Wind power plant intelligent prediction and control method based on lightweight time sequence hybrid model
By combining lightweight time-series hybrid models (TTMs) with the physical laws of wind turbines, a smart prediction and control method for wind farms has been developed, which solves the problem that traditional models cannot be deployed on edge controllers and achieves efficient and accurate wind farm prediction and fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- RUICHENG NINGSHENG NEW ENERGY CO LTD
- Filing Date
- 2025-11-19
- Publication Date
- 2026-04-24
AI Technical Summary
Traditional LSTM/Transformer wind farm prediction models require hundreds of megabytes of memory, making them unsuitable for deployment on wind turbine edge controllers. This results in poor real-time performance and large prediction errors under conditions exceeding rated wind speeds.
A lightweight temporal hybrid model (TTMs) is adopted, and high-frequency and low-frequency features are extracted through adaptive block layer. The wind turbine power physical equation is combined as the loss function to output the wind turbine state probability value and power prediction value, which are then deployed in the edge controller.
The wind farm prediction model was deployed on the edge controller, which reduced prediction errors, improved real-time performance and prediction accuracy, and supported fault diagnosis and control strategies.
Smart Images
Figure CN121923078A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent prediction technology for wind farms, and in particular to an intelligent prediction and control method for wind farms based on a lightweight time-series hybrid model. Background Technology
[0002] Intelligent wind farm forecasting refers to the use of advanced algorithms (such as neural networks and support vector machines) to model and predict key meteorological parameters such as wind speed and wind direction, thereby predicting the wind energy supply situation in advance and providing data support for wind farm operation and scheduling.
[0003] Traditional LSTM / Transformer wind farm prediction models require hundreds of megabytes of memory, making them unsuitable for deployment on wind turbine edge controllers. Furthermore, they ignore resource constraints, resulting in poor real-time performance and large prediction errors under conditions exceeding rated wind speeds. Summary of the Invention
[0004] This invention provides a wind farm intelligent prediction and control method based on a lightweight time-series hybrid model, which solves the shortcomings of existing wind farm intelligent prediction models that cannot be deployed on edge controllers and have large prediction accuracy errors.
[0005] In a first aspect, the present invention provides a method for intelligent prediction and control of wind farms based on a lightweight time-series hybrid model, comprising:
[0006] Real-time acquisition of multi-resolution wind turbine operation data;
[0007] The multi-resolution wind turbine operating data is input into the adaptive block layer of the TTMs wind farm prediction model to perform high-frequency feature extraction and low-frequency feature extraction to obtain the resolution prefix;
[0008] The resolution prefix is encoded into an embedding vector and concatenated to the input layer of the TTMs wind farm prediction model;
[0009] The spliced data from the input layer is output to the backbone network, and combined with the wind turbine power physical equation as a physical constraint loss, wind farm prediction is performed, and the wind turbine state probability value and power prediction value are output.
[0010] Based on the wind turbine state probability value and power prediction value, the wind turbine operating parameters are adjusted.
[0011] According to the present invention, a wind farm intelligent prediction and control method based on a lightweight time-series hybrid model is provided, wherein the high-frequency feature extraction and low-frequency feature extraction are performed to obtain a resolution prefix, including:
[0012] Set the high-frequency block length for high-frequency wind speed and direction data, and set the low-frequency block length for low-frequency temperature and humidity data;
[0013] Feature extraction is performed on the multi-resolution wind turbine operation data based on the high-frequency block length and the low-frequency block length. The receptive field is automatically adjusted by a deformable convolutional layer to control the high-frequency block to capture transient fluctuations and the low-frequency block to extract long-term trends.
[0014] According to the present invention, a wind farm intelligent prediction and control method based on a lightweight time-series hybrid model is provided, wherein the wind farm prediction is performed by combining the wind turbine power physical equation as a physical constraint loss, including:
[0015] Power coefficient is dynamically calculated based on real-time wind speed;
[0016] Using the power coefficient, real-time wind speed, swept area, and air density, the theoretical power is calculated;
[0017] A physical constraint term is added to the cross-entropy loss as a loss function, wherein the physical constraint term is determined by the optimal weight coefficient, the theoretical power, and the predicted power.
