Wind power prediction method based on wake flow perception condition residual neural network

Abstract

The invention discloses the technical field of wind turbine generator power prediction and intelligent modeling, and particularly relates to a wind power prediction method based on a wake flow sensing condition residual error neural network, which fully utilizes monitoring data of upstream and downstream units and geometric arrangement of a wind power plant to construct a condition residual error neural network structure with wake flow sensing capability. The wake residual error is conditionally corrected through the gating signal with clear physical significance, the prediction accuracy and robustness of the wake working condition, especially the severe wake working condition, are improved on the premise of ensuring that the prediction precision of the wake-free working condition is not reduced, and the method can be suitable for minute-level and shorter-time-scale wind turbine generator power prediction scenes.

Description

Wind power prediction method based on wake-sensing conditional residual neural network Technical Field

[0001] This invention relates to the field of wind turbine power prediction and intelligent modeling technology, and in particular to a wind power prediction method based on wake perception conditional residual neural network. Background Technology

[0002] With the continuous expansion of wind power grid connection, the volatility and uncertainty of wind farm output power have an increasingly prominent impact on the safe and stable operation of the power system. In order to improve the wind power absorption level, reduce reserve costs and active power regulation pressure, wind power prediction systems based on unit monitoring data and meteorological data are generally deployed on the grid side and the owner side to estimate the output power of wind turbine units in the future period for power generation planning, unit operation optimization and dispatch decision-making. However, the existing technology has the following shortcomings: 1) Most existing wind power prediction models treat the units as independent individuals, and only use the local wind speed, power and other characteristics for modeling. They do not make full use of upstream and downstream unit monitoring data and wake information such as unit geometry. The wake effect is mostly handled in the form of external engineering models or additional features. There is a lack of conditional residual structure and gating mechanism that are closely coupled with upstream and downstream unit monitoring data, unit geometry and wake intensity indicators. Under complex wake conditions, the accuracy and robustness of downstream unit power prediction are reduced, making it difficult to accurately and timely characterize the power deviation of downstream units under wake conditions.

[0003] 2) Within large wind farms, the extraction of incoming wind energy by upstream turbines creates a significant wake effect at downstream turbines, causing the downstream turbines to exhibit multimodal and strongly nonlinear power-wind speed relationships under different wind direction sectors and wake intensities. Existing single network structures that directly superimpose wake-related features onto the input layer typically use a uniform loss function for overall optimization of all samples. This lacks differentiated modeling and residual correction mechanisms for wakeless and wake-based operating conditions, easily leading to problems such as systematic underfitting under severe wake-based conditions or a significant decrease in prediction accuracy under wakeless conditions after increasing the weight of wake samples.

[0004] 3) Existing methods generally lack gating signals and conditional residual channels constructed based on physical quantities such as wake intensity and wind direction sector in model structure and training design. It is difficult to form a wake perception and prediction framework that has both physical meaning and is easy to train stably, thus limiting the engineering application of achieving refined wind power prediction under complex wake conditions.

[0005] For example, the Chinese patent with authorization announcement number CN110188939B uses historical data of adjacent units to calculate wake model parameters and obtains the predicted values ​​of power of each unit and wind power at the field level through parameter prediction model. The wake is mainly reflected in the external engineering model parameters and does not explicitly distinguish different wake conditions in the neural network structure.

[0006] For example, the Chinese patent application with publication number CN114841077A improves the accuracy of single-unit power prediction under complex terrain and wind conditions by downscaling numerical weather forecasts in time and space and combining them with data-driven models. It focuses on improving the resolution of meteorological input, but does not adequately consider the wake coupling of upstream and downstream units within the same wind farm and the gating of operating conditions.

[0007] For example, the Chinese patent with authorization announcement number CN112270454B uses a quantile LSTM equal probability prediction model in the short-term load forecasting of power systems under extreme factors, which focuses on characterizing the distribution characteristics of load under extreme conditions, but does not combine the geometric layout of wind farms and wake intensity characteristics to construct a conditional residual structure for downstream units.

[0008] For example, Chinese patent application CN112329979A proposes an ultra-short-term wind power prediction method based on adaptive deep residual networks. The prediction accuracy is improved by optimizing residual units and adaptive learning rates. However, the residual structure is oriented towards the total power of a single wind farm. It does not distinguish between wake and non-wake conditions, nor does it combine the geometric layout of upstream and downstream units and wake intensity indicators for conditional modeling.

[0009] For example, Chinese patent application CN121072871A proposes a multimodal offshore wind power ultra-short-term prediction method. It uses a multi-scale wake perception graph spatiotemporal prediction model to characterize the spatiotemporal relationship of multiple units and the wake effect. However, the wake information is mainly reflected in the graph structure through a dynamic adjacency matrix. It does not adopt the conditional residual form of baseline subnetwork + wake residual subnetwork + gated subnetwork, nor does it explicitly introduce structural constraints that improve wake conditions and prevent degradation of non-wake conditions.

