A present wind power blade electrothermal film deicing control method based on big data

Abstract

The application discloses a present wind power blade electric heating film deicing control method based on big data, relates to the technical field of wind turbine operation control and deicing, and the frost ice / glaze ice probability output by the ice type discrimination model of the application is no longer only used for alarming, but is directly converted into an increase / decrease coefficient which is subjected to boundary and monotonicity constraints, so that energy can be fully put in the glaze ice scene and overheating and overconsumption can be avoided in the frost ice scene; the application introduces the change trend of wind speed, temperature and relative humidity as a modulation factor, responds to the working condition in which glaze ice will be formed or is turning to glaze ice in advance, and reduces the deicing deficiency caused by the lag of a static threshold; the power deviation correction of the application takes the wind speed-power curve of the unit as a reference, adopts the structure of a dead zone, segmented gain and cumulative saturation, closes the external performance of aerodynamic degradation into a control quantity without relying on an external icing sensor, considers sensitivity and stability, and reduces start-stop oscillation and time length drift.

Description

Technical Field

[0001] This invention relates to the field of wind turbine operation control and de-icing technology, and in particular to a current wind turbine blade electrothermal film de-icing control method based on big data. Background Technology

[0002] In cold climates and under supercooled water droplet conditions, ice easily forms on the blade surface, altering the airfoil surface roughness and pressure distribution, leading to power reduction and structural and safety hazards. Depending on factors such as ambient temperature, liquid water content (LWC), median water droplet diameter (MVD), and wind speed, frost, glaze, or a mixture of ice may form on the blade. Among these, dense, strongly adhesive glaze ice is more likely to form under conditions close to 0°C, with higher LWC and higher wind speeds, while loose frost ice is more likely to form under low temperature and low LWC conditions. Different ice types exhibit significant differences in density, thermal conductivity, and interfacial adhesion characteristics, resulting in varying heat flux per unit area, heating time, and zoned power distribution requirements for de-icing.

[0003] Electrothermal film de-icing is a common solution. In existing projects, offline mapping / lookup table control based on environmental parameters (such as wind speed and temperature) and power supply duration is widely used to avoid the lightning protection and maintenance burden caused by deploying sensitive devices on the outside of the blades. This method can achieve certain results under steady-state weather conditions, but when environmental conditions change and the ice type switches between frost ice and glaze ice, the thermodynamic characteristics of the ice layer exhibit significant nonlinearity. Fixed thresholds or offline mapping cannot reflect the real equivalent heat load demand in a timely manner: in glaze ice scenarios, insufficient heating and residual ice adhesion may occur; in frost ice scenarios, excessive energy consumption may occur. Furthermore, the unit is also affected by operational constraints such as the upper limit of blade zone temperature, shutdown and restart logic, and the risk of ice shedding, making it difficult for simple table lookup methods that expand parameter dimensions or increase sample size to cover complex and time-varying operating conditions.

[0004] Therefore, there is an urgent need for a big data-based de-icing control method for electric heating films on wind turbine blades to improve the de-icing adequacy and energy efficiency under complex meteorological conditions. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a big data-based method for controlling the de-icing of electric heating films on wind turbine blades, which solves the problems of existing lookup-based power supply being difficult to adapt to sudden changes in ice type, insufficient removal of glaze ice, high energy consumption for frost ice, and insufficient stability under extreme operating conditions.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a big data-based method for controlling the de-icing of electric heating films on wind turbine blades, comprising:

[0009] Step S1: Obtain the environmental parameters and operating status parameters of the wind turbine. The environmental parameters include at least wind speed, ambient temperature and relative humidity, and the operating status parameters include at least the active power of the turbine and the blade angle.

[0010] Step S2: Based on historical test and operation data, establish a database of mapping relationships between environmental parameters, operating status parameters and reference power supply duration;

[0011] Step S3: During unit operation, the ice type discrimination model is used to output the probability distribution of each ice type based on the current environmental parameters and operating status parameters;

[0012] Step S4: Dynamically correct the baseline power supply duration given by the mapping relationship database according to the probability distribution to obtain the target power supply duration, and control the start and stop of the electric heating film accordingly.

[0013] As a preferred embodiment of the big data-based wind turbine blade electrothermal film de-icing control method described in this invention, the ice type discrimination model is a machine learning-based model that uses random forest or long short-term memory network. The inputs include wind speed, ambient temperature, relative humidity, and optional atmospheric pressure and liquid water content. The output is a probability distribution of ice types, including at least frost ice and glaze ice.

[0014] As a preferred embodiment of the big data-based wind turbine blade electrothermal film de-icing control method described in this invention, the dynamic correction includes:

[0015] When the probability of glaze ice is not lower than the first threshold, the base power supply duration is multiplied by an adjustment coefficient greater than 1 to extend the power supply.

[0016] When the probability of frost and ice is not lower than the second threshold, the base power supply duration is multiplied by a reduction factor of less than 1 to shorten the power supply time.

[0017] The adjustment coefficient and reduction coefficient are updated online or offline based on historical de-icing effect data and are constrained by the current trend of environmental parameter changes.

[0018] The dynamic correction process includes:

[0019] Step C1: Perform probabilistic activation and normalization.

[0020] , ,

[0021] in, The normalized probability after activation. The original probability of the ice type. As the trigger threshold, Pick or , Indicates glaze ice channel, Indicates a frost and ice passage. Corresponding to the first threshold, Corresponding to the second threshold;

[0022] Step C2, the monotonically increasing mapping of the glaze ice adjustment coefficient is as follows:

[0023] ,

[0024] in, Adjustment coefficient for glaze ice. This is the lower limit of the glaze-ice coefficient, and its value is greater than 1. The upper limit of the glaze ice coefficient and greater than , This is a modulating factor for the glaze ice channel environment. The gain slope of the glaze ice channel is non-negative. The normalized probability of glaze ice obtained in step C1 is represented by the subscript. and These represent the lower and upper limits of the channel, respectively.