[0018] According to the present invention, a smart prediction and control method for wind farms based on a lightweight time-series hybrid model is provided, wherein the loss function is:
[0019] ;
[0020] In the formula, Represents the probability value of the true state. This represents the predicted state probability value. This represents the optimal weighting coefficient. Indicates predicted power. Indicates theoretical power. Indicates air density, Indicates the swept area. Indicates the power factor. This indicates the real-time wind speed.
[0021] According to the present invention, a wind farm intelligent prediction and control method based on a lightweight time-series hybrid model is provided, wherein the wind turbine state probability values include: normal state probability value, warning state probability value, and dangerous state probability value.
[0022] The step of adjusting the operating parameters of the wind turbine based on the wind turbine state probability value includes:
[0023] When the normal state probability value is greater than or equal to the first preset threshold, the maximum power point tracking algorithm is activated to dynamically adjust the pitch angle to optimize the power generation efficiency.
[0024] When the probability value of the warning state is greater than or equal to the second preset threshold, the power output is limited to a preset percentage of the rated value, and the high-frequency vibration monitoring module is activated at the same time.
[0025] When the probability of the dangerous state is greater than or equal to the third preset threshold, an emergency shutdown procedure is triggered to disconnect the power grid connection and initiate mechanical braking.
[0026] The intelligent prediction and control method for wind farms based on a lightweight time-series hybrid model provided by the present invention further includes:
[0027] When entering a warning or danger state, the original data collected during the triggering process is automatically saved.
[0028] The fault characteristics are identified by analyzing the original acquired data through vibration spectrum. When the fault characteristic is a 1kHz sideband, it is determined to be a gearbox malfunction. When the fault characteristic is a 5-10kHz resonance peak, it is determined to be a bearing failure.
[0029] Based on the fault characteristics, a fault diagnosis report is generated and pushed to the cloud platform, which includes a fault location map and maintenance priority suggestions.
[0030] According to the present invention, a wind farm intelligent prediction and control method based on a lightweight temporal hybrid model is provided. The TTMs wind farm prediction model is trained using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 1024. After each round of training, the hybrid loss value is evaluated on the validation set. When the loss decreases by less than 1e-5 for three consecutive rounds, the learning rate decay is initiated.
[0031] The intelligent prediction and control method for wind farms based on a lightweight time-series hybrid model provided by the present invention further includes:
[0032] The TTMs wind farm prediction model is converted into INT8 format using dynamic quantization technology to quantize the linear layer parameters;
[0033] The quantized model is exported to ONNX format, compiled into a dedicated om model, configured with AI Core computing units to handle convolution operations, and ARM CPUs to handle logic control, thus completing the deployment of the edge controller.
[0034] According to the present invention, a wind farm intelligent prediction and control method based on a lightweight time-series hybrid model is provided, wherein the real-time acquisition of multi-resolution wind turbine operating data includes:
[0035] The SCADA sensor network reads the wind turbine's operating parameters in real time, including the anemometer collecting wind speed values once per second, the wind vane collecting wind direction angles once per second, the temperature sensor collecting temperature values once per minute, and the humidity sensor collecting humidity values once per minute.
[0036] The wind speed, wind direction, temperature, and humidity values are timestamped and normalized to form a four-dimensional vector sequence of [wind speed, wind direction, temperature, humidity], which serves as multi-resolution wind turbine operating data.
[0037] Secondly, the present invention also provides a wind farm intelligent prediction and control system based on a lightweight time-series hybrid model, comprising:
[0038] The data acquisition module is used to acquire multi-resolution wind turbine operating data in real time;
[0039] The block module is used to input the multi-resolution wind turbine operating data into the adaptive block layer of the TTMs wind farm prediction model, and to extract high-frequency and low-frequency features to obtain the resolution prefix.
[0040] The splicing module is used to encode the resolution prefix into an embedding vector and splice it to the input layer of the TTMs wind farm prediction model;
[0041] The prediction module is used to output the spliced data of the input layer to the backbone network, combine the wind turbine power physical equation as the physical constraint loss to predict the wind farm, and output the wind turbine state probability value and power prediction value.
[0042] The adjustment module is used to adjust the operating parameters of the wind turbine based on the wind turbine state probability value and power prediction value.
