A soft sensor-based wind farm field-level intelligent collaborative control method and system
Patent Information
- Application Number
- CN202611047597.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-15
- Publication Date
- 2026-10-30
AI Technical Summary
这种以单机感知为基础的分散式控制模式在复杂运行环境中暴露出多重系统性瓶颈：
[0055]1、在风速仪、温度传感器等物理传感器故障时，自动切换至软传感器数据（即预测数据），维持控制回路连续运行，避免因感知缺失导致的降容或停机，直接减少发电损失。
Smart Images
Figure CN123006501A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of intelligent control of wind farms and industrial data integration, and in particular to a wind farm-level intelligent collaborative control method and system based on soft sensors. Background Technology
[0002] Against the backdrop of the global energy structure accelerating its transition to clean and low-carbon energy, wind power, as a pillar of renewable energy, has moved from the early stage of "scale expansion" to a high-quality development stage of "lean operation" and "intelligent collaboration." Currently, onshore and offshore wind farms generally adopt a centralized layout of tens to hundreds of megawatt-class wind turbine units. Their operating efficiency depends not only on the performance of individual units but also on the dynamic collaboration capabilities between all units in the farm.
[0003] However, existing wind farm control systems still primarily employ the traditional architecture of "single-unit autonomy and centralized monitoring." Each wind turbine independently relies on local physical sensors (such as anemometers, wind vanes, and temperature probes) to perceive the environment and equipment status, and accordingly executes basic control actions such as pitch, yaw, and power regulation. This distributed control model based on single-unit sensing exposes multiple systemic bottlenecks in complex operating environments:
[0004] First, the reliability of sensors is insufficient. In harsh environments such as icing, sandstorms, and lightning strikes, anemometers, wind vanes, and temperature sensors are prone to failure, forcing wind turbines to reduce capacity or shut down, resulting in significant power generation losses that are difficult to recover quickly.
[0005] Second, wake losses are difficult to avoid. The wake generated by upstream wind turbines significantly reduces the efficiency of downstream wind turbines, but individual turbine control cannot actively coordinate yaw or power distribution to mitigate the wake's impact.
[0006] Third, there is a lack of a global coordination mechanism. Existing centralized control systems mostly focus on data acquisition and status monitoring, and have not yet established the ability to generate dynamic coordination strategies that integrate status prediction, real-time equipment operating status, and power grid dispatch instructions.
[0007] Fourth, operational benefits cannot be quantified. Existing systems struggle to automatically statistically analyze and attribute the actual effects of collaborative control.
[0008] Although artificial intelligence (AI) technology has been initially applied in scenarios such as wind power prediction and fault early warning, its output is mostly limited to the analysis level and has not yet been deeply integrated with field-level control. Especially under the real-world condition of partial failure of physical sensors, there is a lack of a highly reliable and low-latency "soft sensing" replacement mechanism, as well as a dynamic collaborative control closed loop based on the fusion of information from the entire field.
[0009] Therefore, there is an urgent need for a new wind farm control scheme that can maintain the integrity of the entire field perception through reliable soft sensor data under the real-world conditions of physical sensor anomalies or failures, and drive multiple wind turbines to achieve intelligent, dynamic, and assessable collaborative operation. Summary of the Invention
[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a wind farm-level intelligent collaborative control method based on soft sensors.
[0011] To achieve the above objectives, the technical solution provided by this invention is as follows:
[0012] A wind farm-level intelligent collaborative control method based on soft sensors includes the following steps:
[0013] S1. Real-time collection and aggregation of physical sensor data from all wind turbine units in the field through the wind turbine data acquisition module. This physical sensor data includes operating parameters and environmental data.
[0014] S2. Based on the collected and aggregated physical sensor data, the prediction module continuously performs online rolling predictions of key physical quantities, generating prediction data in parallel with the physical sensors; the prediction module adopts a corresponding differentiated model system for different prediction targets.
[0015] S3. For the physical sensor data of all wind turbines in the field collected and aggregated in step S1 and the prediction data obtained in step S2, the spatiotemporal alignment and data fusion are performed through the collaborative data fusion module, and a field-wide benchmark is constructed. The field-wide benchmark includes the field-level northerly wind direction benchmark and the field-level outdoor ambient temperature benchmark. During the construction of the field-wide benchmark, when the corresponding physical sensor fails, the prediction data is used to construct the field-wide benchmark information.