[0025] Step C3, the monotonically decreasing mapping of the frost reduction coefficient is as follows:

[0026] ,

[0027] in, This is the frost reduction coefficient. The upper limit of the frost coefficient and less than 1. The lower limit of the frost coefficient is greater than 0 and less than 0. , As a modulating factor for the frost-ice channel environment, The slope of the frost-ice channel is non-negative. The normalized probability of frost obtained in step C1 is represented by the index. Indicates a frost passage, subscript and These represent the lower and upper limits of the channel, respectively.

[0028] Step C4, the environmental trend modulation factor and correlation method are as follows:

[0029] ,

[0030] in, It is a composite trend index. The symbol for the hyperbolic tangent function is... These represent the changes in wind speed, ambient temperature, and relative humidity within the prediction time window. To correspond to the reference change, Indicates reference quantity. The channel gain coefficient has a value range of [value range missing]. , These are the environmental modulation factors in steps C2 and C3, respectively;

[0031] Step C5, glaze ice channel in and Time Monotonous and undiminished, and Frost and ice passages in and Time Monotonically non-increasing, and Threshold Take the open interval (0,1);

[0032] Step C6, the target power supply duration and dual-channel decision are as follows:

[0033] ,

[0034] ,

[0035] in, This represents a relative increase in the glaze ice channel. To reduce the amount of frost and ice passages, The coefficients are the result of comprehensive selection. For absolute value, sign Indicates that, For the target power supply duration, To map the baseline power transmission duration given by the database, the subscript... Indicates the target quantity, subscript Indicates the baseline quantity.

[0036] As a preferred embodiment of the big data-based wind turbine blade electrothermal film de-icing control method described in this invention, it further includes power deviation correction.

[0037] The theoretical cleaning blade power is calculated based on the wind speed-power curve, and the power deviation is obtained by comparing it with the real-time active power of the unit.

[0038] When the power deviation exceeds the preset range, the target power supply duration is corrected in segments or levels, and the maximum correction range is limited.

[0039] The power deviation correction process includes:

[0040] Step D1, in the current control cycle Inside, the theoretical cleaning blade power is obtained based on the wind speed-power curve, and the relative deviation is calculated:

[0041] ,

[0042] in, Indicates period Theoretical cleaning blade power, This represents the wind speed-power curve function. Indicates period wind speed, Indicates period The relative deviation of active power, Indicates period The real-time active power of the unit, The lower bound constant of the denominator, This indicates that the larger of the two options should be chosen.

[0043] Step D2, with For piecewise variables, a threshold sequence is set and graded correction is performed on both positive and negative sides:

[0044] , ,

[0045] when hour:

[0046] ,

[0047] when hour:

[0048] ,

[0049] in, The deviation amplitude, The segmentation threshold, This is the instantaneous power deviation correction factor. To extend the upper limit of the side, To shorten the lower limit of the side, To extend the gain of each segment, To shorten the gain of each segment, Denotes the positive part function, and These represent the undersaturation and oversaturation operations, respectively.

[0050] Step D3, using multiplicative accumulation and saturation within the interval:

[0051] ,

[0052] ,

[0053] ,

[0054] in, For the period The cumulative correction factor, and For the cumulative upper and lower bounds, The coefficient for the current period that is actually implemented under cumulative constraints. The target power delivery duration after power deviation correction. The target power supply duration before proceeding to this step. To control the periodic index, This indicates that the power deviation has been corrected.

[0055] Step D4, Set as the upper boundary of the dead zone to suppress small-amplitude noise. As the wind speed zone and icing risk zoning table are adjusted, the extension side Non-descending order can be used to obtain progressive weighting, shortening the side It can be set independently to reflect asymmetric strategies. Set based on the model and heat load boundary;

[0056] Step D5, when the wind speed is below the cut-in or above the cut-out zone, let To avoid misjudgment, a threshold is set using a small hysteresis pair. Uplink and downlink are separated to reduce the frequency of round-trip handover; updates will not be performed temporarily in case of packet loss or abnormal power measurement. .

[0057] As a preferred embodiment of the wind turbine blade electrothermal film de-icing control method based on big data described in this invention, the method employs a time series model to predict the changing trend of environmental parameters within a preset time window, and obtains the baseline power supply duration by weighted fusion of the predicted and measured values ​​when querying the mapping database.

[0058] As a preferred embodiment of the wind turbine blade electrothermal film de-icing control method based on big data described in this invention, the blade is divided into multiple aerodynamic zones, and ice type discrimination and dynamic correction are performed at the zone granularity.

[0059] The partitions include at least a leading edge partition, a middle section partition, and a trailing edge partition. The leading edge partition is determined based on a preset arc length range at the front of the blade, while the middle section and trailing edge partitions are preset according to the chord length ratio.

[0060] As a preferred embodiment of the big data-based wind turbine blade electrothermal film de-icing control method described in this invention, the determination of the target power supply duration is also based on the weighted calculation of the weights of each partition, and the weight of the leading edge partition is higher than that of other partitions.

[0061] The weight increases with the probability of glaze ice or the incoming wind speed, and is adaptively adjusted between the upper and lower safety limits.

