Search

CN-121983960-A - Offshore wind power prediction method and system

CN121983960ACN 121983960 ACN121983960 ACN 121983960ACN-121983960-A

Abstract

The invention provides a method and a system for predicting offshore wind power, comprising the steps of 1, collecting time series data of wind turbines or wind power plants and multisource external variables, performing time alignment to define a predicted object and causality constraint, 2, preprocessing the data, 3, decomposing time domain multiscale, 4, performing self-adaptive frequency domain decomposition and three-frequency band reconstruction on power signals, constructing two-dimensional time-frequency coupling strength, 5, outputting a main prediction result, 6, modulating an error correction model through risk gating weight, and 7, updating related parameters of the risk gating weight based on historical prediction errors to obtain final prediction output. The method can remarkably improve the precision, stability and long-term adaptability of offshore wind power prediction under strong non-stable and sudden disturbance conditions on the premise of meeting strict causality constraint and online deployment requirements.

Inventors

  • GUO ZHIHANG
  • PENG TIAN
  • NI HAO
  • WANG YANG
  • HAO ZHUBING
  • JIANG LEI
  • ZHANG CHU
  • LI HUA
  • HUANG FENGZHI

Assignees

  • 淮阴工学院

Dates

Publication Date
20260505
Application Date
20260123

Claims (10)