[0010] The above-mentioned solutions either treat the wake effect as an external physical model parameter or adopt residual or probabilistic structures in the field of load forecasting. However, none of them combine the SCADA data, geometric relationships and wake intensity indicators of upstream and downstream units within the same wind farm to construct a gated conditional residual neural network to specifically constrain the power forecasting performance of downstream units under the condition of "wake condition improvement and non-wake condition no degradation". Summary of the Invention

[0011] The purpose of this invention is to address the problems in existing wind turbine power prediction models, such as insufficient wake perception capability, lack of conditional residual correction structure, and difficulty in balancing prediction accuracy under different operating conditions. This invention provides a wind power prediction method based on a wake perception conditional residual neural network. This method can fully utilize upstream and downstream turbine monitoring data and wind farm geometry to construct a conditional residual neural network structure with wake perception capability. It conditionally corrects the wake residual through physically meaningful gating signals, improving prediction accuracy and robustness under wake-free conditions, especially severe wake conditions, without compromising prediction accuracy. This method is applicable to wind turbine power prediction scenarios with timescales of minutes and shorter.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: a wind power prediction method based on wake perception conditional residual neural network, comprising: 1) Data acquisition and processing: acquiring full-condition operation data of a floating wind-wave combined power generation system through sensors and a high-fidelity model; performing preprocessing such as normalization, denoising, and feature extraction on the raw data; and dividing the data into training and testing sets to provide a high-quality data foundation for subsequent modeling.

[0013] 2) Acquisition and alignment of upstream and downstream unit data: Acquire monitoring data such as wind speed, wind direction, active power, and pitch angle of the target downstream unit and its main upstream units, as well as geometric parameters such as distance between units, relative orientation, and rotor diameter; Align the upstream and downstream data according to a unified time axis to construct baseline feature vector, wake feature vector, and operating condition feature vector; The baseline feature vector includes basic input features such as historical power, wind speed, and wind direction of the downstream units.

[0014] 3) Wake condition identification and intensity feature extraction: Divide the main wake sector according to the wind direction and the relative position of the unit, and identify whether the sample is in the wake sector; Combine the upstream and downstream wind speeds after advection alignment to calculate the normalized wind speed attenuation index, divide the sample into no wake, general wake and severe wake, and smooth the wake intensity to a continuous score of 0 to 1 as the gate monitoring signal.

[0015] 4) Baseline prediction subnetwork construction: Using the baseline feature vector as input, a multi-layer feedforward neural network is constructed to output the baseline power prediction value at the corresponding time. This is used to characterize the typical power-wind speed relationship under conditions of no wake or weak wake. A multi-head structure can be set in the output layer to achieve multi-step prediction.

[0016] 5) The wake residual subnetwork is constructed based on the baseline feature vector input. It introduces the wake feature vector composed of upstream unit wind speed, unit spacing, wind direction offset, and lateral overlap to construct the wake residual subnetwork. A Gaussian wake physical layer is embedded before or in the middle layer of the wake residual subnetwork to approximate the wind speed attenuation or power loss, and the wake residual is output to make additional corrections to the baseline prediction results.

[0017] 6) The wake gating subnetwork and conditional residual fusion select a small number of operating condition features, such as sector marking, wake intensity score and geometric features, as the operating condition feature vector input to construct a lightweight wake gating subnetwork. The output is a gating coefficient in the range of 0 to 1, which represents the weight of the wake residual in the correction. The outputs of the three subnetworks are fused in the form of conditional residual "predicted power = baseline prediction + gating coefficient × wake residual", and the gating coefficient is shared for multi-step prediction scenarios.

[0018] 7) Design and training of joint loss function: The joint loss function is designed, which consists of the overall prediction error loss, the consistency loss of the gating coefficient and the wake intensity score, and the wakeless consistency loss where the predicted value is close to the baseline value on the wakeless sample. The three losses are weighted according to the preset weights, and the stochastic gradient descent algorithm is used to perform end-to-end joint training on the baseline prediction subnetwork, the wake residual subnetwork and the gating subnetwork.

[0019] 8) Online prediction and application: Deploy the trained network in the wind farm operation environment, input the updated upstream and downstream unit monitoring data and wake condition characteristics in real time, calculate the single-step or multi-step power prediction results of the target downstream unit, and use them for power generation planning, unit operation optimization and coordinated control with energy storage, hydrogen production and other devices.

[0020] Compared with the prior art, the beneficial effects of this invention are: 1) Explicitly introducing a wake sensing conditional residual structure: Through the structure of "baseline sub-network + wake residual sub-network + gating sub-network", the predicted power of the downstream unit is expressed as... Automatically reduces f when there is no wake or a weak wake. w Contribution, enhancing f under severe wake conditions w Compared to existing methods that directly output single-path predictions, this approach is more advantageous in simultaneously ensuring prediction accuracy under both normal and wake conditions.