[0062] As a preferred embodiment of the wind turbine blade electrothermal film de-icing control method based on big data described in this invention, the dynamic execution is subject to constraints, which include any one or more of the following constraints:

[0063] The upper limit of surface temperature of the heating zone, the upper limit of heating power, the unit's anti-ice-throwing operation strategy, and the minimum shutdown / restart interval are all constraints. When any of these constraints is triggered, the protection logic is executed first.

[0064] As a preferred embodiment of the big data-based wind turbine blade electrothermal film de-icing control method described in this invention, the mapping relationship database and the ice type discrimination model are jointly updated according to the season or meteorological cycle. During the update, the mapping parameters and model parameters are retrained or recalibrated using the latest operating data.

[0065] As a preferred embodiment of the big data-based wind turbine blade electrothermal film de-icing control method described in this invention, the joint update adopts a federated learning framework:

[0066] Each unit trains an ice type discrimination model locally and generates parameter updates. The cloud uses weighted aggregation to obtain a global model before distributing it.

[0067] The update cycle is adaptively adjusted according to seasonal or site climate change strategies, and it reverts to the previous stable version when aggregation fails.

[0068] The beneficial effects of this invention are as follows: This invention establishes a closed-loop control path from table lookup to adaptive control by mapping environmental and operating state parameters to the baseline power supply duration and combining it with the online correction of ice type probability.

[0069] The ice type discrimination model of this invention no longer outputs frost / glaze ice probabilities merely as alarms, but directly converts them into increase / decrease coefficients constrained by boundaries and monotonicity. This ensures sufficient energy delivery in glaze ice scenarios and avoids overheating and excessive energy consumption in frost ice scenarios. This invention introduces the changing trends of wind speed, temperature, and relative humidity as modulation factors to proactively respond to conditions that are forming or transitioning to glaze ice, reducing insufficient de-icing caused by static threshold lag. The power deviation correction of this invention uses the unit's wind speed-power curve as a reference, employing a dead zone, segmented gain, and cumulative saturation structure, without relying on external icing sensors. The invention incorporates the external manifestations of aerodynamic degradation into the control variables, balancing sensitivity and stability, and reducing start-stop oscillations and duration drift. The partitioned weighting strategy of this invention emphasizes leading edge priority and allows weights to be adaptively adjusted according to the probability of glaze ice and the incoming wind speed, so that the heat distribution matches the actual icing sensitive area. The safety constraints of this invention cover the upper limit of partition temperature, upper limit of power and anti-icing strategy, ensuring that the control behavior is limited to the operating red line under extreme weather conditions. The seasonal and federated update mechanism of this invention is oriented towards site differences and climate evolution to periodically recalibrate the mapping and model, improving the generalization and maintainability across sites and across years.

[0070] Therefore, under conditions of rapid ice formation changes and complex weather, it can improve the adequacy and energy efficiency of de-icing, reduce the risk of continuous power generation interruption due to insufficient de-icing, and at the same time suppress energy consumption and material thermal fatigue caused by overheating, thereby enhancing the continuity and predictability of unit operation. Attached Figure Description

[0071] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0072] Figure 1 This is a flowchart illustrating the big data-based de-icing control method for wind turbine blade electrothermal film in this embodiment. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0074] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0075] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0076] This application proposes a method for de-icing existing wind turbine blade electrothermal film based on big data, combined with... Figure 1 As shown, the method includes:

[0077] Step S1: Obtain the environmental parameters and operating status parameters of the wind turbine. The environmental parameters include at least wind speed, ambient temperature and relative humidity, and the operating status parameters include at least the active power of the turbine and the blade angle.

[0078] In this embodiment, wind speed, ambient temperature, and relative humidity are measured from the nacelle or hub, respectively, while the unit's active power and blade angle are measured from standard measurement points of the main control and pitch systems. The default sampling period is 2 seconds, which can be adjusted within the range of 1 to 5 seconds according to the site noise level and communication load, based on the rule of not over-sampling within the autocorrelation time of the wind speed and power measurement points. Relative humidity is taken from the actual sensor value, in percentage; unit active power is taken from the real-time power of the control system; blade angle is taken from the average value of the three blades or a uniform angle provided by the main control. Optionally, when a single sensor is missing a measurement, the previous valid value is used to maintain a sampling period and marked as missing; if the continuous missing measurement exceeds 30 seconds, subsequent calculations are paused until recovery.

[0079] Step S2: Based on historical test and operation data, establish a database of mapping relationships between environmental parameters, operating status parameters and reference power supply duration;

[0080] Specifically, the mapping database is constructed by gridding the environmental and operational parameter space and recording the corresponding baseline power supply duration. The default grid step size is 1 meter per second for wind speed, 1 degree Celsius for ambient temperature, 5% for relative humidity, and 2 degrees for blade angle. The step size can be adjusted within half to twice the above values, based on coverage conditions while avoiding sparseness. The default historical data covers the most recent complete icing season (no less than three months), with priority given to data from the last two years. Interpolation uses bilinear or trilinear interpolation to return continuous baseline durations without changing the physical meaning of the database input and output. Optionally, when the number of samples in a grid is less than 5, the weighted average of adjacent grids is used for backfilling. If necessary, database entries are updated monthly and the previous version is retained for rollback.

[0081] Step S3: During unit operation, the ice type discrimination model is used to output the probability distribution of each ice type based on the current environmental parameters and operating status parameters;

[0082] For example, the input data for the ice type discrimination model uses parameter sequences from the most recent sampling periods, with a default time window of 10 minutes, which can be set within a range of 5 to 20 minutes according to the local weather change rate; the output is the probability of frost ice and glaze ice, with the probability summing to one. Ice type labels for training samples can be generated based on historical maintenance records, downtime inspection records, or through a joint criterion of power degradation and meteorological conditions. The labeling rules are fixed in the instruction manual and are reviewed in batches each season. To avoid learning bias caused by differences in input scale, each input channel uses zero-mean normalization, and the normalization parameters are updated monthly.

