Fan load prediction method and system based on gaussian process regression and storage medium

Abstract

The present application relates to the technical field of industrial energy saving and intelligent control, in particular to a fan load prediction method and system based on Gaussian process regression and a storage medium, comprising: obtaining fan operation data, constructing a standard data set, the standard data set comprising time sequence characteristics, process characteristics and state characteristics; performing feature extraction on the standard data set to obtain a multi-source feature vector; performing feature fusion on the multi-source feature vector to obtain a training data set; constructing a Gaussian process regression prediction model and training the model using the training data set; and using the trained Gaussian process regression prediction model to predict the fan load. Through the above technical solution, the present application embodiment constructs a Gaussian process regression prediction model to adapt to the complex characteristics of the multi-factor coupling of the fan load, capture the dynamic correlation between the fan load and the multi-coupling factors, and achieve accurate prediction of the process air conditioning fan load.

Description

Technical Field

[0001] This invention relates to the field of industrial energy conservation and intelligent control technology, specifically to a wind turbine load prediction method, system, and storage medium based on Gaussian process regression. Background Technology

[0002] Process air conditioning is a key auxiliary system in modern industrial production, playing a vital role in maintaining the stability of the production environment, ensuring product processing quality, and optimizing energy consumption. In cigarette factories, the process air conditioning system is responsible for precisely controlling the temperature and humidity environment of production workshops, storage areas, and other critical areas to ensure that cigarettes and finished products are stored and processed under optimal conditions, thus ensuring their quality stability. As the core actuator, the load changes of the fans directly determine the accuracy of environmental control and energy utilization efficiency. Accurate fan load prediction is a crucial prerequisite for achieving intelligent operation and maintenance of process air conditioning and reducing energy consumption.

[0003] In cigarette factories, existing process air conditioning fan load prediction technologies are ill-suited to the multi-factor coupling and long-term time-series dependence of loads. Individual Gaussian process regression models have limitations in time-series modeling, and single kernel functions cannot accurately capture the nonlinearity, intra-day periodicity, and local fluctuations driven by outdoor temperature and humidity of the load, resulting in low prediction accuracy. At the same time, traditional prediction models often only output a single predicted value, lacking the quantification of prediction uncertainty, making it difficult to support the reliability requirements of operation and maintenance decisions in industrial scenarios. Furthermore, some models introduce redundant features, increasing computational costs and reducing industrial adaptability.

[0004] Therefore, there is an urgent need for a fan load prediction method and system based on Gaussian process regression to achieve accurate prediction of process air conditioning fan load. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting fan load based on Gaussian process regression, in order to solve the problems of multi-factor coupling and long-term time-series dependence characteristics that are difficult to adapt to the load in the existing technology, and to achieve accurate prediction of process air conditioning fan load.

[0006] To achieve the above objectives, embodiments of the present invention provide a wind turbine load prediction method based on Gaussian process regression, comprising: Acquire wind turbine operation data and construct a standard dataset, which includes time-series features, process features, and status features; Feature extraction is performed on the standard dataset to obtain multi-source feature vectors; The multi-source feature vectors are fused to obtain a training dataset; A Gaussian process regression prediction model was constructed and trained using the aforementioned training dataset; The wind turbine load is predicted using a trained Gaussian process regression prediction model.

[0007] Optionally: Feature extraction on the standard dataset to obtain multi-source feature vectors includes: Based on the aforementioned temporal features, an LSTM feature extraction network is used to extract temporal features in order to obtain temporal feature vectors and temporal probability prior parameters. The process features are normalized to obtain a process feature vector; Discrete feature encoding is performed on the state features to obtain a state feature vector.

[0008] Optionally: Based on the aforementioned temporal features, an LSTM feature extraction network is used to extract temporal features to obtain a temporal feature vector, including: The temporal features are input into a preset attention enhancement module, and the attention weights of the temporal features are calculated by the attention enhancement module. The temporal features are updated by weighting based on the attention weights to obtain enhanced temporal features; The enhanced temporal features are input into a bidirectional LSTM network to extract the forward hidden state sequence and the backward hidden state sequence of the enhanced temporal features. The forward hidden state sequence and the backward hidden state sequence are then concatenated to obtain the final hidden state sequence. Extract the hidden state of the last time step of the time window of the final hidden state sequence to obtain the temporal feature vector; The final hidden state sequence is input into the fully connected layer to obtain the temporal probability prior parameters.