[0021] 2) To ensure that non-wake flows do not degrade, gating supervision and consistency constraints are used. Both the gating coefficient and the wake intensity ΔU are added to the loss function. * Consistency constraints, and on wakeless samples The consistency constraint enables the model to learn the ability to correct wake residuals while keeping the prediction of non-wake samples close to the baseline. Compared with the residual structure that only adds wake features at the input or lacks explicit constraints, it is easier to meet the requirement of "strengthening wake prediction without significantly deteriorating non-wake accuracy" in engineering.

[0022] 3) Embedding upstream and downstream unit geometry and wake physical quantities into the network structure: In the feature and structural design, upstream and downstream SCADA data, along with geometric quantities such as unit spacing, lateral offset, and wind direction deviation angle, are used simultaneously to incorporate wake intensity ΔU. *The system directly inputs gating and residual subnetworks to the operating sector, rather than relying solely on external wake models or field-level meteorological downscaling. Compared to existing technologies, this improves the model's adaptability and transferability to different station layouts and wake intensity levels without requiring fine calibration of complex wake models. Attached Figure Description

[0023] Figure 1 is a schematic diagram of the wind power prediction model based on the wake perception conditional residual neural network; Figure 2 is a flowchart of the wind power prediction method based on the wake perception conditional residual neural network; Figure 3 is a bar chart comparing the MAE of different prediction models under full sample conditions; Figure 4 is a bar chart comparing the MAE of different prediction models under the no-wake condition; Figure 5 is a bar chart comparing the MAE of different prediction models under the general wake condition; Figure 6 is a bar chart comparing the MAE of different prediction models under the severe wake condition. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] The structures, proportions, and sizes illustrated in the accompanying drawings are merely for illustrative purposes and to aid those skilled in the art in understanding and reading the invention. They are not intended to limit the scope of the invention and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, provided they do not affect the effectiveness or purpose of the invention, should still fall within the scope of the technical content disclosed herein. Furthermore, the terms "upper," "lower," "left," "right," "middle," and "one" used in this specification are merely for clarity and not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention's implementation.

[0026] The wind power prediction model structure based on a wake-sensing conditional residual neural network is shown in Figure 1. The model consists of a baseline layer (NN), a wake layer (NN), and a soft-gated (switching) layer. Input features are divided into two parts: general data and wake data. General data includes historical power of downstream turbines, downstream turbine wind speed, wind direction, and the pitch angles of the three blades, and is input to both the baseline layer (NN) and the wake layer (NN). Wake data includes upstream turbine wind speed, wind direction, and geometric information such as the distance and relative offset between upstream and downstream turbines, and is input to the Gaussian wake model sublayer within the wake layer (NN). The baseline layer (NN) outputs the baseline power P. base The wake layer NN outputs wake power P with the assistance of physical characteristics generated by the Gaussian wake model. wake Operating characteristics (such as whether it is in the wake sector, wake intensity ΔU*, etc.) are input into the soft gating layer to obtain the gating coefficient g(z), which is then applied to P. wake After weighting, and Pbase The predicted power output of the downstream wind turbine is obtained by superimposing the results at the adder node.

[0027] Figure 2 shows a wind power prediction method based on a wake perception conditional residual neural network. The method mainly includes minute-level data processing, wake physical characteristic calculation, baseline power prediction, wake power prediction, gating coefficient calculation, and conditional residual superposition. First, SCADA data from upstream and downstream units are aggregated, cleaned, and time-aligned at 1-minute intervals to obtain minute-level data. Then, a Gaussian wake layer with learnable parameters is used to calculate wake physical characteristics based on the minute-level data and geometric information, and these characteristics are input into the wake layer neural network along with general data to obtain P. wake Simultaneously, the general data is input into the baseline layer NN to obtain P. base Then, the sector marker, ΔU*, x / D, y / D, and other operating condition characteristics are input into the gating layer to obtain the gating coefficient g(z). At the "×" node, g(z) is correlated with P. wake The product of P at the "+" node. base The results are summed to output the predicted power output of the downstream wind turbine at time t+Δ.

[0028] The detailed steps of the wind power prediction method are as follows: 1. Data acquisition and preprocessing 1.1 SCADA data acquisition Obtain the raw SCADA data of upstream and downstream units from the existing monitoring system of the wind farm. The typical sampling interval is 1 second, and includes at least the following quantities: Wind speed signal: upstream unit wind speed U upraw Downstream unit wind speed U downraw Power signal: Active power P of upstream unit upraw Downstream unit active power P downraw Wind direction signal: Upstream unit wind direction Dir upraw Downstream unit wind direction Dir downraw Blade pitch angles: Pitch angles of the three blades of the downstream unit: Pitch1, Pitch2, and Pitch3; Other optional operating parameters: relative wind direction of the nacelle, unit operating status code, etc.

[0029] 1.2 Data Cleaning and Aggregation: The original SCADA data undergoes coarse cleaning, removing abnormal samples such as those indicating shutdown, faults, disconnection, obvious power limitations, negative power, and negative wind speeds. The cleaned data is then aggregated according to time windows to obtain minute-level data sequences with a time resolution of 1 minute. The aggregation method can be either arithmetic mean or median. After aggregation, U... up (t), U down (t), P up (t), P down (t), Dir up (t), Dir downPitch1(t), Pitch2(t), Pitch3(t), etc., where t represents a minute-level time index.