[0083] Step S4: Dynamically correct the baseline power supply duration given by the mapping relationship database according to the probability distribution to obtain the target power supply duration, and control the start and stop of the heating film accordingly. Similarly, the execution of the target power supply duration is constrained by the minimum duration and minimum pause time to reduce frequent start and stop. The default minimum duration is 60 seconds and the minimum pause time is 30 seconds, which can be adjusted within the range of 30 to 120 seconds, based on the thermal inertia of the heating film and the allowable temperature rise of the contact material. The control cycle is triggered according to the sampling period. If the difference between the new target value and the current executed value is less than 5%, it remains unchanged as a dead zone to suppress jitter. If necessary, when the safety constraint is triggered or the critical measurement is missing for more than the set duration, a rollback strategy is executed, that is, the baseline power supply duration given by the database is adopted and maintained until the next effective control cycle.

[0084] In one embodiment, the ice type discrimination model is a machine learning-based model that uses a random forest or long short-term memory network. The input includes wind speed, ambient temperature, relative humidity, and optional atmospheric pressure and liquid water content. The output is a probability distribution of ice types that includes at least frost ice and glaze ice.

[0085] Furthermore, the random forest model defaults to 200 trees, which can be adjusted from 100 to 500 trees depending on the amount of data and the risk of overfitting. The time step of the Long Short-Term Memory network defaults to the aforementioned 10-minute time window, and the hidden layer size defaults to 64, which can be adjusted from 32 to 128 according to the sample size. Training and validation are split by time, without cross-leakage; the update cycle defaults to once per quarter, but earlier updates are allowed in the event of abnormal weather events. Input missing values ​​are kept within the most recent value, no more than two minutes, and samples older than two minutes are not used for training.

[0086] In one embodiment, dynamic correction includes:

[0087] When the probability of glaze ice is not lower than the first threshold, the base power supply duration is multiplied by an adjustment coefficient greater than 1 to extend the power supply.

[0088] When the probability of frost and ice is not lower than the second threshold, the base power supply duration is multiplied by a reduction factor of less than 1 to shorten the power supply time.

[0089] The adjustment and reduction coefficients are updated online or offline based on historical de-icing performance data and are constrained by the current environmental parameter trends. Optionally, historical de-icing performance is statistically analyzed based on the actual power recovery magnitude and duration after de-icing, with the average power increase within 30 minutes after de-icing completion as the default performance indicator, and calculated in groups by wind speed and temperature zones. The online fine-tuning update step is limited to no more than 5% per hour by default, and the offline recalibration cycle is once a month by default; the single update magnitude of any coefficient is limited to no more than one-tenth of its operating range to ensure stability.

[0090] The dynamic correction process includes:

[0091] Step C1: Perform probabilistic activation and normalization.

[0092] , ,

[0093] in, The normalized probability after activation. The original probability of the ice type. As the trigger threshold, Pick or , Indicates glaze ice channel, Indicates a frost and ice passage. Corresponding to the first threshold, Corresponding to the second threshold;

[0094] In this embodiment, the first and second thresholds are set to around 0.3 by default, and can be configured between 0.2 and 0.5 according to the site characteristics, based on the trade-off between false alarms and missed alarms under historical tags. Probabilities below the trigger threshold are considered noise and do not participate in subsequent mapping; when the probabilities of both ice types are lower than their respective thresholds, the database baseline power supply duration is directly adopted. It is recommended that the seasonal adjustment of the thresholds be set discretely for three periods: early winter, severe winter, and late winter.

[0095] Step C2, the monotonically increasing mapping of the glaze ice adjustment coefficient is as follows:

[0096] ,

[0097] in, Adjustment coefficient for glaze ice. This is the lower limit of the glaze-ice coefficient, and its value is greater than 1. The upper limit of the glaze ice coefficient and greater than , This is a modulating factor for the glaze ice channel environment. The gain slope of the glaze ice channel is non-negative. The normalized probability of glaze ice obtained in step C1 is represented by the subscript. and These represent the lower and upper limits of the channel, respectively.

[0098] Specifically, the lower limit of the glaze ice coefficient is 1.10 by default and can be set between 1.05 and 1.30; the upper limit of the glaze ice coefficient is 1.80 by default and can be set between 1.30 and 2.00, based on the rated heat load of the heating system and the temperature resistance of the material. The gain slope of the glaze ice channel is 0.5 by default and can be adjusted between 0.2 and 1.0; the amplitude of the environmental modulation coefficient is 0.3 by default and can be set between 0.1 and 0.7 to ensure that it amplifies appropriately without exceeding the upper limit when the trend increases.

[0099] Step C3, the monotonically decreasing mapping of the frost reduction coefficient is as follows:

[0100] ,

[0101] in, This is the frost reduction coefficient. The upper limit of the frost coefficient and less than 1. The lower limit of the frost coefficient is greater than 0 and less than 0. , As a modulating factor for the frost-ice channel environment, The slope of the frost-ice channel is non-negative. The normalized probability of frost obtained in step C1, with subscript... Indicates a frost passage, subscript and These represent the lower and upper limits of the channel, respectively.

[0102] In this embodiment, the upper limit of the frost-ice coefficient is 0.95 by default and can be set between 0.90 and 0.99; the lower limit of the frost-ice coefficient is 0.70 by default and can be set between 0.50 and 0.85, based on the principle of reducing energy consumption without affecting re-icing and removal in the case of loose ice. The loss slope of the frost-ice channel is 0.4 by default and can be adjusted between 0.1 and 0.8; the amplitude of the environmental modulation coefficient is 0.3 by default and can be set between 0.1 and 0.7.