[0009] Optionally: Feature fusion is performed on the multi-source feature vectors to obtain a training dataset, including: The fusion features are obtained according to formula (1). (1) in, As a feature of fusion, For the weights of the gating network, For the final time series features, As a technological feature, As state characteristics, For the bias of the gated network; Based on the fusion features, a feature cross-network is used, and after L layers of cross-interaction, enhanced features are obtained; A training dataset is constructed based on the temporal probability prior parameters of the enhanced features and the multi-source feature vectors.

[0010] Optionally: Construct a Gaussian process regression prediction model and train it using the aforementioned training dataset: Select multiple kernel functions to construct a hybrid kernel function; A Gaussian process regression prediction model is constructed based on the aforementioned hybrid kernel function; Input the training dataset into the Gaussian process regression prediction model to obtain the predicted wind turbine load value and confidence interval; Based on the predicted and actual values ​​of the wind turbine load, the performance evaluation index of the Gaussian process regression prediction model is obtained. The parameters of the Gaussian process regression prediction model are optimized based on the performance evaluation index.

[0011] Optionally: Select multiple kernel functions to construct a hybrid kernel function, including: Construct the RBF kernel function; Construct a periodic kernel function; The kernel function of the heat load mechanism is constructed using formula (2). (2) in, For the heat load mechanism kernel function, The signal variance of the kernel function of the heat load mechanism. Due to the temperature difference between indoors and outdoors, For the indoor-outdoor temperature difference under another eigenvector, For process capacity heat load, The process capacity heat load under another eigenvector; Construct a perturbation-sensitive local kernel function; Adaptive weighted fusion processing is performed on the RBF kernel function, periodic kernel function, heat load mechanism kernel function, and disturbance-sensitive local kernel function to obtain a hybrid kernel function.

[0012] Optionally: Inputting the training dataset into the Gaussian process regression prediction model to obtain the predicted wind turbine load value and confidence interval includes: Based on the training dataset, the Gaussian process regression prediction model is used to obtain the prediction mean and prediction standard deviation; The predicted mean is inversely normalized to obtain the predicted wind turbine load value; Based on the predicted wind turbine load value and the predicted standard deviation, the confidence interval of the wind turbine load is obtained.

[0013] Optionally: Based on the training dataset, using the Gaussian process regression prediction model, obtaining the predicted mean and predicted standard deviation includes: The covariance matrix is ​​constructed using formula (3). (3) in, Let covariance matrix be the variance matrix. It is a mixed kernel function matrix. To enhance features, For dynamic noise diagonal matrix, This is prior noise; Using the Gaussian process regression prediction model, the predicted mean is obtained according to formula (4). (4) in, To predict the mean, It is the prior mean function. For the new enhanced features, This represents the actual load value of the wind turbine.

[0014] On the other hand, the present invention also provides a wind turbine load prediction system based on Gaussian process regression, the system comprising: The data acquisition and preprocessing module is used to acquire wind turbine operating data and construct a standard dataset; The feature extraction module is used to extract features from the standard dataset to obtain multi-source feature vectors; The feature fusion module is used to fuse the multi-source feature vectors to obtain a training dataset. The prediction model building module is used to build a Gaussian process regression prediction model and train it using the training dataset. The wind turbine load prediction module is used to predict the wind turbine load using a trained Gaussian process regression prediction model. The processor is connected to the data acquisition and preprocessing module, the feature extraction module, the feature fusion module, the prediction model construction module, and the wind turbine load prediction module, and the processor is configured to perform any of the methods described above.

[0015] In another aspect, the present invention provides a machine-readable storage medium storing instructions that cause a machine to perform any of the methods described above in this application.

[0016] The beneficial effects of this invention are: The embodiments of this invention construct a Gaussian process regression prediction model, which effectively adapts to the complex characteristics of fan load under the coupling effect of multiple factors, accurately captures the dynamic correlation between fan load and factors such as hourly humidity and temperature, external environment, and production shifts, thereby effectively improving the prediction accuracy of process air conditioning fan load and achieving accurate prediction of fan load.

