Multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement
By employing dynamic time warping and a hybrid attention mechanism, the timing discrepancy between NWP and measured wind farm data was resolved, improving the accuracy and stability of multi-step wind power prediction and achieving accurate wind power prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-24
- Publication Date
- 2026-05-15
AI Technical Summary
In existing technologies, the time-series deviation between NWP and wind farm measured data is not corrected, and the traditional seq2seq attention mechanism lacks positional constraints, resulting in low accuracy of multi-step wind power prediction.
The temporal misalignment between high-frequency inertial sequences and low-frequency meteorological sequences is quantified by dynamic time warping algorithm. An elastic constraint mechanism based on path linearity is constructed, and a hybrid attention mechanism is introduced into the heterogeneous seq2seq model. Adaptive feature extraction and fusion are performed using modified slope factor and dynamic gating coefficient.
It improves the accuracy and stability of multi-step wind power prediction, solves the time-series deviation problem between NWP and measured data, and achieves accurate multi-step wind power prediction.
Smart Images
Figure CN122046252A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, specifically to a multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement. Background Technology
[0002] Wind power generation is characterized by strong randomness and volatility, making accurate multi-step wind power forecasting crucial for grid dispatch planning and operational safety. To extend forecast lead time and capture long-term trends, numerical weather prediction (NWP) data is typically introduced as future environmental information input, using the wind speed, wind direction, and other meteorological forecast values provided to assist models in perceiving future weather conditions.
[0003] Currently, deep learning-based sequence-to-sequence (seq2seq) models are widely used in such time series prediction tasks. This architecture typically utilizes an encoder to extract time-dependent features from historical sequences and a decoder to progressively generate future predictions. When processing multi-source inputs containing NWP (Near-Power-Wide) parameters, existing techniques often employ heterogeneous data concatenation or dual-channel processing, using attention mechanisms to calculate the contribution weights of historical hidden states to the current decoding step, attempting to establish a mapping relationship between meteorological features and power output.
[0004] However, NWP data and measured wind farm data are heterogeneous data sources. Due to limitations in the physical mechanisms and update frequencies of forecast models, there is often a nonlinear temporal phase deviation between the two. Most existing methods assume strict alignment of the two on the time axis, lacking adaptive correction for this temporal misalignment, leading to biases in the model input features. Furthermore, traditional seq2seq attention mechanisms often calculate weights across the entire sequence when performing multi-step predictions, lacking positional constraints for specific prediction step sizes. This makes it difficult for the model to accurately focus on the meteorological time step logically corresponding to the current moment in long sequences, resulting in blurred feature extraction and interference from irrelevant information, thus limiting the improvement of multi-step wind power prediction accuracy.
[0005] Therefore, this invention proposes a multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement to address the shortcomings of existing technologies. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides a multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement, which solves the problem that existing technologies have low multi-step wind power prediction accuracy due to the lack of correction for the time sequence deviation between NWP and measured data and the lack of position constraints in the attention mechanism.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement, comprising the following steps: High-frequency inertial sequences from wind farms and low-frequency meteorological sequences from meteorological service providers are collected, and the low-frequency meteorological sequences are interpolated to obtain aligned meteorological sequences. A historical time window is selected, and the dynamic time warping path between the measured wind speed in the high-frequency inertial sequence and the predicted wind speed in the aligned meteorological sequence is calculated. The original path slope and global alignment cost are extracted from the dynamic time warping path. Linear regression analysis is performed on the dynamic time warping path to calculate path linearity. Confidence factors are generated based on the path linearity, and the original path slope is corrected using the confidence factors to obtain a corrected slope factor. A heterogeneous neural network containing a first encoder and a second encoder is constructed to process the high-frequency... An inertial sequence is input into the first encoder to extract a high-frequency hidden state vector, and an aligned meteorological sequence is input into the second encoder to extract a low-frequency hidden state vector. A virtual mapping index is calculated based on the current prediction step size and the correction slope factor. Attention features are then extracted from the low-frequency hidden state vector using the virtual mapping index to obtain a meteorological channel context vector, and attention features are also extracted from the high-frequency hidden state vector to obtain an inertial channel context vector. Dynamic gating coefficients are generated based on the global alignment cost. These dynamic gating coefficients are used to weight and control the meteorological channel context vector, which is then fused with the inertial channel context vector to output the predicted wind power value for future time moments.
[0008] Preferably, the steps of collecting high-frequency inertial sequences from wind farms and low-frequency meteorological sequences from meteorological service providers, and interpolating the low-frequency meteorological sequences to obtain aligned meteorological sequences include: defining the high-frequency inertial sequence of the wind farm as a feature vector containing active power, measured wind speed, and measured wind direction, and decomposing the measured wind direction into sine and cosine components; reconstructing the low-frequency meteorological sequence using a cubic spline interpolation algorithm to construct a piecewise cubic polynomial function, such that the function value of the piecewise cubic polynomial function is equal at all known data points, and the first and second derivatives at the nodes are continuous; generating an aligned meteorological sequence with the same time resolution as the high-frequency inertial sequence based on the piecewise cubic polynomial function; wherein, before constructing the high-frequency inertial sequence and the aligned meteorological sequence, a standardization process is also performed.
[0009] Preferably, the steps for calculating the dynamic time warping path between the measured wind speed in the high-frequency inertial sequence and the forecast wind speed in the aligned meteorological sequence include: extracting the measured wind speed sequence from the high-frequency inertial sequence and the forecast wind speed sequence from the aligned meteorological sequence; constructing a Euclidean distance matrix, where the elements represent the distance metric between the time points of the measured wind speed sequence and the time points of the forecast wind speed sequence; calculating a cumulative cost matrix, where the elements represent the minimum cumulative path cost from the starting point to the current position; and using a dynamic programming algorithm to search for the path with the minimum cumulative distance in the cumulative cost matrix as the dynamic time warping path, where the dynamic time warping path consists of a series of coordinate points representing the best alignment between the time points of the measured wind speed sequence and the time points of the forecast wind speed sequence.
[0010] Preferably, the steps of extracting the original path slope and global alignment cost from the dynamic time warping path include: calculating the normalized value of the sum of the local Euclidean distances of all nodes on the dynamic time warping path, and using the normalized value as the global alignment cost; for path points on the dynamic time warping path, calculating the ratio of the predicted time index increment to the measured time index increment, and using the ratio as the original path slope; the method further includes: calculating the average vertical distance of the dynamic time warping path relative to the geometric diagonal to obtain the time phase drift.
[0011] Preferably, the steps of performing linear regression analysis on a dynamic time warped path to calculate path linearity include: extracting the measured time index of path points in the dynamic time warped path as independent variables and extracting the predicted time index as dependent variables; constructing a linear regression model and using the least squares method to solve for the regression slope and intercept of the linear regression model to fit a best-fit line; calculating the coefficient of determination between the dynamic time warped path and the best-fit line, and using the coefficient of determination as the path linearity, which is used to characterize the linear temporal correspondence between the measured time index and the predicted time index.
