particularly evident in situations with scarce or insufficiently labeled data, significantly improving prediction accuracy. The specific formula is as follows:

[0160] C1. The MMD expression is as follows:

[0161] (18);

[0162] Formula (18) uses kernel mean embedding to measure the difference in the source high-order distribution and the difference in the target high-order distribution, adapting to nonlinear domain shift.

[0163] The C2. CORAL expression is as follows:

[0164] (19);

[0165] Formula (19) uses second-order statistics to align the source covariance and the target covariance, which is simple to calculate and complementary to MMD.

[0166] C3. The weighted combination expression is as follows:

[0167] (20);

[0168] The fusion representation layer is used to jointly align the distribution and covariance, and combined with layer-by-layer unfreezing and domain-specific normalization to achieve robust generalization.

[0169] In summary, transfer learning simultaneously measures and compresses the higher-order distributional differences and second-order statistical differences between the source and target domains at the fusion representation layer, incorporating them into a unified objective through weighted combination, and is further enhanced by layer-by-layer unfreezing and domain-specific normalization; it achieves cross-site generalization, significantly reduces domain bias under the condition of limited target domain samples, and stabilizes transfer performance; it achieves complementary alignment: MMD adapts to nonlinear differences, CORAL is low-cost and robust, and the weighting of the two improves alignment quality; and it achieves controllable transfer, with the alignment layer and weights explicitly set, facilitating the trade-off between bias and variance under different domain differences and data volume scenarios.

[0170] Table 5 shows the joint loss function and its components, corresponding weight parameters, learning rate, and standardized gradient-related notation. This is to illustrate the objective function composition, weight coefficient settings, and optimization-related notation during the joint training process, providing a unified symbolic basis for training process description and stability analysis.

[0171] The joint loss function is expressed as follows:

[0172] (twenty one);

[0173] Formula (21) minimizes the conditional likelihood on the data side, the assimilation consistency on the physical side, and the cross-domain representation alignment simultaneously in the same objective, avoiding redundant loss calculation, and enabling the three gradients to converge collaboratively under the same optimizer: the first term is the conditional likelihood of MDA-LSTM; the second term is the physical assimilation of MP-DT, including observation consistency, dynamic consistency, and prediction consistency; and the third term is the cross-domain alignment of transfer learning.

[0174]

[0175] The joint training and stability testing are detailed below:

[0176] D1. The block update expression is as follows:

[0177] (twenty two);

[0178] Formula (22) adopts Subject to the combined gradient of the three terms, Block-based first-order updates dominated by physical assimilation are used to improve numerical stability and separation of duties.

[0179] D2. The standardized expression is as follows:

[0180] (twenty three);

[0181] Standardization is achieved through training set statistics to prevent information leakage and ensure consistent standards for verification and deployment.

[0182] (c) The step-weighted expression is as follows:

[0183] (twenty four);

[0184] Formula (24) allows for emphasizing proximal or distal step sizes based on operational preferences without altering the overall target architecture. Optimal prediction is achieved through training and stability testing. This embodiment of the invention employs weighted average variance step size weighting for training, maintaining consistent step size weighting caliber.

[0185] In summary, during training and stability testing, Updated by the combined gradient of the joint loss, Updates are dominated by the physical assimilation term; input sequences and output labels are uniformly standardized according to training set statistics; and the following methods are used. Adjusting the importance of different prediction steps, supplemented by lightweight regularization (entropy and time smoothing, weight decay, warm-up to cosine annealing, gradient pruning). Achieving separation of responsibilities and numerical stability, block updates alleviate gradient conflicts between the data side and the physical side, improving convergence; achieving strict leakage prevention, standardization uses only training statistics to ensure consistency between evaluation and deployment; enabling customized objectives, step-size weighting allows the model to be optimized for short-term scheduling or medium-term rolling plans; achieving engineering reproducibility, unified optimization and regularization strategies ensure comparability and reproducibility across different data batches and sites.

[0186] The wind turbine group multi-parameter collaborative prediction method provided in this embodiment of the invention takes formula (21) as the core, and forms an end-to-end reproducible optimization chain by combining the multi-source interpretable modeling of MDA-LSTM, the physical assimilation constraint of MP-DT and the domain alignment mechanism of transfer learning. The stability and implementation are ensured by the engineering approach of formulas (22)-(24).

[0187] Example 2

[0188] To verify the collaborative prediction capability of a multi-parameter digital twin (MP-DT) of a wind turbine cluster based on a multi-dimensional adaptive long short-term memory network (MDA-LSTM) combined with transfer learning, this Example 2 uses SCADA data from two wind farms, A and B, with a time resolution of 10 minutes, totaling over 50,000 valid samples. A cross-site transfer learning framework is employed, with wind farm A as the source domain for model pre-training and wind farm B as the target domain for fine-tuning and testing, verifying the model's generalization ability across heterogeneous wind farms.