[0017] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a wind turbine load prediction method based on Gaussian process regression according to an embodiment of the present invention; Figure 2 This is a flowchart of a specific method for extracting features from a standard dataset to obtain multi-source feature vectors according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating a specific method for constructing a Gaussian process regression prediction model and training it using the training dataset, according to an embodiment of the present invention. Figure 4 This is a structural block diagram of a wind turbine load prediction system based on Gaussian process regression according to an embodiment of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0020] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0021] like Figure 1 The diagram shows a flowchart of a wind turbine load prediction method based on Gaussian process regression according to an embodiment of the present invention. Figure 1 In this context, the prediction method may include the following steps: In step S10, wind turbine operating data is acquired, and a standard dataset is constructed; In step S11, feature extraction is performed on the standard dataset to obtain multi-source feature vectors; In step S12, feature fusion is performed on the multi-source feature vectors to obtain the training dataset; In step S13, a Gaussian process regression prediction model is constructed and trained using the training dataset; In step S14, the trained Gaussian process regression prediction model is used to predict the wind turbine load.

[0022] In such Figure 1 In the wind turbine load prediction method based on Gaussian process regression shown, step S10 is used to acquire wind turbine operating data and construct a standard dataset. In this embodiment, the standard dataset may include time-series features, process features, and state features.

[0023] In this embodiment, the specific method for acquiring wind turbine operating data and constructing a standard dataset in step S10 can be of various forms known to those skilled in the art. In this embodiment, the input feature set for wind turbine load prediction is constructed. ,in, For time-series feature datasets, For process feature datasets, Given a state feature dataset, the prediction target is... .

[0024] Furthermore, specific methods for constructing a time-series feature dataset can be: .in, The time window length, This represents a single-time-step temporal feature vector. In this embodiment, It can contain multiple variables, such as defined as In this defined formula, For the historical load of the wind turbine, The air supply temperature of the fan. The return air temperature of the fan is affected by the cooling capacity of the reaction system and thermal buoyancy. The ambient temperature is used to reflect the current actual thermal state of the controlled area. The main external heat load drivers and the number of operating equipment in the workshop are considered. These characteristic variables reflect the short-term time-series dependence of the load. (Process characteristic dataset) The construction can include variables strongly correlated with production, such as workshop capacity and production scheduling, to reflect the linear and nonlinear effects of process fluctuations on load. State feature dataset The construction can include discrete value variables such as fan start / stop markers and production shift markers, to reflect the sudden impact of the equipment's own state on the load.

[0025] Step S11 is used to extract features from the standard dataset to obtain multi-source feature vectors. In this embodiment, the specific method for extracting features from the standard dataset to obtain multi-source feature vectors in step S11 can be of various forms known to those skilled in the art. In this embodiment, step S11 can be as follows: Figure 2 The steps are shown. In this Figure 2 In this context, step S11 may include: In step S20, based on the temporal features, an LSTM feature extraction network is used to extract the temporal features in order to obtain the temporal feature vector and the temporal probability prior parameters. In step S21, the process features are normalized to obtain the process feature vector; In step S22, discrete feature encoding is performed on the state features to obtain the state feature vector.

[0026] In such Figure 2 In the method shown, step S20 is used to extract temporal features based on temporal features using an LSTM feature extraction network to obtain temporal feature vectors and prior temporal probability parameters. In this embodiment, firstly, the temporal features are preprocessed, which includes at least outlier handling, normalization, and the construction and enhancement of lag features.

[0027] The specific methods for handling outliers can take various forms known to those skilled in the art. In this embodiment, the Laida criterion is used to determine whether a feature value in the dataset is an outlier. If it is determined to be an outlier, it is replaced with the neighborhood mean. Subsequently, to eliminate differences in feature dimensions, a linear transformation is performed on the time-series feature dataset, mapping it to the [0,1] interval to ensure that the scales of each feature are consistent. The normalized time-series features can be denoted as... .

[0028] Furthermore, a multi-scale lag feature is constructed for each variable in the time-series features, expanding the 5-dimensional features to cover short-term, medium-term, and long-term dependencies. This lag feature can be defined as: .