[0103] Step C4, the environmental trend modulation factor and correlation method are as follows:

[0104] ,

[0105] in, It is a composite trend index. The symbol for the hyperbolic tangent function is... These represent the changes in wind speed, ambient temperature, and relative humidity within the prediction time window. To correspond to the reference change, Indicates reference quantity. The channel gain coefficient has a value range of [value range missing]. , These are the environmental modulation factors in steps C2 and C3, respectively;

[0106] For example, the default forecast window is 15 minutes, which can be set within a range of 10 to 30 minutes according to the rate of weather change at the site. The default wind speed reference change is 1.5 meters per second, the default ambient temperature reference change is 1 degree Celsius, and the default relative humidity reference change is 5 percentage points. These three can be adjusted within the ranges of 0.5 to 3, 0.5 to 3, and 3 to 10, respectively, to ensure that the trend index has moderate sensitivity and is not saturated within common fluctuations. The trend index maintains a smooth change during continuous rises or falls to avoid overmodulation during single anomalous jumps.

[0107] Step C5, glaze ice channel in and Time Monotonous and undiminished, and Frost and ice passages in and Time Monotonically non-increasing, and Threshold Take the open interval (0,1);

[0108] Furthermore, the monotonicity is checked offline using a monotonicity check table in the software implementation. The check frequency is set to be executed immediately after each parameter update, and the pass result is recorded. If any channel mapping goes out of bounds or monotonicity is violated, a rollback to the previous version of parameters is initiated and an alarm is issued.

[0109] Step C6, the target power supply duration and dual-channel decision are as follows:

[0110] ,

[0111] ,

[0112] in, This represents a relative increase in the glaze ice channel. To reduce the amount of frost and ice passages, The coefficients are the result of comprehensive selection. For absolute value, sign Indicates that, For the target power supply duration, To map the baseline power transmission duration given by the database, the subscript... Indicates the target quantity, subscript Indicates a reference quantity;

[0113] Optionally, when the relative amplitudes of the two channels are close and the difference is below the set resolution threshold, the extended side is selected first, or the coefficient of the previous cycle is kept unchanged. The default resolution threshold is 5% of the target power-on duration. The rate of change of the target value is limited to no more than 10% per cycle to avoid thermal shock caused by excessively rapid changes. The execution results, input parameters, final coefficients, and duration are recorded together for subsequent recalibration.

[0114] The system is periodically recalibrated offline based on historical de-icing performance data, and fine-tuned online when necessary under stability constraints. (Threshold) Climate archives are zoned according to season and site;

[0115] In this embodiment, offline recalibration is performed monthly by default and can be adjusted within a range of two weeks to two months; online fine-tuning allows for variations within a range not exceeding one-tenth of their respective working intervals. Seasonal zoning is recommended to use three levels: early winter, severe winter, and late winter. Station climate files are divided according to multi-year averages, and parameters automatically switch and record version numbers when crossing levels for traceability.

[0116] Specifically, this section presents a computable, verifiable, and bounded probability mapping to the power transmission coefficient; activation and normalization isolate noise below the threshold, avoiding frequent start-stop cycles caused by low-probability disturbances; a saturated monotonically increasing function is used to extend power transmission for glaze ice paths, while a saturated monotonically decreasing function is used to shorten power transmission for frost ice paths, both consistent with engineering intuition; by introducing a composite trend index, the direction and amplitude of changes in wind speed, temperature, and humidity are transformed into modulation factors, thereby changing the mapping slope and achieving a sensitive response to meteorological changes without exceeding upper and lower bounds; when dual channels are running concurrently, selection is made based on relative amplitude, outputting a single coefficient, resulting in simple control logic and avoiding oscillations under contradictory instructions; finally, the target value is obtained by multiplying the comprehensive coefficient by the baseline power transmission duration, facilitating integration with the database and operational logic;

[0117] In one embodiment, power deviation correction is also included:

[0118] The theoretical cleaning blade power is calculated based on the wind speed-power curve, and the power deviation is obtained by comparing it with the real-time active power of the unit.

[0119] Similarly, the theoretical clean blade power is derived from the unit's wind speed-power curve, calibrated on-site, and then fixed in the control system. If necessary, a one-time correction is made based on long-term deviations at the site. Real-time active power uses the instantaneous value or short-window average provided by the control system, with the default being the average of the most recent 10 seconds to suppress instantaneous fluctuations. Wind speed is taken as the value of the same cycle synchronized with the power measurement point.

[0120] When the power deviation exceeds the preset range, the target power supply time is corrected in segments or levels, and the maximum correction range is limited.

[0121] Furthermore, the preset range is set to ±5% by default, and can be configured between ±3% and ±8% according to the unit's sensitivity; the single limit for the maximum correction amplitude is limited to no more than 20% of the target power supply duration by default, and the cumulative limit will be given in subsequent steps. Correction will not be triggered if the deviation calculation fails or the wind speed is not within the valid range.

[0122] The power deviation correction process includes:

[0123] Step D1, in the current control cycle Inside, the theoretical cleaning blade power is obtained based on the wind speed-power curve, and the relative deviation is calculated:

[0124] ,

[0125] in, Indicates period Theoretical cleaning blade power, This represents the wind speed-power curve function. Indicates period wind speed, Indicates period The relative deviation of active power, Indicates period The real-time active power of the unit, The lower bound constant of the denominator, This indicates taking the larger of the two.