[0030] 1.3 Training and Test Set Division: While maintaining the chronological order, minute-level data is divided into training, validation, and test sets based on dates or time periods. For example, the first few days of a month are selected as the training set, the middle few days as the validation set, and the last few days as the test set to avoid information leakage.

[0031] 1.4 Normalization Process: The numerical features used as network input are normalized using standardization or interval scaling methods. Mean-variance standardization can be used. Where ξ represents any input feature to be normalized, μ ξ s ξ These represent the mean and standard deviation of the feature on the training set, respectively, and remain consistent across the validation and test sets. Power output can be calculated based on the unit's rated power P. r Normalization is performed to obtain This improves training stability.

[0032] 2. Upstream and Downstream Unit Geometric Information and Wake Condition Classification 2.1 Geometric Information Acquisition Based on the wind farm design data, the plane coordinates (X, Y, F) of the upstream and downstream units are obtained. up ,Y up ), (X down ,Y down ), hub height H and impeller diameter D. Calculate the azimuth angle θ of the line connecting the centers of the upstream and downstream units. geo and the projected distance x of the upstream and downstream units along the main wake axis ud Horizontal offset y ud : ; lateral offset y ud The value can be calculated from the (X,Y) difference using coordinate rotation; for a selected single upstream and downstream unit pair, if the line connecting the centers of the two units is taken as the geometric principal axis, the lateral offset distance y ud It can be set to 0. The normalized distance is defined as x. D =x ud / D、y D =y ud / D.

[0033] 2.2 Wake main sector division based on geometric connection azimuth angle θ geo Centered on a symmetrical sector half-width Δθ, for example, Δθ = 15°, the geometric wake main sector is defined as [θ]. geo -Δθ,θ geo +Δθ]. Based on the wind direction measured by the downstream unit, Dir down(t), determine whether it is in the geometric wake sector: insector(t) = 1, when Dir down (t)∈[θ geo -Δθ,θ geo +Δθ]; otherwise inector(t)=0.

[0034] 2.3 Advection Alignment and Wake Intensity Calculation Considering the wake propagation delay, the projected distance x between upstream and downstream units along the main wake axis is calculated. ud and upstream wind speed U up (t) Estimate the advection time τ(t) = x ud / U up (t), and round it to the nearest minute integer k to align the upstream signal with the downstream time: U upshift (t)=U up (tk); P upshift (t)=P up (tk).

[0035] When only data from both upstream and downstream is available after alignment, the normalized wind speed attenuation ΔU* is calculated for the downstream unit: ΔU*(t) = 1 - U down (t) / U upshift (t) (1) When U upshift If (t) is too small or anomalies exist, the corresponding sample is discarded. Several operating condition ranges can be set according to ΔU*, for example: ΔU*<0.05 is considered nonwake; 0.05≤ΔU*<0.20 is considered insector; ΔU*≥0.20 is considered severe.

[0036] To facilitate learning by the gating network, ΔU* is linearly or nonlinearly normalized to the interval of 0 to 1 to obtain the wake intensity score s. wake (t).

[0037] 3. Feature Vector Construction 3.1 Baseline Feature Vector At each time t, a feature vector x is constructed for the baseline subnetwork. base (t), including but not limited to: the current wind speed U of the downstream unit down (t); Current wind direction of downstream units Dir down (t); Current blade pitch angles of downstream units: Pitch1(t), Pitch2(t), Pitch3(t); Current or historical power P of downstream units. down (t-1), P down (t-2)... to characterize short-term inertia and hysteresis effects; optional air density estimates, wind speed change rates, etc.

[0038] 3.2 Wake eigenvectors in x baseBased on (t), the wake feature vector x is extended to form the wake residual subnetwork. wake (t), additionally includes: the wind speed U after the upstream wind turbine is aligned with the advection current. upshift (t); Power P after upstream wind turbine advection alignment upshift (t); Upstream and downstream geometric quantities x D ,y D Geometric wind direction error θ err (t)=Dir down (t)-θ geo (Normalized by 360°); Physical characteristics calculated from the Gaussian wake model, such as wake wind speed ratio and overlap coefficient.

[0039] 3.3 Construction of the operating condition feature vector z(t) for the gated subnetwork includes: geometric sector label insector(t); wake intensity score swake(t); x D ,y D ,θ err (t) and other geometric quantities; optional wind speed range markings (e.g., before / after rated).

[0040] 4. Neural Network Structure Design The wake sensing conditional residual neural network in this embodiment mainly consists of a baseline prediction subnetwork f0 and a wake residual subnetwork f... w It consists of four parts: the Gaussian wake physical layer, and the gated subnetwork g.