[0189] Based on the multi-parameter collaborative prediction method for wind turbine clusters provided in Example 1, an MP-DT model based on MDA-LSTM and a transfer learning strategy (hereinafter referred to as this model or MDA-LSTM) is adopted to construct a multi-parameter prediction system using seven physical quantities: wind speed, rotor speed, pitch angle, active power, main bearing temperature, generator temperature, and vibration intensity. The data preprocessing workflow includes timestamp alignment and resampling, outlier removal, and robust normalization. A sliding window is used to construct samples, with an input sequence length of 48 time steps (8 hours of historical data) and an output sequence of 6 time steps (1-hour predicted values), reflecting the multi-dimensional input design concept. Evaluation metrics include MSE and RMSE, and all evaluation metrics are calculated in a robust z-score space to ensure cross-parameter comparability. The Wasserstein distance is also used to test distribution consistency. Traditional time series models ARIMA, Support Vector Regression (SVR), and the traditional LSTM model are selected as baseline methods for comparison. This model employs a two-stage transfer learning training strategy: the pre-training stage trains on the source data for 14 rounds, and the fine-tuning stage trains on the target data for 10 rounds; the optimizer used is AdamW.

[0190] Multi-physical quantity prediction and results analysis

[0191] like Figure 5 As shown, this is the loss change curve during the training process based on transfer learning. (Left side) Figure 5 (a) represents the MSE loss during the pre-training phase, showing that both the training loss and validation loss during the pre-training phase exhibit a stable decreasing trend, and the gap between them gradually narrows, essentially converging after 14 rounds; (right side) Figure 5 (b) represents the MSE loss during the fine-tuning phase, showing that the loss decreases rapidly during this phase, plateaus after 10 rounds, and eventually stabilizes at a low level, verifying that the loss does not rebound. This indicates that the transfer learning strategy provided by this embodiment of the invention exhibits good convergence characteristics and generalization ability in both training phases, without overfitting, providing a reliable guarantee for subsequent prediction performance.

[0192] like Figure 6As shown, the prediction performance of MDA-LSTM based on multi-dimensional input is demonstrated on seven physical quantities. MDA-LSTM (cal.) represents the calibrated sequence after fine-tuning by transfer learning. MDA-LSTM (cal.) and the measured sequence maintain a high degree of consistency in overall trend, stage pattern, and major turning points. The two basically overlap in the steady and slow change stages. Including the transient stages of peaks, cliff-like drops, or rapid rebounds, MDA-LSTM can correctly give the inflection point timing and magnitude. Slight smoothing and phase lag are occasionally seen in the extreme value neighborhood, but the performance across physical quantities is consistent, demonstrating good cross-variable generalization and engineering usability. Figure 6 In (a), the rotor speed response to the rise and fall driven by wind speed is well characterized, and the dynamic lag caused by the moment of inertia is reasonably reproduced; when a sudden change occurs, the model phase lags slightly by 1 to several sampling steps, but the overall envelope is consistent. Figure 6 In (b), the pitch angle can stably track the step and small-amplitude adjustments caused by commands and adjustments; in the sharp pitch adjustment section, the prediction has a slight tendency to smooth out noise, but the turning point matches the platform level well, which characterizes the main rhythm of power regulation. Figure 6 In (c), the active power nonlinear segment (from rising to a wide peak and from plateau to fall) has a high degree of consistency in shape, and the peak value and plateau range are closely aligned; slight underestimation and overestimation are occasionally seen in the neighborhood of rapid rise and fall, but this does not affect the judgment of the main trend. Figure 6 In (d), the gradual change process of the main bearing temperature thermal inertia is accurately characterized, and the steady-state section almost overlaps; the temperature rise and fall slopes after the load change are fully reproduced with the time delay, indicating that the model has excellent memory and noise filtering capabilities for slow variables. Figure 6 In (e), the correlation between generator temperature and power level is well captured, and the temperature peak value has a reasonable time lag relative to the power peak value; the steady state and long-tail decay segment fit well, demonstrating the stability of thermal process modeling. Figure 6 In (f), the wind speed can accurately reproduce the slow fluctuations and gust outlines of the background wind; the peak amplitude of short-term strong gusts is slightly conservative, and the peak edge is smoothed, but the timing and width of the main peak are well matched. Figure 6 In (g), the vibration intensity reproduces multi-peak fluctuations and transient enhancements well, and can distinguish between baseline and event-triggered increments; the amplitude of individual sharp peaks is slightly conservative, but the peak time and duration are accurately matched, demonstrating sensitivity to abnormal dynamics. MDA-LSTM achieves high-fidelity trend, accurate peak and valley positioning, and strong cross-variable consistency in time-domain fitting across seven physical quantities, specifically as follows: 1. High overlap between plateaus and long tail segments, with no divergence in predictions; 2. It can still grasp inflection points and amplitude magnitudes under conditions with strong transients, maintaining sufficient temporal consistency in engineering; 3. It is robust to both slow variables (such as temperature) and fast variables (such as wind speed, power, and vibration), showing good generalization and robustness, providing a reliable time-domain prediction basis for unit operation assessment, health monitoring, and feedforward and closed-loop control.