[0012] Preferably, the steps of generating a confidence factor based on path linearity and using the confidence factor to correct the original path slope to obtain a corrected slope factor include: constructing a mapping relationship between path linearity and confidence factor using a Logistic activation function, so that the confidence factor exhibits a non-linear continuous change in the interval from zero to one; constructing a correction formula based on elastic constraints, and using the confidence factor to perform a weighted summation of the original path slope and the unit slope to calculate the corrected slope factor; wherein, when the confidence factor value is large, the corrected slope factor is dominated by the original path slope, and when the confidence factor value is small, the corrected slope factor tends to be the unit slope.
[0013] Preferably, the steps of constructing a heterogeneous neural network including a first encoder and a second encoder, inputting a high-frequency inertial sequence into the first encoder to extract a high-frequency hidden state vector, and inputting an aligned meteorological sequence into the second encoder to extract a low-frequency hidden state vector include: configuring the first encoder and the second encoder as parallel bidirectional gated recurrent unit network structures, wherein the first encoder and the second encoder do not share weight parameters; inputting the high-frequency inertial sequence into the first encoder, calculating the forward hidden state and the backward hidden state using the gating mechanism of the first encoder, and concatenating them to obtain a high-frequency hidden state vector; inputting the aligned meteorological sequence into the second encoder, calculating the forward meteorological hidden state and the backward meteorological hidden state using the independent weight matrix of the second encoder, and concatenating them to obtain a low-frequency hidden state vector.
[0014] Preferably, the steps of calculating the virtual mapping index based on the current prediction step size and the correction slope factor, and extracting attention features from the low-frequency hidden state vector based on the virtual mapping index to obtain the meteorological channel context vector include: calculating the statistical representative value of the correction slope factor, and superimposing the product of the statistical representative value and the current prediction step size to the benchmark alignment position to obtain the virtual mapping index; constructing a hybrid attention mechanism that incorporates position Gaussian bias, and calculating the content relevance score between the decoder state and the low-frequency hidden state vector; constructing a position bias score using the virtual mapping index, wherein the position bias score adopts the form of a Gaussian kernel function to penalize the time step that deviates from the virtual mapping index; superimposing the content relevance score and the position bias score and normalizing them to obtain the attention weight, and using the attention weight to perform a weighted summation on the low-frequency hidden state vector to obtain the meteorological channel context vector.
[0015] Preferably, the step of generating dynamic gating coefficients based on the global alignment cost and using the dynamic gating coefficients to weight the meteorological channel context vector includes: inputting the meteorological channel context vector into the gating perception layer, and mapping the global alignment cost to the same dimensional space as the meteorological channel context vector through a learnable weight vector; processing the linear transformation result of the meteorological channel context vector and the mapped global alignment cost using the Sigmoid activation function to generate a dynamic gating coefficient vector with values ranging from zero to one; establishing a negative feedback mechanism through the dynamic gating coefficient vector so that the value of the dynamic gating coefficient vector decreases when the global alignment cost increases; and calculating the Hadamard product of the dynamic gating coefficient vector and the meteorological channel context vector to obtain the weighted meteorological feature vector.
[0016] Preferably, the step of generating dynamic gating coefficients based on the global alignment cost and using the dynamic gating coefficients to weight the meteorological channel context vector includes: inputting the meteorological channel context vector into the gating perception layer, and mapping the global alignment cost to the same dimensional space as the meteorological channel context vector through a learnable weight vector; processing the linear transformation result of the meteorological channel context vector and the mapped global alignment cost using the Sigmoid activation function to generate a dynamic gating coefficient vector with values ranging from zero to one; establishing a negative feedback mechanism through the dynamic gating coefficient vector so that the value of the dynamic gating coefficient vector decreases when the global alignment cost increases; and calculating the Hadamard product of the dynamic gating coefficient vector and the meteorological channel context vector to obtain the weighted meteorological feature vector.
[0017] This invention provides a multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement. It has the following beneficial effects: 1. This invention quantifies the temporal misalignment between high-frequency inertial sequences and low-frequency NWP data by introducing a dynamic time warping algorithm, and adaptively corrects the original path slope using an elastic constraint mechanism based on path linearity. This method suppresses noise through constraint correction when NWP forecast bias is large, and preserves time phase drift characteristics by using a correction slope factor when alignment quality is high. It effectively solves the nonlinear drift problem on the time axis of multi-source heterogeneous data, improves the temporal consistency of input data, and provides an accurate multi-step wind power prediction data foundation for heterogeneous seq2seq models.
[0018] 2. This invention constructs a dual-channel heterogeneous seq2seq network to extract high-frequency measured features and low-frequency meteorological features respectively, and combines a modified slope factor to calculate a virtual mapping index, establishing a hybrid attention mechanism that includes location Gaussian bias. This mechanism enables the model to concentrate attention weights on highly correlated meteorological time steps according to the prediction step size when decoding wind power at future times, overcoming the extraction ambiguity problem caused by feature divergence in traditional long-sequence prediction, and realizing accurate temporal mapping of NWP meteorological information in the multi-step wind power prediction process.
[0019] 3. This invention generates dynamic gating coefficients by quantifying the global alignment cost, constructing a negative feedback adjustment mechanism for the meteorological channel context vector. When there is a significant systematic deviation between the NWP forecast data and the measured data, this mechanism can automatically reduce the weight of meteorological features in the decoding and fusion stage, reducing the interference of low-quality external meteorological information on the prediction results. This adaptive fusion strategy enhances the robustness of the model under fluctuating meteorological data conditions, ensuring the stability of multi-step wind power prediction results based on the heterogeneous seq2seq architecture. Attached Figure Description
[0020] Figure 1This is a flowchart of the multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement of the present invention; Figure 2 This is a diagram of the multi-step wind power prediction system architecture based on heterogeneous seq2seq and NWP refinement according to the present invention.