[0016] S4. Based on the data fusion information and the overall benchmark obtained in step S3, the wind farm collaborative control module combines the grid dispatch instructions and the real-time operating status of the equipment to dynamically generate a collaborative strategy using the Automatic Generation Control (AGC) algorithm, and sends instructions to the wind turbine controller for execution through the industrial communication protocol.
[0017] S5. Through the collaborative status visualization and benefit evaluation module, the execution process of collaborative strategies is transparently monitored, recorded, statistically analyzed, and traced, forming a management closed loop.
[0018] Furthermore, the prediction module employs a differentiated model system for different prediction objectives, including:
[0019] For wind speed and wind direction control parameters, a spatial correlation-based modeling method, such as graph neural networks, is adopted. The spatial topological distance matrix is constructed based on the latitude and longitude of the units, dynamic time warping, and mutual information algorithms. The spatial topological relationship is used to reconstruct missing data and make real-time predictions.
[0020] For parameters with strong time-series correlation, such as generator winding temperature and gearbox temperature, a time-series prediction model is adopted, such as a long short-term memory network based on time attention mechanism, which uses historical second-level operating data as input to predict the temperature change trend within a set time period in the future.
[0021] For parameters with strong time-series correlation, such as generator winding temperature and gearbox temperature, a data-driven prediction model, such as an ensemble learning model, is used. This model combines feature importance screening and cross-validation optimization to output real-time predicted values.
[0022] Furthermore, the field-level northerly wind direction benchmark is obtained by converting the northerly wind direction angle vector of each unit into two-dimensional coordinates, removing outliers using a joint method of K-means clustering and single-class support vector machine, and then performing vector averaging; the field-level outdoor ambient temperature benchmark is obtained by using 3 The criterion is to remove outliers from the outdoor ambient temperature data and then take the median.
[0023] Furthermore, when the corresponding physical sensor fails, the predicted data is used to construct full-field reference information, including:
[0024] Temperature-related soft sensing data employs multiple modes, including a redundant mode: the predictive mode uses historical operating data and operating condition characteristics to predict the temperature rise trend of key components in advance, and triggers feedforward intervention when the predicted value approaches the safety threshold; the redundant mode seamlessly switches to the predictive data when the physical sensor detecting temperature fails, automatically takes over the control logic, ensures that the temperature protection logic is not interrupted, and guarantees continuous system operation.
[0025] Control-type soft sensor data is used to protect the core control loop. When the main control anemometer or wind vane fails, the corresponding predictive data is immediately used as the control input to ensure the continuous execution of yaw and pitch commands.
[0026] The field-level reference soft sensor data is integrated with the effective northerly wind direction and outdoor temperature data of the entire field to dynamically generate a unified field-level northerly wind direction reference and a field-level outdoor ambient temperature reference, providing a consistent reference system for yaw coordination, power distribution, and unit-to-northerly wind direction mutual verification strategies.
[0027] Furthermore, the fault determination conditions for the physical sensor are as follows:
[0028] The fault determination conditions for the component temperature sensor are: the detected component temperature data exceeds the preset temperature threshold range corresponding to the component type, or a component temperature sensor fault code is detected, or an abnormal feature of no change in the data for a preset time is detected.
[0029] The fault determination conditions for the outdoor ambient temperature sensor are: the detected outdoor ambient temperature data exceeds the preset normal temperature range of -20℃ to 60℃, or an outdoor ambient temperature sensor fault code is detected, or an abnormal feature of no change in the data for a preset time is detected.
[0030] The fault determination conditions for anemometers or wind vanes are: the detected wind speed and wind direction data exceed the preset normal range, or a fault code of the anemometer or wind vane is detected, or an abnormal feature of no change in the preset time of the data is detected; the preset normal range of wind speed is 0~80m / s, and the preset normal range of wind direction is 0~360°.
[0031] Furthermore, in step S3, when the collaborative data fusion module performs data fusion, it switches between physical sensor data and predicted data through a soft and hard sensor switching mechanism.