[0041] 4.1 Baseline Prediction Subnetwork The baseline prediction subnetwork f0 is implemented using a multi-layer feedforward neural network, with x... base (t) is the input, and the output is the baseline power p of the downstream unit at the predicted time t+Δ. base (t+Δ). Let the input dimension be d. base The network contains L0 fully connected hidden layers, each with a width of H0. Activation functions can include ReLU, LeakyReLU, etc. Its forward propagation form is: h0(0) = x base (t) (2)h0(l)=σ(W0(l)h0(l-1)+b0(l)),l=1,2,…,L0(3)p base (t+Δ)=W0(out)h0(L0)+b0(out) (4) where, x base(t) represents the baseline input feature vector of the downstream wind turbine at time t, such as the historical power, wind speed, wind direction, and pitch angle of the wind turbine; h0(0) is the input layer output, h0(l) is the output vector of the l-th hidden layer; W0(l) is the weight matrix connecting the (l-1)-th layer and the l-th layer, b0(l) is the corresponding bias vector, l=1,2,…,L0; σ(·) is the element-wise nonlinear activation function, which can be ReLU, tanh, or other common activation functions; L0 is the number of hidden layers in the baseline subnetwork. W0(out) is the output layer weight matrix, b0(out) is the output layer bias term, p base (t+Δ) represents the downstream wind turbine power prediction value given by the baseline subnetwork at time t+Δ (a scalar for single-step prediction, and a vector composed of multiple prediction step sizes for multi-step prediction). It can be extended to a multi-head structure in the output layer as needed, for example, simultaneously outputting multi-step prediction sequences of Δ=1~5.

[0042] 4.2 Gaussian Wake Physical Layer The Gaussian wake physical layer is embedded before or in the middle layer of the wake residual subnetwork input, with x... wake U in (t) upshift (t), x D ,y D Using these parameters as inputs, the downstream incoming wind speed attenuation and velocity distribution are approximated using the Gaussian wake model. Taking the classic Bastankhah Gaussian model as an example, let the upstream turbine thrust coefficient be C. T If the half-width of the Gaussian wake is σ(x), then the normalized velocity deficit at a distance x can be expressed as: (5), where x represents the downstream position coordinate along the main axis of the wake model; when calculating the wake loss at the target downstream unit location, x = x ud When the wheel hub heights are approximately uniform, the radial distance r can be approximated as the absolute value of the lateral offset distance, i.e., r ≈ y. ud ;δ U (x,r) represents the normalized velocity loss at the downstream location (x,r); U(x,r) represents the actual wind speed at that location; U upshift The upstream wind speed is oriented according to the advection time delay; r is the radial distance relative to the wake centerline, obtained by combining the lateral and vertical offsets; e is the base of the natural constant. The local wind speed ratio can be further constructed based on equation (5). Speed ​​loss δ U The wake physical characteristics, such as (x,r), are denoted as φ. wake (t), and x wake (t) Concatenation forms the enhanced wake input vector .

[0043] 4.3 Wake Residual Subnetwork Wake residual subnetwork f wAccept the enhanced wake input vector As input, a multi-layer feedforward network structure is used to output the wake correction power component p. wake (t+Δ). Its structural form is similar to f0, but the number of layers L w and width h w Different: (6)h w (l)=σ(W w (l)h w (l-1)+b w (l)), l=1,2,…,L w (7)p wake (t+Δ)=W w (out)h w (L g )+b w (out) (8) where, This represents the wake-related input feature vector constructed at time t, including upstream and downstream unit wind speeds, wind directions, unit geometry, and physical features φ obtained from the Gaussian wake layer. wake (t); h w (0) represents the input layer output of the wake residual subnetwork, h w (l) represents the output of the l-th hidden layer; W w (l), b w (l) represents the weight matrix and bias vector of the l-th layer, respectively; σ(·) is the element-wise nonlinear activation function; L w L represents the number of hidden layers in the wake residual subnetwork. g The index of the last layer connected to the output layer (usually L). g =L w );W w (out), b w (out) represents the output layer weight matrix and bias term, respectively; p wake (t+Δ) represents the power correction component output by the wake residual subnetwork at time t+Δ. It is a scalar for single-step prediction and can be expanded into a vector for multi-step prediction, used in conjunction with the baseline power p. base The final predicted power is obtained by adding (t+Δ).

[0044] Like the baseline subnetwork, it can be expanded to a multi-head output for multi-step prediction.

