Method for predicting ultra-short-term wind power based on improved UKF algorithm

Abstract

The invention belongs to the technical field of wind power prediction, and particularly relates to a method for predicting ultra-short-term wind power based on an improved UKF algorithm. According to the method, a square root UKF, normalized innovation square gating and self-adaption, multi-information joint updating and multi-model interaction technologies are fused on the basis of the UKF. The method comprises the following steps: establishing a nonlinear state space model, generating a propagation sigma point by SR-UKF to complete time updating, acquiring, measuring and executing NIS gating, fusing and innovating a multi-information strategy and self-adapting to a noise covariance, constraining a projection to ensure that a physical boundary and the covariance are positive definite, fusing multiple working condition probabilities by IMM, and outputting a prediction result and uncertainty information. According to the method, linearization errors are avoided, and prediction robustness is improved; the problems of phase lag, unstable numerical value and the like are solved through multiple technologies respectively, second-to-minute level accurate prediction is achieved, and meanwhile deliverable capacity information capable of being directly used for control and performance is output.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, specifically involving a method for predicting ultra-short-term wind power based on an improved UKF algorithm. Background Technology

[0002] In actual power grid operation, especially when wind farms participate in ancillary services such as frequency regulation and peak shaving, the system has extremely high requirements for the accuracy and timeliness of dynamic response to wind power. To meet the fulfillment requirements of ancillary services and ensure grid security, the dispatch center needs to issue power adjustment commands on a timescale of seconds to several seconds to wind farms. Wind farms, in turn, need to quickly decide on the magnitude of power increase or decrease within a timescale of seconds to minutes. This process is highly dependent on the accurate prediction of wind power output in the very short term, i.e., ultra-short-term wind power forecasting.

[0003] Currently, various technical solutions have been developed in the field of ultra-short-term wind power forecasting. Common methods can be broadly categorized as follows: first, statistical baselines based on historical sequences; second, physical post-processing centered on numerical weather prediction; third, machine learning and deep learning, and their hybrid frameworks with numerical weather prediction or nacelle wind speed; and fourth, state-space and filtering methods. These methods have their own engineering applications, but they also share common and unique shortcomings. Furthermore, from a regulatory perspective, these methods generally suffer from the problem of providing good-looking predictions but insufficient information on deliverable capabilities. They struggle to directly provide information on uphill / downhill turning points, ramp limits, and minute-level energy budgets, to smoothly transition during periods of missing data, and to integrate multi-source measurements into the same time domain in a structured manner. However, the basic unscented Kalman filter algorithm is prone to several failure mechanisms in ultra-short-term realistic predictions of wind power from seconds to minutes. These include innovation bias and variance underestimation caused by model and object mismatch and nonlinear saturation, colored innovation and phase lag caused by asynchronous numerical weather prediction and first-order power inertia, numerical instability or non-positive covariance, statistical inconsistency, and single-model likelihood degradation and erroneous updates caused by discrete operating condition switching. Summary of the Invention

[0004] This invention aims to provide a method for predicting ultra-short-term wind power based on an improved UKF algorithm. It improves upon the Unscented Kalman Filter (UKF) by using a Square-Root Unscented Kalman Filter (SR-UKF) to maintain the square root of the covariance matrix. It incorporates Normalized Innovation Squared (NIS) gating and adaptation, Multiple Innovation (MI) joint updates, and Interacting Multiple Model (IMM) to address the shortcomings of the basic UKF algorithm in four dimensions: dynamic tracking, numerical stability, statistical calibration, and operational robustness. Furthermore, it significantly improves the reliability of ultra-short-term wind power prediction and delivery capability assessment with controllable computational costs.

[0005] To achieve the above objectives, the present invention provides a method for predicting ultra-short-term wind power based on an improved UKF algorithm, comprising the following steps: Step S1: Establish a nonlinear state-space model that includes effective wind speed, short-term trend, wind speed deviation, and power; Step S2: Generate and propagate sigma points using square root unscented Kalman filtering at a preset second-level sampling period to complete the time update; Step S3: Collect at least one on-site measurement of SCADA power, nacelle wind speed, and lidar, and perform normalized innovative square gating; Step S4: In the measurement update, a multi-innovation strategy is adopted to fuse the most recent L-step innovations, and the process and observation noise covariances are adapted online through covariance matching and / or fading factors; Step S5: Apply constraint projection to the update results to satisfy physical boundaries such as power and ramp rate, while maintaining positive covariance; Step S6: Construct an interactive multi-model based on discrete working conditions and fuse the mode probabilities; Step S7: Output the current filtered power and the predicted power at h steps, as well as the corresponding variance or confidence interval.

[0006] The working principle and beneficial effects of this scheme are as follows: The core of unscented Kalman filtering is to explicitly write the nonlinear mechanism of wind speed, power curves, and the first-order inertia of the power station into a state-space model, and to directly propagate the mean and covariance on the original nonlinearity using unscented transform (UT), thus avoiding the Jacobian linearization error of extended Kalman. Compared with extended Kalman, it is more stable and closer to the true response near the cut-in, rated plateau, and cut-out points. At the same time, the inherent uncertainty of output state and power allows point predictions to be directly mapped to a computable quantity of deliverable capability. Compared to purely data-driven methods, UKF incorporates implicit elements such as trends, drift, and biases into the state vector, enabling online self-calibration of cabin wind speed deviations, wakes, and wind direction changes. Compared to post-processing in numerical weather prediction, UKF focuses on phase correction and noise suppression in the last mile at the minute level, providing second- to minute-level delivery trajectories within the coarser spatiotemporal framework of numerical weather prediction. Compared to traditional statistical baselines, UKF escapes the persistence trap more quickly during abrupt changes and turning points, while possessing a smooth degradation mechanism that advances only the prediction step for missing data. This allows UKF to not only improve the robustness of ultra-short-term point predictions but, more importantly, transform the prediction results into variables that can be directly used for control and market fulfillment.

[0007] (1) To address the issues of innovation and phase lag in non-ferrous metals, the multiple innovation theory is used to reflect the most recent... Step innovation is to update the block diagonal covariance together, pre-whiten the equivalent time domain and improve the utilization rate of effective information, thereby accelerating the convergence of the transition segment and significantly reducing the second-level jitter; (2) Due to the numerical instability and boundary shearing caused by the unscented Kalman filter algorithm during prediction, square root UKF and constraint projection are added, and trigonometric factors are propagated in the time and measurement channels using QR or Cholesky. After the update, feasible domain projection with Mahalanobis distance minimization and covariance consistency are performed to ensure positive definiteness of the matrix and achieve boundary stability; (3) When encountering statistical inconsistency, time-varying noise and abnormal measurement that may drag the state off, resulting in inaccurate prediction, NI is added. S-gating and covariance matching or fading factor are used to perform statistical consistency monitoring and abnormal measurement gating with NIS, and Q and R are adjusted online according to the NIS deviation to calibrate the coverage to the standard level, which is more robust to abnormal measurement; (4) Since wind power often switches between discrete modes, while UKF only has a single continuous model, it will equate institutional jumps to process noise, resulting in deviations, etc. IMM-UKF is used to perform Markov interaction and likelihood weighted fusion between normal, limited or cut-out multiple models to achieve hybrid prior, parallel filtering and likelihood weighted multimodal Bayesian model selection, avoiding mistaking institutional jumps as process noise.