[0193] like Figure 7 (a) and Figure 8 As shown in (b), MDA-LSTM achieves the minimum error on the parameter-by-parameter RMSE of all seven physical quantities, and the error bars converge significantly, demonstrating higher prediction stability and reproducibility. Compared with baselines such as ARIMA, SVR, and LSTM, MDA-LSTM's advantages are particularly prominent on high dynamic parameters (such as wind speed, power, and vibration), with errors decreasing by orders of magnitude; it also maintains the minimum error on slowly varying parameters (such as main bearing temperature and generator temperature). Figure 8 (a) is a comparison chart of overall RMSE. Figure 8 (b) The parameter-by-parameter RMSE comparison chart shows that MDA-LSTM comprehensively outperforms the baseline method across all physical quantities, with improvements exceeding 30% on key physical quantities. Figure 8 As shown in (a), MDA-LSTM maintains a significant advantage in overall RMSE performance, and its error bar range is narrower, demonstrating better prediction stability. Quantitative analysis shows that the average RMSE of MDA-LSTM is 0.312, which is 23.7% higher than the optimal baseline method. This verifies the effectiveness of the multi-dimensional input fusion and adaptive weight adjustment mechanism proposed in this embodiment of the invention, which can effectively suppress noise and reduce cross-variable coupling errors, thereby obtaining more accurate and robust multi-parameter synchronous prediction results.

[0194] Cross-condition adaptability verification

[0195] like Figure 9 As shown, the time-domain behavior of seven physical quantities is compared under different wind conditions. Figure 9 (a) On the full sample without different working conditions, the model exhibits high trend consistency, low bias and narrow prediction discrepancy, indicating that the global calibration of the model is effective and has the ability to suppress noise. Figure 9 (a) Overall comparison of the whole sample shows that, in the whole sample data without wind speed, the predicted dashed line and the measured solid line are generally in phase and shape with no systematic deviation in all parameters; the peak and valley times are basically aligned with the platform intervals, and the discrete band of the predicted trajectory is narrow, showing good global robustness and noise suppression effect. Figure 9 (b) Under low wind speed conditions, small fluctuations in power and rotor speed are effectively suppressed, while variables such as pitch angle and temperature remain stable, and the prediction closely matches the actual measurement in the entire time domain. Figure 9 (c) Under near-rated operating conditions, the model accurately reconstructs the strong coupling relationship between power, pitch and speed. The power plateau and pitch transition process are continuous and smooth. The time delay and amplitude of speed adjustment are consistent with the actual measurement. No abnormal peaks were observed in the vibration. Figure 9(d) Under high wind speed conditions, the model is in the active power-limiting zone, with power maintaining a plateau characteristic near the cutoff boundary, and pitch increasing with wind speed. The prediction shows no visible overshoot or oscillation in the boundary region, and the trends of slow variables such as temperature and vibration are consistent with the actual measurements. This model can stably reproduce the key dynamic characteristics of the wind turbine under the full sample and three typical operating conditions, taking into account the modeling accuracy of fast variables (such as wind speed, power, and vibration) and slow variables (such as temperature), demonstrating strong adaptability and robustness across operating conditions.

[0196] Distribution Consistency and Error Distribution

[0197] like Figure 10 As shown, kernel density estimation is used to compare the probability distributions of predicted and measured values. The curves of this model almost coincide with the measured distributions, and can simultaneously align modal positions, skewness, and tail mass. Baseline methods exhibit distribution collapse or shifts in multiple physical quantities, ARIMA shows narrow peak pulses at the boundaries, and SVR and LSTM fail to adequately characterize the tail distribution. This model accurately reproduces the right-skewed long tail of wind speed, the narrow peak concentration of rotor speed, and the limiting and plateau characteristics of power. The density changes near the peak position and truncation boundary are consistent with the measurements. The probability mass distribution of the control parameter (pitch angle) in the neighborhood of the truncation boundary is accurately captured, without the one-sided accumulation or spurious pulses common in baseline methods. The thermodynamic parameters (main bearing temperature, generator temperature) exhibit a narrow peak and light tail shape consistent with thermal inertia, and the model's fitting of peak width and decay rates on both sides is highly consistent with the measurements. The mechanical parameters (vibration intensity) have their heavy tail and tail probabilities of rare events reasonably preserved, avoiding excessive smoothing of extreme values. The matching at the distribution level demonstrates that this model can not only accurately predict the mean trend but also effectively characterize higher-order statistical properties (skewness, kurtosis, and tail probability), improving usability in risk-sensitive scenarios. Further quantitative evaluation shows that the average Wasserstein distance is 0.023, significantly better than the baseline, validating the advantage of distribution consistency. Figure 11 As shown in (a), the residual distribution (violin plot) is generally symmetrical around zero with a sharp peak and short tails. The median is close to 0, and the interquartile range is narrow, indicating that there is no significant systematic bias in this model, and the error is mainly concentrated in a small range. A few high dynamic parameters (such as wind speed, vibration, and power) have relatively wider tails, but the tail probability is still low. Figure 11 As shown in (b), the empirical cumulative distribution function of the absolute error rises rapidly in the interval near zero and quickly approaches 1, indicating that the absolute error of the vast majority of samples is very small and the probability of extreme error events is extremely low. The distribution level is consistent with the aforementioned RMSE results, further demonstrating that this model possesses unbiasedness, error concentration, and robustness, and can effectively suppress abnormal errors.