[0021] Among them, 101 is the data acquisition module; 102 is the feature extraction module; 103 is the slope constraint module; 104 is the heterogeneous coding module; 105 is the attention decoding module; and 106 is the dynamic fusion module. Detailed Implementation
[0022] See attached document Figure 1 This invention provides a multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement. This method runs on a computer device including a processor and memory, and includes: Step S100: Collect the high-frequency inertial sequence of the wind farm and the low-frequency meteorological sequence of the meteorological service provider, and perform interpolation processing on the low-frequency meteorological sequence to align the time resolution to obtain the aligned meteorological sequence. Step S200: Select a historical time window, calculate the dynamic time warping path between the measured wind speed in the high-frequency inertial sequence and the forecast wind speed in the aligned meteorological sequence, and extract the original path slope and global alignment cost from the dynamic time warping path; Step S300: Perform linear regression analysis on the dynamic time warped path to calculate the path linearity, generate confidence factors based on the path linearity, and use the confidence factors to perform elastic constraint correction on the original path slope to obtain the corrected slope factor. Step S400: Construct a heterogeneous neural network containing a first encoder and a second encoder. Input the high-frequency inertial sequence into the first encoder to extract the high-frequency hidden state vector, and input the aligned meteorological sequence into the second encoder to extract the low-frequency hidden state vector. Step S500: In the decoding stage, the virtual mapping index is calculated based on the current prediction step size and the correction slope factor, and attention features are extracted from the low-frequency hidden state vector based on the virtual mapping index to obtain the meteorological channel context vector. Attention features are also extracted from the high-frequency hidden state vector to obtain the inertial channel context vector. Step S600: Generate dynamic gating coefficients based on global alignment costs, use the dynamic gating coefficients to weight and control the meteorological channel context vector, and fuse them with the inertial channel context vector to output the wind power prediction value for future time.
[0023] See attached document Figure 2The computer equipment is equipped with a multi-step wind power prediction system based on heterogeneous seq2seq and NWP refinement. This system has a logic processing module for performing the above steps. The system includes a data acquisition module 101, a feature extraction module 102, a slope constraint module 103, a heterogeneous encoding module 104, an attention decoding module 105, and a dynamic fusion module 106.
[0024] The data acquisition module 101 is used to establish communication connections with the wind farm monitoring and control data acquisition system and the numerical weather prediction (NWP) server. The data acquisition module 101 acquires high-frequency inertial sequences from the wind farm and low-frequency meteorological sequences from the meteorological service provider, and performs interpolation processing on the low-frequency meteorological sequences to align the time resolution and obtain aligned meteorological sequences. The feature extraction module 102 is used to quantify the nonlinear spatiotemporal deviation between measured data and forecast data. The feature extraction module 102 selects a historical time window, calculates the dynamic time warping path between the measured wind speed in the high-frequency inertial sequence and the forecast wind speed in the aligned meteorological sequence, and extracts the original path slope and global alignment cost from the dynamic time warping path. The slope constraint module 103 is used to improve the robustness of the features and eliminate computational noise. The slope constraint module 103 performs linear regression analysis on the dynamic time warping path to calculate the path linearity, generates a confidence factor based on the path linearity, and uses the confidence factor to elastically constrain and correct the original path slope to obtain a corrected slope factor. The heterogeneous coding module 104 is used to extract spatiotemporal hidden features of different frequency sequences in parallel. The heterogeneous encoding module 104 constructs a heterogeneous neural network containing a first encoder and a second encoder. High-frequency inertial sequences are input into the first encoder to extract high-frequency hidden state vectors, and aligned meteorological sequences are input into the second encoder to extract low-frequency hidden state vectors. The attention decoding module 105 is used to achieve nonlinear feature retrieval during multi-step prediction. In the decoding stage, the attention decoding module 105 calculates a virtual mapping index based on the current prediction step size and correction slope factor, and extracts attention features from the low-frequency hidden state vectors based on the virtual mapping index to obtain the meteorological channel context vector. It also extracts attention features from the high-frequency hidden state vectors to obtain the inertial channel context vector. The dynamic fusion module 106 is used to perform feature filtering and output based on data quality. The dynamic fusion module 106 generates dynamic gating coefficients based on the global alignment cost, uses these dynamic gating coefficients to weight the meteorological channel context vectors, and fuses them with the inertial channel context vectors to output the predicted wind power value for future time moments.
[0025] See attached document Figure 1Step S100 mainly involves the construction and cleaning of the data foundation, namely, collecting high-frequency inertial sequences from wind farms and low-frequency meteorological sequences from meteorological service providers, and interpolating the low-frequency meteorological sequences to align the time resolution. This process aims to eliminate the physical differences in sampling frequency between multi-source heterogeneous data and map physical quantities of different dimensions to the same numerical space, providing a unified spatiotemporal scale for subsequent feature extraction from heterogeneous neural networks.
[0026] Specifically, the high-frequency inertial sequence originates from the wind farm's monitoring, control, and data acquisition system. In this embodiment, the high-frequency inertial sequence... This is defined as a multivariate time series matrix reflecting the real-time state of the wind turbine and its surrounding microenvironment. Because changes in wind power exhibit high physical inertia—meaning the current state strongly depends on the operating state of the wind turbine in previous moments—this type of measured data is defined as an inertial sequence. The sampling frequency of this high-frequency inertial sequence is denoted as... This serves as the reference time granularity for this embodiment (e.g., one sampling point every 15 minutes). Each time step in the high-frequency inertial sequence Corresponding feature vector It includes the following physical quantities: the actual active power generated by the wind turbine, the measured wind speed at the hub height, and the measured wind direction.
[0027] For the physical quantity of measured wind direction, the original angle values (from 0° to 360°) exhibit discontinuous jumps in numerical space (e.g., 359° and 1° are physically close but numerically different), directly using them as input to a neural network would lead to instability in the model's gradient calculation. Therefore, this implementation uses trigonometric function transformation to decompose the wind direction into sine and cosine components. At this time, time... eigenvectors It can be represented as: ; in, Active power; This is the actual measured wind speed; This is the measured wind direction angle. The time span covers a past historical observation window, and its historical time step length is denoted as... .
[0028] Low-frequency weather sequences originate from numerical weather prediction (NWP) servers provided by meteorological service providers. This type of data, acting as a strong boundary condition of the external environment, determines the upper limit of wind turbine power output at a macroscopic level. Low-frequency weather sequences It contains weather forecast data for a period of time in the future, and its sampling frequency is denoted as . In practical applications in this field, the update frequency of numerical weather forecasts is typically lower than the sampling frequency of monitoring, control, and data acquisition systems (i.e., ...). For example, if NWP is forecasted every hour, the raw NWP data will be sparse on the time axis. Each time step in a low-frequency weather series... Corresponding feature vector It includes forecast wind speed, forecast air pressure, forecast temperature, and forecast relative humidity. The time span not only covers past historical windows but also extends to future periods to be predicted, with a total length corresponding to a time range of [missing information]. ,in To predict the step size.
[0029] In order to eliminate and The frequency mismatch between data and measured data necessitates point-by-point correspondence between meteorological data and measured data on a micro-timescale. A multi-step wind power prediction system based on heterogeneous seq2seq and NWP refinement requires interpolation processing of low-frequency meteorological sequences to align with the time resolution. Considering that meteorological parameters (especially temperature and pressure) typically exhibit continuous and smooth trends over short periods, and that subsequent step S200 requires calculating the first derivative (slope), simple linear interpolation would introduce non-smooth inflection points in a broken line, resulting in spurious impulse noise. Therefore, this embodiment preferably employs a cubic spline interpolation algorithm. Reconstruction is performed to ensure the continuity of the first and second derivatives of the interpolation curve.