[0008] Optionally, a state-space model is established according to step S1, including the following steps: Step S1.1: Align and preprocess the data, setting the discrete time step. Discrete-time index The original measurement was power. Cabin wind speed LIDAR / NWP: Resample all channels to a uniform step size. Perform external control on outliers first, and only perform prediction steps for missing steps. Step S1.2: For continuous state variables Control input Observable measurement Define: , , (1) In equation (1), For discrete-time indexing, Let be the system state vector, where For effective wind speed status, For the short-term trend or acceleration state of wind speed, For wind speed deviation or scale correction status, This refers to the active power status of the power station. To control or exogenous input vectors, only the power upper limit for this step is included here. This represents the maximum power available in this step. Let be the measurement vector, where Power measurements obtained by SCADA For cabin wind speed measurement, Optional auxiliary measurement; Step S1.3: Provide the time evolution model, give the nonlinear function for UKF time update, and the nonlinear state equation is, according to the mode : (2) in, For the process noise of each state, see equation (5). For pattern Downtrend The autoregressive coefficient, and , For pattern The first-order inertial time constant of the lower power The power curve from wind speed to steady-state power is shown in equation (3). This is the set of parameters for the power curve. For pattern Lower deviation state The autoregressive coefficient, and ; (2a) in, For the scalar input that is being limited, this is usually the case here. The output is shown in equation (3). As the lower limit, when Then output , As the upper limit, when Then output , This is a bilateral limiting function used to clamp the target power within... Inside; (2b) in, For pattern The noiseless state evolution function is consistent with the right-hand side of equation (2). For the state vector, see equation (1). For control or exogenous inputs, see equation (1). For power state components, according to first-order inertia Converging towards the target power after limiting, parameters (or ), See (2a), For the steady-state mapping of the power curve to wind speed, see equation (3), where The effective wind speed after deviation correction. Power curve parameter set See equation (3). This is the maximum permissible power for this step; (2c) in, Let be the random state vector for the next time step. For pattern For noiseless propagation, see (2b), at time use Evaluate, For pattern process noise vector See equation (2). It follows a zero-mean Gaussian distribution with covariance of See equation (5). For pattern The process noise covariance matrix is ​​given in equation (5), with subscripts... , To represent the time index and the discrete pattern index respectively, consistent with equations (2) and (5); Step S1.4: Map wind speed to a steady-state power target and define the power curve: (3) in, Wind speed The corresponding steady-state power target, The cut-in wind speed for the fan. Rated wind speed, To cut off the wind speed, The scaling factor for the cubic segment. Rated power of the equipment; Step S1.5: Map the state to the sensor measurement space. The measurement equation is: (4) in, For nonlinear mapping from state to measurement, For the observed noise vector, see equation (5). For the fixed offset of the nacelle wind speed channel; Step S1.6: Specify the statistical characteristics of the observed noise, and the noise model: (5) in, For process noise, the components are shown in equation (2). To observe the noise, It follows a zero-mean Gaussian distribution. For pattern The process noise covariance matrix is ​​as follows. For model The observation noise covariance matrix is ​​given below, where The covariance is semidefinite.

[0009] Power on a second to minute scale Effective wind speed Short-term trend Wind measurement deviation Operating limits Joint influence on continuous state quantities Control input Observable measurement This definition provides a variable interface for the state or measurement equations, laying the foundation for the dimensions and meanings of subsequent nonlinear mapping and noise injection; wind speed can be approximated as being driven by trends in the short term, and power exhibits first-order dynamic inertia. Furthermore, given the limitations imposed by the quota and the slow-varying bias in wind measurement, a time evolution model is presented to describe the evolution of the state over time, and a nonlinear function for UKF time updates is provided. This lays the physical foundation for subsequent prediction step sigma point propagation and extrapolation; the wind turbine power exhibits strong nonlinear segmentation with respect to wind speed, such as cut-in, cubic segment, rated, or cut-out, mapping wind speed to a steady-state power target, improving the accuracy of the nonlinear region, and laying the foundation for accurate characterization of saturation or boundary behavior; what is actually measurable is power, nacelle wind speed, and possible auxiliary quantities, mapping the state to the sensor measurement space, laying the foundation for subsequent innovation and likelihood calculation; UKF requires a noise model with covariance injection and consistency verification, laying the foundation for subsequent adaptive and adjustable NIS.

[0010] Optionally, according to step S2, sigma points are generated and propagated using square root unscented Kalman filtering within a preset second-level sampling period, including the following steps: Step 2.1: Set the discrete scale of the sigma points, and perform an unscented transformation: Regarding the dimension of the state ,Pick , , (6) in , For UT scaling parameters, For UT secondary scaling parameters, Let be the state dimension. This is the combined scaling factor for UT. The offset scale constant of UT ; Step 2.2: Calculate the mean and covariance of the propagated sigma points using weighted averages. The UT weights are: (7) in, For the first The average weight of each sigma point. For the first Covariance weights at sigma points The parameters are the second-order corrections made by UT to the Gaussian distribution. The value is usually 2, and the index is... ,and Center point; Step 2.3: Sigma point based on square root covariance S: (8) in, For the first A state sigma point, This is the mean of the current state used for sampling. Let Cholesky's factor be the state covariance, such that... , For matrix The List, For the UT offset scale, see equation (6). Let be the state dimension, see equation (6).

[0011] The discrete scale of the sigma points is set to ensure matching of the second moment under Gaussian conditions and control numerical stability. The weighted mean and covariance of the propagated sigma points are calculated to ensure the second-order accuracy and unbiasedness of the unscented transformation, laying the foundation for the accuracy of subsequent numerical integration.

[0012] Optionally, the IMM algorithm's pattern interaction includes the following steps: Step 1: Calculate the pattern from the previous step To this step mode The conditional weights, the mixed probability is (9) in, Mode transition matrix elements, For the previous mode The posterior probability, For pattern The normalization constant, To give the pattern From pattern Conditional mixed weights, index For pattern index; Step 2: Generate the prior mean and covariance for each pattern, and synthesize them using square roots: (10) in, For pattern The mixture of prior means, For pattern Mixed prior covariance, For the previous step mode The posterior mean, For the previous step mode The posterior covariance, For the conditional mixing weight, see Equation (9), where the subscript 0 indicates the prior state at the start of this filtering step.