[0198] Mechanism consistency and interpretability

[0199] like Figure 12As shown, the evaluation of the importance of the seven input features yields a long-tailed distribution. The vertical reference line is around 0.10 (average importance of all features), indicating that pitch, wind speed, and rotor are significantly above average, while the remaining features are below average. This suggests that pitch angle, as an active control variable, directly determines power limitation and load distribution, hence its highest weight. Wind speed-rotor reflects aerodynamic input and dynamic response, making a significant contribution. Power and vibration reflect current load and structural response, at a moderate level. Temperature-related features are dominated by thermal inertia, primarily describing slow timescale processes, with a smaller marginal contribution to short-term prediction. The feature ranking is consistent with the physical mechanism of wind turbine pitch control-aerodynamics-transmission, verifying the interpretability and physical consistency of the model in feature utilization. This indicates that the model relies more on effective modeling of control and fast dynamic information while suppressing the interference of redundant slow variables.

[0200] like Figure 13 As shown, a correlation heatmap illustrating the coupling relationships among seven physical quantities is presented, with coefficients labeled. The dominant correlation in the heatmap (wind speed-rotor-power-generator temperature) aligns with the aerodynamic-transmission-electrical-thermal chain of the wind turbine. The significant negative correlation between pitch and rotor validates the pitch-limiting power strategy. The directional differences in temperature variables (positive correlation with generator temperature, negative correlation with main bearing temperature) reflect the heating mechanisms and time-scale differences of different components. These results corroborate the aforementioned conclusions regarding feature importance and time-domain fitting, supporting the physical consistency and rationality of multi-physical-quantity collaborative modeling.

[0201] like Figure 14 As shown, the contribution of local features is evaluated for a selected time period. The bar length represents the most useful information source for prediction during that period. The results show that vibration response dominates during this period. This model adaptively focuses its attention on fast structural dynamic signals, followed by incoming flow and rotational links (wind speed, rotor, pitch), while slow variables such as temperature (generator temperature, main bearing temperature) contribute the least. This demonstrates that the adaptive weighting mechanism can dynamically redistribute information sources according to changes in operating conditions. When there are vibration events and enhanced structural response, this model increases the weight of the features most relevant to them, strengthening the characterization of anomalies and short-term dynamics; in other time periods, it allocates the main driving forces such as steerable wind speed, and... Figure 8 The global importance of these factors is mutually corroborated. This model improves sensitivity to short-term anomalies and robustness to steady state, supporting accuracy, stability, and interpretability in complex operating environments.