[0126] In this embodiment, the control period is consistent with the aforementioned sampling period; the lower bound constant of the denominator is defaulted to 2% of the rated power, and can be set between 1% and 3%, based on the principle of avoiding amplified deviations caused by an excessively small denominator in low-wind areas. Theoretical and measured values ​​are synchronized with a timestamp, and the maximum allowed time difference does not exceed one sampling period.

[0127] For example, the three thresholds are set to 3%, 7%, and 12% by default, and can be set within the ranges of 2% to 5%, 6% to 9%, and 10% to 15%. The three-segment gain on the extended side is set to 0.05, 0.08, and 0.12 by default, and the three-segment gain on the shortened side is set to 0.05, 0.06, and 0.08 by default. The upper limit of the instantaneous correction coefficient is 1.30 by default, and the lower limit is 0.80 by default. The upper and lower limits can be configured according to the model and heat load capacity within the ranges of 1.10 to 1.50 and 0.60 to 0.95, respectively.

[0128] Step D2, with For piecewise variables, a threshold sequence is set and graded correction is performed on both positive and negative sides:

[0129] , ,

[0130] when (When power is lower than theoretical, tending to extend power transmission time):

[0131] ,

[0132] when (When power is higher than theoretical, tending to shorten power transmission):

[0133] ,

[0134] in, The deviation amplitude, The segmentation threshold, This is the instantaneous power deviation correction factor. To extend the upper limit of the side, To shorten the lower limit of the side, To extend the gain of each segment, To shorten the gain of each segment, Denotes the positive part function, and These represent the undersaturation and oversaturation operations, respectively.

[0135] Step D3: To limit the cumulative correction magnitude, multiplicative accumulation is used and saturation is achieved within the interval.

[0136] ,

[0137] ,

[0138] ,

[0139] in, For the period The cumulative correction factor, and For the cumulative upper and lower bounds, The coefficient for the current period that is actually implemented under cumulative constraints. The target power delivery duration after power deviation correction. The target power supply duration before proceeding to this step. To control the periodic index, This indicates that the power deviation has been corrected.

[0140] Optionally, the upper bound of the cumulative correction coefficient is 1.80 by default, and the lower bound is 0.60 by default. It can be configured within the range of 1.30 to 2.00 and 0.50 to 0.90. The initial cumulative value is reset to 1 when the unit is powered on or the control is reset to prevent historical accumulation from affecting the new round of control. If the instantaneous correction is not triggered for three consecutive control cycles, the cumulative value slowly returns to 1 at a rate of 5% per cycle.

[0141] In this embodiment, the field tuning sequence for thresholds and gains is as follows: first determine the dead zone, then determine the segmented thresholds, and finally determine the gains on both sides. When the model is more sensitive to overheating, the gain on the shortened side can be appropriately increased by one level. All parameter changes must be performed in a non-icing window to avoid execution jitter during the tuning process.

[0142] Step D4, Set as the upper boundary of the dead zone to suppress small-amplitude noise. As the wind speed zone and icing risk zoning table are adjusted, the extension side Non-descending order can be used to obtain progressive weighting, shortening the side It can be set independently to reflect asymmetric strategies. Set based on the model and heat load boundary;

[0143] Optionally, the wind speed cut-in and cut-out sections are executed according to the predetermined values ​​for the model; the hysteresis of the uplink and downlink separation is 10% to 20% of the threshold by default. When any key measurement is lost for more than 30 seconds or is judged to be abnormal, the deviation correction is paused and the cumulative coefficient remains unchanged. After recovery, the normal procedure is continued.

[0144] Step D5, when the wind speed is below the cut-in or above the cut-out zone, can be set as follows: To avoid misjudgment, a threshold is set using a small hysteresis pair. Uplink and downlink are separated to reduce the frequency of round-trip handover; updates will not be performed temporarily in case of packet loss or abnormal power measurement. ;

[0145] Specifically, this section defines power deviation correction rules; the deviation is split into amplitude and sign, with the amplitude used to enter the segmented channel and the sign corresponding to the extension or shortening of the action; deviation calculation uses the curve model output as a reference and introduces a lower bound for the denominator to avoid unreasonable ratios in low-wind areas and facilitate unified measurement; the threshold sequence provides a clear level division, with small deviations entering the dead zone without triggering adjustment, and medium and large deviations forming a linear-segmented adjustable slope through segmented gain to adapt to the strategy differences of different aircraft models and sites; to control the duration drift caused by long-term superposition, the accumulation coefficient is saturated within the upper and lower bounds, and the current action is closed by the ratio of the actual execution coefficient to the cumulative amount, resulting in a simple structure and stable behavior; the strategy uses independent gains on both positive and negative sides to facilitate the construction of asymmetric responses to cope with the two risk scenarios of icing and overheating; combined with the effective range, hysteresis, and anomaly shielding, smooth control can be maintained even under unstable field measurement conditions;

[0146] In one embodiment, a time series model is used to predict the changing trend of environmental parameters within a preset time window, and the predicted value and the measured value are weighted and fused together when querying the mapping database to obtain the baseline power transmission duration.

[0147] In this embodiment, the prediction time window is set to 30 minutes by default, but can be set within the range of 15 to 60 minutes according to weather variability; the real-time weight of the weighted fusion is set to 0.7 by default, but can be adjusted within the range of 0.5 to 0.9, based on ensuring that the latest measurements dominate while utilizing trend information. The prediction model is automatically trained once a day, and if there is insufficient effective data for the day, the parameters of the previous version are used.

[0148] In one embodiment, the blades are divided into multiple aerodynamic zones, and ice type discrimination and dynamic correction are performed at the zone granularity.