[0043] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the wind farm intelligent prediction and control method based on a lightweight time-series hybrid model as described above.
[0044] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the wind farm intelligent prediction and control method based on a lightweight time-series hybrid model as described above.
[0045] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the wind farm intelligent prediction and control method based on a lightweight time-series hybrid model as described above.
[0046] The present invention provides a wind farm intelligent prediction and control method based on a lightweight temporal hybrid model, comprising: real-time acquisition of multi-resolution wind turbine operating data; inputting the multi-resolution wind turbine operating data into the adaptive block layer of the TTMs wind farm prediction model, performing high-frequency feature extraction and low-frequency feature extraction to obtain a resolution prefix; encoding the resolution prefix into an embedding vector and concatenating it into the input layer of the TTMs wind farm prediction model; outputting the concatenated data from the input layer to the backbone network, combining the wind turbine power physical equation as a physical constraint loss to perform wind farm prediction, and outputting wind turbine state probability values and power prediction values; adjusting the wind turbine operating parameters based on the wind turbine state probability values and power prediction values. Since the TTMs wind farm prediction model is a lightweight model, it can be better deployed on edge controllers, and the physical law of the wind turbine power physical equation is incorporated into the loss function during the model's operation, enabling the model to better adapt to wind farm prediction under conditions exceeding the rated wind speed, thereby reducing prediction errors. Attached Figure Description
[0047] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0048] Figure 1 This is a flowchart illustrating the intelligent prediction and control method for wind farms based on a lightweight time-series hybrid model provided in this embodiment.
[0049] Figure 2 This is a schematic diagram of the TTMs wind farm prediction model provided in this embodiment;
[0050] Figure 3 This is a schematic diagram of the structure of the wind farm intelligent prediction and control system based on a lightweight time-series hybrid model provided in this embodiment;
[0051] Figure 4 This is a schematic diagram of the structure of the electronic device provided in this embodiment. Detailed Implementation
[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0053] Figure 1This is a flowchart illustrating the intelligent prediction and control method for wind farms based on a lightweight time-series hybrid model provided in this embodiment.
[0054] like Figure 1 As shown in the figure, the intelligent prediction and control method for wind farms based on a lightweight time-series hybrid model provided by this invention mainly includes the following steps:
[0055] 101. Real-time acquisition of multi-resolution wind turbine operation data.
[0056] Specifically, the first step is to build a data acquisition system that includes a SCADA sensor network. The sensor types and sampling characteristics are strictly matched to the wind farm monitoring requirements. High-frequency sensors include anemometers (1Hz sampling, m / s), wind vanes (1Hz sampling, °), and vibration monitors (10kHz high-frequency sampling, monitoring equipment vibration status) to capture transient changes during wind turbine operation. Low-frequency sensors include temperature sensors (1 / 60Hz sampling, °C) and humidity sensors (1 / 60Hz sampling, %) to record slow changes in environmental parameters.
[0057] During data collection, each set of data needs to be labeled with a millisecond-level timestamp to ensure that the data accurately corresponds to the operating time of the wind turbine. At the same time, it is necessary to consider that the data of multiple wind turbines in the same wind farm have strong spatiotemporal correlation (the data of adjacent wind turbines will interfere with each other). During the collection process, the wind turbine number needs to be recorded to reserve an identifier for subsequent data correlation analysis.
[0058] Due to differences in sensor sampling frequencies, the raw data exhibits resolution heterogeneity, requiring a two-step preprocessing approach. Timestamp alignment: A dynamic linear interpolation algorithm is used to process low-frequency data (temperature, humidity). Based on the timestamps and values of the original data, supplementary data is generated at each second (e.g., for one temperature data point per minute, 60 interpolation points need to be evenly generated between adjacent original points), ultimately unifying all parameters to a 1Hz sampling frequency. During interpolation, extreme points of the original data must be preserved to avoid over-smoothing that could lead to the loss of key operating condition features.
[0059] Anti-interference normalization involves calculating the mean (μ) and standard deviation (σ) of historical data for each parameter type (wind speed, wind direction, temperature, and humidity), and processing the data according to the formula "normalized value = (original value - μ) / (5σ)". Specifically, the typical μ value for wind speed is 8.2 m / s and the σ value is 3.1; the typical μ value for wind direction is 152° and the σ value is 45; and the typical μ value for temperature is 18℃ and the σ value is 5. After normalization, outliers within ±5σ are retained without truncation to ensure the integrity of extreme condition data. This process retains outliers within ±5σ (without truncation) to avoid the loss of critical information from extreme conditions (such as exceeding rated wind speed).