[0202] In summary, this embodiment 2 verifies the effectiveness of the MP-DT model based on MDA-LSTM and transfer learning from multiple dimensions, including prediction accuracy, working condition adaptability, statistical reliability, mechanism consistency and engineering practicality. Figure 6 Multi-physical quantity prediction comparison Figure 9 Cross-condition adaptability and Figure 10The distribution consistency analysis intuitively demonstrates the comprehensive advantages of the model in multi-parameter digital twins of wind turbine clusters. In terms of prediction performance and robustness, MDA-LSTM achieves an average RMSE of 0.312 across the seven physical quantities, a 23.7% reduction compared to the best deep learning baseline model; and reduces the overall error by approximately 30.7%, 23.9%, and 17.9% compared to traditional methods ARIMA, SVR, and LSTM, respectively. The results show that the collaborative prediction method provided in this invention exhibits significant performance advantages and statistical robustness in multi-parameter joint modeling and under different operating conditions. Regarding cross-condition adaptability, under typical operating conditions such as low wind, near-rated, and high wind, the predicted trajectories of the seven physical quantities are highly consistent with the measured curves in terms of phase, waveform, and trend. The residual distribution is concentrated, the discrete band is narrow, and no obvious systematic bias appears, demonstrating the model's ability to collaboratively characterize multi-timescale features such as sudden wind speed changes, load fluctuations, and slow evolution of unit temperature, as well as its good noise suppression and operating condition adaptability. Regarding digital twin and physical consistency, this model integrates multi-source meteorological and turbine operation data, incorporating physical priors and learnable residuals into the state equation. Under three types of assimilation constraints—observational consistency, dynamic consistency, and prediction consistency—it achieves high-fidelity multi-parameter mapping between the physical wind farm and the virtual model. Experimental results show that the introduction of the digital twin layer effectively suppresses the phenomenon of purely data-driven models having numerically reasonable but physically questionable results under extreme conditions, providing reliable physical constraint support and engineering interpretability for collaborative prediction. Regarding transfer learning and domain alignment, the fusion representation layer jointly employs MMD and CORAL to align the feature distributions and covariance structures of the source and target wind farms. Through source field pre-training and target field fine-tuning, layer-by-layer unfreezing, and domain normalization strategies, it effectively reduces domain shifts caused by differences in wind resource conditions, terrain features, and operational strategies. Regarding the evaluation system and result reliability, this embodiment of the invention conducts error evaluation in a robust z-score normalized space and simultaneously uses multi-dimensional indicators such as MAE, RMSE, and Wasserstein distance to evaluate model performance from both point prediction error and distribution consistency perspectives. The conclusions of different evaluation indicators remain consistent, and the range of indicator changes is limited under multiple random initializations and data partitioning. This further verifies the reliability of the results and the robustness of the experimental conclusions of the collaborative prediction method provided in this embodiment of the invention. It is a unified modeling path that takes into account accuracy, stability and transferability. It can serve as a general foundation for digital twins in wind power scenarios, providing reliable support for applications such as health assessment, preventive maintenance and group control optimization. It lays the foundation for subsequent engineering applications in intelligent operation and maintenance and cluster scheduling of actual wind turbine groups.

[0203] To quantify the contribution to performance, this embodiment 2 also conducted ablation experiments to verify the results, constructing a five-level ablation sequence as follows: (i) Uni-LSTM (wind speed as the only physical quantity); (ii) Multi-LSTM (no adaptation); (iii) Multi-LSTM+TL (adding transfer learning TL to (ii)); (iv) MDA-LSTM (no TL, introducing multi-scale or dynamic adaptation); (v) MDA-LSTM+TL (Full, i.e., the model provided in this embodiment). Figures 15 to 18 The figure shown is a comparison of overall performance and parameter-by-parameter performance. Figure 16 As shown in the figure, the overall RMSE comparison plot shows that the RMSE decreased from 0.452 (Uni-LSTM) to 0.312 (Full), a decrease of approximately 31.0%. The contributions of each stage are as follows: from Uni-LSTM to Multi-LSTM, 15.7%; from Multi-LSTM to Multi-LSTM +TL, 8.4%; from Multi-LSTM +TL to MDA-LSTM (without TL), 6.8%; and from MDA-LSTM (without TL) to MDA-LSTM +TL, 5.7%. This demonstrates that multi-parameter coupling, adaptive mechanisms, and transfer learning all have significant and additive gains. Figure 17 As shown in the figure, this is a comparison with traditional baseline methods. The overall RMSE of ARIMA, SVR, and traditional LSTM are approximately 0.450, 0.410, and 0.380, respectively. The overall RMSE of MDA-LSTM+TL (approximately 0.312) shows error reductions of approximately 30.7%, 23.9%, and 17.9%, respectively, validating the advantages of this model in terms of statistical significance and engineering usability. Figure 15 , Figure 16 As shown in Table 6, the performance improvements for each physical quantity are as follows: wind speed decreased from 0.423 to 0.295, rotor from 0.398 to 0.288, pitch angle from 0.385 to 0.276, active power from 0.485 to 0.315, main bearing from 0.394 to 0.285, generator from 0.401 to 0.292, and vibration from 0.412 to 0.301. The performance improvements for each physical quantity are stably distributed between approximately 26% and 35%, and the RMSE of the complete MDA-LSTM+TL model relative to other ablation sequences shows a monotonically decreasing trend. The improvement in active power is the most significant, at approximately 35%, while wind speed and pitch angle also show significant decreases. This indicates that the complete MDA-LSTM+TL model not only improves the prediction of single physical quantities but also stably improves the digital twin representation of multiple physical quantities. The improvement is not concentrated in a single physical quantity but rather multiple physical quantities benefit synergistically.