[0045] 4.4 Gated Subnetwork The gated subnetwork g takes the working condition feature z(t) as input and outputs the gate coefficient g(t+Δ)∈[0,1]. It adopts a lightweight fully connected structure with L hidden layers. g Fewer (e.g., 1-2 layers), each layer width H g It is relatively small. Its forward propagation form is: h g(0)=z(t) (9)h g (l)=σ(W g (l)h g (l-1)+b g (l)), l=1,2,…,L g (10) g(t+Δ)=σ gate (W g (out)h g (L g )+b g (out) ) (11) where z(t) represents the operating condition feature vector constructed at time t, including the indication of whether it is in the wake sector, the wake intensity ΔU*, the wind direction sector number and other operating condition-related features; h g (0) represents the input layer output of the gated subnetwork, h g (0) represents the output of the l-th hidden layer; W g (l), b g (l) represents the weight matrix and bias vector of the l-th layer, respectively; σ(·) is the element-wise nonlinear activation function; L g W represents the number of hidden layers in the gated subnetwork. g (out), b g (out) represents the output layer weight matrix and bias term, respectively; σ gate (·) represents the gated output activation function, preferably the sigmoid function, used to restrict the gate coefficients to the interval [0,1]; g(t+Δ) is the gate coefficient at time t+Δ, the closer the value is to 1, the greater the weight of the wake residual correction, and the closer the value is to 0, the more the prediction results of the baseline subnetwork are used. In the case of multi-step prediction, the output of the gated subnetwork can be extended into a vector of length Horizon, giving the corresponding gate coefficients for different prediction step sizes.

[0046] 4.5 Conditional residual output of the final predicted normalized power It is given by the following formula: (12); When restoring to the actual power, multiply by the rated power P. r, Equation (12) reflects the conditional residual structure: when g approaches 0, the model degenerates into a single baseline prediction; when g approaches 1 and p wake When the value is negative, the model provides a significant power correction under severe wake conditions.

[0047] 5. Joint Loss Function and Training Method 5.1 Prediction Error Loss For each sample (t, Δ), the prediction error loss L is defined. pred L1 loss or L2 loss can be used, for example: (13) Among them, P is the predicted power of downstream units output by the model. down (t+Δ) represents the measured power at the corresponding time, where t is the current sample time and Δ is the prediction step size. Formula (13) represents the magnitude of the error between the predicted value and the measured value, which is used to constrain the overall fitting accuracy of the model on the entire sample. The smaller the error, the closer the model's prediction result is to the true power; the larger the error, the more obvious the prediction deviation of the model on the corresponding sample. During training, this loss is averaged over all samples and used as an important component of the total loss function.

[0048] 5.2 The gating supervision loss is introduced as the gating coefficient that guides the output of the gating network to match the actual wake intensity. gate The wake intensity score can be calculated. wake (t+Δ) normalized to [0,1] as a soft label, using L2 or cross-entropy form: L gate =(g(t+Δ)-s wake (t+Δ)) 2 (14) Based on this, weighted calculation can be performed only on samples within the geometric sector or samples whose ΔU* exceeds a certain threshold.

[0049] 5.3 Wakeless Consistency Loss To ensure that the prediction of non-wake samples does not deteriorate, for samples with small ΔU*(t+Δ) (e.g., <0.05) and insert(t+Δ)=0, constraints are applied... With baseline subnetwork output P base Approximately. Let the set of wakeless samples be S. cons ={(t,Δ)∣ΔU*(t+Δ)<0.05,insector(t+Δ)=0}, for S cons Define a single-sample loss l for each sample. cons : Then there is no wake consistency loss L cons for (15) Wherein, P base (t+Δ) represents the predicted power output of the baseline subnetwork f0. This loss is calculated only on samples with no or weak wake, so the gating coefficient g(t+Δ) approaches 0 in such cases, and the wake residual channel does not participate in the correction.

[0050] 5.4 The total loss function, during training, assigns the above three types of losses to the weighted coefficient λ. pred , λ gate , λ cons Weighted summation yields the total loss L. total :L total =λ pred L pred +λ gate L gate +λcons L cons (16) where λ pred >0、λ gate ≥0, λ cons ≥0 can be determined through empirical or validation set adjustments, for example, λ pred =1, λ gate =0.2~0.5, λ cons =0.5~1.

[0051] 5.5 Training Steps (1) Initialize network parameters W0, b w W w b w W g b g The learnable parameters in the Gaussian wake layer can be initialized using Xavier or Kaiming; (2) Construct batch samples in chronological order on the training set, each sample containing x base (t), z(t), label p down (t+Δ) and wake intensity score s wake (t+Δ); (3) x base (t), z(t) is fed into the baseline subnetwork f0 and the wake residual subnetwork f, respectively. w The baseline prediction p is obtained by forward computation of the gated subnetwork g according to equations (2)-(12). base (t+Δ), wake correction term p wake (t+Δ), g(t+Δ), and obtain the total predicted output. (4) Calculate L according to formulas (13) to (16). total (5) Use Adam and other optimization algorithms for backpropagation and parameter update; monitor L on the validation set. pred Alternatively, partition the MAE and select the optimal training round based on the early stopping strategy; (6) Save the optimal model parameters for subsequent online prediction.

[0052] 6. Online Prediction and Application Deployment 6.1 Online Data Access In the actual operating environment of a wind farm, the real-time, second-level data output by the SCADA system is accessed to a server or embedded computing device. Following the rules in the embodiment, coarse cleaning and minute-level aggregation are performed to generate real-time updated U... up (t), U down (t), P down (t), Dir up (t), Dir down Signals such as Pitch1(t), Pitch1(t), etc.