[0013] Optionally, scaling parameters are first set in the unscented transform, and sigma points are generated using the square root trigonometric factor of the current covariance. Mixed priors for each mode are obtained through IMM. Subsequently, in the time update phase, the sigma points are propagated through the nonlinear state equation, and the square root of the prior covariance is directly recursively derived using QR / Cholesky. In the measurement update phase, measurement predictions and innovations are first calculated, and statistical tests are performed on the arriving measurements using NIS gating. Then, covariance matching and fading factors are combined to adaptively adjust the process noise and observation noise covariances. A multi-innovation strategy is introduced, where the most recent L-step innovations are adjusted according to the attenuation coefficient. The steps are as follows: Step 1: Time update, advance the sigma point to the next time step through the state equation: by , Based on the equation (8), generate the Sigma point. ,by (11) Propagate Sigma points, where, For the first time after the spread 1 Sigma point, For pattern The state evolution function is shown in Equation 2. For the sigma point before propagation, see equation (8). For control or exogenous inputs, see equation (1). For pattern index; Step 2: Calculate the prior mean from the propagated sigma points: (12) in, For pattern The mean of the prior states, For the mean weight, see equation (7). For the prior sigma point, see equation (11). For the state dimension, see equation (6); Step 3: Synthesize the square root of the prior covariance And inject process noise: (13) Among them, The square root of the process noise covariance matrix. For the covariance weights, see equation (7), for Perform QR decomposition and take the upper triangle R, so that , The Cholesky factor for the prior covariance; Step 4: Extract the prediction and uncertainty of the power component from the state prior. The prior power is: definition (14) in, Choose a vector for power. For pattern The current step prior power prediction, Let Variance be the variance of the prior power prediction. It is the square of the second norm of the vector; Step 5: Map the prior sigma points to the measurement space and calculate their expectation. The measurement is then updated as follows: , (15) in, For the first The mapping result of a prior sigma point in the measurement space. For the measurement mapping of equation (4), For pattern The prior mean of the measurement, For the mean weight, see equation (7). For the prior sigma point, see equation (11); Step Six: Synthesize the Square Root of the Innovative Covariance : (16) Among them, The square root of the observation noise covariance matrix, for Perform QR decomposition and take the upper triangle R, so that , The Cholesky factor for innovative covariance; Step 7: Measure the covariance between the state and the measurement, where the cross-covariance is: (17) in, See equations (11), (12), (15), (15), and (7) respectively; Step 8: Computational Innovation With normalized innovation square : untie (18) in, These are the actual measurements and the predicted measurements, respectively, see equations (1) and (15), trigonometric solutions. Avoid explicit inversion. for The numerically stable notation uses a threshold. upper valve and The lower valve is used for gating and consistency monitoring, where m is the measurement dimension. Step 9: When NIS is too high, increase the process noise and adjust it using the fading factor. : (19) in, This is the dimension normalization statistic for NIS. To adjust the step size parameter for process noise, To Adaptive amplification factor Indicates using Update process noise covariance; Step 10: Match the observed noise to the actual innovation statistics, adjust R: (20) in, To observe the step size parameters that adapt to noise, The nearest positive semidefinite projection operator projects the matrix into a positive semidefinite form. For innovation, see equation (18). To innovate the covariance, refer to Equation (16), and use covariance matching to update the observation noise covariance. ; Step 11: Combine the most recent L-step innovations and their covariances according to decay, and explicitly utilize time correlation: , (twenty one) in, For length is The multi-information stack vector, The decay coefficient of multiple innovations, The value ranges from 0 to 1, but is typically 0.7 to 0.9. For multiple information window lengths, It is usually 3 to 6. To construct operators for block diagonal matrices, For the first The innovation covariance Cholesky factor for the step is shown in equation (16). For innovation, see equation (18); (twenty two) in, For the first The cross-variance between step state and measurement is given in equation (17); Step 12: Calculate the Kalman gain: (twenty three) in, For the multi-information covariance, see equation (22). For the block diagonal triangular factor of the multi-innovation covariance, see Equation (21); Step Thirteen: Update the Mean and SR Covariance: (twenty four) in, For pattern The posterior mean of the states. For pattern The posterior covariance Cholesky factor The Cholesky update operator for reducing the order of trigonometric factors, For the Kalman gain, see equation (23). For the multiple innovations and their covariance factors, see equation (21). For the prior mean and factors, see equations (12) and (13); Step Fourteen: Extract the current power estimate and uncertainty from the posterior state: (25) in, For pattern The current power posterior estimate, The uncertainty of this posterior estimate, For selecting a power vector, see equation (14). For the posterior mean and factor, see Equation (24).

[0014] Optionally, the updated results are subjected to constraint projection according to step S5 to satisfy physical boundaries such as power and ramp rate, while maintaining positive covariance, including the following steps: Step S5.1: Define the physically feasible region for power, wind speed, or ramp rate, and the set of wind power physical boundaries and operational constraints. for: (26) in, This represents the change in power between adjacent steps. Maximum climbing ability per unit time; Step S5.2: Project the posterior mean back into the feasible region using the covariance measure: (27) in, This represents the variable that minimizes the objective function. It is the inverse of the posterior covariance. is the unprojected posterior mean (see Equation (24)). The feasible region is (see equation (26)). For the posterior covariance, by get See equation (24); Step S5.3: Make the covariance consistent with the projected mean: (28) in, To constrain the projection mapping on Jacobian matrix at the location, The Cholesky decomposition of a symmetric positive definite matrix. The Cholesky factor for uniformizing covariance. For the unprojected posterior covariance, see Equation (24).

[0015] Wind power has hard boundaries, clearly defining the physical feasible region for power, wind speed, or ramp rate, laying the foundation for a feasible region consistent with projection and covariance. Projecting the posterior mean back into the feasible region under covariance measurement is crucial; hard shearing can disrupt statistical consistency, so minimizing Mahalanobis distance is the most reasonable approach, providing feasible and stable initial values ​​for subsequent extrapolations. Projection alters local geometry, necessitating the use of Jacobian. Correcting uncertainties and achieving covariance uniformity lays the foundation for long-term numerical stability.

[0016] Optionally, according to step S6, an interactive multi-model is constructed based on discrete working conditions and the mode probabilities are fused, including the following steps: Step S6.1: Use innovative statistical measures to measure the pattern The compatibility of the two components, with a likelihood value of: (29) in, Let be the Gaussian normalization constant. , It is a triangular factor determinant. For innovation, see equation (18). For the innovative covariance factor, see equation (16); Step S6.2: Update the posterior probability of the pattern at the current time step: , (30) in, For pattern The likelihood value is given by equation (29). Let be the normalization constant of equation (9). The subscript is the sign for summing over all patterns. The current moment; Step S6.3: Fuse the current power estimates from multiple models into a single output and factor in the inter-model discretization: (31) in, For the current power estimate after fusion, To estimate the variance of the current power after fusion, For the current power estimates and variances of each model, see Equation (25). For the posterior probability of the pattern, see equation (30). The variance term contributing to the discrete inter-model contribution; Step S6.4: The current step prior power prediction for fusion is: , (32) in, For the current step prior power prediction of fusion, Its predicted variance, including the inter-model dispersion, For the prior power and variance of each model, see Equation (14). For the posterior probability of the pattern, see equation (30).

[0017] Using innovative statistical measures to measure patterns The compatibility of the model is automatically identified as normal, limited, or switched off, avoiding single-model bias and serving as the basis for calculating the posterior probability of the model. The model posterior probability at the current moment is obtained, and the output is automatically weighted to the most compatible model, laying the foundation for a single power output. The current power estimates of multiple models are merged into a single output and the discrepancies between models are taken into account, improving the robustness and interpretability under switching conditions and laying the foundation for initial extrapolation values.