[0060] The preprocessed data is organized in chronological order as a four-dimensional vector sequence of [wind speed, wind direction, temperature, humidity], which is used as input for subsequent models.
[0061] Data preprocessing addresses the resolution heterogeneity issue of multi-source SCADA data while preserving transient changes and extreme operating conditions, providing the model with high-quality, time-consistent input data and avoiding prediction errors caused by chaotic data formats.
[0062] 102. Input multi-resolution wind turbine operating data into the adaptive block layer of the TTMs wind farm prediction model, perform high-frequency feature extraction and low-frequency feature extraction, and obtain the resolution prefix.
[0063] Specifically, the preprocessed four-dimensional vector sequence is input into the adaptive block layer (AP) of the TTMs wind farm prediction model, and a differentiated block strategy is adopted to address the feature differences of data at different resolutions.
[0064] Segment length setting: Based on the data sampling frequency and variation characteristics, set the segment length pl=8 (covering 8 seconds of data) for 1Hz high-frequency data (wind speed, wind direction) to focus on capturing transient fluctuations in wind speed and direction. Set the segment length pl=64 (covering 64 seconds of data) for equivalent 1Hz low-frequency interpolated data (temperature, humidity) to focus on extracting long-term trends in environmental parameters.
[0065] Within the layer, deformable convolutional layers automatically adapt to the feature scale of the segmented data and dynamically adjust the receptive field. For example, the receptive field is reduced for high-frequency segments to accurately capture wind speed fluctuations every second; the receptive field is expanded for low-frequency segments to integrate minute-level temperature and humidity trends and avoid mutual interference between different frequency features.
[0066] By using a block-based layer, differential feature extraction of multi-resolution data is achieved, which neither loses the transient details of high-frequency data nor ignores the long-term trends of low-frequency data, thus laying the foundation for subsequent cross-frequency feature fusion.
[0067] 103. Encode the resolution prefix into an embedding vector and concatenate it to the input layer of the TTMs wind farm prediction model.
[0068] Specifically, to avoid model confusion of feature data from different sources, the data source attributes need to be clearly defined through Resolution Prefix Tuning (RPT). First, the original sampling frequencies of each parameter (e.g., "wind speed: 1s, temperature: 5min") are encoded into embedding vectors. During encoding, the numerical characteristics of the frequencies must be preserved (e.g., "1s" corresponds to "1", "5min" corresponds to "300") to ensure the vectors accurately reflect resolution differences. Then, the encoded embedding vectors are directly concatenated to the input layer of the TTMs model, forming a fused input with the high-frequency and low-frequency features extracted by the adaptive block layer. During concatenation, the vector dimension must match the feature data to avoid feature misalignment.
[0069] By clearly distinguishing data sources of different resolutions, prediction biases caused by confusion between transient wind speed characteristics and long-term temperature and humidity characteristics are avoided, thereby improving the accuracy of correlation analysis of multi-source features.
[0070] 104. Output the spliced data from the input layer to the backbone network, combine it with the wind turbine power physical equation as a physical constraint loss to predict the wind farm, and output the wind turbine state probability value and power prediction value.
[0071] Specifically, the fused input data is fed into the model backbone network (TSMixer backbone network), and constraints are constructed by incorporating the physical laws of wind turbines to achieve accurate prediction.
[0072] First, construct the physical constraints (power equation and dynamics). (Coefficient). Based on the aerodynamic principles of wind turbines, the power equation is... Embedding the model loss function, specific steps:
[0073] Dynamic calculation of power factor Based on real-time wind speed Adjustment The formula is (1):
[0074] (1)
[0075] in, This indicates the fan's cut-off wind speed. Exceeding this speed requires the fan to shut down for protection. The higher the wind speed, the better. The closer it is to 0.5, the better it conforms to the actual energy conversion law of wind turbines.