[0030] Suppose that a variable of a certain dimension in the original low-frequency meteorological sequence (e.g., forecast wind speed) is at discrete time points The value above is .in, This is the set of discrete time points of the original low-frequency meteorological sequence. For the set of variable values at corresponding discrete time points, This represents the total number of discrete sampling points in the original low-frequency meteorological sequence. The goal of cubic spline interpolation is to construct a piecewise cubic polynomial function. , making In each sub-interval (in The index of the time sub-interval, with a value range of 1. to The following conditions must be met: ; The piecewise cubic polynomial function The following condition must be met simultaneously: the function value is equal at all known data points, i.e. The first derivative at the node and second derivative It is continuous; and it satisfies the natural boundary condition at the boundary (i.e., the second derivative is zero). The coefficient of the constant term, The coefficient of the linear term, The coefficient of the quadratic term, The coefficient of the cubic term. The coefficients to be determined can be found by solving a system of linear equations containing the aforementioned constraints. For example, based on the continuity condition of the second derivative of the function at the nodes, a system of three moment equations containing unknown second derivative vectors can be constructed. After solving the second derivatives using the pursuit method, the coefficients of the aforementioned polynomials can then be determined. The specific matrix operations for solving the system of linear equations and cubic spline interpolation can be implemented using well-known mathematical toolkits by those skilled in the art, and will not be elaborated upon here.
[0031] After interpolation, the time resolution and Aligned meteorological sequences .at this time, The sequence length is expanded to (Assuming the frequency ratio is an integer) this ensures that there is a corresponding meteorological feature input at each high-frequency time step.
[0032] Furthermore, to eliminate the influence of different physical dimensions (such as megawatts of power and hectopascals of air pressure) on the neural network weight updates, standardization processing must be performed on the raw data of all dimensions (such as megawatt-level power values and meter-per-second-level wind speed values) before constructing the high-frequency inertial sequence and aligned meteorological sequence. For any single physical quantity (such as wind speed), the feature sequence must be standardized. The Z-Score standardization formula is used: ; in, These are the original observations; This represents the mean of a single physical quantity over the historical training dataset. The standard deviation of a single physical quantity on the historical training dataset; These are the standardized input values. This processing ensures that the high-frequency inertial sequence and the low-frequency meteorological sequence are on the same order of magnitude in terms of numerical distribution, which is beneficial for the stable descent of the gradient of the subsequent heterogeneous encoder.
[0033] See attached document Figure 1Step S200 mainly involves the dynamic evaluation and feature extraction of the accuracy of meteorological forecast data. Specifically, it involves selecting a historical time window, calculating the dynamic time warping path between the measured wind speed in the high-frequency inertial sequence and the predicted wind speed in the aligned meteorological sequence, and extracting the original path slope and global alignment cost from the dynamic time warping path. This embodiment introduces the Dynamic Time Warping (DTW) algorithm to address the problem that point-to-point error calculation at a single moment cannot quantify the time phase shift of the meteorological field movement process, that is, to quantify whether the predicted wind speed peaks and troughs are ahead or behind the measured data.
[0034] In practical implementation, the multi-step wind power prediction system based on heterogeneous seq2seq and NWP refinement first extracts univariate subsequences for alignment analysis from the dataset processed in step S100. The system then extracts high-frequency inertial sequences... Extract the measured wind speed component and cut off the length of the segment. The historical time window is recorded as the measured wind speed sequence. Accordingly, from the aligned meteorological sequences Extract the forecast wind speed components within the same time period and denot them as the forecast wind speed sequence. The wind speed component is chosen as the normalization object here because in the physics of wind power generation, wind speed is the dominant variable that determines power output, and its spatiotemporal evolution characteristics best represent the laws of meteorological field movement.
[0035] Subsequently, a Euclidean distance matrix is constructed to quantize the distance between sequences. Euclidean distance matrix The size is , of which Line number Column elements Indicates the number of times in the measured wind speed sequence The value at the nth moment and the nth time point in the forecast wind speed sequence The distance between values at each point in time is measured. In this embodiment, the squared Euclidean distance is used for calculation: ; in, This is the time index for the measured wind speed sequence. For the time index of the forecast wind speed sequence, and The Euclidean distance matrix represents the degree of matching between values at any two time points, ignoring temporal order constraints.
[0036] Based on Euclidean distance matrix We use dynamic programming to search for the dynamically time warped path with the minimum cumulative distance. We define the cumulative cost matrix. Its elements This indicates the distance from the starting point (1, 1) to the position. The minimum cumulative path cost. The recursive calculation formula is as follows: ; This recursive process must satisfy the following boundary conditions: Furthermore, the cumulative cost becomes infinite when the index goes out of bounds. The calculation continues recursively until the endpoint is reached. A dynamic time-warped path is obtained by using backtracking. This dynamic time warping path It consists of a series of coordinate points, represented as , of which Path points Represents the first of the measured wind speed sequence The time interval and the forecast wind speed sequence The optimal alignment was achieved at that moment. Let be the total length of the dynamically time-warped path, and satisfy . .
[0037] Based on dynamic time warping path This study extracts key features that characterize the quality and spatiotemporal migration of numerical weather prediction data. First, it calculates the global alignment cost. This is used to quantify the overall reliability of forecast data within a historical time window. The global alignment cost is defined as the normalized value of the sum of distances to all nodes on the dynamic time warping path, calculated as follows: ; in, path point The corresponding local Euclidean distance. The smaller the value, the closer the waveform shape of the measured wind speed is to the forecast wind speed, and the higher the overall quality of the forecast data; conversely, the larger the value, the greater the forecast deviation.
[0038] Secondly, from the dynamic time warping path The original path slope is extracted to quantify the relative movement rate of the meteorological field. In a coordinate system composed of the measured time index and the forecast time index, if the dynamic time warped path extends along the diagonal (slope of 1), it indicates that the forecast time progress is consistent with the actual time progress; if the dynamic time warped path deviates from the diagonal, it implies time phase drift information. This embodiment obtains local slope characteristics by calculating the first difference of the path point coordinates. For the first... points (of which) ), the corresponding original path slope Defined as the ratio of the predicted time index increment to the measured time index increment: ; in, To prevent small constants with a denominator of zero (e.g., 10) -6 The original path slope sequence constructed from this. This reflects the local temporal scaling of the weather forecast relative to the observed data at each alignment node. For example, when When the forecast time is longer than the actual time, it means that the forecast waveform is stretched, implying that the actual weather changes lag behind the forecast.
[0039] In addition, as a supplement to the original path slope, this embodiment also calculates the time phase drift. , defined as the average perpendicular distance of the dynamically time-warped path relative to the geometric diagonal. Its calculation formula is: ; The time phase drift feature intuitively reflects the average number of offset steps (advance or lag steps) of the forecast data on the time axis. Through the above steps, the original measured and forecast sequences are transformed into a feature set containing amplitude bias (global alignment cost) and temporal bias (original path slope, time phase drift), providing a quantitative basis for the fine-grained correction of numerical weather forecast data in the subsequent step S300.