[0029] By constructing and enhancing lag features on time-series feature datasets, historical information at multiple time scales is introduced to help LSTM networks better capture the impact of different lag times. This includes constructing short-term lags to capture the instantaneous effects of equipment start-up and shutdown and control response; constructing medium-term lags to capture the lag effects of daily production fluctuations; and constructing long-term lags to capture daily cycle similar patterns.

[0030] Secondly, for the three types of time-series information including short-term, medium-term, and long-term lagging features, a scale-channel attention unit is added to the network input layer according to the following formula to enhance the feature contribution of the dominant dependency under the current working condition: ( [ ; ; ] + ); in, The scale attention weight vector is consistent with the input dimension, and its value range is [0,1]. This is the attention layer weight matrix. For the corresponding bias term, This is the Sigmoid activation function.

[0031] Enhanced temporal features are obtained by weighting the lagged features using attention weights according to the following formula. : .

[0032] Furthermore, the enhanced temporal features are input into the bidirectional LSTM network, and a condition gating unit is embedded in the cell state update link. The cell state update rate is dynamically adjusted using discrete state features such as fan start / stop status and production shift switching as trigger signals to adapt to load step change scenarios.

[0033] Extract the forward hidden state sequence using the following formula: ; in, For the first The hidden state of the layer-forward LSTM at time t. This is the hidden state from the previous moment. This represents the cell state at the previous moment.

[0034] Extract the backward hidden state sequence using the following formula: ; in, For the first The hidden state of the layer backward LSTM at time t. To hide the state in the next moment. This represents the cell state at the next moment.

[0035] The forward hidden state sequence and the backward hidden state sequence are concatenated using the following formula to obtain the final hidden state: .

[0036] Finally, the final hidden state of the last time step of the time window is extracted as the temporal feature vector. This is used to characterize the temporal evolution of the load; the final hidden state is input into two independent fully connected layers, and the preliminary predicted mean and temporal fluctuation amplitude of the load are obtained by mapping according to the following formula, which are used to set the prior distribution of the subsequent Gaussian process regression prediction model: ; in, This is the preliminary forecast average of the load. The weights of the mean-mapping layer, For the bias of the mean mapping layer, This represents the amplitude of time-series fluctuations. The weights of the fluctuation mapping layer, This is the bias for the fluctuation mapping layer. In this embodiment, To output the trainable parameters of the fully connected layer, the following method is used: The activation function ensures that the fluctuation amplitude is non-negative.

[0037] Step S21 is used to normalize the process features to obtain the process feature vector. In this embodiment, since the process features are continuous features, outlier handling and normalization are performed on the process feature dataset. First, the mean and standard deviation within the sliding window are calculated. The Raida criterion is used to determine whether the feature values ​​in the dataset are outliers. If an outlier is identified, it is replaced with the following formula: ; in, This is an outlier. The size of the neighborhood window ensures that outlier replacements conform to temporal continuity.

[0038] Subsequently, in order to eliminate the differences in feature dimensions, the process features after outlier processing are linearly transformed and mapped to the [0,1] interval according to the following formula to ensure that the scales of each feature are consistent: ; in, For feature vectors, For eigenvalues, The minimum value of the feature. The maximum value of the feature is denoted as . The normalized process feature dataset is defined as the process feature vector. .

[0039] Step S22 is used to perform discrete feature encoding on the state features to obtain a state feature vector. In this embodiment, binary vectors of corresponding lengths are generated for the discrete variables in the state features. For example, if there are morning, afternoon, and night shifts in the production shifts, then the afternoon shift corresponds to [0,1,0], and the fan status is [1,0] when it is running, ensuring that the discrete features can participate in the kernel function calculation. The preprocessed state feature vector is defined as follows: .

[0040] Step S12 is used to perform feature fusion on the multi-source feature vectors to obtain the training dataset. In this implementation, the initial predicted mean of the load and the amplitude of the time-series fluctuation are first directly used as the prior mean function and prior noise basis of the Gaussian process regression prediction model using the following formula, so that the Gaussian process regression prediction model only performs nonlinear correction and uncertainty calibration on the LSTM prediction residuals: ; in, Let be the prior mean function of the Gaussian process regression prediction model. Let be the prior mean function and the prior noise basis.