[0018] Optionally, output the next 0-5 minutes according to step S7. Power prediction and variance at each step: (33) in, To predict the number of steps, To cover the maximum predicted steps in the 1-5 minute range, This is obtained by updating only the time from the current time. Step state prediction mean The corresponding state prediction covariance Cholesky factor. for Step power prediction mean for Step power prediction variance. Attached Figure Description

[0019] Figure 1This is a flowchart of an embodiment of the present invention; Figure 2 This is a comparison chart of the actual power output and short-term power prediction of a wind farm in a certain location on a certain day, as shown in an embodiment of the present invention. Figure 3 This is a performance comparison table of different prediction methods in the embodiments of the present invention. Detailed Implementation

[0020] The following detailed description illustrates the specific implementation method: Example This embodiment is basically as follows: Figure 1 , Figure 2 As shown: The method for predicting ultra-short-term wind power based on the improved UKF algorithm has the following specific implementation steps: Step S1: Establish a state-space model. Because the mapping from wind speed to power is highly nonlinear and there are station response inertia and wind measurement bias, explicit physics plus low-dimensional state characterization is used, which can be small-sample, interpretable, and online self-calibrated.

[0021] Step S1.1: Data alignment and preprocessing to ensure all measurements participate in filtering at the same discrete time, enabling innovative statistics to be used for NIS consistency and adaptation. Let the discrete time step be... Discrete-time index The original measurement was power. Cabin wind speed LIDAR / NWP resamples all channels to a uniform step size. Outliers are first controlled externally, and missing steps are only used for prediction.

[0022] Step S1.2: For continuous state variables Control input Observable measurement Define: , , (1) In equation (1), For discrete-time indexing, Let be the system state vector, where For effective wind speed status, For the short-term trend or acceleration state of wind speed, For wind speed deviation or scale correction status, This refers to the active power status of the power station. To control or exogenous input vectors, only the power upper limit for this step is included here. This represents the maximum power available in this step. Let be the measurement vector, where Power measurements obtained by SCADA For cabin wind speed measurement, This is an optional auxiliary measurement.

[0023] Due to power on a second to minute scale Effective wind speed Short-term trend Wind measurement deviation Operating limits The combined effects of this step on continuous state variables Control input Observable measurement Defined as follows: the solution quantity is the state to be predicted subsequently. With the future This provides a variable interface for the state or measurement equations, laying the foundation for the dimensions and meaning of subsequent nonlinear mapping and noise injection.

[0024] Step S1.3: Provide the time evolution model, give the nonlinear function for UKF time update, and the nonlinear state equation is, according to the mode , (2) in, For the process noise of each state, see equation (5). For pattern Downtrend The autoregressive coefficient, and , For pattern The first-order inertial time constant of the lower power The power curve from wind speed to steady-state power is shown in equation (3). The set of parameters for the power curve See equation (3). For pattern Lower deviation state The autoregressive coefficient, and .

[0025] (2a) in, For the scalar input that is being limited, this is usually the case here. The output is shown in equation (3). As the lower limit, when Then output , As the upper limit, when Then output , This is a bilateral limiting function used to clamp the target power within... Inside; (2b) in, For pattern The noiseless state evolution function is consistent with the right-hand side of equation (2). For the state vector, see equation (1). For control or exogenous inputs, see equation (1). For power state components, according to first-order inertia Converging towards the target power after limiting, parameters (or ), See (2a), For the steady-state mapping of the power curve to wind speed, see equation (3), where The effective wind speed after deviation correction. Power curve parameter set See equation (3). This is the maximum permissible power for this step; (2c) in, Let be the random state vector for the next time step. For pattern For noiseless propagation, see (2b), at time use Evaluate, For pattern process noise vector See equation (2). It follows a zero-mean Gaussian distribution with covariance of See equation (5). For pattern The process noise covariance matrix is ​​given in equation (5), with subscripts... , To represent the time index and the discrete pattern index respectively, consistent with equations (2) and (5).

[0026] Because wind speed can be approximated by a trend in the short term, and power exhibits first-order dynamic inertia. Furthermore, given the limitations imposed by the quota and the slow-varying bias in wind measurement, this step provides a time evolution model to describe the evolution of the state over time. The solution is the prior distribution of the next state. The purpose of this step is to provide a nonlinear function for UKF time updates. This laid the physical foundation for subsequent prediction step sigma point propagation and extrapolation.

[0027] Step S1.4: Map wind speed to a steady-state power target and define the power curve: (3) in, Wind speed The corresponding steady-state power target, The cut-in wind speed for the fan. Rated wind speed, To cut off the wind speed, The scaling factor for the cubic segment. This refers to the rated power of the equipment.

[0028] Since the wind turbine power exhibits strong nonlinear segmentation with respect to wind speed (such as cut-in, cubic segment, rated, or cut-out), this step maps the wind speed to a steady-state power target, and the solution is the power target term in equation (2). The purpose of this step is to embed physical priors into the model, improve the accuracy of the nonlinear region, and lay the foundation for accurate characterization of saturation or boundary behavior in subsequent steps.

[0029] Step S1.5: Map the state to the sensor measurement space. The measurement equation is: (4) in, For nonlinear mapping from state to measurement, For the observed noise vector, see equation (5). This is a fixed offset for the nacelle wind speed channel.

[0030] The purpose of this step is to map the state to the sensor measurement space, and to solve for the quantity, which is the measurement prediction. This leads to subsequent innovations. This step serves as a prerequisite for the update step, laying the foundation for subsequent innovations and likelihood calculations.

[0031] Step S1.6: Specify the statistical characteristics of the observed noise, and the noise model: (5) in, For process noise, the components are shown in equation (2). To observe the noise, It follows a zero-mean Gaussian distribution. For pattern The process noise covariance matrix is ​​as follows. For model The observation noise covariance matrix is ​​given below, where The covariance is semidefinite.

[0032] UKF requires a noise model that includes covariance injection and consistency testing. This step specifies the statistical characteristics of the process or observed noise; the solution variables are not directly generated, but the covariance to be injected is determined. The purpose of this step is to provide... This laid the foundation for the subsequent adaptive and adjustable NIS.

[0033] Step S2: Generate and propagate sigma points using a square root unscented Kalman filter at a preset sampling period of seconds, including the following steps: Step 2.1: Set the discrete scale of the sigma points, and perform an unscented transformation: Regarding the dimension of the state ,Pick , , (6) in , For UT scaling parameters, For UT secondary scaling parameters, Let be the state dimension. This is the combined scaling factor for UT. The offset scale constant of UT ; This step sets the discrete scale of the sigma points to ensure matching of the second moment under Gaussian conditions and control numerical stability. The solution quantity is... , This step is used to generate sigma points, which lays the foundation for the UT sampling scale in subsequent steps.

[0034] Step 2.2: Calculate the mean and covariance of the propagated sigma points using weighted averages. The UT weights are: (7) in, For the first The average weight of each sigma point. For the first Covariance weights at sigma points The parameters are the second-order corrections made by UT to the Gaussian distribution. The value is usually 2, and the index is... ,and Center point; The purpose of this step is to calculate the mean and covariance of the propagated sigma points using weighted averages, ensuring the second-order accuracy and unbiasedness of the unscented transformation. The solution is the weights of each point in the mean and covariance, laying the foundation for the accuracy of subsequent numerical integration.