[0040] See attached document Figure 1 Step S300 mainly involves validating and adaptively correcting the features extracted in the preceding steps. Specifically, it involves performing linear regression analysis on the dynamically time-warped path to calculate path linearity, generating confidence factors based on the path linearity, and using these confidence factors to elastically constrain and correct the original path slope, resulting in a corrected slope factor. The key point of this step is to use statistical methods to determine whether the alignment between the measured and predicted data is driven by real physical laws or by spurious matches caused by random noise, thereby introducing a weighting mechanism to prevent overcorrection.
[0041] In practical implementation, the multi-step wind power prediction system based on heterogeneous seq2seq and NWP refinement first obtains the dynamic time warping path generated in step S200. Each path point Includes measured time index and forecast time index To evaluate the overall trend stability of the dynamic time warping path, linear regression analysis was performed on the path. Specifically, the measured time index sequences of each path point were extracted. Extract the forecast time index sequence as the independent variable. As the dependent variable, construct a linear regression model: ; in, These are the measured time index coordinates of path points in a dynamically time-warped path. These are the predicted time index coordinates of path points in a dynamically time-warped path. The regression slope; The intercept is given. The parameters of this equation are solved using the least squares (OLS) method, aiming to find a best-fitting straight line that minimizes the sum of squared residuals from all points on the path to that line. The specific calculation process of the least squares method is well-known in this field and will not be elaborated upon here.
[0042] Based on this, the path linearity is calculated. In this embodiment, path linearity is specifically represented by the coefficient of determination in the linear regression analysis described above. The coefficient of determination is a statistical measure used to characterize the proportion of fluctuations in the dependent variable (here, the forecast time index) that can be explained by the independent variable (here, the measured time index) through the regression model. The formula used to quantify the degree of fit between a dynamic time warped path and the best-fit line is as follows: ; in, This is the index of the actual forecast time corresponding to the path point; The predicted value calculated for the linear regression model; This is the arithmetic mean of all forecast time indices. The value range is [0, 1]. When When the value approaches 1, it indicates a significant linear temporal correspondence between the measured time index and the predicted time index, suggesting that the dynamic time warping path is reliable. When the value approaches 0, it indicates that the path exhibits highly disordered random walk characteristics. In this case, the DTW algorithm may be matching the sound rather than the physical characteristics.
[0043] A confidence factor is generated based on path linearity, and the original path slope is corrected using this confidence factor to obtain a corrected slope factor. To achieve adaptive elastic constraints based on data quality—that is, rapidly decreasing the confidence in the dynamic time warping result to preserve the original physical time series when path linearity is low, and maintaining high confidence to adopt the spatiotemporal alignment result when path linearity is high—this embodiment constructs a mapping relationship based on a Logistic activation function. The specific calculation formula is as follows: ; in, is the confidence factor, and its value range is (0, 1); The steepness coefficient (e.g., a value of 10) controls the sensitivity of the confidence factor to changes in linearity. This is a linearity threshold (e.g., 0.5), representing the critical point at which to begin trusting a dynamically time-warped path. When the path linearity is high, A value close to 1 indicates a high degree of trust in the spatiotemporal offset calculated by DTW; when the path linearity is low, a value close to 0 indicates a tendency to believe that there is no obvious systematic bias and a conservative strategy should be maintained.
[0044] Finally, the original path slope is elastically constrained and corrected using a confidence factor. The original path slope here is the original path slope sequence extracted in step S200. The physical essence of the correction is to weight and shrink the calculated local slope towards a "unit slope" (i.e., a slope of 1, representing no time scaling). The correction formula is as follows: ; in, For the first Correction slope factor corresponding to each path point; For the first The formula shows that the original path slope corresponds to each path point. This indicates that when the data matching quality is high ( When the data matching quality is poor (large), the correction slope factor is mainly dominated by the original path slope, allowing the model to capture sharp temporal phase drift; when .... When the slope factor is small, it is forced back to 1, which effectively suppresses unreliable noise interference, causing the model to degenerate into a standard unaligned input pattern. Finally, the calculated values for all time points are... The combination output includes a corrected slope factor. The feature set.
[0045] See attached document Figure 1 Step S400 mainly involves constructing a heterogeneous neural network including a first encoder and a second encoder. High-frequency inertial sequences are input into the first encoder to extract high-frequency hidden state vectors, and low-frequency meteorological sequences are input into the second encoder to extract low-frequency hidden state vectors. In this embodiment, the heterogeneous neural network is a heterogeneous Seq2Seq model (deep learning network) built based on the Sequence-to-Sequence (Seq2Seq) model framework. To adapt to the input characteristics of multi-source heterogeneous data, the single encoder structure in the traditional Seq2Seq model is improved into a parallel dual-channel encoding structure, thus forming a heterogeneous Seq2Seq architecture. The heterogeneous Seq2Seq architecture solves the feature overload and gradient conflict problems that easily occur when traditional single encoders process mixed frequency data by establishing a physically decoupled parallel dual-branch feature extraction structure in the encoding stage, which respectively connects to input data with different frequency domain characteristics.
[0046] In practical implementation, the multi-step wind power prediction system based on heterogeneous seq2seq and NWP refinement constructs the encoding layer of the heterogeneous neural network by configuring two independent recurrent neural network modules. These two independent modules are configured as the first encoder and the second encoder, respectively. They are set up in parallel in the network topology and do not share weight parameters, but logically they together constitute the encoding end of the heterogeneous Seq2Seq model.
[0047] The first encoder is configured as a feature extractor for processing high-frequency time-series data, and its input port is connected to the high-frequency inertial sequence constructed in step S100. In this embodiment, the first encoder employs a bidirectional gated cyclic unit (BiGRU) network structure. This structure utilizes a gating mechanism to maintain the ability to remember instantaneous fluctuations and short-term inertia of wind power in long sequences. For high-frequency inertial sequences... Each time step in (in The first encoder internally performs the following state update operation: ; ; ; ; in: The update gate controls the proportion of state information from the previous time step that is retained in the current time step. The reset gate controls the degree to which the state information from the previous moment is ignored when calculating the current candidate state. The candidate hidden state at the current moment represents the fusion feature of the current input and the filtered historical information; For the current moment The forward-hidden state; For the previous moment The forward-hidden state; For the current moment The high-frequency inertial sequence input vector; This is the Sigmoid activation function, used to generate a gated signal in the (0, 1) interval; This is the hyperbolic tangent activation function, used to generate state values in the interval (-1, 1); This represents the Hadamard product, which is the element-wise multiplication of vectors. These are the input weight matrices for the update gate, reset gate, and candidate state, respectively. These are the cyclic weight matrices corresponding to the update gate, reset gate, and candidate state, respectively; These are the corresponding bias vectors.