[0204] Ablation experiments show that multi-parameter coupled modeling, adaptive attention mechanisms, and transfer learning have a significant synergistic effect on improving prediction performance. This model maintains good generalization accuracy and stability even with a small sample size of target sites: 1. The complete MDA-LSTM+TL model achieves stable and considerable error reductions on both the overall model and individual physical quantities, with an overall reduction of 31.0% and parameter-level reductions of approximately 26% to 35%, with the largest improvement in power prediction (approximately 35.1%). Prediction accuracy is effectively improved. 2. The combination of adaptive, multi-scale mechanisms, and transfer learning ensures consistent improvements across variables and time periods, such as... Figure 18 The heat map shown is predominantly green, and as shown in Table 6, each column is monotonically decreasing, reflecting the adaptability to operating conditions. 3. From Figure 15 As shown in Table 6, the multi-parameter coupling, adaptation, and transfer learning introduced into each ablation sequence all produce reproducible gains, consistent with the mechanism of multi-scale information fusion plus time-varying weight allocation. 4. As Figure 17 As shown, the complete MDA-LSTM+TL model demonstrates significant advantages over ARIMA, SVR, and LSTM. The quantitative improvements shown in Table 6 demonstrate that this model meets the accuracy and stability requirements of online deployment of multi-parameter digital twins for wind turbine clusters, and possesses engineering practicality.

[0205]

[0206] Example 3

[0207] This embodiment 3 also provides a multi-parameter collaborative prediction system for wind turbine groups, including:

[0208] The reference dataset module is used to collect historical operating sequence data of wind turbines in a reference wind farm, including physical quantities and corresponding feature vectors within the total number of time steps. A sliding window mechanism is used to extract continuous time segments from the historical operating sequence data; each time segment includes the preceding... The input time window and the subsequent time step The predicted horizon is calculated at each time step; the feature vectors within the input time window are combined to form the input sequence, and the physical quantities within the predicted horizon are combined to form the output label; the input sequence and the output label constitute a time segment sample; all time segment samples within the total number of time steps are combined to form a reference dataset.

[0209] The target dataset module is used to obtain the target dataset by repeating S1 based on the target wind farm.

[0210] A pre-trained model module is used to construct a deep learning model, including a data-driven module and a physical assimilation constraint module. The data-driven module receives the input sequence of the reference dataset, extracts the reference fusion representation, and outputs the conditional probability distribution and predicted physical quantity values ​​within the predicted horizon. Based on the output labels of the reference dataset and the conditional probability distribution within the predicted horizon, the deep learning model is pre-trained using a preset objective function. The physical assimilation constraint module applies physical assimilation constraints to the predicted physical quantity values ​​to obtain the pre-trained model.

[0211] A collaborative prediction model module is used to input the target dataset into the pre-trained model to obtain a target fusion representation; the reference fusion representation and the target fusion representation are aligned in a domain; based on the target dataset, the pre-trained model is fine-tuned through a layer-by-layer unfreezing strategy to minimize the constructed joint loss function, thereby obtaining a collaborative prediction model; wherein, the joint loss function is the sum of the preset objective function loss, the physical assimilation constraint loss, and the domain alignment loss;

[0212] The real-time prediction module is used to collect real-time operating sequence data of wind turbines in the target wind farm, and input the real-time operating sequence data into the collaborative prediction model to obtain the predicted physical quantities.

[0213] This embodiment 3 also provides an electronic device, which includes: a memory, a processor, and a computer program or instructions stored in the memory. The processor executes the computer program or instructions to implement the multi-parameter collaborative prediction method for wind turbine groups in embodiment 1.

[0214] Although not shown, the electronic device includes a processor that can perform various appropriate operations and processes based on programs and / or data stored in read-only memory (ROM) or loaded from a storage portion into random access memory (RAM). The processor can be a multi-core processor or may contain multiple processors. In some embodiments, the processor may include a general-purpose main processor and one or more specialized coprocessors, such as a central processing unit, graphics processing unit (GPU), neural network processor (NPU), digital signal processor (DSP), etc. Various programs and data required for device operation are also stored in RAM. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0215] The above embodiments should be understood as being used only to illustrate the present invention more clearly, and not to limit the scope of the present invention. After reading the present invention, any modifications of the present embodiments by those skilled in the art will fall within the scope defined by the appended claims.