  1. 1. The marine wind power prediction method is characterized by comprising the following steps of: Step 1, collecting time sequence data of wind turbine generator or wind farm power and multisource external variables, performing time alignment, and defining a prediction object and causality constraint; Step 2, data preprocessing, namely performing anomaly correction, noise suppression, deletion complement and normalization on the power time sequence, and constructing a combined observation and sliding window; Step 3, time domain multi-scale decomposition, namely performing causal learning lifting decomposition on the purifying power sequence in a sliding window and reconstructing to obtain a long-term trend component, a mesoscale fluctuation component and a short-time disturbance component; Step 4, carrying out frequency domain decomposition and three-frequency band reconstruction based on the self-adaptive GS on the power signal, and constructing two-dimensional time-frequency coupling strength; Step 5, constructing a power prediction model taking a time domain component as an input on the basis of the multidimensional decomposition result, and outputting a main prediction result; step 6, introducing a risk gating error correction mechanism on the basis of a main prediction result, constructing a first risk index based on the two-dimensional time-frequency coupling strength of the power signal, further constructing a working condition disturbance risk factor based on the expansion observed quantity variation amplitude, fusing the first risk index and the working condition disturbance risk factor to form a comprehensive risk index, and modulating ARIMAX an error correction model through a risk gating weight; and 7, establishing a parameter optimization target and a recursive updating strategy under the fusion prediction structure, carrying out self-adaptive optimization updating on key parameters of the gating function, and updating related parameters of the risk gating weight based on the historical prediction error to obtain final prediction output.
  2. 2. The system of claim 1, wherein step 1 comprises the steps of: step 1.1, power time series data acquisition and modeling object definition: active output power data of a wind turbine generator or a wind farm SCADA system are collected to form a discrete time sequence: , Wherein, the For the kth sampling point time Is used for the control of the active power of the power source, N is the total number of sampling points; Sampling interval The definition is as follows: , Introducing predictive advance step size And define the number of discrete steps : , Wherein the method comprises the steps of Representing a positive integer set; the main prediction target is written as: , wherein Pred represents the prediction result of the main test model, For the mapping of the primary prediction model, For the moment of time A set of available input information; Causality constraint at any predicted time Only the historical observed data is allowed to be used, ; Step 1.2, multisource enhanced data acquisition is aligned with time; Synchronously acquiring multiple source variables and uniformly aligning to a time axis while acquiring power ; Meteorological variables include wind speed at time Wind direction Air pressure Ambient temperature Humidity of air Wave height ; The unit operation state variables comprise the pitch angle at the moment And the rotational speed of the unit ; Time limited power running sign ; Time of day fault status flag ; Constructing extended observation vectors : , Wherein the method comprises the steps of Represents the transpose operator and, Representing real space; let the original sampling set of the mth exogenous variable be Aligned to the time axis Is defined as: , Wherein, the Represents the value of the mth exogenous observation variable at the time k, Time pair Ji Suanzi, which is deterministic, → is a mapped symbol.
  3. 3. The system of claim 2, wherein step 2 comprises the steps of: Step 2.1, preprocessing power data; The preprocessing process comprises outlier correction, noise suppression, missing data completion and normalization processing to obtain a preprocessed power sequence : , Wherein, the Representing a comprehensive pretreatment mapping function; Step 2.2, constructing a joint feature matrix; fusing the preprocessing power and the extended observation vector into a joint observation: ; Wherein, the Representing an initial input state or initial set of observations at time k, A preprocessed power signal representing time k; an extended observation vector representing time k; step 2.3, sliding history window constraint; at the time of prediction A sliding history window of length L is introduced: , Wherein the method comprises the steps of ; A long-term trend power component representing time k; Step 2.4, overall output definition of the multi-dimensional decomposition framework; performing time domain, frequency domain and two-dimensional time-frequency coupling decomposition on the power signal, wherein the overall output is defined as: , Wherein, the Representing the coupling characteristics at time k; respectively a long-term trend time domain component, a mesoscale fluctuation time domain component and a short-time disturbance time domain component; Respectively representing a low-band reconstruction component, a middle-band reconstruction component, and a high-band reconstruction component based on the generalized S-transform GS; To index in frequency band The two-dimensional time-frequency coupling strength of the upper layer, ; The main predictive model uses only And A prediction model constructed based on the multi-scale time domain decomposition result and used for outputting the basic prediction result is referred to as a main prediction model.