[0048] To fully capture the bidirectional dependencies of the time series, the first encoder also includes a reverse processing channel, which processes the input sequence in reverse time to calculate the reverse hidden state. The forward hidden state and the backward hidden state are concatenated to generate the time step. High-frequency hidden state vector : ; After performing the above calculations on all historical time steps, the first encoder outputs a high-frequency hidden state vector sequence. .
[0049] The second encoder is configured as a feature extractor for processing low-frequency trend data, with its input port connected to the aligned meteorological sequence generated in step S100. The second encoder also uses a BiGRU network structure, but it is initialized with a set of weight parameters independent of the first encoder. For aligned meteorological sequences... The corresponding time step The second encoder utilizes its independent weight matrix (denoted as...). (etc.) and the cyclic weight matrix (denoted as (etc.), and the forward weather concealment state is obtained by logical calculation using the same logic as the aforementioned formula. Reverse weather concealment state .
[0050] Furthermore, time is generated through splicing operations. Low-frequency hidden state vector : ; Finally, the second encoder outputs a low-frequency hidden state vector sequence. .
[0051] Through the above construction process, the heterogeneous neural network in this embodiment achieves the splitting of multi-source data. (High-frequency hidden state vector) The microscopic dynamic characteristics of the wind turbine were characterized in a high-dimensional feature space, while the low-frequency hidden state vector... This characterizes the evolution of the macroscopic meteorological environment. This design based on a heterogeneous Seq2Seq architecture ensures that high-frequency measured data and low-frequency meteorological data retain their complete spatiotemporal characteristics before entering the decoding and prediction stage, laying the foundation for the subsequent decoder to perform adaptive feature fusion using the corrected slope factor obtained in step S300. For the initialization of network parameters (e.g., using Xavier uniform distribution initialization) and the backpropagation algorithm during training, those skilled in the art can use standard implementations within deep learning frameworks, which are well-known techniques in the field and will not be elaborated upon here.
[0052] See attached document Figure 1Step S500 mainly involves calculating the virtual mapping index based on the current prediction step size and the correction slope factor during the decoding stage, and extracting attention features from the low-frequency hidden state vector based on the virtual mapping index to obtain the meteorological channel context vector. This embodiment introduces a physically constrained attention mechanism at the decoding end, namely, an attention decoding strategy based on virtual index mapping. This attention decoding strategy aims to solve the problem that traditional soft attention mechanisms rely solely on the similarity of hidden layer vector content for matching, ignoring the physical drift characteristics of the meteorological field in the time dimension. By introducing the correction slope factor calculated in step S300 as prior knowledge, the effective physical alignment position of the meteorological data on the time axis can be calculated, thereby constraining the decoder to focus on feature regions that are truly relevant in terms of physical evolution logic.
[0053] In specific implementation, the multi-step wind power prediction system based on heterogeneous seq2seq and NWP refinement first initializes the state of the decoding stage. In this embodiment, the main structure of the decoder adopts a gated loop unit, which is referred to as the decoder to clarify its functional position. The high-frequency hidden state vector sequence output by the first encoder in step S400 is then used. The final moment of the historical window The low-frequency hidden state vector sequence output by the second encoder final moment state Perform nonlinear fusion to construct the initial hidden state of the decoder. The fusion calculation is shown in the following formula: ; in, To initialize the weight matrix; To initialize the bias vector; This represents a vector concatenation operation; It is the hyperbolic tangent activation function.
[0054] This then enters a multi-step prediction loop. Assume the total prediction step size is... For each current prediction step size (in ), and perform the operation of calculating the virtual mapping index based on the current prediction step size and the correction slope factor.
[0055] Virtual mapping index Defined as within the low-frequency meteorological feature space, with the current prediction step size The continuous-time index position with the strongest physical correspondence. This is due to the correction slope factor generated in step S300. (This feature set contains multiple slope values calculated based on the DTW path within the historical time window) reveals the temporal scaling of the forecast data relative to the observed data. First, the statistical representative value of the corrected slope factor is calculated. (The weighted arithmetic mean of all elements in the modified slope factor feature set is taken) to characterize the average time elapsed rate ratio at future times. Then, the virtual mapping index is calculated using the following formula. : ; in: This represents the current relative time step of the decoder (e.g., step 1, step 2, etc.). To correct the statistical representative value of the slope factor. When When this indicates a lag in the meteorological process, the growth step of the virtual mapping index will be greater than the natural time step; when This indicates that the meteorological process is occurring earlier than expected, and the growth step of the virtual mapping index will be smaller than the natural time step. The reference alignment position is typically taken as the end index of the input sequence. Or based on the time phase drift in step S200 Make adjustments (e.g.) This is done to compensate for the initial phase deviation. The calculation process maps discrete prediction time steps to continuous virtual mapping indices, achieving physical time realignment.
[0056] Based on this, an attention feature extraction operation is performed on the low-frequency hidden state vector according to the virtual mapping index. This embodiment employs a hybrid attention mechanism that incorporates positional Gaussian bias to extract key information from the low-frequency hidden state vector.
[0057] First, calculate the decoder hidden state at the previous time step. With low-frequency hidden state vector sequence Each vector in Content relevance score between : ; in, Represents the transpose operation of a matrix or vector; For the context score vector parameters in the attention mechanism; This is the learnable weight matrix used to transform the decoder state; This is the learnable weight matrix used to transform the encoder state.
[0058] To introduce physical timing constraints, the calculated virtual mapping index is used. Construct position deviation score The positional deviation score is used to penalize time steps that deviate from the virtual mapping index, and it takes the form of a Gaussian kernel function: ; in, This is a learnable parameter for the attention window width, used to control the model's tolerance to temporal bias. This term forces the attention weight distribution to follow a bell-shaped curve, with the center aligned with the virtual mapping index. .
[0059] The content relevance score and the location deviation score are combined and then normalized using the Softmax function to obtain the final attention weight. : ; Finally, attention weights are utilized. The meteorological channel context vector is obtained by weighted summation of the low-frequency hidden state vector sequence. : ; in, It represents the total length of the low-frequency hidden state vector sequence (including historical time steps and future prediction time steps).
[0060] Meteorological Channel Context Vector By aggregating meteorological features after spatiotemporal correction, the decoder ensures that when generating prediction values, it refers to meteorological information that is truly relevant at physical time, rather than erroneous information obtained solely from waveform similarity matching.