[0149] The partitioning includes at least a leading edge partition, a middle section partition, and a trailing edge partition. The leading edge partition is determined based on the preset arc length range of the front part of the blade, while the middle section and trailing edge partitions are preset according to the chord length ratio.

[0150] Specifically, the leading edge partition is assumed to have a radius of 8% to 12% of the surface arc length at the front, while the middle and trailing edges are assumed to have two ranges based on the chord length ratio: 0.2 to 0.6 and 0.6 to 1.0, respectively. The basic parameters and weights of each partition are maintained independently, and the execution time is calculated separately for each partition and written to the corresponding loop. If a measurement of a partition is abnormal, the correction result of the adjacent partition is temporarily used, and the event is recorded.

[0151] In one embodiment, the determination of the target power supply duration is also based on a weighted calculation of the weights of each partition, with the leading edge partition having a higher weight than other partitions.

[0152] The weight increases with the probability of glaze ice or the incoming wind speed, and is adaptively adjusted between the upper and lower safety limits;

[0153] For example, the initial weights of the three zones can be 0.5, 0.3, and 0.2 when there is no trend and low risk. When the probability of ice formation or the incoming wind speed increases, the weight of the leading edge is increased linearly, with a maximum of 0.7; the weights of the other zones are reduced proportionally to ensure that the sum of the weights is one. The time resolution of the weight update is consistent with the control cycle.

[0154] In one embodiment, dynamic execution is subject to constraints, which include one or more of the following:

[0155] The upper limit of surface temperature of heating zone, the upper limit of heating power, the unit's anti-ice-throwing operation strategy and the minimum shutdown / restart interval, and the protection logic is executed first when any of the constraints is triggered.

[0156] In this embodiment, the upper limit of the surface temperature of the heating zone is 65 degrees Celsius by default, and can be set within the range of 55 to 75 degrees Celsius according to the temperature resistance of the material and coating; the upper limit of the heating power is executed according to the rated power of the zone; the minimum shutdown and restart intervals are both 10 minutes by default, and can be configured within the range of 5 to 20 minutes. When the anti-icing strategy is triggered, the target power supply duration and zone weight are frozen until the protection ends.

[0157] In one embodiment, the mapping database and the ice type discrimination model are jointly updated according to the season or meteorological cycle. During the update, the mapping parameters and model parameters are retrained or recalibrated using the latest running data.

[0158] Optionally, joint updates are performed monthly by default, but can be shortened to two weeks during extreme weather seasons; the data participating in the update must be complete and free of major faults. After the update is completed, the previous version is retained as a rollback version, valid for at least one month.

[0159] In one embodiment, the joint update employs a federated learning framework:

[0160] Each unit trains an ice type discrimination model locally and generates parameter updates. The cloud uses weighted aggregation to obtain a global model before distributing it.

[0161] The update cycle is adaptively adjusted according to seasonal or site climate change strategies, and will revert to the previous stable version if aggregation fails.