[0035] Step 2.3: Sigma point based on square root covariance S: (8) in, For the first A state sigma point, This is the mean of the current state used for sampling. Let Cholesky's factor be the state covariance, such that... , For matrix The List, For the UT offset scale, see equation (6). Let be the state dimension, see equation (6).

[0036] This step uses the square root of covariance. Taking deterministic sample points on an uncertain ellipsoid, and approximating the nonlinear Gaussian propagation with a finite-point approximation, the solution quantity is... sigma points The purpose of this step is to provide input samples for time / measurement propagation, laying the foundation for the subsequent realization of SR-UKF positive definiteness.

[0037] Step 3: Perform pattern interaction for the IMM algorithm, including the following steps: Step 3.1: Calculate the pattern from the previous step To this step mode The conditional weights, the mixed probability is (9) in, Mode transition matrix elements, For the previous mode The posterior probability, For pattern The normalization constant, To give the pattern From pattern Conditional mixed weights, index For pattern index; This step calculates from the previous pattern. To this step mode The condition weights, the solution quantity is The purpose of this step is to guide the weighted mixing of model priors, laying the foundation for a smooth transition between multiple models in the subsequent process.

[0038] Step 3.2: Generate the prior mean and covariance for each pattern, and synthesize them using square roots: (10) in, For pattern The mixture of prior means, For pattern Mixed prior covariance, For the previous step mode The posterior mean, For the previous step mode The posterior covariance, For the conditional mixing weight, see Equation (9), where the subscript 0 indicates the prior state at the start of this filtering step.

[0039] This step generates the prior mean and covariance for each pattern, and the solution is... This step serves as the initial value for the model-level UKF, laying the foundation for a numerically stable starting point for subsequent parallel time or measurement propagation.

[0040] First, scaling parameters are set in the unscented transform, and sigma points are generated using the square root trigonometric factor of the current covariance. Mixed priors for each mode are obtained through IMM. Subsequently, in the time update phase, the sigma points are propagated through the nonlinear state equation, and the square root of the prior covariance is directly recursively derived using QR and Cholesky. In the measurement update phase, measurement predictions and innovations are first calculated, and statistical tests are performed on the arriving measurements using NIS gating. Then, covariance matching and fading factors are combined to adaptively adjust the process noise and observation noise covariances. A multi-innovation strategy is introduced, which adjusts the most recent L-step innovations according to the attenuation coefficient. ; Step S4.1: Time update, advance the sigma point to the next time step through the state equation: by , Based on the equation (8), generate the Sigma point. ,by (11) Propagate Sigma points, where, For the first time after the spread 1 Sigma point, For pattern The state evolution function is shown in Equation 2. For the sigma point before propagation, see equation (8). For control or exogenous inputs, see equation (1). For pattern index; This step advances the sigma point to the next time step through the state equation. It is a necessary input for prior statistics and lays the foundation for subsequent calculations of prior mean or covariance.

[0041] Step S4.2: Calculate the prior mean from the propagated sigma points: (12) in, For pattern The mean of the prior states, For the mean weight, see equation (7). For the prior sigma point, see equation (11). For the state dimension, see equation (6); Step S4.3: Synthesize the square root of the prior covariance And inject process noise: (13) Among them, The square root of the process noise covariance matrix. For the covariance weights, see equation (7), for Perform QR decomposition and take the upper triangle R, so that , The Cholesky factor for the prior covariance; This step stably synthesizes the square root of the prior covariance. Introducing process noise avoids non-positive definiteness caused by direct subtraction on P, resulting in a solution quantity of... The purpose of this step is to provide a reliable covariance for innovation and gain, laying the foundation for the stability of the prior power variance and update step values.

[0042] Step S4.4: Extract the prediction and uncertainty of the power component from the state prior, where the prior power is: definition (14) in, Choose a vector for power. For pattern The current step prior power prediction, Let Variance be the variance of the prior power prediction. It is the square of the second norm of the vector; The task output is a power prediction, which requires a clearly defined extraction method. This step extracts the prediction and uncertainty of the power component from the state prior and solves for the quantity. The purpose of this step is to provide the predicted output of the current step, laying the foundation for subsequent comparisons with the posterior estimate and for multi-step extrapolation of initial values.

[0043] Step S4.5: Map the prior sigma points to the measurement space and calculate their expectation. The measurement is then updated as follows: , (15) in, For the first The mapping result of a prior sigma point in the measurement space. For the measurement mapping of equation (4), For pattern The prior mean of the measurement, For the mean weight, see equation (7). For the prior sigma point, see equation (11); Before constructing an innovation, prior measurements must be obtained. This step maps the prior sigma points to the measurement space and calculates their expectation. This step serves as a precondition for the innovation and measurement covariance.

[0044] Step S4.6: Synthesize the square root of the innovative covariance : (16) Among them, The square root of the observation noise covariance matrix, for Perform QR decomposition and take the upper triangle R, so that , The Cholesky factor for innovative covariance; Gain and NIS calculations require... For inverse calculation, SR is more stable, and this step stably synthesizes the innovative square root of covariance. The purpose of this step is to provide a scale for innovative statistics, laying the foundation for the numerical stability of NIS and gain in subsequent steps.

[0045] Step S4.7: Measure the covariance between the state and the measurement, where the cross-covariance is: (17) in, See equations (11), (12), (15), (15), and (7) respectively; The Kalman gain essentially depends on this covariance. This step measures the co-variance between the state and the measurement, and serves as a necessary input for the gain calculation, laying the foundation for the direction and magnitude of subsequent updates.

[0046] Step S4.8: Computational Innovation With normalized innovation square : untie (18) in, These are the actual measurements and the predicted measurements, respectively, see equations (1) and (15), trigonometric solutions. Avoid explicit inversion. for The numerically stable notation uses a threshold. upper valve and The lower valve is used for gating and consistency monitoring, where m is the measurement dimension. This step calculates the innovation and the normalized squared innovation, used for gated anomaly measurement and adaptive noise adjustment. The solution is: The purpose of this step is to provide criteria for gating and adaptation, laying the foundation for robust updates in the future.

[0047] Step S4.9: When NIS is too high, increase the process noise and adjust it using the fading factor. : (19) in, This is the dimension normalization statistic for NIS. To adjust the step size parameter for process noise, To Adaptive amplification factor Indicates using Update process noise covariance; Fixed under time-varying scenarios Mismatch will occur. This step increases process noise and improves tracking performance when NIS is too high. The solution complexity is... and the updated The purpose of this step is to alleviate overconfidence and lag, laying the groundwork for a more reasonable gain magnitude in the subsequent steps.

[0048] Step S4.10: Match the observed noise to the actual innovation statistics, adjust R: (20) in, To observe the step size parameters that adapt to noise, The nearest positive semidefinite projection operator projects the matrix into a positive semidefinite form. For innovation, see equation (18). To innovate the covariance, refer to Equation (16), and use covariance matching to update the observation noise covariance. ; The sensor status and observation error during the limited-release phase change over time. This step matches the observation noise with the actual innovation statistics, and the solution is the updated... This step serves to maintain consistency and gating effectiveness, laying the foundation for long-term stability.