  4. 4. A system according to claim 3, wherein step 3 comprises the steps of: Step 3.1, modeling causal learnable lifting decomposition CLLW; at the time of prediction Constructing a discrete sequence of length L for a sequence of clean power within a window : , Splitting the sequence into parity samples: , , Wherein, the Representing even sample sub-sequences in a layer 0 lifting decomposition, Representing an odd sampling sub-sequence in a layer 0 lifting decomposition; If L is odd, 1 sample is added to the tail end to make L become even and then decomposed; In the lifting decomposition of the j-th layer, Constructing a causal predictor And cause and effect update operator And get the detail component And approximate component : , , Wherein, the Indicating the number of layers or decomposition depth of the lifting decomposition; And (3) with Causal expansion convolution FIR with limited memory, predictor: , Updating operators: , Wherein the method comprises the steps of The causal memory length of the predictor and the causal memory length of the update operator are respectively; Is the first Index of layer expansion ratio, actual expansion step length is ; Representing a power time sequence or a reconstructed power sequence; Representing a detail component; Is a learnable parameter; For all indexes Definition of the term(s) 、 ; Reversible reconstruction constraint, arbitrary layer The previous layer of subsequences are reconstructed as follows: , , Wherein, the Representing that is obtained by parity interleaving Recursive implementation For a pair of Is a reversible reconstruction of (2); representing the result of the detailed component after obtaining the even subsequence estimate And by predictors Based on the addition of the prediction terms calculated by the even subsequence estimation values, the odd subsequence estimation value of the j-1 th layer is restored ; Step 3.2, adaptively determining the number of decomposition layers; given the maximum number of decomposition layers For each candidate layer number Executing the step 3.1 to obtain a corresponding reconstruction sequence Defining normalized reconstruction errors : , Wherein, the Representing the original sequence of input powers, Representing the power sequence estimate after the lifting decomposition and reconstruction is completed at layer J; constructing an evaluation criterion comprising a complexity penalty: , Wherein, the Representing an evaluation function or a cost function corresponding to the number of decomposition layers J; for complexity trade-off coefficients, choose: ; Wherein, the An optimal number of decomposition layers for minimizing the evaluation function Q (J); Step 3.3, on-line self-adaptive adjustment of operator parameter lifting; Definition of reconstruction loss : , , , Wherein, the Regularized weight coefficients representing the reconstruction loss, Indicating the number of optimal decomposition layers The final reconstructed power sequence below is then used, 、 Respectively represent the first The layer promotes the predictor parameters and the update operator parameters in the decomposition, Respectively represent the first End parameters and first of layer predictors The end parameters of the layer update operator; adopting recursive gradient update and combining projection constraint: , Wherein, the Representing the updated parameter vector at time k+1; Representing the gradient operator with respect to the parameter theta, ; Is the learning rate; Is a projection operator; removable constraint set Wherein 、 As an upper bound of the amplitude value, ; Step 3.4, reconstructing the multi-scale time domain component; the decomposition results are combined into three time domain components and passed through up-sampling and time pair Ji Suanzi Mapping back to window length L: Long term trend component ; Mesoscale fluctuation component ; Short-term disturbance component Wherein, the Representing and optimizing the number of decomposition layers The corresponding parameters of the approximation components, Representing the detail component corresponding to the optimal layer, Representing detail components obtained by decomposing the j-th layer; defining three time domain components at a time The value of (2) is the window end sample: , Thereby each of Are all obtained ; Wherein, the Representing a set of multi-scale power components at time k, Representing a set of long-term scale power components, The power component of the trend is represented, Representing a mesoscale fluctuating power component, Representing short-term disturbance power components, M, S representing trend scale, mesoscale and short-term scale, respectively.
  5. 5. The system of claim 4, wherein step 4 comprises the steps of: step 4.1, frequency domain decomposition and frequency band reconstruction; the output definition holds: ; step 4.2, constructing two-dimensional time-frequency coupling strength; the time-frequency energy map obtained based on the generalized S-transform, GS, is defined as: , Wherein, the For a time-frequency representation of the power signal at time k and frequency v, Representing the adaptive generalized S transformation result adopting the optimal parameters; Dividing the frequency axis into three sections of frequency bands which are not overlapped with each other: , Satisfy the following requirements And covers the target analysis frequency band; Wherein, the 、 、 V 1 ⁻、ν 1 + represents the lower limit frequency and the upper limit frequency of the low frequency band, v 2 ⁻、ν 2 + represents the lower limit frequency and the upper limit frequency of the middle frequency band, and v 3 ⁻、ν 3 + represents the lower limit frequency and the upper limit frequency of the high frequency band; defining two-dimensional time-frequency coupling strength: , Wherein, the And S (k, v) is the time-frequency representation of the power signal at the time k and the frequency v, and dv frequency bins are used for frequency domain integration.