[0032] The switching mechanism between hardware and software sensors includes:
[0033] A1. Determine whether the physical sensor is faulty by comprehensively checking for invalid parameters, abnormal data, and fan fault codes. If the physical sensor is determined to be faulty, proceed to the next step; otherwise, continue to determine whether the physical sensor is faulty.
[0034] A2. Calculate the transition value between physical sensor data and predicted data using a linear weighted transition formula within the switching window, and use this transition value as the control input for smooth transition;
[0035] A3. After the physical sensor data returns to normal, perform reverse smooth recovery to avoid immediate switchback that could cause a sudden change in control.
[0036] Furthermore, the linear weighted transition formula is as follows:
[0037] ;
[0038] in, This represents the transition value between physical sensor data and predicted data. This is the accumulated cycle number since the start of the switchover; This represents the total number of sampling periods within the switching window; This is the last valid value before the physical sensor failed. The alternative predicted value.
[0039] Furthermore, to achieve the above objectives, the present invention also provides a wind farm-level intelligent collaborative control system based on soft sensors, used to implement the above-mentioned wind farm-level intelligent collaborative control method based on soft sensors, the system comprising:
[0040] The wind turbine data acquisition module is used to collect and aggregate physical sensor data from all wind turbine units in the field in real time. This physical sensor data includes operating parameters and environmental data.
[0041] The prediction module is used to continuously predict key physical quantities online based on the collected and aggregated physical sensor data, generating prediction data in parallel with the physical sensors; the prediction module adopts a corresponding differentiated model system for different prediction targets;
[0042] The collaborative data fusion module is used to perform spatiotemporal alignment and data fusion on the physical sensor data of all wind turbines in the field and the obtained prediction data, and to construct a field-wide benchmark. The field-wide benchmark includes a field-level benchmark for northerly wind direction and a field-level benchmark for outdoor ambient temperature. During the construction of the field-wide benchmark, when the corresponding physical sensor fails, the prediction data is used to construct the field-wide benchmark information.
[0043] The wind farm collaborative control module is used to dynamically generate collaborative strategies based on the data fusion information output by the collaborative data fusion module and the whole field benchmark, combined with the grid dispatch instructions and the real-time operating status of the equipment, using the AGC algorithm, and sending instructions to the wind turbine controller for execution through the industrial communication protocol.
[0044] The collaborative status visualization and benefit evaluation module obtains real-time data and execution status from the wind turbine data acquisition module, prediction module, collaborative data fusion module, and wind farm collaborative control module through the data service interface. It also transparently monitors, classifies, statistically analyzes, and traces events during the execution of collaborative strategies, forming a management closed loop.
[0045] Furthermore, the collaborative data fusion module includes:
[0046] The physical sensor fault diagnosis unit is used to determine whether the physical sensor is faulty by comprehensively judging the parameter invalidity detection, data abnormality detection and fault code;
[0047] The smooth switching unit is used to calculate the transition value between physical sensor data and predicted data using a linear weighted transition method within the switching window after confirming a physical sensor fault.
[0048] The reverse smooth recovery unit is used to perform reverse smooth recovery after the physical sensor data returns to normal, avoiding immediate switchback that could cause a sudden change in control.
[0049] Furthermore, the collaborative state visualization and benefit evaluation module includes:
[0050] The real-time status monitoring and highlighting unit is used to highlight the wind turbines in the coordinated state with highlighted icons and dynamic styles in the centralized control interface, and to display the physical sensor data and predicted data of the corresponding measuring points in a floating comparison.
[0051] The core indicator dynamic summary unit is used to dynamically summarize key indicators such as "current number of coordinated wind turbines" and "total number of coordinated events on the day".
[0052] The benefit quantification and visualization unit is used to present the trigger frequency of different wind turbines and different coordination types through stacked bar charts, and to reflect the distribution of the number of units of various coordination strategies through a pie chart.
[0053] The structured event recording and tracing unit is used to record all collaborative events in a structured form. Each record contains at least five fields: triggering conditions, involved units, activated model, issued instructions, and execution results. It supports retrieval, filtering, and statistical analysis based on three dimensions: time, event type, and unit number.