[0162] Furthermore, the aggregation uses version number management. If communication is abnormal or the data provided by any site is substandard, the current round of aggregation will skip the data from that site and continue to ensure that the global model is generated on schedule. When two consecutive rounds of aggregation fail, the usage period of the local model is automatically extended and the event is recorded. Aggregation will be performed again after communication is restored.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0164] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for controlling the de-icing of electric heating films on wind turbine blades based on big data, characterized in that, include: Step S1: Obtain the environmental parameters and operating status parameters of the wind turbine. The environmental parameters include at least wind speed, ambient temperature and relative humidity, and the operating status parameters include at least the active power of the turbine and the blade angle. Step S2: Based on historical test and operation data, establish a database of mapping relationships between environmental parameters, operating status parameters and reference power supply duration; Step S3: During unit operation, the ice type discrimination model is used to output the probability distribution of each ice type based on the current environmental parameters and operating status parameters; Step S4: Dynamically correct the baseline power supply duration given by the mapping relationship database according to the probability distribution to obtain the target power supply duration, and control the start and stop of the electric heating film accordingly. Dynamic corrections include: When the probability of glaze ice is not lower than the first threshold, the base power supply duration is multiplied by an adjustment coefficient greater than 1 to extend the power supply. When the probability of frost and ice is not lower than the second threshold, the base power supply duration is multiplied by a reduction factor of less than 1 to shorten the power supply time. The adjustment coefficient and reduction coefficient are updated online or offline based on historical de-icing effect data and are constrained by the current trend of environmental parameter changes. The dynamic correction process includes: Step C1: Perform probabilistic activation and normalization. ； in, The normalized probability after activation. The original probability of the ice type. As the trigger threshold, Pick or , Indicates glaze ice channel, Indicates a frost and ice passage. Corresponding to the first threshold, Corresponding to the second threshold; Step C2, the monotonically increasing mapping of the glaze ice adjustment coefficient is as follows: in, Adjustment coefficient for glaze ice. This is the lower limit of the glaze-ice coefficient, and its value is greater than 1. The upper limit of the glaze ice coefficient and greater than , This is a modulating factor for the glaze ice channel environment. The gain slope of the glaze ice channel is non-negative. The normalized probability of glaze ice obtained in step C1 is represented by the subscript. and These represent the lower and upper limits of the channel, respectively. Step C3, the monotonically decreasing mapping of the frost reduction coefficient is as follows: in, This is the frost reduction coefficient. The upper limit of the frost coefficient is less than 1. The lower limit of the frost coefficient is greater than 0 and less than 0. , As a modulating factor for the frost-ice channel environment, The slope of the frost-ice channel is non-negative. The normalized probability of frost obtained in step C1 is represented by the index. Indicates a frost passage, subscript and These represent the lower and upper limits of the channel, respectively. Step C4, the environmental trend modulation factor and correlation method are as follows: in, It is a composite trend index. The symbol for the hyperbolic tangent function is... These represent the changes in wind speed, ambient temperature, and relative humidity within the prediction time window. To correspond to the reference change, Indicates reference quantity. The channel gain coefficient has a value range of [value range missing]. , These are the environmental modulation factors in steps C2 and C3, respectively; Step C5, glaze ice channel in and Time Monotonous and undiminished, and Frost and ice passages in and Time Monotonic and does not increase, and Threshold Take the open interval (0,1); Step C6, the target power supply duration and dual-channel decision are as follows: in, This represents a relative increase in the glaze ice channel. To reduce the amount of frost and ice passages, The coefficients are the result of comprehensive selection. For absolute value, sign Indicates that, For the target power supply duration, To map the baseline power transmission duration given by the database, the subscript... Indicates the target quantity, subscript Indicates a reference quantity; It also includes power deviation correction: The theoretical cleaning blade power is calculated based on the wind speed-power curve, and the power deviation is obtained by comparing it with the real-time active power of the unit. When the power deviation exceeds the preset range, the target power supply duration is corrected in segments or levels, and the maximum correction range is limited. The power deviation correction process includes: Step D1, in the current control cycle Inside, the theoretical cleaning blade power is obtained based on the wind speed-power curve, and the relative deviation is calculated: in, Indicates period Theoretical cleaning blade power, This represents the wind speed-power curve function. Indicates period wind speed, Indicates period The relative deviation of active power, Indicates period The real-time active power of the unit, The lower bound constant of the denominator, This indicates that the larger of the two options should be chosen. Step D2, with For piecewise variables, a threshold sequence is set and graded correction is performed on both positive and negative sides: when hour: when hour: in, The deviation amplitude, The segmentation threshold, This is the instantaneous power deviation correction factor. To extend the upper limit of the side, To shorten the lower limit of the side, To extend the gain of each segment, To shorten the gain of each segment, Denotes the positive part function, and These represent the subsaturation and supersaturation operations, respectively. Step D3, using multiplicative accumulation and saturation within the interval: in, For the period The cumulative correction factor, and For the cumulative upper and lower bounds, The coefficient for the current period that is actually implemented under cumulative constraints. The target power delivery duration after power deviation correction. The target power supply duration before proceeding to this step. To control the periodic index, This indicates that the power deviation has been corrected. Step D4, Set as the upper boundary of the dead zone to suppress small-amplitude noise. As the wind speed zone and icing risk zoning table are adjusted, the extension side Non-descending order can be used to obtain progressive weighting, shortening the side It can be set independently to reflect asymmetric strategies. Set based on the model and heat load boundary; Step D5, when the wind speed is below the cut-in or above the cut-out zone, let To avoid misjudgment, a threshold is set using a small hysteresis pair. Uplink and downlink are separated to reduce the frequency of round-trip handover; updates will not be performed temporarily in case of packet loss or abnormal power measurement. .

2. The method for controlling the de-icing of wind turbine blades based on big data as described in claim 1, characterized in that, The ice type discrimination model is a machine learning-based model that uses random forest or long short-term memory network. The input includes wind speed, ambient temperature, relative humidity, atmospheric pressure and liquid water content, and the output is a probability distribution of ice types that includes at least frost ice and glaze ice.

3. The method for controlling the de-icing of wind turbine blades using electrothermal film based on big data as described in claim 1, characterized in that, A time series model is used to predict the changing trend of environmental parameters within a preset time window, and the predicted value and the measured value are weighted and fused together when querying the mapping database to obtain the baseline power transmission duration.

4. The method for controlling the de-icing of wind turbine blades using electrothermal film based on big data as described in claim 1, characterized in that, The blades are divided into multiple aerodynamic zones, and ice type discrimination and dynamic correction are performed at the zone granularity. The partitions include at least a leading edge partition, a middle section partition, and a trailing edge partition. The leading edge partition is determined based on a preset arc length range at the front of the blade, while the middle section and trailing edge partitions are preset according to the chord length ratio.

5. The method for controlling the de-icing of wind turbine blades based on big data as described in claim 4, characterized in that, The determination of the target power transmission duration is also based on a weighted calculation of the weights of each partition, with the leading edge partition having a higher weight than other partitions; The weight increases with the probability of glaze ice or the incoming wind speed, and is adaptively adjusted between the upper and lower safety limits.

6. The method for controlling the de-icing of existing wind turbine blades using electrothermal film based on big data as described in claim 1, characterized in that, Dynamic execution is subject to constraints, which include one or more of the following: The upper limit of surface temperature of the heating zone, the upper limit of heating power, the unit's anti-ice-throwing operation strategy, and the minimum shutdown / restart interval are all constraints. When any of these constraints is triggered, the protection logic is executed first.

7. The method for controlling the de-icing of wind turbine blades using electrothermal film based on big data as described in claim 1, characterized in that, The mapping database and the ice type discrimination model are jointly updated according to seasons or meteorological cycles. During the update, the mapping parameters and model parameters are retrained or recalibrated using the latest operating data.

8. The method for controlling the de-icing of wind turbine blades based on big data as described in claim 7, characterized in that, The joint update employs a federated learning framework: Each unit trains an ice type discrimination model locally and generates parameter updates. The cloud uses weighted aggregation to obtain a global model before distributing it. The update cycle is adaptively adjusted according to seasonal or site climate change strategies, and it reverts to the previous stable version when aggregation fails.

Images