  6. 6. The system of claim 5, wherein step 5 comprises the steps of: step 5.1, constructing multi-channel input; at the time of prediction Introducing sliding windows with the length L, and respectively constructing three-scale channel input: , , , Wherein, the 、 、 Respectively representing state vectors corresponding to the trend, the mesoscale and the short-time channel; Combining into a master predictive model input tensor : ; 5.2, Multi-scale structured state space prediction modeling; single-scale structured state space modeling: For each scale channel Defining a discrete state space: , , Wherein, the Represents the hidden state of the c-th scale at time n +1, 、 、 A state transition matrix, an input matrix and an output matrix respectively representing a c-th scale; Is the first Channel window in the first A plurality of input samples; in order to be in a hidden state, Is a state dimension; for the scale response output, ; , , ; For a pair of Applying a spectral radius constraint: , Wherein, the Representing assignment or update operations; In order to be a radius of the spectrum, A stabilization threshold is preset; constructing scale representation vector to obtain scale output sequence Afterwards, scale characterization was constructed by pooling: , Wherein the method comprises the steps of To deterministic pooling operators, output , Represent the first The output dimensions of the individual scale channels; Performing cross-scale gating fusion on each scale representation, wherein the cross-scale gating fusion is specifically as follows: ; Wherein, the A gating function respectively representing a trend, a mesoscale gating function and a gating function of a short-time channel; 、 The scale representation vector of the trend scale, the scale representation vector of the mesoscale fluctuation channel and the scale representation vector of the short-time disturbance channel are respectively represented; characterization of each scale Extracting statistical mapping characteristics: , Wherein the method comprises the steps of For a fixed defined statistical mapping, the dimensions are output Fixing in the implementation; Is the dimension of the statistical feature; And (3) constructing fusion weights: , Wherein, the Splice vectors which are three scale statistical feature vectors; mapping a matrix for cross-scale gating weights; Bias terms for gating functions; Representing vector stitching; , ; And is also provided with Softmax is a normalized exponential function; fusion characterization: ; Wherein the method comprises the steps of Is a unified representation of multi-scale information; Main prediction output: Wherein And (3) with Training target is to minimize mean square error: , Wherein the method comprises the steps of ; A parameter set for a multi-scale structured state space model; representing the true power value at time k+τ; Step 6.1, constructing a predicted outbreak risk index; Based on two-dimensional time-frequency coupling strength Defining the strongest coupling scale : , And constructing a predicted burst risk index based on the power signal structural characteristics through normalization and nonlinear mapping: , Wherein, the Respectively representing the mean value and standard deviation of the risk index R; Representing a random noise or residual term, A risk indicator constructed based on the power signal; Representing the non-linear mapping function, Normalizing the original risk amount; Introducing a working condition disturbance risk factor constructed based on the variation amplitude of the extended observation vector, and setting at the moment Extended observation vector of (a) The method comprises the following steps: , Wherein, the Representing a mesoscale extended observation vector; By calculating the variation amplitude of the expansion observed quantity at adjacent moments, the working condition disturbance intensity is constructed : , Wherein, the Representing the first in the extended observation vector The individual observed variables being at the moment Is a value of (2); Representing the same observation variable at the previous adjacent moment M represents the dimension or variable number of the extended observation vector; and obtaining the working condition disturbance risk index through statistical normalization and nonlinear mapping of the sliding window : , Wherein, the Representing the amount of risk of the original observation, 、 Respectively representing the mean value and standard deviation of the observed risk quantity; then fusion writing: , Wherein, the Is a trade-off coefficient; the dynamic weights are determined by a risk gating mechanism: , Wherein the method comprises the steps of Is a bias parameter, controls Basic weights of the individual frequency bands; is risk sensitive weight, determines comprehensive risk index The intensity of influence of the variation on the frequency band weight; e is a natural constant; 、 Respectively representing an ith observation or risk characteristic value and a jth observation or risk characteristic value; Representing a weight coefficient corresponding to the ith feature; An exponential smoothing mechanism is introduced: ; Wherein, the Representing the smoothed value of the weight ai, Representing the smoothing coefficient; Step 6.2, risk gating frequency domain prediction error correction; constructing a historical prediction error based on a master prediction output and observed real power : , Exogenous input vector: , Wherein, the A power state vector representing the system at time k; Based on the main prediction output, a ARIMAX prediction error correction model which introduces exogenous variables is constructed: , Wherein, the 、 Autoregressive and moving average parameters of the ARIMAX model are represented; A power state vector for the system at time k; Input vectors for exogenous; Error correction terms: , Wherein, the Representing a prediction error estimate for future k+τ at time k; representing the predicted power after error correction; is a prediction residual; and 6.3, prediction output correction: ; After the risk gating prediction error correction in the step 6 is completed, a historical prediction residual sequence is constructed based on the main prediction output in the sliding history window, the prediction output after the risk gating error correction and the corresponding observed real power; On the premise of keeping the main prediction model structure and the risk gating error correction mechanism structure unchanged, determining bias parameters and risk sensitive weight parameters in the risk gating function as parameters to be optimized, and constructing an optimization objective function comprising a prediction error item and a parameter constraint item; And inputting parameters to be optimized, an optimized objective function and a historical prediction residual sequence into a risk gating parameter self-adaptive optimization module, wherein the risk gating parameter self-adaptive optimization module is used for carrying out recursive updating on the risk gating parameter based on the historical prediction error on the premise of not changing a main prediction structure and a risk gating form, so that the self-adaptive calibration of the relation between the prediction error correction strength and the risk state is realized, and the self-adaptive optimization updating of the risk gating parameter is executed.