[0054] Compared with existing technologies, the principles and advantages of this technical solution are as follows:
[0055] 1. When physical sensors such as anemometers and temperature sensors fail, the system automatically switches to soft sensor data (i.e., predictive data) to maintain continuous operation of the control loop, avoiding derating or shutdown due to loss of sensing, and directly reducing power generation losses.
[0056] 2. Based on a unified field-level northerly direction reference and unit operating status, the wake influence area is effectively identified, the yaw angle and power settings of upstream wind turbines are dynamically optimized, and the yaw angle is actively adjusted to weaken the interference of the wake on downstream units, thereby improving the overall wind energy capture efficiency and achieving synergistic power generation. At the same time, based on the field-level outdoor ambient temperature reference, the deviation of individual unit temperature measurement is eliminated, providing a reliable input for power curve temperature correction and icing risk assessment, further improving the overall wind energy capture efficiency.
[0057] 3. Predictive soft sensors can identify temperature rise and icing risks in advance, triggering feedforward intervention strategies to avoid emergency power rationing caused by sudden failures and extend high-power operation time.
[0058] 4. Construct a complete "perception-decision-execution-feedback" chain, and use a visualization platform to statistically present the trigger frequency and strategy distribution of various collaborative events, supporting the continuous iteration and performance optimization of collaborative strategies.
[0059] 5. Entirely software-based, requiring no additional physical sensors or other hardware. It can be seamlessly integrated into existing SCADA systems or deployed independently, significantly reducing hardware procurement and engineering costs for intelligent upgrades of wind farms, and possessing plug-and-play industrial applicability. Attached Figure Description
[0060] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0061] Figure 1 This is a flowchart illustrating the principle of a wind farm-level intelligent collaborative control method based on soft sensors, according to an embodiment of the present invention.
[0062] Figure 2 This is a connection block diagram of a wind farm-level intelligent collaborative control system based on soft sensors, according to an embodiment of the present invention. Detailed Implementation
[0063] The present invention will be further described below with reference to specific embodiments:
[0064] like Figure 1 As shown in this embodiment, a wind farm-level intelligent collaborative control method based on soft sensors includes the following steps:
[0065] S1. The wind turbine data acquisition module (reuses the data acquisition channel of the existing SCADA system of the wind farm, without adding any acquisition hardware) collects and aggregates the physical sensor data of all wind turbines in the field in real time. The physical sensor data includes operating parameters and environmental data.
[0066] The operating parameters include active power, speed, pitch angle, generator temperature, and gearbox temperature. Environmental data includes wind speed, wind direction, and icing status.
[0067] S2. Based on the collected and aggregated physical sensor data, a prediction module continuously performs online rolling predictions of key physical quantities, generating prediction data in parallel with the physical sensors; the prediction module adopts a corresponding differentiated model system for different prediction targets, including:
[0068] a) For control parameters such as wind speed and wind direction that directly affect the yaw and pitch control loops, a spatial correlation-based modeling method, such as graph neural network (GNN), is adopted. The spatial topological distance matrix is constructed based on the latitude and longitude of the units in the field, dynamic time warping and mutual information algorithm, and the spatial topological relationship is used to realize the reconstruction and real-time prediction of missing data.
[0069] b) For the generator winding temperature and gearbox temperature, which are strongly time-series correlated parameters, a time-series prediction model is adopted, such as a long short-term memory network based on the time attention mechanism. The model uses historical second-level operating data as input to predict the temperature change trend within the next minute. The trend prediction results are mainly used for feedforward early warning control.
[0070] c) For parameters with strong time-series correlation, such as generator winding temperature and gearbox temperature, a data-driven prediction model, such as an ensemble learning model, is used, which combines feature importance screening and cross-validation optimization to output real-time predicted values.
[0071] S3. For the physical sensor data of all wind turbines in the field collected and aggregated in step S1 and the prediction data obtained in step S2, the spatiotemporal alignment and data fusion are performed through the collaborative data fusion module, and a field-wide benchmark is constructed. The field-wide benchmark includes the field-level northerly wind direction benchmark and the field-level outdoor ambient temperature benchmark. During the construction of the field-wide benchmark, when the corresponding physical sensor fails, the prediction data is used to construct the field-wide benchmark information.