[0041] Subsequently, the time-series feature vector, process feature vector, and state feature vector are fused according to the following formula to construct the complete model input features: ; in, As a feature of fusion, For the weights of the gating network, For the final time series features, As a technological feature, As state characteristics, This is the bias of the gated network.

[0042] Based on this fusion feature, a feature cross-network is introduced to capture higher-order interaction effects. After L layers of cross-networking, enhanced features are obtained. .

[0043] Step S13 is used to construct a Gaussian process regression prediction model and train it using the training dataset. In this embodiment, the specific method for constructing the Gaussian process regression prediction model and training it using the training dataset in step S13 can be of various forms known to those skilled in the art, such as using a single kernel function in a traditional Gaussian process regression model. In this embodiment, step S13 can be as follows: Figure 3 The steps are shown. In this Figure 3 In this context, step S13 may include: In step S30, multiple kernel functions are selected to construct a hybrid kernel function; In step S31, a Gaussian process regression prediction model is constructed based on the hybrid kernel function; In step S32, the training dataset is input into the Gaussian process regression prediction model to obtain the predicted wind turbine load value and confidence interval. In step S33, the performance evaluation index of the Gaussian process regression prediction model is obtained based on the predicted and actual values ​​of the wind turbine load. In step S34, the parameters of the Gaussian process regression prediction model are optimized based on the performance evaluation index.

[0044] In such Figure 3In the method shown, step S30 is used to select multiple kernel functions to construct a hybrid kernel function. In this embodiment, since a single kernel function of the traditional Gaussian process regression model is difficult to adapt to the multi-factor coupling characteristics of wind turbine load, the RBF kernel, periodic kernel, and heat load mechanism kernel are selected, and a local perturbation kernel is added. Through weighted fusion, a hybrid kernel function is finally constructed. Specifically, the RBF kernel function is constructed according to the following formula to capture the global nonlinearity brought about by the fusion of multivariate features: ; in, For RBF kernel function, The signal variance of the RBF kernel function. and for The feature vectors of any two different samples in the vector set. is the length scale of the RBF kernel function.

[0045] Furthermore, a periodic kernel function is constructed based on the following formula to capture the periodic fluctuations caused by production shifts: ; in, It is a periodic kernel function. For signal variance, For a period of time, is the length scale of the periodic kernel function.

[0046] Furthermore, a heat load mechanism kernel function is constructed based on formula (2) to capture the correlation between process air conditioning heat load and fan load: (2) in, For the heat load mechanism kernel function, The signal variance of the kernel function of the heat load mechanism. Due to the temperature difference between indoors and outdoors, For the indoor-outdoor temperature difference under another eigenvector, For process capacity heat load, This represents the process capacity heat load under another eigenvector.

[0047] Furthermore, a perturbation-sensitive local kernel function is constructed according to the following formula. In this embodiment, the Matern3 / 2 kernel is selected to capture the perturbation effect of sudden local fluctuations such as outdoor temperature and equipment status on the model: ; in, For perturbation-sensitive local kernel functions, To perturb the signal variance of the sensitive local kernel function, is the length scale of the perturbation-sensitive local kernel function.

[0048] Finally, based on the above kernel functions, learnable soft weights are used to dynamically combine the kernel functions, automatically allocating the weights to the impact of various features on the load. Specifically, the hybrid kernel function is obtained according to the following formula: ; in, For hybrid kernel functions, Let be the weight coefficient of the i-th kernel. For learnable parameters, For temperature coefficient, Let i be the i-th kernel function.

[0049] Step S31 is used to construct a Gaussian process regression prediction model based on the hybrid kernel function. In this embodiment, the model is based on the constructed hybrid kernel function and enhanced features. We construct a Gaussian process regression prediction model with dynamic priors by using the prior distribution parameters passed by LSTM.

[0050] Step S32 is used to input the training dataset into the Gaussian process regression prediction model to obtain the predicted wind turbine load value and confidence interval. In this embodiment, the covariance matrix is ​​first constructed using formula (3): (3) in, Let covariance matrix be the variance matrix. The matrix obtained by calculating the mixed kernel function. For prior noise, It is a dynamic noise diagonal matrix, and the noise level of each sample point changes in real time with the temporal fluctuation amplitude of the LSTM output.