Claims

1. A multi-parameter collaborative prediction method for wind turbine clusters, characterized in that, include: S1. Collect historical operating sequence data of wind turbines in a reference wind farm, including physical quantities and corresponding feature vectors within the total number of time steps; use a sliding window mechanism to extract continuous time segments from the historical operating sequence data; each time segment includes the previous... The input time window and the subsequent time step The predicted horizon is calculated at each time step; the feature vectors within the input time window are combined to form the input sequence, and the physical quantities within the predicted horizon are combined to form the output label; the input sequence and the output label constitute a time segment sample; all time segment samples within the total number of time steps are combined to form a reference dataset. S2. Based on the target wind farm, repeat S1 to obtain the target dataset; S3. Construct a deep learning model, including a data-driven module and a physical assimilation constraint module; the data-driven module receives the input sequence of the reference dataset, extracts the reference fusion representation, and outputs the conditional probability distribution and predicted physical quantity values ​​within the predicted horizon; based on the output labels of the reference dataset and the conditional probability distribution within the predicted horizon, the deep learning model is pre-trained using a preset objective function; the physical assimilation constraint module applies physical assimilation constraints to the predicted physical quantity values ​​to obtain the pre-trained model; the network parameters of the pre-trained model include data-side parameters and physical-side parameters; S4. Input the target dataset into the pre-trained model to obtain the target fusion representation; The reference fusion representation and the target fusion representation are aligned in domains. Based on the target dataset, the pre-trained model is optimized through a layer-by-layer unfreezing strategy to minimize the constructed joint loss function, thereby obtaining a collaborative prediction model; wherein, the joint loss function is the sum of the preset objective function loss, the physical assimilation constraint loss, and the domain alignment loss; S5. Collect real-time operating sequence data of wind turbines in the target wind farm, and input the real-time operating sequence data into the collaborative prediction model to obtain the predicted physical quantities; In step S3, the data-driven module is a multi-dimensional adaptive long short-term memory network; the multi-dimensional adaptive long short-term memory network extracts fusion representations through multi-channel long short-term memory network encoding, temporal attention, and cross-channel attention mechanisms; and outputs the conditional probability distribution and predicted physical quantity values ​​within the predicted horizon through heteroscedastic probability decoding; wherein, the multi-channel long short-term memory network includes multiple long short-term memory networks set in parallel for processing the feature vectors; the fusion representation is the dependency relationship of the physical quantity at different time scales and the nonlinear coupling characteristics between the physical quantities; the preset objective function includes a probabilistic objective function and a point estimation objective function.

2. The multi-parameter collaborative prediction method for wind turbine clusters according to claim 1, characterized in that, In step S1, the physical quantities include wind turbine operating parameters and / or meteorological information; the wind turbine operating parameters include rotor speed, pitch angle, active power, main bearing temperature, generator temperature, and vibration intensity; the meteorological information includes wind speed, air temperature, humidity, and wind direction.

3. The multi-parameter collaborative prediction method for wind turbine clusters according to claim 1, characterized in that, The point estimation objective function is a weighted mean square error function; the probabilistic objective function is a negative log-likelihood function, and the corresponding expression for the negative log-likelihood loss is: ； in, Let be the conditional probability density function of the multidimensional adaptive long short-term memory network under the Gaussian assumption. These are the trainable parameters of the multi-dimensional adaptive long short-term memory network. For the first A channel long short-term memory network in The input sequence within the window, The number of input channels for the multi-channel long short-term memory network. The input channel is numbered. For the current moment Starting from the first The output label of the step, For a moment No. The predicted mean of each output dimension For a moment No. The prediction variance of each output dimension The vector dimension of the predicted physical quantity value. , Output dimension index, Let be the normalization constant term of the Gaussian negative log-likelihood. For a moment The output label of the first One output dimension; The multidimensional adaptive long short-term memory network also incorporates a simplex constraint, expressed as follows: ； in, For the first Channel Long Short-Term Memory Network at the current moment Cross-channel attention weights For the first Channel Long Short-Term Memory Network at time Historical moments Time attention weights .

4. The multi-parameter collaborative prediction method for wind turbine clusters according to claim 1, characterized in that, In step S3, the physical assimilation constraints include observational consistency constraints, dynamic consistency constraints, and prediction consistency constraints, the specific expressions of which are as follows: ； in, This refers to the physical assimilation constraint. The total number of timing steps. For the current moment, For the current moment physical quantity, For the wind turbine at the current moment The hidden state, To map hidden states to observation functions of physical quantities, To observe the inverse of the noise covariance matrix, For the state update function that includes physical priors and learnable residuals, For the wind turbine unit at the previous moment The hidden state, For the wind turbine unit at the previous moment Exogenous input, The inverse of the process noise covariance matrix. For observation function and state update function The parameters, η To predict the consistency weighting coefficient, For the predicted horizon length, For the prediction step index, For the multi-dimensional adaptive long short-term memory network in the first... The predicted values ​​of physical quantities output by the step. The predicted hidden state is obtained by the physical assimilation constraint module through attention rolling extrapolation on the prediction horizon.

5. The multi-parameter collaborative prediction method for wind turbine clusters according to claim 4, characterized in that, In step S4, the specific formula for the joint loss function is as follows: ； in, Let the joint loss function be... These are the trainable parameters of the multi-dimensional adaptive long short-term memory network. The number of input channels for the multi-channel long short-term memory network. The input channel is numbered. The length of the input time window. For the first Each channel is in The input sequence within the window, For the current moment Starting from the first The output label of the step, Let be the conditional probability density function of the multidimensional adaptive long short-term memory network under the Gaussian assumption. For domain alignment distance, For the reference fusion representation set, For the target fusion representation set, These are the weighting coefficients of the physical assimilation constraint. is the weighting coefficient for the domain alignment loss.