[0049] Step S4.11: Combine the most recent L-step innovations and their covariances according to decay, and explicitly utilize time correlation: , (twenty one) in, For length is The multi-information stack vector, The decay coefficient of multiple innovations, The value ranges from 0 to 1, but is typically 0.7 to 0.9. For multiple information window lengths, It is usually 3 to 6. To construct operators for block diagonal matrices, For the first The innovation covariance Cholesky factor for the step is shown in equation (16). For innovation, see equation (18); (twenty two) in, For the first The cross-variance between step state and measurement is given in equation (17); This step calculates the most recent L-step innovation and its covariance in a decaying manner. By combining these methods and explicitly utilizing time correlation, the solution is a "equivalent measurement system" after stacking. This step provides more information for gain calculation, laying the foundation for faster convergence and jitter reduction in the subsequent steps.

[0050] Step S4.12: Calculate the Kalman gain: (twenty three) in, For the multi-information covariance, see equation (22). For the block diagonal triangular factor of the multi-innovation covariance, see Equation (21); This step calculates the Kalman gain to balance prior and measurement, using a trigonometric solution to avoid explicit inversion and improve numerical robustness. The solution is the gain matrix. This step determines the update magnitude and direction, laying the foundation for subsequent calculations of the posterior mean and covariance.

[0051] Step S4.13: Update the mean and SR covariance: (twenty four) in, For pattern The posterior mean of the states. For pattern The posterior covariance Cholesky factor The Cholesky update operator for reducing the order of trigonometric factors, For the Kalman gain, see equation (23). For the multiple innovations and their covariance factors, see equation (21). For the prior mean and factors, see equations (12) and (13); This step completes the Bayes update in square root form and maintains it. ,direct Numerical solutions are often non-positive definite, and the solutions are a posteriori. The purpose of this step is to obtain the optimal estimate at the current moment, laying a reliable initial value for subsequent constrained projection and extrapolation.

[0052] Step S4.14: Extract the current power estimate and uncertainty from the posterior state: (25) in, For pattern The current power posterior estimate, The uncertainty of this posterior estimate, For selecting a power vector, see equation (14). For the posterior mean and factor, see Equation (24).

[0053] This step extracts the current power estimate and uncertainty from the posterior state, and the solution is: This step serves to provide a direct output when an "estimated value" is needed, laying the foundation for subsequent pattern fusion and extrapolation.

[0054] According to step S5, constrained projections are applied to the updated results to satisfy physical boundaries such as power and ramp rate, while maintaining positive covariance. This includes the following steps: Step S5.1: Define the physically feasible region for power, wind speed, or ramp rate, and the set of wind power physical boundaries and operational constraints. for: (26) in, This represents the change in power between adjacent steps. Maximum climbing ability per unit time; Wind power has hard boundaries. This step clarifies the physical feasible region of power, wind speed, or ramp rate, which is a condition for the projection problem and lays the foundation for the feasible region of projection and covariance consistency in the subsequent steps.

[0055] Step S5.2: Project the posterior mean back into the feasible region using the covariance measure: (27) in, This represents the variable that minimizes the objective function. It is the inverse of the posterior covariance. is the unprojected posterior mean (see Equation (24)). The feasible region is (see equation (26)). For the posterior covariance, by get See equation (24); This step projects the posterior mean back to the feasible region using a covariance metric. Hard shearing would destroy statistical consistency, so minimizing the Mahalanobis distance is the most reasonable approach. The solution is the projected mean. The purpose of this step is to ensure that the estimated physical values ​​are compliant, thus laying a feasible and stable initial value foundation for subsequent extrapolations.

[0056] Step S5.3: Make the covariance consistent with the projected mean: (28) in, To constrain the projection mapping on Jacobian matrix at the location, The Cholesky decomposition of a symmetric positive definite matrix. The Cholesky factor for uniformizing covariance. For the unprojected posterior covariance, see Equation (24).

[0057] Projection alters the local geometry, requiring the use of the Jacobian. Correcting uncertainty, this step makes the covariance consistent with the projected mean, and the solution is the projected uniform value. The purpose of this step is to provide a consistent statistical description for the next step of fusion extrapolation, laying the foundation for long-term numerical stability.

[0058] According to step S6, constructing an interactive multi-model based on discrete operating conditions and fusing mode probabilities includes the following steps: Step S6.1: Use innovative statistical measures to measure the pattern The compatibility of the two components, with a likelihood value of: (29) in, Let be the Gaussian normalization constant. , It is a triangular factor determinant. For innovation, see equation (18). For the innovative covariance factor, see equation (16); This step uses innovative statistical measures to measure the pattern. This step ensures compatibility and avoids single-model bias. It serves as the basis for calculating the posterior probability of the model, laying the foundation for subsequent updates to the fusion weights.

[0059] Step S6.2: Update the posterior probability of the pattern at the current time step: , (30) in, For pattern The likelihood value is given by equation (29). Let be the normalization constant of equation (9). The subscript is the sign for summing over all patterns. The current moment; This step obtains the posterior probability of the pattern at the current time, and outputs the most compatible model with automatic weighting. The solution complexity is... This step serves as the weighting source for the fused output, laying the foundation for subsequent single power output.

[0060] Step S6.3: Fuse the current power estimates from multiple models into a single output and factor in the inter-model discretization: (31) in, For the current power estimate after fusion, To estimate the variance of the current power after fusion, For the current power estimates and variances of each model, see Equation (25). For the posterior probability of the pattern, see equation (30). The variance term contributing to the discrete inter-model contribution; This step fuses the current power estimates from multiple models into a single output and incorporates the discretization between models, improving robustness and interpretability under switching conditions. The solution complexity is... This step estimates the final point at the current moment, providing initial values ​​for subsequent extrapolations.

[0061] Step S6.4: The current step prior power prediction for fusion is: (32) in, For the current step prior power prediction of fusion, Its predicted variance, including the inter-model dispersion, For the prior power and variance of each model, see Equation (14). For the posterior probability of the pattern, see equation (30).

[0062] According to step S7, the output will cover the next 0-5 minutes. Power prediction and variance at each step: (33) in, To predict the number of steps, To cover the maximum predicted steps in the 1-5 minute range, This is obtained by updating only the time from the current time. Step state prediction mean The corresponding state prediction covariance Cholesky factor. for Step power prediction mean for Step power prediction variance.

[0063] In the specific steps Use history Piecewise fitting; Use step segment fitting; use playback to make the NIS coverage close to the nominal value for Q and R.

[0064] The improved unscented Kalman filter ultra-short-term wind power prediction method proposed in this embodiment introduces strategies such as square root SR-UKF, constrained projection, NIS gating and covariance matching and fading factor, multi-innovation residual injection, and IMM-UKF with discrete operating conditions, while maintaining the interpretability of conventional models and engineering constraints.