[0051] Subsequently, for the new input feature vector Using a Gaussian process regression prediction model, the predicted mean and predicted standard deviation are obtained according to the following formulas: ; in, The predicted value of the wind turbine load is to be predicted. This represents the normalized true value of the wind turbine load. To predict the mean, This represents the predicted standard deviation.

[0052] The predicted mean is obtained according to formula (4): (4) in, It is the prior mean function.

[0053] Furthermore, since the output forecast mean is the normalized load forecast value, it can be restored to the actual physical value of the wind turbine load forecast using the following inverse normalization formula: ; in, The actual physical value of the predicted wind turbine load. This represents the historical maximum load value. This represents the historical minimum load value.

[0054] Furthermore, based on the actual physical values ​​and standard deviation of the predicted wind turbine load, a 95% confidence interval is constructed using the following formula to reflect the reliability of the prediction results: ; in, This represents the confidence interval.

[0055] By constructing a confidence interval, 95% of the actual load will fall within this interval, and during sudden fluctuations... The increase in confidence interval indicates a higher level of uncertainty in the prediction at that moment, and maintenance personnel should combine this with their own experience.

[0056] Step S33 is used to obtain performance evaluation metrics for the Gaussian process regression prediction model based on the predicted and actual wind turbine load values. In this implementation, a comprehensive model performance evaluation system is established, providing quantitative basis for model updates, version management, and deployment decisions by obtaining prediction accuracy metrics, uncertainty quality metrics, and computational efficiency metrics. Specifically, the prediction accuracy metrics include mean absolute percentage. and coefficient of determination The load is calculated using the predicted and actual values ​​of the wind turbine load; the uncertainty quality index is obtained through the KL divergence of the data distribution; and the computational performance index is obtained through the memory ratio.

[0057] To address time-varying factors such as changing operating conditions and equipment aging, the Gaussian process regression prediction model parameters are optimized based on the performance evaluation index in step S34. In this implementation, an online adaptive update mechanism is established, triggering an update when any of the following conditions occur, according to the aforementioned model evaluation index: Prediction error continues to exceed the limit: ,in, This is the error threshold; The coefficient of determination decreased significantly: ,in, for Decrease threshold; Significant data distribution shift: ,in, KL divergence measures the distribution of new data. Distribution of old data Differences This is the distribution drift threshold; Excessive memory usage: Memory consumption ,in, This is a limit on memory usage.

[0058] Furthermore, an incremental learning strategy is adopted to update the model. Specifically, the incremental update method for the LSTM part can be as follows: adopt an optimized elastic weight solidification strategy and perform incremental training in layers; fine-tune the feature extraction layer to initially adapt to the new data distribution; update the core gating parameters to achieve limited adjustment under strong constraints; and finally perform overall lightweight optimization to find the optimal balance point.

[0059] The specific method for incremental updates to the Gaussian process regression prediction model can be as follows: A gradient-based mini-batch stochastic optimization method is used to fine-tune hyperparameters such as signal variance and length scale of each sub-kernel in the mixture kernel function. Simultaneously, to address the computational cubic complexity issue caused by the continuous growth of data volume, the system employs a variational sparse Gaussian process framework to dynamically maintain a set of representative induced points. When new data arrives, the positions of the induced points and their variational distribution parameters are optimized to enable the sparse model to approximate a complete posterior distribution. This reduces the computational complexity of prediction and updates to a linear relationship with the number of induced points while maintaining accuracy.

[0060] Step S14 is used to predict the wind turbine load using the trained Gaussian process regression prediction model. In this embodiment, the data to be predicted is input into the trained Gaussian process regression prediction model to obtain the predicted value and confidence interval of the wind turbine load, thereby achieving accurate prediction and uncertainty quantification of the wind turbine load.