[0072] In this step, the field-level northerly wind direction benchmark is obtained by converting the northerly wind direction angle vector of each unit into two-dimensional coordinates, removing outliers using a combination of K-means clustering and single-class support vector machine, and then performing vector averaging; the field-level outdoor ambient temperature benchmark is obtained by removing outliers from the outdoor ambient temperature data using the 3σ criterion and then taking the median.
[0073] In this step, the fault determination conditions for various types of physical sensors are as follows:
[0074] The fault determination conditions for the component temperature sensor are: the detected component temperature data exceeds the preset temperature threshold range corresponding to the component type, or a component temperature sensor fault code is detected, or an abnormal feature of no change in data for a preset time (10 minutes) is detected.
[0075] The fault determination conditions for the outdoor ambient temperature sensor are: the detected outdoor ambient temperature data exceeds the preset normal temperature range of -20℃ to 60℃, or an outdoor ambient temperature sensor fault code is detected, or an abnormal feature of no change in the data for a preset time is detected.
[0076] The fault determination conditions for anemometers or wind vanes are: the detected wind speed and wind direction data exceed the preset normal range, or a fault code of the anemometer or wind vane is detected, or an abnormal feature of no change in the preset time of the data is detected; the preset normal range of wind speed is 0~80m / s, and the preset normal range of wind direction is 0~360°.
[0077] In this step, if the component temperature data exceeds a preset threshold, the corresponding physical sensor is deemed to have failed. The criterion for this determination is based on data validity, aiming to identify unreliable sensor inputs and trigger soft sensor switching. For safety protection actions triggered by actual physical overheating of the component (such as shutdown, derating, etc.), the fan controller executes the actions according to independent safety protection logic, without relying on the sensor fault determination results of this invention.
[0078] In this step, when the corresponding physical sensor fails, the predicted data is used to construct full-field baseline information, including:
[0079] Temperature-related soft sensing data employs multiple modes, including a redundant mode: the predictive mode uses historical operating data and operating condition characteristics to predict the temperature rise trend of key components in advance, and triggers feedforward intervention when the predicted value approaches the safety threshold; the redundant mode seamlessly switches to the predictive data when the physical sensor detecting temperature fails, automatically takes over the control logic, ensures that the temperature protection logic is not interrupted, and guarantees continuous system operation.
[0080] Control-type soft sensor data is used to protect the core control loop. When the main control anemometer or wind vane fails due to icing, lightning strikes, or exceeding limits, the corresponding predictive data is immediately used as the control input to ensure the continuous execution of yaw and pitch commands.
[0081] The field-level reference soft sensor data is integrated with the effective northerly wind direction and outdoor temperature data of the entire field to dynamically generate a unified field-level northerly wind direction reference and a field-level outdoor ambient temperature reference, providing a consistent reference system for yaw coordination, power distribution, and unit-to-northerly wind direction mutual verification strategies.
[0082] In this step, when the collaborative data fusion module performs data fusion, it switches between physical sensor data and prediction data through a soft and hard sensor switching mechanism.
[0083] The switching mechanism between hardware and software sensors includes:
[0084] A1. Determine whether the physical sensor is faulty by comprehensively checking for invalid parameters, abnormal data, and fault codes. If the physical sensor is determined to be faulty, proceed to the next step; otherwise, continue to determine whether the physical sensor is faulty.
[0085] A2. Calculate the transition value between physical sensor data and predicted data using a linear weighted transition formula within the switching window, and use this transition value as the control input for smooth transition;
[0086] The linear weighted transition formula is:
[0087] ;
[0088] in, This represents the transition value between physical sensor data and predicted data. This is the accumulated cycle number since the start of the switchover; This represents the total number of sampling periods within the switching window; This is the last valid value before the physical sensor failed. The alternative predicted value.
[0089] A3. After the physical sensor data returns to normal, perform reverse smooth recovery to avoid immediate switchback that could cause a sudden change in control.
[0090] S4. Based on the data fusion information and the overall benchmark obtained in step S3, the wind farm collaborative control module combines grid dispatch instructions and real-time equipment operating status to dynamically generate collaborative strategies using the AGC algorithm, and sends instructions to the wind turbine controller for execution via industrial communication protocol.