[0076] Calculate the theoretical power by substituting the air density. (Standard atmospheric conditions), fan swept area (Determined according to the fan model, in this embodiment) Real-time wind speed Calculate theoretical power according to the power equation This ensures that the theoretical values conform to physical laws.
[0077] The loss function incorporates physical constraints. A hybrid loss function combining classification loss and physical constraint loss is constructed, as shown in formula (2):
[0078] (2)
[0079] In the formula, Cross-entropy loss reflects the model's probability of three states: normal, warning, and danger. Represents the probability value of the true state. The accuracy of the prediction (representing the probability value of the predicted state). " represents physical constraint terms, (Optimal weighting coefficients verified by experiments) The power prediction value output by the model is constrained to ensure that the power prediction result does not deviate from the physical laws. Indicates air density, Indicates the swept area. Indicates the power factor. This indicates the real-time wind speed.
[0080] The loss function takes into account both state classification accuracy and physical consistency of power prediction, avoiding a surge in prediction error under conditions exceeding rated wind speed.
[0081] The backbone network ultimately outputs four node results: the first three are the wind turbine status probability values, and the normal state probability is... Warning status probability Probability of dangerous state The fourth one is the predicted power value. It simultaneously meets the requirements for status judgment and power prediction.
[0082] By constraining by physical laws, this method solves the problem of large prediction errors in super wind speed caused by traditional models deviating from physical principles. At the same time, it enables dual-dimensional prediction of state and power, providing a comprehensive basis for subsequent control strategies.
[0083] 105. Adjust the operating parameters of the wind turbine based on the wind turbine state probability value and power prediction value.
[0084] Specifically, based on the state probability value and power prediction value output by the TTMs model, and combined with the optimization requirements of three objectives—safety, power generation efficiency, and maintenance cost—the wind turbine controller performs hierarchical control.
[0085] Different control modes are triggered based on the threshold of the state probability value. The specific correspondence is as follows:
[0086] When the probability of normal state At this time, the control mode is mode 0 (optimized power generation), and the action is to start the maximum power point tracking (MPPT) algorithm, dynamically adjust the pitch angle, and enable the wind turbine to capture wind energy to the maximum extent according to the power prediction value.
[0087] When the probability of a warning state When the control mode is mode 1 (derating operation), the action is to limit the power output to 80% of the rated value, and at the same time activate the 10kHz high-frequency vibration monitor to monitor abnormal vibration of the equipment in real time.
[0088] When the probability of a dangerous state When the control mode is mode 2 (emergency protection), the action is to immediately trigger the emergency shutdown procedure, disconnect the power grid connection, start the mechanical brake, and record the operating condition data at the time of shutdown.
[0089] The control strategy design must balance three main objectives: safety, prioritizing immediate shutdown in dangerous situations to prevent equipment damage caused by overspeeding or excessive vibration; power generation efficiency, maximizing power generation under normal conditions using the MPPT algorithm to avoid energy waste due to overly conservative approaches; and maintenance costs, combining capacity reduction under warning conditions with monitoring to minimize power generation losses caused by frequent shutdowns and proactively identify potential hazards.
[0090] By using adaptive hierarchical control, efficiency and cost can be balanced under the premise of safety, avoiding equipment risks caused by a single objective (such as only pursuing power generation).
[0091] The data processing flow of the TTMs wind farm prediction model is as follows: Figure 2 As shown, the biggest difference from existing models lies in the ability to uniformly process data of multiple resolutions through adaptive block layers, and the addition of a physical equation library to the loss function to ensure the accuracy of model predictions and effectively reduce errors. The entire process is supported by two dedicated databases providing data and theoretical backing, ensuring the reliability of the model and control.
[0092] State Database. Functionality: Stores over 100,000 historical operating condition data entries, covering normal operating conditions (such as operation at rated wind speed) and abnormal operating conditions (such as exceeding wind speed limits or vibration exceeding limits). The data includes complete SCADA parameters, status labels, and control results. Application: Supports few-shot fine-tuning of TTMs models. When wind farm operating conditions change (such as seasonal wind speed differences), there is no need for full retraining. Only a small amount of new operating condition data (such as a few hundred entries) is used to fine-tune the model based on the state database, quickly adapting to new scenarios.