  7. 7. The system of claim 6, wherein step 7 comprises the steps of: Step 7.1 optimization variable definition Taking the risk gating mechanism parameter as an optimization variable: , Wherein, the , Is the first Offset of individual band gating; , Is the first Individual band gating pair comprehensive risk Sensitive weights of (2); A parameter set which is a risk gating function; step 7.2, constructing an optimization objective function; Constructing a risk gating parameter optimization target: , Wherein, the Is a regular coefficient; step 7.3, a recursive updating mechanism of risk gating parameters; Defining transient losses : , After taking into account the regularization term, pair The gradient update write of (1) is: , , Wherein, the The learning rate of the gating parameter; Representing bias leads; 、 Respectively represent the gating network in Weight and bias of moment; From the following components And (2) and Independent of Therefore: , Wherein, the The term for error correction is indicated as, Representing the actual power value of the power, Representing the predicted loss function and, Representing the first in a risk-gated network The number of weight parameters to be used in the process, Representing the first in a risk-gated network A bias parameter; the derivative of the gating portion is determined: , , , , Wherein, the Represent the first The linear activation values of the individual gating cells, Representation of representation No Gating scoring values corresponding to the gating units; R represents a risk index or a risk intensity; And exponential smoothing satisfies the recursive derivative: , ; step 7.4, recursive updating and stability constraint Defining the parameter change amplitude of adjacent update time: , Wherein the method comprises the steps of Representing the updated gating parameter set; imposing stability constraints: , Wherein, the Indicating the variation of the gating parameter, A maximum allowable change threshold, and if not, correcting projection as follows: , step 7.5, constructing a perception prediction closed loop feedback mechanism; Defining corrected prediction error feedback: , Wherein, the Representing the prediction error after the risk gating correction; And uses the coupling strength at the corresponding time Updating the risk statistics reference amount: , , Wherein, the 、 Respectively representing the updated risk statistical mean and standard deviation; representing the risk statistics update coefficients, Updating coefficients for statistics smoothing; is shown at the moment The amount of original coupling strength obtained by time-frequency coupling analysis of the power signal.
  8. 8. An offshore wind power prediction system implemented by the method of any one of claims 1-7, comprising: the data acquisition module is used for acquiring the time sequence data of the power of the wind turbine generator or the wind power plant and multisource external variables and performing time alignment to define a prediction object and causality constraint; the multi-source preprocessing and sliding window construction module is used for executing data preprocessing, carrying out anomaly correction, noise suppression, deletion complement and normalization on the power time sequence, and constructing a combined observation and sliding window; The causal learnable lifting decomposition module CLLW is configured to perform time domain multi-scale decomposition, perform causal learnable lifting decomposition on the purification power sequence in the sliding window, and reconstruct to obtain a long-term trend component, a mesoscale fluctuation component, and a short-term disturbance component; the time-frequency coupling characteristic construction module is used for carrying out frequency domain decomposition and three-frequency band reconstruction based on the self-adaptive GS on the power signal and constructing two-dimensional time-frequency coupling strength; the main prediction module is used for constructing a power prediction model taking a time domain component as an input on the basis of the multidimensional decomposition result and outputting a main prediction result; The risk gating error correction module is used for introducing a risk gating error correction mechanism on the basis of a main prediction result, constructing a first risk index based on the two-dimensional time-frequency coupling strength of the power signal, further constructing a working condition disturbance risk factor based on the variation amplitude of the expansion observation quantity, fusing the first risk index and the working condition disturbance risk factor to form a comprehensive risk index, and modulating ARIMAX an error correction model through a risk gating weight; The closed-loop self-adaptive optimization module is used for introducing a self-adaptive optimization thought based on historical data on the premise of keeping a main prediction structure and a gating mechanism unchanged, and providing structural support for parameter optimization; The risk gating parameter self-adaptive optimization module is used for establishing a parameter optimization target and a recursive updating strategy under the fusion prediction structure, carrying out self-adaptive optimization updating on key parameters of the gating function, and updating related parameters of the risk gating weight based on historical prediction errors to obtain final prediction output.
  9. 9. An electronic device comprising a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method of any of claims 1 to 7.
  10. 10. A storage medium storing a computer program or instructions which, when run on a computer, performs the steps of the method of any one of claims 1 to 7.