[0091] S5. The collaborative status visualization and benefit evaluation module monitors, records, statistically analyzes, and traces events during the execution of collaborative strategies, forming a closed-loop management system.
[0092] like Figure 2 As shown, this embodiment also includes a wind farm-level intelligent collaborative control system based on soft sensors, used to implement the above-mentioned wind farm-level intelligent collaborative control method based on soft sensors. This system includes:
[0093] The wind turbine data acquisition module is used to collect and aggregate physical sensor data from all wind turbine units in the field in real time. This physical sensor data includes operating parameters and environmental data.
[0094] The prediction module is used to continuously predict key physical quantities online based on the collected and aggregated physical sensor data, generating prediction data in parallel with the physical sensors; the prediction module adopts a corresponding differentiated model system for different prediction targets;
[0095] The collaborative data fusion module is used to perform spatiotemporal alignment and data fusion on the physical sensor data of all wind turbines in the field and the obtained prediction data, and to construct a field-wide benchmark. The field-wide benchmark includes a field-level benchmark for northerly wind direction and a field-level benchmark for outdoor ambient temperature. During the construction of the field-wide benchmark, when the corresponding physical sensor fails, the prediction data is used to construct the field-wide benchmark information.
[0096] The wind farm collaborative control module is used to dynamically generate collaborative strategies based on the data fusion information output by the collaborative data fusion module and the whole field benchmark, combined with the grid dispatch instructions and the real-time operating status of the equipment, using the AGC algorithm, and sending instructions to the wind turbine controller for execution through the industrial communication protocol.
[0097] The collaborative status visualization and benefit evaluation module obtains real-time data and execution status from the wind turbine data acquisition module, prediction module, collaborative data fusion module, and wind farm collaborative control module through the data service interface. It also transparently monitors, classifies, statistically analyzes, and traces events during the execution of collaborative strategies, forming a management closed loop.
[0098] Specifically, the collaborative data fusion module includes:
[0099] The physical sensor fault diagnosis unit is used to comprehensively determine whether the physical sensor is faulty by detecting invalid parameters, abnormal data, and wind turbine fault codes.
[0100] The smooth switching unit is used to calculate the transition value between physical sensor data and predicted data using a linear weighted transition method within the switching window after confirming a physical sensor fault.
[0101] The reverse smooth recovery unit is used to perform reverse smooth recovery after the physical sensor data returns to normal, avoiding immediate switchback that could cause a sudden change in control.
[0102] The collaborative status visualization and benefit assessment module includes:
[0103] The real-time status monitoring and highlighting unit is used to highlight the wind turbines in the coordinated state with highlighted icons and dynamic styles in the centralized control interface, and to display the physical sensor data and predicted data of the corresponding measuring points in a floating comparison.
[0104] The core indicator dynamic summary unit is used to dynamically summarize key indicators such as "current number of coordinated wind turbines" and "total number of coordinated events on the day".
[0105] The benefit quantification and visualization unit is used to present the trigger frequency of different wind turbines and different coordination types through stacked bar charts, and to reflect the distribution of the number of units of various coordination strategies through a pie chart.
[0106] The structured event recording and tracing unit is used to record all collaborative events in a structured form. Each record contains at least five fields: triggering conditions, involved units, activated model, issued instructions, and execution results. It supports retrieval, filtering, and statistical analysis based on three dimensions: time, event type, and unit number.
[0107] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.
Claims
1. A wind farm-level intelligent collaborative control method based on soft sensors, characterized in that, Includes the following steps: S1. Real-time collection and aggregation of physical sensor data from all wind turbine units in the field through the wind turbine data acquisition module. This physical sensor data includes operating parameters and environmental data. S2. Based on the collected and aggregated physical sensor data, the prediction module continuously performs online rolling predictions of key physical quantities, generating prediction data in parallel with the physical sensors; the prediction module adopts a corresponding differentiated model system for different prediction targets. S3. For the physical sensor data of all wind turbines in the field collected and aggregated in step S1 and the prediction data obtained in step S2, the spatiotemporal alignment and data fusion are performed through the collaborative data fusion module, and a field-wide benchmark is constructed. The field-wide benchmark includes the field-level northerly wind direction benchmark and the field-level outdoor ambient temperature benchmark. During the construction of the field-wide benchmark, when the corresponding physical sensor fails, the prediction data is used to construct the field-wide benchmark information. S4. Based on the data fusion information and the overall benchmark obtained in step S3, the wind farm collaborative control module combines grid dispatch instructions and real-time equipment operating status to dynamically generate a collaborative strategy using an automatic generation control algorithm, and sends instructions to the wind turbine controller for execution via an industrial communication protocol. S5. The collaborative status visualization and benefit evaluation module monitors, records, statistically analyzes, and traces events during the execution of collaborative strategies, forming a closed-loop management system.