[0093] A physical equation library. Contents: Contains core physical models of wind turbines, specifically including: wind turbine aerodynamic equations (such as power equations and lift / drag coefficient formulas); torque-speed models of the transmission system (reflecting the mechanical relationship between the gearbox and main shaft); and grid fluctuation constraints (such as grid-connected power fluctuation range and voltage stability thresholds). Applications: Provides a theoretical basis for "physical constraint loss calculation" in TTMs models, and also provides safety boundaries for control strategies (such as adjusting the upper limit of power output during grid fluctuations).
[0094] With the support of two databases, the problems of insufficient model generalization ability and weak physical theory support are effectively solved. The database enables rapid model adaptation and control safety safeguards.
[0095] This embodiment achieves breakthroughs in three dimensions—prediction accuracy, computational efficiency, and safety control—by innovatively integrating lightweight time-series hybrid models (TTMs) with wind power physics, effectively addressing the technical pain points in wind power forecasting. The TTMs wind farm prediction model has only 1-5M parameters, supports domestically produced edge computing chips, and effectively overcomes computational efficiency bottlenecks. Adaptive Blocking (AP) dynamically adjusts the receptive field for data with different sampling rates, and Resolution Prefix (RPT) encodes the sampling frequency into an embedding vector. The model explicitly distinguishes data sources, enabling the full splicing of second-level / minute-level data, fusing high- and low-frequency features, and improving the accuracy of cross-frequency feature correlation. Hard constraints on the power equation and a dynamic coefficient mechanism enhance the reliability of predictions under ultra-high wind speed conditions.
[0096] Furthermore, to ensure the model's prediction accuracy and lightweight characteristics, the model needs to be trained according to the following process:
[0097] Training prerequisites: Hardware: Equipped with an NVIDIA GPU (to meet the computing power requirements for large-sample training); Data: Use a pre-processed 1Hz aligned time-series dataset, which must contain at least 72 hours of historical data (covering normal / abnormal operating conditions); Pre-trained model: Load the TTMs-Base basic model (1 million parameter version), which has been pre-trained with 100 million wind power time-series samples and has basic time-series feature extraction capabilities.
[0098] Training parameter configuration. Optimizer: Adam optimizer is used, with an initial learning rate of 0.001; Batch size: 1024 (to balance GPU memory usage and training convergence speed); Learning rate decay: After each training round, the mixed loss value is evaluated on the validation set. If the loss decreases by less than 1e-5 for three consecutive rounds, learning rate decay (decay factor 0.5) is initiated; Training metric requirements: State classification accuracy must be ≥95%, and power prediction error in rated wind speed areas (e.g., 12-25 m / s) must be <3%.
[0099] While ensuring the model is lightweight (1-5M parameters, far lower than the hundreds of megabytes of memory required by traditional LSTM / Transformer), the prediction accuracy meets the actual needs of wind farms through pre-training and fine-tuning.
[0100] Furthermore, to address the issue that traditional models cannot be deployed on wind turbine edge controllers, deployment optimization is required for domestically produced edge hardware.
[0101] Deployment prerequisites: Hardware: Domestic edge devices are used to adapt to the harsh environment of wind power sites and support low-power operation; Toolchain: GD32Keil MDK v5.37+ (edge controller development tool) and Ascend CANN Toolkit 6.0+ (model compilation and optimization tool) are used.
[0102] Deployment optimization steps. Model quantization: Dynamic quantization technology is used to convert the FP32 precision model to INT8 precision, quantizing only the linear layer parameters and retaining the output layer in FP32 precision (ensuring the accuracy of state probability and power prediction). After quantization, the accuracy loss needs to be verified: the F1 value for fault state identification decreases by ≤0.5%, and the power prediction error increases by ≤1%. Format conversion and compilation: First, the quantized model is exported to the ONNX universal format, and then compiled into a device-specific om model using the Ascend ATC tool to adapt to Huawei Ascend's AI Core computing unit. Hardware resource allocation: The AI Core dedicated computing unit handles the model's convolution operations (computationally intensive tasks), while the ARM CPU handles control logic and data transmission (lightweight tasks). A double buffering mechanism is also set up to ensure that data transfer (from SCADA to edge device) and model inference are completed within 50ms, meeting real-time control requirements.