6. The multi-parameter collaborative prediction method for wind turbine clusters according to claim 5, characterized in that, The domain alignment distance is obtained by weighted combination of the higher-order distribution difference and second-order statistical difference between the reference fusion representation and the target fusion representation, and the calculation formula is as follows: ； ； ； in, The distance between the reference fusion representation and the target fusion representation is the maximum mean difference. For kernel function, The number of samples in the reference dataset. The number of samples in the target dataset. For the reference dataset of the first The sample number, The first reference dataset The sample number, For the target dataset of the first The sample number, The first of the target dataset The sample number, The second-order statistical difference distance is defined as the distance between the covariance matrix of the reference fusion representation and the covariance matrix of the target fusion representation. Let be the covariance matrix of the reference fusion representation. Let be the covariance matrix of the target fusion representation. The maximum mean difference distance The weighting coefficients, The second-order statistical difference distance The weighting coefficients.

7. The multi-parameter collaborative prediction method for wind turbine clusters according to claim 6, characterized in that, In step S4, the optimization of the collaborative prediction model is specifically implemented using the following formula: ； ； ； in, For the data-side parameters, For the physical side parameters, Updated during optimization process learning rate, Updated during optimization process The learning rate The joint loss function is relative to The estimated gradient, For the physical assimilation constraint relative to The estimated gradient, The input sequence is standardized. For the standardized output labels, For the first Channel Long Short-Term Memory Networks in Historical Moments The input feature vector, The input sequence of the reference dataset is in the th... The mean of the channel, The input sequence of the reference dataset is in the th... Standard deviation of the channel The output labels for the reference dataset are in the 1st... The mean of the channel, The output labels for the reference dataset are in the 1st... Standard deviation of the channel These are the weighting coefficients for different prediction step sizes. .

8. A multi-parameter collaborative prediction system for wind turbine clusters, characterized in that, include: The reference dataset module is used to collect historical operating sequence data of wind turbines in a reference wind farm, including physical quantities and corresponding feature vectors within the total number of time steps. A sliding window mechanism is used to extract continuous time segments from the historical operating sequence data; each time segment includes the preceding... The input time window and the subsequent time step The predicted horizon is calculated at each time step; the feature vectors within the input time window are combined to form the input sequence, and the physical quantities within the predicted horizon are combined to form the output label; the input sequence and the output label constitute a time segment sample; all time segment samples within the total number of time steps are combined to form a reference dataset. The target dataset module is used to obtain the target dataset by repeating S1 based on the target wind farm. A pre-trained model module, used to construct a deep learning model, includes a data-driven module and a physical assimilation constraint module. The data-driven module receives the input sequence of the reference dataset, extracts a reference fusion representation, and outputs the conditional probability distribution and predicted physical quantity values ​​within the predicted horizon. Based on the output labels of the reference dataset and the conditional probability distribution within the predicted horizon, the deep learning model is pre-trained using a preset objective function. The physical assimilation constraint module applies physical assimilation constraints to the predicted physical quantity values ​​to obtain the pre-trained model. The data-driven module is a multi-dimensional adaptive long short-term memory network. The multi-dimensional adaptive long short-term memory network extracts fusion representations through multi-channel long short-term memory network encoding, temporal attention, and cross-channel attention mechanisms. Through heteroscedastic probability decoding, it outputs the conditional probability distribution and predicted physical quantity values ​​within the predicted horizon. The multi-channel long short-term memory network includes multiple parallel long short-term memory networks for processing the feature vectors. The fusion representation represents the dependence of the physical quantities at different time scales and the nonlinear coupling characteristics between the physical quantities. The preset objective function includes a probabilistic objective function and a point estimation objective function. A collaborative prediction model module is used to input the target dataset into the pre-trained model to obtain a target fusion representation; the reference fusion representation and the target fusion representation are aligned in a domain; based on the target dataset, the pre-trained model is fine-tuned through a layer-by-layer unfreezing strategy to minimize the constructed joint loss function, thereby obtaining a collaborative prediction model; wherein, the joint loss function is the sum of the preset objective function loss, the physical assimilation constraint loss, and the domain alignment loss; The real-time prediction module is used to collect real-time operating sequence data of wind turbines in the target wind farm, and input the real-time operating sequence data into the collaborative prediction model to obtain the predicted physical quantities.

9. An electronic device comprising a memory, a processor, and a computer program / instructions stored in the memory, characterized in that, The processor executes the computer program / instructions to implement the multi-parameter collaborative prediction method for wind turbine groups as described in any one of claims 1 to 7.

Images