2. The wind farm-level intelligent collaborative control method based on soft sensors according to claim 1, characterized in that, The prediction module employs a differentiated model system for different prediction objectives, including: For wind speed and wind direction control parameters, a spatial correlation-based modeling method is adopted. The spatial topological distance matrix is constructed using all the units in the field as nodes, based on the latitude and longitude of the units, dynamic time warping and mutual information algorithm. The spatial topological relationship is used to reconstruct missing data and make real-time predictions. For the generator winding temperature and gearbox temperature, which are strongly correlated by time, a time-series prediction model is adopted. Based on the long short-term memory network with time attention mechanism, the model uses historical second-level operating data as input to predict the temperature change trend within a set time period in the future. For the generator winding temperature and gearbox temperature, which are strongly correlated with time series, a data-driven prediction model is adopted, which combines feature importance screening and cross-validation optimization to output real-time predicted values.
3. The wind farm-level intelligent collaborative control method based on soft sensors according to claim 1, characterized in that, The field-level northerly wind direction benchmark is obtained by converting the northerly wind direction angle vector of each unit into two-dimensional coordinates, removing outliers using a combination of K-means clustering and single-class support vector machine, and then performing vector averaging. The field-level outdoor ambient temperature benchmark is obtained by removing outliers from the outdoor ambient temperature data using the 3σ criterion and then taking the median.
4. The wind farm-level intelligent collaborative control method based on soft sensors according to claim 1, characterized in that, When the corresponding physical sensor fails, the predicted data is used to construct full-field baseline information, including: Temperature-related soft sensing data employs multiple modes, including a redundant mode: the predictive mode uses historical operating data and operating condition characteristics to predict the temperature rise trend of key components in advance, and triggers feedforward intervention when the predicted value approaches the safety threshold; the redundant mode seamlessly switches to the predictive data when the physical sensor detecting temperature fails, automatically takes over the control logic, ensures that the temperature protection logic is not interrupted, and guarantees continuous system operation. Control-type soft sensor data is used to protect the core control loop. When the main control anemometer or wind vane fails, the corresponding predictive data is immediately used as the control input to ensure the continuous execution of yaw and pitch commands. The field-level reference soft sensor data is integrated with the effective northerly wind direction and outdoor temperature data of the entire field to dynamically generate a unified field-level northerly wind direction reference and a field-level outdoor ambient temperature reference, providing a consistent reference system for yaw coordination, power distribution, and unit-to-northerly wind direction mutual verification strategies.
5. The wind farm-level intelligent collaborative control method based on soft sensors according to claim 4, characterized in that, The fault determination conditions for the physical sensor are as follows: The fault determination conditions for the component temperature sensor are: the detected component temperature data exceeds the preset temperature threshold range corresponding to the component type, or a component temperature sensor fault code is detected, or an abnormal feature of no change in the data for a preset time is detected. The fault determination conditions for the outdoor ambient temperature sensor are: the detected outdoor ambient temperature data exceeds the preset normal temperature range of -20℃ to 60℃, or an outdoor ambient temperature sensor fault code is detected, or an abnormal feature of no change in the data for a preset time is detected. The fault determination conditions for anemometers or wind vanes are: the detected wind speed and wind direction data exceed the preset normal range, or a fault code of the anemometer or wind vane is detected, or an abnormal feature of no change in the data for a preset time is detected. The preset normal range for wind speed is 0~80m / s, and the preset normal range for wind direction is 0~360°.