[0065] Figure 2 The actual power output of a wind farm on a certain day in a certain location is selected as the true curve, and short-term power prediction is compared using ARIMA, LSTM, unmodified UKF and the improved UKF proposed in this invention. Figure 3 The table below compares the performance of the four different prediction methods. Compared with the comparison methods ARIMA, LSTM, and the unmodified UKF, the prediction method in this embodiment exhibits stronger look-ahead tracking capability and smaller overshoot during large takeoffs, drops, and climbs. This method significantly outperforms existing schemes in both mean relative error (MAPE) and mean absolute error (RMSE), and the average prediction error is less than 3%.

[0066] Because the improved UKF can "sense" trends in advance, especially in rapidly changing time-varying sections such as power rise or fall or large slope, it still maintains low latency and low overshoot tracking performance, with smaller overall jitter, lower overshoot and undershoot, thus improving the accuracy and stability of online tracking and extrapolation of wind power at the second to minute level, thereby effectively verifying the engineering practicality and reliability of the method.

[0067] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness or practicality of the present invention. The specific embodiments described in the specification can be used to interpret the claims.

Claims

1. A method for predicting ultra-short-term wind power based on an improved UKF algorithm, characterized in that: Includes the following steps: Step S1: Establish a nonlinear state-space model that includes effective wind speed, short-term trend, wind speed deviation, and power; Step S2: Generate and propagate sigma points using square root unscented Kalman filtering at a preset second-level sampling period to complete the time update; Step S3: Collect at least one on-site measurement of SCADA power, nacelle wind speed, and lidar, and perform normalized innovative square gating; Step S4: In the measurement update, a multi-innovation strategy is adopted to fuse the most recent L-step innovations, and the process and observation noise covariances are adapted online through covariance matching and / or fading factors; Step S5: Apply constraint projection to the update results to satisfy physical boundaries such as power and ramp rate, while maintaining positive covariance; Step S6: Construct an interactive multi-model based on discrete working conditions and fuse the mode probabilities; Step S7: Output the current filtered power and the predicted power at h steps, as well as the corresponding variance or confidence interval.

2. The method for predicting ultra-short-term wind power based on the improved UKF algorithm according to claim 1, characterized in that: The state-space model is established according to step S1, including the following steps: Step S1.1: Align and preprocess the data, setting the discrete time step. Discrete-time index The original measurement was power. Cabin wind speed LIDAR / NWP: Resample all channels to a uniform step size. Perform external control on outliers first, and only perform prediction steps for missing steps. Step S1.2: For continuous state variables Control input Observable measurement Define: ， ， （1） In equation (1), For discrete-time indexing, Let be the system state vector, where For effective wind speed status, For the short-term trend or acceleration state of wind speed, For wind speed deviation or scale correction status, This refers to the active power status of the power station. To control or exogenous input vectors, only the power upper limit for this step is included here. This represents the maximum power available in this step. Let be the measurement vector, where Power measurements obtained by SCADA For cabin wind speed measurement, Optional auxiliary measurement; Step S1.3: Provide the time evolution model, give the nonlinear function for UKF time update, and the nonlinear state equation is, according to the mode : （2） in, For the process noise of each state, see equation (5). For pattern Downtrend The autoregressive coefficient, and , For pattern The first-order inertial time constant of the lower power The power curve from wind speed to steady-state power is shown in equation (3). This is the set of parameters for the power curve. For pattern Lower deviation state The autoregressive coefficient, and ; （2a） in, For the scalar input that is being limited, this is usually the case here. The output is shown in equation (3). As the lower limit, when Then output , As the upper limit, when Then output , This is a bilateral limiting function used to clamp the target power within... Inside; （2b） in, For pattern The noiseless state evolution function is consistent with the right-hand side of equation (2). For the state vector, see equation (1). For control or exogenous inputs, see equation (1). For power state components, according to first-order inertia Converging towards the target power after limiting, parameters (or ), See (2a), For the steady-state mapping of the power curve to wind speed, see equation (3), where The effective wind speed after deviation correction. Power curve parameter set See equation (3). This is the maximum permissible power for this step; (2c) in, Let be the random state vector for the next time step. For pattern For noiseless propagation, see (2b), at time use Evaluate, For pattern process noise vector See equation (2). It follows a zero-mean Gaussian distribution with covariance of See equation (5). For pattern The process noise covariance matrix is ​​given in equation (5), with subscripts... , To represent the time index and the discrete pattern index respectively, consistent with equations (2) and (5); Step S1.4: Map wind speed to a steady-state power target and define the power curve: (3) in, Wind speed The corresponding steady-state power target, The cut-in wind speed for the fan. Rated wind speed, To cut off the wind speed, The scaling factor for the cubic segment. Rated power of the equipment; Step S1.5: Map the state to the sensor measurement space. The measurement equation is: (4) in, For nonlinear mapping from state to measurement, For the observed noise vector, see equation (5). For the fixed offset of the nacelle wind speed channel; Step S1.6: Specify the statistical characteristics of the observed noise, and the noise model: (5) in, For process noise, the components are shown in equation (2). To observe the noise, It follows a zero-mean Gaussian distribution. For pattern The process noise covariance matrix is ​​as follows. For model The observation noise covariance matrix is ​​given below, where The covariance is semidefinite.

3. The method for predicting ultra-short-term wind power based on the improved UKF algorithm according to claim 2, characterized in that: According to step S2, sigma points are generated and propagated using square root unscented Kalman filtering within a preset second-level sampling period, including the following steps: Step 2.1: Set the discrete scale of the sigma points, and perform an unscented transformation: Regarding the dimension of the state ,Pick , , (6) in , For UT scaling parameters, For UT secondary scaling parameters, Let be the state dimension. This is the combined scaling factor for UT. The offset scale constant of UT ; Step 2.2: Calculate the mean and covariance of the propagated sigma points using weighted averages. The UT weights are: (7) in, For the first The average weight of each sigma point. For the first Covariance weights at sigma points The parameters are the second-order corrections made by UT to the Gaussian distribution. The value is usually 2, and the index is... ,and Center point; Step 2.3: Sigma point based on square root covariance S: (8) in, For the first A state sigma point, This is the mean of the current state used for sampling. Let Cholesky's factor be the state covariance, such that... , For matrix The List, For the UT offset scale, see equation (6). Let be the state dimension, see equation (6).

4. The method for predicting ultra-short-term wind power based on the improved UKF algorithm according to claim 3, characterized in that: The pattern interaction of the IMM algorithm includes the following steps: Step 1: Calculate the pattern from the previous step To this step mode The conditional weights, the mixed probability is （9） in, Mode transition matrix elements, For the previous mode The posterior probability, For pattern The normalization constant, To give the pattern From pattern Conditional mixed weights, index For pattern index; Step 2: Generate the prior mean and covariance for each pattern, and synthesize them using square roots: （10） in, For pattern The mixture of prior means, For pattern Mixed prior covariance, For the previous step mode The posterior mean, For the previous step mode The posterior covariance, For the conditional mixing weight, see Equation (9), where the subscript 0 indicates the prior state at the start of this filtering step.