[0061] As an aid, simultaneously processing high-frequency hidden state vector sequences Perform standard attention feature extraction to obtain the inertial channel context vector. To ensure sufficient feature extraction, this embodiment employs a standard dot product attention mechanism to capture inertial dependencies. First, the inertial correlation score between the current decoder state and high-frequency historical states is calculated. : ; in, , , These are the learnable parameters for inertial channel attention. The inertial attention weights are then calculated using the Softmax function. : ; Note that the upper bound for the summation of the denominator here is... This means that weights are assigned only within the historical measurement window. Finally, a weighted sum is obtained to obtain the inertial channel context vector. : ; The decoder will use the weather channel context vector Inertial channel context vector and the output value of the previous time step The concatenated state is used as input, and the current hidden state is obtained by updating it according to the standard state update formula of GRU. The system then outputs the predicted wind power value for the current moment through a fully connected layer. For the specific internal control calculation logic of the decoder and the linear transformation operations of the fully connected layer, those skilled in the art can refer to the general technical specifications of recurrent neural networks, which will not be elaborated upon here.
[0062] See attached document Figure 1 Step S600 mainly involves generating dynamic gating coefficients based on the global alignment cost, using these dynamic gating coefficients to weight the meteorological channel context vector, and fusing them with the features of the high-frequency hidden state vector to output the predicted wind power value for future moments. This embodiment constructs a quality-aware dynamic gating fusion mechanism to address the interference of numerical weather forecast data quality fluctuations on prediction results at different times. In actual power grid operation, the alignment degree between meteorological forecast data and measured power data objectively reflects the reliability of meteorological data during that period. By introducing the global alignment cost calculated in step S200 as an explicit indicator to measure the quality of meteorological data, the system can adaptively adjust the weight of meteorological information in the final prediction logic, preventing erroneous meteorological guidance information from disrupting the inertia-based prediction trend when the meteorological forecast deviation is large (i.e., the alignment cost is high).
[0063] In practical implementation, the multi-step wind power prediction system based on heterogeneous seq2seq and NWP refinement obtains the meteorological channel context vector output in step S500. and inertial channel context vector Then, the operation of generating dynamic gating coefficients based on global alignment costs is performed.
[0064] Global alignment cost It is the minimum cumulative distance normalized value calculated by the Dynamic Time Warping (DTW) algorithm in step S200. Its value directly quantifies the degree of matching difference between the meteorological sequence pattern and the measured power sequence pattern within the historical time window. The larger the value, the lower the physical consistency of the meteorological data, and the weaker its guiding significance for the future. In order to transform this scalar indicator into a control signal that can participate in neural network operations, a gated sensing layer was constructed.
[0065] For each prediction step (in ), calculate dynamic gating coefficient This coefficient is a vector ranging from (0, 1) and is used to determine the pass rate of each dimension of meteorological characteristics. The calculation follows the formula: ; in, For the current moment The dynamic gating coefficient vector; A learnable weight matrix for context vectors, used to perceive the current feature content; A learnable weight vector for alignment cost, used to scalar global alignment cost. Mapping to In the same dimensional space, the weight vector is typically optimized into negatively correlated parameters during training, that is, as... As the value increases, the gated output value decreases. This is the gated bias vector. This is the Sigmoid activation function, used to map the result of a linear transformation to the interval (0, 1).
[0066] Based on the above calculations, a negative feedback mechanism is established: when the global alignment cost... When significantly increased, The overall value decreased, thereby suppressing the transmission of low-quality meteorological information.
[0067] Based on this, an operation is performed to weight the meteorological channel context vector using dynamic gating coefficients, generating a weighted meteorological feature vector. : ;in, It represents the Hadamardi (or Hadama) stack.
[0068] Subsequently, the process is executed, fusing the data with the inertial channel context vector to output the predicted wind power for future timeframes. This inertial channel context vector aggregates autoregressive inertial information from high-frequency measured data. The weighted meteorological feature vector is then... Inertial channel context vector and the predicted output value of the previous time step. Deep fusion is performed. The fusion process uses a concatenation method to construct the decoder's input vector. : ; in, For the previous moment The predicted output value (Teacher Forcing can be performed using real labels during the training phase). This indicates a vector concatenation operation.
[0069] Using the decoder based on the decoder hidden state of the previous time step and the current input vector Update the current decoder hidden state : ; in, This represents the state update function of a standard gated loop unit.
[0070] Finally, a fully connected readout layer maps the decoder state to the wind power prediction for future time steps. : ; in, and These are the weight matrix and bias term of the output layer, respectively. For multi-step prediction tasks, the system repeats steps S500 to S600 until a complete future prediction sequence is generated.
[0071] During the training phase of the heterogeneous Seq2Seq model, this system employs the backpropagation algorithm to jointly optimize all learnable parameters (including encoder weights, decoder weights, attention weights, and fully connected layer weights). The loss function is defined as the sequence of predicted values. With the true power sequence Mean squared error (MSE) between: ; By minimizing this loss function, the model can automatically learn how to align the inputs according to the cost. By dynamically adjusting the gating coefficient, the effective use of meteorological information can be maximized while ensuring forecast accuracy.
[0072] Through step S600, the dynamic gating fusion mechanism in this embodiment ensures robustness under different meteorological data qualities. When the weather forecast is accurate ( When the weather forecast is small, make full use of meteorological characteristics to capture long-term trends; when the weather forecast deviation is large ( When the weather is large, the system automatically reduces its reliance on meteorological features and instead relies more on high-frequency inertial information for conservative predictions, thereby reducing the risk of extreme value errors while ensuring prediction accuracy. The specific mathematical operations of the fully connected layer and decoder are well-known technologies in this field and will not be elaborated upon here.
Claims
1. A multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement, characterized in that, Includes the following steps: High-frequency inertial sequences from wind farms and low-frequency meteorological sequences from meteorological service providers are collected, and the low-frequency meteorological sequences are interpolated to obtain aligned meteorological sequences. Select a historical time window, calculate the dynamic time warping path between the measured wind speed in the high-frequency inertial sequence and the predicted wind speed in the aligned meteorological sequence, and extract the original path slope and global alignment cost from the dynamic time warping path; Linear regression analysis is performed on the dynamic time warped path to calculate the path linearity. A confidence factor is generated based on the path linearity, and the original path slope is corrected using the confidence factor to obtain a corrected slope factor. A heterogeneous neural network comprising a first encoder and a second encoder is constructed. The high-frequency inertial sequence is input into the first encoder to extract the high-frequency hidden state vector, and the aligned meteorological sequence is input into the second encoder to extract the low-frequency hidden state vector. The virtual mapping index is calculated based on the current prediction step size and the correction slope factor. Attention features are extracted from the low-frequency hidden state vector based on the virtual mapping index to obtain the meteorological channel context vector. Attention features are also extracted from the high-frequency hidden state vector to obtain the inertial channel context vector. Dynamic gating coefficients are generated based on the global alignment cost. The dynamic gating coefficients are then used to weight and control the meteorological channel context vector, and fused with the inertial channel context vector to output the wind power prediction value for future time.