[0103] By stably deploying TTMs models on domestically produced edge devices, the bottlenecks of traditional models, such as large memory consumption and poor real-time performance, are solved, making them suitable for wind farm edge computing scenarios.
[0104] Furthermore, based on the above embodiments, this embodiment also includes fault diagnosis and linkage processing. When the wind turbine enters a warning / dangerous state, the fault diagnosis process needs to be initiated to quickly locate the problem and push maintenance suggestions. Fault data retention: The original SCADA data (including wind speed, vibration, temperature and humidity) of the 120 seconds prior to triggering the warning / dangerous state is automatically saved to provide complete data support for diagnosis. Vibration spectrum analysis: Spectrum analysis is performed using 10kHz high-frequency data from the vibration monitor to identify fault characteristics. If a 1kHz sideband appears in the spectrum, it is judged to be a gearbox abnormality; if a 5-10kHz resonance peak appears, it is judged to be a bearing damage. Diagnostic report push: A report containing the fault type (gearbox / bearing), fault location map (specific wind turbine number, component location), and maintenance priority (dangerous state is prioritized for maintenance) is generated and pushed to the wind farm cloud platform via the network to facilitate rapid response by operation and maintenance personnel.
[0105] By quickly locating faults and providing precise repair guidance, equipment downtime can be reduced, thus lowering wind farm maintenance costs.
[0106] Figure 3 This is a schematic diagram of the structure of the wind farm intelligent prediction and control system based on a lightweight time-series hybrid model provided in this embodiment.
[0107] like Figure 3 As shown in the figure, this embodiment provides a wind farm intelligent prediction and control system based on a lightweight time-series hybrid model, including:
[0108] The data acquisition module 301 is used to acquire multi-resolution wind turbine operating data in real time;
[0109] Block module 302 is used to input multi-resolution wind turbine operation data into the adaptive block layer of the TTMs wind farm prediction model, perform high-frequency feature extraction and low-frequency feature extraction, and obtain resolution prefix;
[0110] The splicing module 303 is used to encode the resolution prefix into an embedding vector and splice it to the input layer of the TTMs wind farm prediction model;
[0111] Prediction module 304 is used to output the spliced data of the input layer to the backbone network, combine the wind turbine power physical equation as the physical constraint loss to predict the wind farm, and output the wind turbine state probability value and power prediction value.
[0112] The adjustment module 305 is used to adjust the operating parameters of the wind turbine based on the wind turbine state probability value and power prediction value.
[0113] Figure 4 This is a schematic diagram of the structure of the electronic device provided in this embodiment.
[0114] like Figure 4 As shown, the electronic device may include a processor 401, a communication interface 402, a memory 403, and a communication bus 404. The processor 401, communication interface 402, and memory 403 communicate with each other via the communication bus 404. The processor 401 can call logic instructions from the memory 403 to execute a wind farm intelligent prediction and control method based on a lightweight time-series hybrid model.
[0115] Furthermore, the logical instructions in the aforementioned memory 403 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0116] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the wind farm intelligent prediction and control method based on a lightweight time-series hybrid model provided by the above methods.
[0117] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the wind farm intelligent prediction and control method based on a lightweight time-series hybrid model provided by the above methods.
[0118] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0119] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for intelligent prediction and control of wind farms based on a lightweight time-series hybrid model, characterized in that, include: Real-time acquisition of multi-resolution wind turbine operation data; The multi-resolution wind turbine operating data is input into the adaptive block layer of the TTMs wind farm prediction model to perform high-frequency feature extraction and low-frequency feature extraction to obtain the resolution prefix; The resolution prefix is encoded into an embedding vector and concatenated to the input layer of the TTMs wind farm prediction model; The spliced data from the input layer is output to the backbone network, and combined with the wind turbine power physical equation as a physical constraint loss, wind farm prediction is performed, and the wind turbine state probability value and power prediction value are output. Based on the wind turbine state probability value and power prediction value, the wind turbine operating parameters are adjusted.
2. The method according to claim 1, characterized in that, The process of extracting high-frequency and low-frequency features to obtain a resolution prefix includes: Set the high-frequency block length for high-frequency wind speed and direction data, and set the low-frequency block length for low-frequency temperature and humidity data; Feature extraction is performed on the multi-resolution wind turbine operation data based on the high-frequency block length and the low-frequency block length. The receptive field is automatically adjusted by a deformable convolutional layer to control the high-frequency block to capture transient fluctuations and the low-frequency block to extract long-term trends.