5. The method for predicting ultra-short-term wind power based on the improved UKF algorithm according to claim 4, characterized in that: First, scaling parameters are set in the unscented transform, and sigma points are generated using the square root trigonometric factor of the current covariance. Mixed priors for each mode are obtained through IMM. Then, in the time update phase, the sigma points are propagated through a nonlinear state equation, and the square root of the prior covariance is directly recursively derived using QR / Cholesky. In the measurement update phase, measurement predictions and innovations are first calculated, and statistical tests are performed on the arriving measurements using NIS gating. Then, covariance matching and fading factors are combined to adaptively adjust the process noise and observation noise covariances. A multi-innovation strategy is introduced, where the most recent L-step innovations are adjusted according to the decay coefficient. This includes the following steps: Step 1: Time update, advance the sigma point to the next time step through the state equation: by , Based on the equation (8), generate the Sigma point. ,by （11） Propagate Sigma points, where, For the first time after the spread 1 Sigma point, For pattern The state evolution function is shown in Equation 2. For the sigma point before propagation, see equation (8). For control or exogenous inputs, see equation (1). For pattern index; Step 2: Calculate the prior mean from the propagated sigma points: （12） in, For pattern The mean of the prior states, For the mean weight, see equation (7). For the prior sigma point, see equation (11). For the state dimension, see equation (6); Step 3: Synthesize the square root of the prior covariance And inject process noise: (13) Among them, The square root of the process noise covariance matrix. For the covariance weights, see equation (7), for Perform QR decomposition and take the upper triangle R, so that , The Cholesky factor for the prior covariance; Step 4: Extract the prediction and uncertainty of the power component from the state prior. The prior power is: definition (14) in, Choose a vector for power. For pattern The current step prior power prediction, Let Variance be the variance of the prior power prediction. It is the square of the second norm of the vector; Step 5: Map the prior sigma points to the measurement space and calculate their expectation. The measurement is then updated as follows: ， （15） in, For the first The mapping result of a prior sigma point in the measurement space. For the measurement mapping of equation (4), For pattern The prior mean of the measurement, For the mean weight, see equation (7). For the prior sigma point, see equation (11); Step Six: Synthesize the Square Root of the Innovative Covariance : (16) Among them, The square root of the observation noise covariance matrix, for Perform QR decomposition and take the upper triangle R, so that , The Cholesky factor for innovative covariance; Step 7: Measure the covariance between the state and the measurement, where the cross-covariance is: （17） in, See equations (11), (12), (15), (15), and (7) respectively; Step 8: Computational Innovation With normalized innovation square : untie (18) in, These are the actual measurements and the predicted measurements, respectively, see equations (1) and (15), trigonometric solutions. Avoid explicit inversion. for The numerically stable notation uses a threshold. upper valve and The lower valve is used for gating and consistency monitoring. m For measurement dimensions; Step 9: When NIS is too high, increase the process noise and adjust it using the fading factor. : (19) in, This is the dimension normalization statistic for NIS. To adjust the step size parameter for process noise, To Adaptive amplification factor Indicates using Update process noise covariance; Step 10: Match the observed noise to the actual innovation statistics, adjust R: (20) in, To observe the step size parameters that adapt to noise, The nearest positive semidefinite projection operator projects the matrix into a positive semidefinite form. For innovation, see equation (18). To innovate the covariance, refer to Equation (16), and use covariance matching to update the observation noise covariance. ; Step 11: Combine the most recent L-step innovations and their covariances according to decay, and explicitly utilize time correlation: ， (21) in, For length is The multi-information stack vector, The decay coefficient of multiple innovations, The value ranges from 0 to 1, but is typically 0.7 to 0.

9. For multiple information window lengths, It is usually 3 to 6. To construct operators for block diagonal matrices, For the first The innovation covariance Cholesky factor for the step is shown in equation (16). For innovation, see equation (18); （22） in, For the first The cross-variance between step state and measurement is given in equation (17); Step 12: Calculate the Kalman gain: (23) in, For the multi-information covariance, see equation (22). For the block diagonal triangular factor of the multi-innovation covariance, see Equation (21); Step Thirteen: Update the Mean and SR Covariance: (24) in, For pattern The posterior mean of the states. For pattern The posterior covariance Cholesky factor The Cholesky update operator for reducing the order of trigonometric factors, For the Kalman gain, see equation (23). For the multiple innovations and their covariance factors, see equation (21). For the prior mean and factors, see equations (12) and (13); Step Fourteen: Extract the current power estimate and uncertainty from the posterior state: (25) in, For pattern The current power posterior estimate, The uncertainty of this posterior estimate, For selecting a power vector, see equation (14). For the posterior mean and factor, see Equation (24).

6. The method for predicting ultra-short-term wind power based on the improved UKF algorithm according to claim 5, characterized in that: According to step S5, constrained projections are applied to the updated results to satisfy physical boundaries such as power and ramp rate, while maintaining positive covariance. This includes the following steps: Step S5.1: Define the physically feasible region for power, wind speed, or ramp rate, and the set of wind power physical boundaries and operational constraints. for: (26) in, This represents the change in power between adjacent steps. Maximum climbing ability per unit time; Step S5.2: Project the posterior mean back into the feasible region using the covariance measure: (27) in, This represents the variable that minimizes the objective function. It is the inverse of the posterior covariance. is the unprojected posterior mean (see Equation (24)). The feasible region is (see equation (26)). For the posterior covariance, by get See equation (24); Step S5.3: Make the covariance consistent with the projected mean: （28） in, To constrain the projection mapping on Jacobian matrix at the location, The Cholesky decomposition of a symmetric positive definite matrix. The Cholesky factor for uniformizing covariance. For the unprojected posterior covariance, see Equation (24).

7. The method for predicting ultra-short-term wind power based on the improved UKF algorithm according to claim 6, characterized in that: According to step S6, constructing an interactive multi-model based on discrete operating conditions and fusing mode probabilities includes the following steps: Step S6.1: Use innovative statistical measures to measure the pattern The compatibility of the two components, with a likelihood value of: （29） in, Let be the Gaussian normalization constant. , It is a triangular factor determinant. For innovation, see equation (18). For the innovative covariance factor, see equation (16); Step S6.2: Update the posterior probability of the pattern at the current time step: ， （30） in, For pattern The likelihood value is given by equation (29). Let be the normalization constant of equation (9). The subscript is the sign for summing over all patterns. The current moment; Step S6.3: Fuse the current power estimates from multiple models into a single output and factor in the inter-model discretization: （31） in, For the current power estimate after fusion, To estimate the variance of the current power after fusion, For the current power estimates and variances of each model, see Equation (25). For the posterior probability of the pattern, see equation (30). The variance term contributing to the discrete inter-model contribution; Step S6.4: The current step prior power prediction for fusion is: ， （32） in, For the current step prior power prediction of fusion, Its predicted variance, including the inter-model dispersion, For the prior power and variance of each model, see Equation (14). For the posterior probability of the pattern, see equation (30).

8. The method for predicting ultra-short-term wind power based on the improved UKF algorithm according to claim 7, characterized in that: According to step S7, the output will cover the next 0-5 minutes. Power prediction and variance at each step: (33) in, To predict the number of steps, To cover the maximum predicted steps in the 1-5 minute range, This is obtained by updating only the time from the current time. Step state prediction mean The corresponding state prediction covariance Cholesky factor. for Step power prediction mean for Step power prediction variance.

Images