[0053] 6.2 Operating Condition Identification and Feature Generation For each latest time step t, insector(t) is calculated using geometric relationships and wind direction, and ΔU*(t) is calculated using historically aligned upstream and downstream wind speeds. The wake intensity score s is obtained according to equation (1) and the threshold values ​​for each interval. wake (t). Simultaneously construct x base (t), z(t) is normalized using the same normalization method as during the training phase.

[0054] 6.3 Network forward inference will x base (t), z(t) is input to the trained baseline subnetwork f0 and the wake residual subnetwork f0. w And the gated network g, to obtain p base (t+Δ), p wake (t+Δ), g(t+Δ), and then obtain the normalized predicted value according to equation (12). And multiply by the rated power P r get .

[0055] 6.4 Multi-step Rolling Forecasting: Based on actual needs, multiple forecast step sizes Δ∈{1,2,3,4,5}min can be set. A multi-head output or cyclic network call method is used to obtain the power forecast sequence for the next few minutes. The forecast results are output to the wind farm dispatching system, the unit's local control system, or the superior dispatching center for rolling power generation plan preparation, ramp constraint checking, and power coordination control with energy storage, hydrogen production, and other equipment.

[0056] 6.5 Model Updates During the long-term operation of the system, newly added SCADA data can be periodically added to the training set, and the network parameters can be retrained or fine-tuned according to the above steps to adapt to factors such as seasonal changes, equipment aging, and adjustments to operating strategies, thereby improving the long-term stability and generalization ability of the model.

[0057] Figure 3 shows a bar chart comparing the MAE of different prediction models under full-sample conditions. The horizontal axis represents different prediction models, including the method of this invention (conditional residual neural network), wake-only structure, wakeless structure, full wakeless structure, as well as RF, SVR, TCN, LSTM, RNN, GRU, CNN, TFT, etc. The wake-only structure means that only the wake layer output is retained as the final prediction result; the wakeless structure means that only the baseline layer output is retained as the final prediction result, its equivalent form being soft-gated and always 0; the full wakeless structure means that the baseline layer and the wake layer are directly added to obtain the final prediction result, its equivalent form being soft-gated and always 1. The vertical axis represents the mean absolute error (MAE) (kW) on the full-sample test set under the condition Δ=1. The red bars on the left represent the conditional residual model of this invention, the green columns represent the ablation experiment models under different structural settings, and the blue columns on the right represent the comparison models between traditional machine learning and deep learning, used to demonstrate the predictive performance advantage of this invention on the overall sample.

[0058] Figure 4 shows a bar chart comparing the MAE of different prediction models under the wakeless condition. The MAE of the proposed method, wake-only structure, wakeless structure, full wake structure, and models such as RF, SVR, TCN, LSTM, RNN, GRU, CNN, and TFT are compared on wakeless condition samples. In the figure, the red bars still correspond to the proposed model, the green columns represent the ablation results of different wake processing methods, and the blue columns represent the performance of various comparison models, illustrating that the proposed method maintains or outperforms existing methods in prediction accuracy under the wakeless condition.

[0059] Figure 5 shows a bar chart comparing the MAE of different prediction models under typical wake conditions. The MAE of each model is compared on a sample of typical wake conditions. This figure reflects that under moderate wake disturbances, the conditional residual structure of this invention fully utilizes wake information while suppressing unnecessary downward corrections, resulting in lower errors compared to using only the baseline model or a simple wake correction model.

[0060] Figure 6 shows a bar chart comparing the MAE of different prediction models under severe wake conditions. The MAE of each model is compared on a sample of models operating under severe wake conditions. It can be seen that when ΔU* is large and the wake effect is significant, the method of this invention, corresponding to the red bar, has the lowest error among the various comparison models. This indicates that introducing a Gaussian wake physical layer in the form of gated residuals can more accurately capture the power loss caused by severe wakes and improve the prediction reliability of downstream units under extreme conditions.