3. The method according to claim 1, characterized in that, The method of combining the physical equations of wind turbine power as physical constraint losses for wind farm prediction includes: Power coefficient is dynamically calculated based on real-time wind speed; Using the power coefficient, real-time wind speed, swept area, and air density, the theoretical power is calculated; A physical constraint term is added to the cross-entropy loss as a loss function, wherein the physical constraint term is determined by the optimal weight coefficient, the theoretical power, and the predicted power.
4. The method according to claim 3, characterized in that, The loss function is: ； In the formula, Represents the probability value of the true state. This represents the predicted state probability value. This represents the optimal weighting coefficient. Indicates predicted power. Indicates theoretical power. Indicates air density, Indicates the swept area. Indicates the power factor. This indicates the real-time wind speed.
5. The method according to claim 1, characterized in that, The wind turbine status probability values include: normal status probability value, warning status probability value, and dangerous status probability value; The step of adjusting the operating parameters of the wind turbine based on the wind turbine state probability value includes: When the normal state probability value is greater than or equal to the first preset threshold, the maximum power point tracking algorithm is activated to dynamically adjust the pitch angle to optimize the power generation efficiency. When the probability value of the warning state is greater than or equal to the second preset threshold, the power output is limited to a preset percentage of the rated value, and the high-frequency vibration monitoring module is activated at the same time. When the probability of the dangerous state is greater than or equal to the third preset threshold, an emergency shutdown procedure is triggered to disconnect the power grid connection and initiate mechanical braking.
6. The method according to claim 5, characterized in that, Also includes: When entering a warning or danger state, the original data collected during the triggering process is automatically saved. The fault characteristics are identified by analyzing the original acquired data through vibration spectrum. When the fault characteristic is a 1kHz sideband, it is determined to be a gearbox malfunction. When the fault characteristic is a 5-10kHz resonance peak, it is determined to be a bearing failure. Based on the fault characteristics, a fault diagnosis report is generated and pushed to the cloud platform, which includes a fault location map and maintenance priority suggestions.
7. The method according to claim 1, characterized in that, The TTMs wind farm prediction model is trained using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 1024. After each training round, the mixed loss value is evaluated on the validation set. When the loss decreases by less than 1e-5 for three consecutive rounds, the learning rate decay is initiated.
8. The method according to claim 1, characterized in that, Also includes: The TTMs wind farm prediction model is converted into INT8 format using dynamic quantization technology to quantize the linear layer parameters; The quantized model is exported to ONNX format, compiled into a dedicated om model, configured with AI Core computing units to handle convolution operations, and ARM CPUs to handle logic control, thus completing the deployment of the edge controller.
9. The method according to any one of claims 1-8, characterized in that, The real-time acquisition of multi-resolution wind turbine operating data includes: The SCADA sensor network reads the wind turbine's operating parameters in real time, including the anemometer collecting wind speed values once per second, the wind vane collecting wind direction angles once per second, the temperature sensor collecting temperature values once per minute, and the humidity sensor collecting humidity values once per minute. The wind speed, wind direction, temperature, and humidity values are timestamped and normalized to form a four-dimensional vector sequence of [wind speed, wind direction, temperature, humidity], which serves as multi-resolution wind turbine operating data.
10. A wind farm intelligent prediction and control system based on a lightweight time-series hybrid model, characterized in that, include: The data acquisition module is used to acquire multi-resolution wind turbine operating data in real time; The block module is used to input the multi-resolution wind turbine operating data into the adaptive block layer of the TTMs wind farm prediction model, and to extract high-frequency and low-frequency features to obtain the resolution prefix. The splicing module is used to encode the resolution prefix into an embedding vector and splice it to the input layer of the TTMs wind farm prediction model; The prediction module is used to output the spliced data of the input layer to the backbone network, combine the wind turbine power physical equation as the physical constraint loss to predict the wind farm, and output the wind turbine state probability value and power prediction value. The adjustment module is used to adjust the operating parameters of the wind turbine based on the wind turbine state probability value and power prediction value.