[0061] On the other hand, embodiments of the present invention also provide a wind turbine load prediction system based on Gaussian process regression. The structural block diagram of this system is shown below. Figure 4 As shown. In this Figure 4 The system comprises a data acquisition and preprocessing module, a feature extraction module, a feature fusion module, a prediction model building module, and a wind turbine load prediction module. The data acquisition and preprocessing module acquires wind turbine operating data and constructs a standard dataset. The feature extraction module extracts features from this standard dataset to obtain multi-source feature vectors. The feature fusion module fuses these multi-source feature vectors to obtain a training dataset. The prediction model building module constructs a Gaussian process regression prediction model and trains it using the training dataset. The wind turbine load prediction module uses the trained Gaussian process regression prediction model to predict wind turbine load.

[0062] In this embodiment, the wind turbine load prediction algorithm is encapsulated into a dedicated industrial edge computing device for deployment, achieving deep integration of the algorithm and hardware. Specifically, the edge computing device uses a dual-port industrial Ethernet communication interface. One port connects to the lower-level PLC / DCS system to collect real-time data such as temperature, while the other connects to the system ring network to obtain data such as production capacity and number of operating devices from systems like MES / EMS. The two network interfaces are physically isolated, supporting virtual LAN partitioning and different network policy configurations to ensure the security of the control system network. Two RS-485 ports are also configured for connecting to third-party devices. Furthermore, the device has a built-in data mapping and protocol conversion module, which can be configured to parse data points from devices from different manufacturers.

[0063] The software architecture adopts a microservice containerized architecture, with all services running in Docker containers and orchestrated and managed through Kubernetes. Resources are dynamically allocated based on CPU load, and task priority scheduling is determined. It also supports rolling upgrades and automatic fault recovery.

[0064] In another aspect, the present invention provides a machine-readable storage medium storing instructions that cause a machine to perform any of the methods described above in this application.

[0065] The beneficial effects of this invention are: This invention constructs a Gaussian process regression prediction model and selects multiple hybrid kernel functions. In particular, based on the correlation between process air conditioning heat load and fan load, a heat load mechanism kernel is designed. This breaks through the limitation of the traditional Gaussian process regression model, where a single kernel function is difficult to adapt to the multi-factor coupling characteristics of fan load. At the same time, it combines a Long Short-Term Memory (LSTM) network to extract dependency features from time-series data sequences, accurately capturing the dynamic correlation between fan load and factors such as hourly temperature and humidity, external environment, and production shifts. This improves the prediction accuracy of fan load and achieves accurate prediction of fan load.

[0066] The embodiments of the present invention output load prediction values ​​and confidence intervals through a Gaussian process regression prediction model to quantify the uncertainty of prediction. This breaks the limitation of existing traditional prediction technologies that only output a single prediction value, and solves the problem that existing technologies cannot provide reliable uncertainty basis for operation and maintenance decisions, making operation and maintenance decisions more scientific and reasonable.

[0067] This invention, through the construction of a result evaluation and dynamic optimization mechanism, can adjust parameters based on operating conditions and data accumulation, while also flexibly adapting to the temperature and humidity process requirements of different scenarios. By employing core evaluation indicators of regression tasks, the model's prediction results are comprehensively evaluated. Simultaneously, by combining actual industrial operation and maintenance needs, the feasibility of model implementation is assessed, and issues such as insufficient accuracy and unreasonable parameters are identified in a timely manner, ensuring the long-term stable operation of the model.

[0068] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0070] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0071] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0072] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0073] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0074] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0075] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0076] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A wind turbine load prediction method based on Gaussian process regression, characterized in that, The method includes: Acquire wind turbine operation data and construct a standard dataset, which includes time-series features, process features, and status features; Feature extraction is performed on the standard dataset to obtain multi-source feature vectors; The multi-source feature vectors are fused to obtain a training dataset; A Gaussian process regression prediction model was constructed and trained using the aforementioned training dataset; The wind turbine load is predicted using a trained Gaussian process regression prediction model.

2. The method according to claim 1, characterized in that, Feature extraction from the standard dataset to obtain multi-source feature vectors includes: Based on the aforementioned temporal features, an LSTM feature extraction network is used to extract temporal features in order to obtain temporal feature vectors and temporal probability prior parameters. The process features are normalized to obtain a process feature vector; Discrete feature encoding is performed on the state features to obtain a state feature vector.