[0061] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A wind power prediction method based on a wake-sensing conditional residual neural network, characterized in that it includes: (1) Data acquisition and processing; Acquire full-condition operating data of the floating wind and wave combined power generation system and preprocess the raw data; (1) Divide the training set and the test set; (2) Acquire and align the upstream and downstream unit data, acquire the monitoring data and geometric parameters of the target downstream unit and its main upstream unit; perform advection alignment of the upstream and downstream unit data according to the unified time axis, and construct the baseline feature vector, wake feature vector and operating condition feature vector; (3) Identify the wake operating condition and extract the intensity feature, divide the wake main sector according to the wind direction and the relative position of the unit, and identify whether the sample is in the wake sector; Combined with the wind speed of the upstream and downstream units after advection alignment, the normalized wind speed attenuation index is calculated. The samples are divided into no wake, general wake and severe wake, and the wake intensity is smoothly mapped to a continuous score of 0 to 1 as the gated supervision signal; (4) Baseline prediction sub-network construction: with the baseline feature vector as input, a multi-layer feedforward neural network is constructed to output the baseline power prediction value at the corresponding time, which is used to characterize the typical power-wind speed relationship under no wake or weak wake conditions. A multi-head structure is set in the output layer to realize multi-step prediction; (5) Wake residual sub-network construction: based on the baseline feature vector input, the wake feature vector is introduced to construct the wake residual sub-network; a Gaussian wake physical layer is embedded in the wake residual sub-network before or in the middle layer to approximate the wind speed attenuation or power loss, and the wake residual is output to make additional corrections to the baseline prediction results; (6) Wake gate sub-network and condition residual are fused: with the operating condition feature vector For input, output gating coefficient, adopt a lightweight fully connected structure, construct a lightweight wake gating sub-network, output gating coefficient in the range of 0 to 1, representing the weight of wake residuals participating in correction; merge the outputs of the three sub-networks according to the conditional residual form of "predicted power = baseline prediction + gating coefficient × wake residual", and share the gating coefficient for multi-step prediction scenarios; (7) Joint loss function design and training, design a joint loss function, weight the three losses according to the preset weights, and use the stochastic gradient descent algorithm to perform end-to-end joint training of the baseline prediction sub-network, wake residual sub-network and gating sub-network; (8) Online prediction and application, deploy the trained network in the wind farm operation environment, input the updated upstream and downstream unit monitoring data and wake condition characteristics in real time, calculate the single-step or multi-step power prediction results of the target downstream unit, and use them for power generation planning, unit operation optimization and collaborative control with energy storage and hydrogen production devices.

2. The wind power prediction method based on wake-sensing conditional residual neural network as described in claim 1, characterized in that, The data acquisition and processing in step (1) specifically involves acquiring full-condition operating data of the floating wind and wave combined power generation system through sensors and high-fidelity models; performing preprocessing such as normalization, denoising, and feature extraction on the raw data; and dividing the data into training and testing sets to provide a data foundation for subsequent modeling.

3. The wind power prediction method based on wake-sensing conditional residual neural network as described in claim 1, characterized in that, The monitoring data in step (2) includes wind speed, wind direction, active power and pitch angle; the geometric parameters include the distance between units, relative orientation and impeller diameter; the basic input characteristics of the downstream units include historical power, wind speed, pitch angle and wind direction.

4. The wind power prediction method based on wake-sensing conditional residual neural network as described in claim 1, characterized in that, The advection alignment in step (2) specifically involves taking into account the wake propagation delay, estimating the advection time based on the upstream and downstream distances and the upstream wind speed, and rounding it to the nearest minute to align the upstream signal with the downstream time.

5. The wind power prediction method based on wake-sensing conditional residual neural network as described in claim 1, characterized in that, The specific steps in step (3) of dividing the wake main sector are as follows: taking the azimuth angle of the geometric connection line as the center, setting the half-width of the symmetrical sector, defining the geometric wake main sector, and determining whether it is in the geometric wake sector based on the wind direction measured by the downstream unit.

6. The wind power prediction method based on wake-sensing conditional residual neural network as described in claim 1, characterized in that, In step (5), a wake residual subnetwork is constructed, and a multi-layer feedforward network structure is adopted to output the wake correction power component, which is expanded into a multi-head output for multi-step prediction.

7. The wind power prediction method based on wake-sensing conditional residual neural network as described in claim 1, characterized in that, The joint loss function in step (7) consists of the overall prediction error loss, the consistency loss between the gating coefficient and the wake intensity score, and the wakeless consistency loss where the predicted value is close to the baseline value on the wakeless sample.

8. The wind power prediction method based on wake-sensing conditional residual neural network as described in claim 7, characterized in that, The total loss function of the joint loss function in step (7) is obtained by weighting and summing the prediction error loss, the consistency loss of the gating coefficient and the wake intensity score, and the consistency loss without wake according to the weight coefficients during the training process.

9. The wind power prediction method based on wake-sensing conditional residual neural network as described in claim 7, characterized in that, The end-to-end joint training in step (7) includes: (a) initializing network parameters and learnable parameters in the Gaussian wake layer; (b) constructing batch samples in chronological order on the training set; (c) feeding the baseline input feature vector, enhanced wake features, and operating condition features into the baseline sub-network, wake residual sub-network, and gating sub-network, respectively, and performing forward calculations to obtain the baseline prediction, wake correction term, and gating coefficients, and obtaining the total prediction output; (d) calculating the total loss, and using the Adam optimization algorithm for backpropagation and parameter updates; (e) monitoring the prediction error loss or partitioned mean absolute error on the validation set, and selecting the optimal training rounds according to the early stopping strategy; (f) saving the optimal model parameters for subsequent online prediction.

10. The wind power prediction method based on wake-sensing conditional residual neural network as described in claim 1, characterized in that, The multi-step rolling prediction in step (8) specifically involves setting multiple prediction step lengths according to actual needs, and using a multi-head output or cyclic network call method to obtain a power prediction sequence for the next few minutes; the prediction results are output to the wind farm dispatch system, the unit local control system or the upper-level dispatch center for rolling power generation plan preparation, ramp constraint check and power coordination control with energy storage and hydrogen production equipment.

Images