6. The wind farm-level intelligent collaborative control method based on soft sensors according to claim 1, characterized in that, In step S3, when the collaborative data fusion module performs data fusion, it switches between physical sensor data and predicted data through a soft and hard sensor switching mechanism. The switching mechanism between hardware and software sensors includes: A1. Determine whether the physical sensor is faulty by comprehensively checking for invalid parameters, abnormal data, and fan fault codes. If the physical sensor is determined to be faulty, proceed to the next step; otherwise, continue to determine whether the physical sensor is faulty. A2. Calculate the transition value between physical sensor data and predicted data using a linear weighted transition formula within the switching window, and use this transition value as the control input for smooth transition; A3. After the physical sensor data returns to normal, perform reverse smooth recovery to avoid immediate switchback that could cause a sudden change in control.
7. The wind farm-level intelligent collaborative control method based on soft sensors according to claim 6, characterized in that, The linear weighted transition formula is: ； in, This represents the transition value between physical sensor data and predicted data. This is the accumulated cycle number since the start of the switchover; This represents the total number of sampling periods within the switching window; This is the last valid value before the physical sensor failed. The alternative predicted value.
8. A wind farm-level intelligent collaborative control system based on soft sensors, used to implement the wind farm-level intelligent collaborative control method based on soft sensors as described in any one of claims 1-7, characterized in that, include: The wind turbine data acquisition module is used to collect and aggregate physical sensor data from all wind turbine units in the field in real time. This physical sensor data includes operating parameters and environmental data. The prediction module is used to continuously predict key physical quantities online based on the collected and aggregated physical sensor data, generating prediction data in parallel with the physical sensors; the prediction module adopts a corresponding differentiated model system for different prediction targets; The collaborative data fusion module is used to perform spatiotemporal alignment and data fusion on the physical sensor data of all wind turbines in the field and the obtained prediction data, and to construct a field-wide benchmark. The field-wide benchmark includes a field-level benchmark for northerly wind direction and a field-level benchmark for outdoor ambient temperature. During the construction of the field-wide benchmark, when the corresponding physical sensor fails, the prediction data is used to construct the field-wide benchmark information. The wind farm collaborative control module is used to dynamically generate collaborative strategies based on the data fusion information output by the collaborative data fusion module and the whole field benchmark, combined with the grid dispatch instructions and the real-time operating status of the equipment, using the AGC algorithm, and sending instructions to the wind turbine controller for execution through the industrial communication protocol. The collaborative status visualization and benefit evaluation module obtains real-time data and execution status from the wind turbine data acquisition module, prediction module, collaborative data fusion module, and wind farm collaborative control module through the data service interface. It also transparently monitors, classifies, statistically analyzes, and traces events during the execution of collaborative strategies, forming a management closed loop.
9. A wind farm-level intelligent collaborative control system based on soft sensors according to claim 8, characterized in that, The collaborative data fusion module includes: The physical sensor fault diagnosis unit is used to determine whether the physical sensor is faulty by comprehensively judging the parameter invalidity detection, data abnormality detection and fault code; The smooth switching unit is used to calculate the transition value between physical sensor data and predicted data using a linear weighted transition method within the switching window after confirming a physical sensor fault. The reverse smooth recovery unit is used to perform reverse smooth recovery after the physical sensor data returns to normal, avoiding immediate switchback that could cause a sudden change in control.
10. A wind farm-level intelligent collaborative control system based on soft sensors according to claim 8, characterized in that, The collaborative status visualization and benefit evaluation module includes: The real-time status monitoring and highlighting unit is used to highlight the wind turbines in the coordinated state with highlighted icons and dynamic styles in the centralized control interface, and to display the physical sensor data and predicted data of the corresponding measuring points in a floating comparison. The core indicator dynamic summary unit is used to dynamically summarize key indicators such as "current number of coordinated wind turbines" and "total number of coordinated events on the day". The benefit quantification and visualization unit is used to present the trigger frequency of different wind turbines and different coordination types through stacked bar charts, and to reflect the distribution of the number of units of various coordination strategies through a pie chart. The structured event recording and tracing unit is used to record all collaborative events in a structured form. Each record contains at least five fields: triggering conditions, involved units, activated model, issued instructions, and execution results. It supports retrieval, filtering, and statistical analysis based on three dimensions: time, event type, and unit number.