2. The multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement according to claim 1, characterized in that, The steps of collecting high-frequency inertial sequences from wind farms and low-frequency meteorological sequences from meteorological service providers, and interpolating the low-frequency meteorological sequences to obtain aligned meteorological sequences include: The high-frequency inertial sequence of the wind farm is defined as a feature vector containing active power, measured wind speed and measured wind direction, and the measured wind direction is decomposed into sine and cosine components. The low-frequency meteorological sequence is reconstructed using a cubic spline interpolation algorithm to construct a piecewise cubic polynomial function, such that the function value of the piecewise cubic polynomial function is equal at all known data points, and the first and second derivatives at the nodes are continuous. The aligned meteorological sequence is generated based on the piecewise cubic polynomial function, and its time resolution is consistent with that of the high-frequency inertial sequence.
3. The multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement according to claim 1, characterized in that, The step of calculating the dynamic time warping path between the measured wind speed in the high-frequency inertial sequence and the predicted wind speed in the aligned meteorological sequence includes: The measured wind speed sequence is extracted from the high-frequency inertial sequence, and the forecast wind speed sequence is extracted from the aligned meteorological sequence; Construct a Euclidean distance matrix, wherein the elements of the Euclidean distance matrix represent the distance measure between the values of time points in the measured wind speed sequence and the values of time points in the predicted wind speed sequence; Calculate the cumulative cost matrix, where each element represents the minimum cumulative path cost from the starting point to the current position; The dynamic programming algorithm is used to search for the path with the minimum cumulative distance in the cumulative cost matrix as the dynamic time warping path. The dynamic time warping path consists of a series of coordinate points representing the best alignment between the measured wind speed sequence time and the predicted wind speed sequence time.
4. The multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement according to claim 1, characterized in that, The steps of extracting the original path slope and global alignment cost from the dynamic time warping path include: Calculate the normalized value of the sum of local Euclidean distances of all nodes on the dynamic time warping path, and use the normalized value as the global alignment cost; For each path point on the dynamic time warping path, the ratio of the predicted time index increment to the measured time index increment is calculated, and the ratio is used as the slope of the original path. The method further includes: calculating the average vertical distance of the dynamic time warping path relative to the geometric diagonal to obtain the time phase drift.
5. The multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement according to claim 1, characterized in that, The step of performing linear regression analysis on the dynamic time warped path to calculate the path linearity includes: The measured time index of the path point in the dynamic time warping path is extracted as the independent variable, and the predicted time index is extracted as the dependent variable. Construct a linear regression model, and use the least squares method to solve for the regression slope and intercept of the linear regression model in order to fit an optimal fitting line; Calculate the coefficient of determination between the dynamic time warping path and the best-fit line, and use the coefficient of determination as the path linearity. The path linearity is used to characterize the linear time series correspondence between the measured time index and the predicted time index.
6. The multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement according to claim 1, characterized in that, The step of generating a confidence factor based on the path linearity and using the confidence factor to correct the original path slope to obtain a corrected slope factor includes: A mapping relationship between the path linearity and the confidence factor is constructed using a Logistic activation function, such that the confidence factor exhibits a non-linear continuous change in the interval from zero to one. A correction formula based on elastic constraints is constructed, and the original path slope and the unit slope are weighted and summed using the confidence factor to calculate the correction slope factor; wherein, when the confidence factor value is large, the correction slope factor is dominated by the original path slope, and when the confidence factor value is small, the correction slope factor tends to the unit slope.
7. The multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement according to claim 1, characterized in that, The steps of constructing a heterogeneous neural network including a first encoder and a second encoder, inputting the high-frequency inertial sequence into the first encoder to extract a high-frequency hidden state vector, and inputting the aligned meteorological sequence into the second encoder to extract a low-frequency hidden state vector include: The first encoder and the second encoder are configured as a parallel bidirectional gated cyclic unit network structure, and the first encoder and the second encoder do not share weight parameters. The high-frequency inertial sequence is input into the first encoder, and the forward hiding state and the reverse hiding state are calculated using the gating mechanism of the first encoder. The high-frequency hiding state vector is then obtained by concatenating them. The aligned meteorological sequence is input into the second encoder, and the forward meteorological hidden state and the reverse meteorological hidden state are calculated using the independent weight matrix of the second encoder. The two are then concatenated to obtain the low-frequency hidden state vector.
8. The multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement according to claim 1, characterized in that, The steps of calculating the virtual mapping index based on the current prediction step size and the correction slope factor, and extracting attention features from the low-frequency hidden state vector based on the virtual mapping index to obtain the meteorological channel context vector include: Calculate the statistical representative value of the corrected slope factor, and superimpose the product of the statistical representative value and the current prediction step size onto the benchmark alignment position to obtain the virtual mapping index; A hybrid attention mechanism incorporating positional Gaussian bias is constructed to calculate the content relevance score between the decoder state and the low-frequency hidden state vector; A position deviation score is constructed using the virtual mapping index, and the position deviation score is in the form of a Gaussian kernel function to penalize time steps that deviate from the virtual mapping index; The attention weight is obtained by superimposing and normalizing the content relevance score and the position deviation score, and then using the attention weight to perform a weighted summation of the low-frequency hidden state vector to obtain the meteorological channel context vector.
9. The multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement according to claim 1, characterized in that, The step of generating dynamic gating coefficients based on the global alignment cost and using the dynamic gating coefficients to perform weighted control on the meteorological channel context vector includes: The meteorological channel context vector is input into the gated perception layer, and the global alignment cost is mapped to the same dimensional space as the meteorological channel context vector through a learnable weight vector. The linear transformation result of the meteorological channel context vector and the mapped global alignment cost is processed by the Sigmoid activation function to generate the dynamic gating coefficient vector with values ranging from zero to one. A negative feedback mechanism is established through the dynamic gating coefficient vector, so that the value of the dynamic gating coefficient vector decreases when the global alignment cost increases; The Hadamard product of the dynamic gating coefficient vector and the meteorological channel context vector is calculated to obtain the weighted meteorological feature vector.
10. The multi-step wind power prediction method based on heterogeneous seq2seq and NWP refinement according to claim 9, characterized in that, The step of generating dynamic gating coefficients based on the global alignment cost and using the dynamic gating coefficients to perform weighted control on the meteorological channel context vector includes: The meteorological channel context vector is input into the gated perception layer, and the global alignment cost is mapped to the same dimensional space as the meteorological channel context vector through a learnable weight vector. The linear transformation result of the meteorological channel context vector and the mapped global alignment cost is processed by the Sigmoid activation function to generate the dynamic gating coefficient vector with values ranging from zero to one. A negative feedback mechanism is established through the dynamic gating coefficient vector, so that the value of the dynamic gating coefficient vector decreases when the global alignment cost increases; The Hadamard product of the dynamic gating coefficient vector and the meteorological channel context vector is calculated to obtain the weighted meteorological feature vector.