Description

Offshore wind power prediction method and system Technical Field The invention relates to the technical fields of new energy generated power prediction, power system operation analysis and intelligent information processing, in particular to a method and a system for predicting offshore wind power. Background With the continuous increase of the installed scale of offshore wind power, wind power prediction has become an important technical support for power system safety dispatching, spare capacity allocation and new energy consumption. However, the wind power is affected by the complex weather and marine environment, and is closely related to external weather factors such as wind speed, wind direction, air pressure, temperature, humidity, wave height and the like, and is also affected by the combined action of internal operating factors such as a set pitch angle, a rotating speed, a limited power operating state, a fault working condition and the like, so that the power time sequence shows remarkable strong non-stationarity, multi-scale fluctuation and sudden disturbance characteristics. The existing offshore wind power prediction method models a power sequence based on a single time scale or a fixed feature space, and is difficult to simultaneously describe the structural evolution rule of a power signal on multiple scales of a time domain and a frequency domain. When the wind condition or the operation condition is suddenly changed, the method cannot identify the power structure change in time, the prediction error is easy to amplify, and even short-time prediction instability is caused. On the other hand, although the prediction model based on deep learning can learn a complex nonlinear mapping relation through large-scale historical data, the modeling process usually depends on offline training, model parameter updating is lagged, and an explicit sensing mechanism for power mutation risks and operation disturbance risks is lacked. In an actual online operation scene, when the wind power statistical characteristic drifts or abnormal working conditions frequently occur, the prediction precision and stability of the model are difficult to ensure for a long time. In addition, in the existing prediction system, a fixed weight or simple statistical compensation mode is adopted for prediction error correction, the evolution characteristics of the power signal on different frequency scales cannot be combined, and the error correction strength cannot be dynamically adjusted according to the running risk state, so that the robustness and the self-adaptive capacity of the system under the complex working condition are limited. Therefore, there is a need for an offshore wind power prediction system that can model a multi-scale structure of a wind power signal, fuse disturbance information of extended observation conditions, sense and quantify power mutation risks, and dynamically adaptively correct prediction errors according to risk states on the premise of meeting strict causality constraint and online deployable requirements, so as to improve accuracy and stability of offshore wind power prediction in a complex operation environment. Disclosure of Invention The invention aims to provide a method and a system for predicting the offshore wind power, aiming at the characteristics of strong non-stationarity, multi-scale fluctuation and sudden disturbance of the offshore wind power under the combined action of complex sea condition and unit operation condition. The system operates under strict causality constraint, and can sense and quantify potential risks reflected by internal structural changes of power signals and risks caused by external operation condition disturbance at the same time on the premise of not introducing any future information, and dynamically adjust and adaptively correct a prediction process and a prediction error based on the risk state, so that stability, robustness and long-term operation adaptability of offshore wind power prediction under a sudden fluctuation condition are improved. The method comprises the following steps: Step 1, collecting time sequence data of wind turbine generator or wind farm power and multisource external variables, performing time alignment, and defining a prediction object and causality constraint; Step 2, data preprocessing, namely performing anomaly correction, noise suppression, deletion complement and normalization on the power time sequence, and constructing a combined observation and sliding window; Step 3, time domain multi-scale decomposition, namely performing causal learning lifting decomposition on the purifying power sequence in a sliding window and reconstructing to obtain a long-term trend component, a mesoscale fluctuation component and a short-time disturbance component; Step 4, carrying out frequency domain decomposition and three-frequency band reconstruction based on the self-adaptive GS on the power signal, and constructing two-dimensional time