3. The method according to claim 2, characterized in that, Based on the aforementioned temporal features, an LSTM feature extraction network is used to extract temporal features to obtain a temporal feature vector, including: The temporal features are input into a preset attention enhancement module, and the attention weights of the temporal features are calculated by the attention enhancement module. The temporal features are updated by weighting based on the attention weights to obtain enhanced temporal features; The enhanced temporal features are input into a bidirectional LSTM network to extract the forward hidden state sequence and the backward hidden state sequence of the enhanced temporal features. The forward hidden state sequence and the backward hidden state sequence are then concatenated to obtain the final hidden state sequence. Extract the hidden state of the last time step of the time window of the final hidden state sequence to obtain the temporal feature vector; The final hidden state sequence is input into the fully connected layer to obtain the temporal probability prior parameters.

4. The method according to claim 1, characterized in that, The multi-source feature vectors are fused to obtain a training dataset, including: The fusion features are obtained according to formula (1). ，（1） in, As a feature of fusion, For the weights of the gating network, For the final time series features, As a technological feature, As state characteristics, For the bias of the gated network; Based on the fusion features, a feature cross-network is used, and after L layers of cross-interaction, enhanced features are obtained; A training dataset is constructed based on the temporal probability prior parameters of the enhanced features and the multi-source feature vectors.

5. The method according to claim 1, characterized in that, Construct a Gaussian process regression prediction model and train it using the aforementioned training dataset: Select multiple kernel functions to construct a hybrid kernel function; A Gaussian process regression prediction model is constructed based on the aforementioned hybrid kernel function; The training dataset is input into the Gaussian process regression prediction model to obtain the predicted wind turbine load value and confidence interval. Based on the predicted and actual values ​​of the wind turbine load, the performance evaluation index of the Gaussian process regression prediction model is obtained. The parameters of the Gaussian process regression prediction model are optimized based on the performance evaluation index.

6. The method according to claim 5, characterized in that, Multiple kernel functions are selected to construct hybrid kernel functions, including: Construct the RBF kernel function; Construct a periodic kernel function; The kernel function of the heat load mechanism is constructed using formula (2). ，（2） in, For the heat load mechanism kernel function, The signal variance of the kernel function of the heat load mechanism. Due to the temperature difference between indoors and outdoors, For the indoor-outdoor temperature difference under another eigenvector, For process capacity heat load, For the process capacity heat load under another eigenvector; Construct a perturbation-sensitive local kernel function; An adaptive weighted fusion process is performed on the RBF kernel function, the periodic kernel function, the heat load mechanism kernel function, and the disturbance-sensitive local kernel function to obtain a hybrid kernel function.

7. The method according to claim 5, characterized in that, Inputting the training dataset into the Gaussian process regression prediction model to obtain the predicted wind turbine load values ​​and confidence intervals includes: Based on the training dataset, the Gaussian process regression prediction model is used to obtain the prediction mean and prediction standard deviation; The predicted mean is inversely normalized to obtain the predicted wind turbine load value; Based on the predicted wind turbine load value and the predicted standard deviation, the confidence interval of the wind turbine load is obtained.

8. The method according to claim 7, characterized in that, Based on the training dataset, the Gaussian process regression prediction model is used to obtain the predicted mean and predicted standard deviation, including: The covariance matrix is ​​constructed using formula (3). ，（3） in, Let covariance matrix be the variance matrix. It is a mixed kernel function matrix. To enhance features, For dynamic noise diagonal matrix, This is prior noise; Using the Gaussian process regression prediction model, the predicted mean is obtained according to formula (4). ，（4） in, To predict the mean, It is the prior mean function. For the new enhanced features, This represents the actual load value of the wind turbine.

9. A wind turbine load prediction system based on Gaussian process regression, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire wind turbine operating data and construct a standard dataset; The feature extraction module is used to extract features from the standard dataset to obtain multi-source feature vectors; The feature fusion module is used to fuse the multi-source feature vectors to obtain a training dataset. The prediction model building module is used to build a Gaussian process regression prediction model and train it using the training dataset. The wind turbine load prediction module is used to predict the wind turbine load using a trained Gaussian process regression prediction model. The processor is connected to the data acquisition and preprocessing module, the feature extraction module, the feature fusion module, the prediction model construction module, and the wind turbine load prediction module, and the processor is configured to perform the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 8.

Images