A wind power prediction method based on multi-objective optimization

Abstract

The application belongs to the technical field of artificial intelligence, and particularly relates to a wind power prediction method based on multi-objective optimization, which comprises wind power data acquisition and training data set construction, adaptive sliding window and dynamic fluctuation decomposition of wind power time series data, short-term power prediction model construction and short-term power prediction. The method can autonomously adjust the historical window length and decomposition scale according to the inherent fluctuation characteristics of the data, realize the multi-scale accurate representation of the non-stationary power sequence, construct a dynamic space-time graph by fusing the geographical distance and the correlation of instantaneous power, design a graph attention network combined with the trend similarity gate, realize the dynamic fine modeling of the spatial correlation, dynamically balance the prediction accuracy and interval prediction reliability through the fluctuation intensity adaptive loss function, synchronously output the deterministic and probabilistic prediction results, and embed the rated power limit and the climbing rate constraint in the model in a soft manner, so that the prediction results conform to the actual operation law of the wind turbine.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a wind power prediction method based on multi-objective optimization. Background Technology

[0002] Currently, with the global energy structure transitioning towards clean and low-carbon energy, the penetration rate of wind power in the power system is rapidly increasing. However, the inherent intermittency, volatility, and uncertainty of wind energy pose serious challenges to the safe and stable operation of the power grid and power dispatch. Accurate short-term wind power forecasting is a key technological support for mitigating these challenges, improving wind power absorption capacity, and reducing system backup costs. For a long time, researchers have conducted extensive research on this topic, forming an evolutionary path from traditional physical modeling and statistical time series analysis to modern machine learning methods.

[0003] In the prior art, Chinese invention patent CN121094207A, published on December 9, 2025, discloses a decoupling evaluation method for wind power prediction errors considering turbine failures and real-time capacity changes. This method includes dividing wind power prediction into three stages: numerical weather prediction, wind-power model conversion, and power correction, based on the actual wind farm prediction process. For the power correction stage, a turbine failure prediction model is constructed to predict capacity reduction caused by failures in advance. Combining a health index and a capacity decay coefficient, faulty turbines are removed in real time, and the available capacity of the turbines is dynamically corrected. The equivalent actual capacity is calculated using a weighted average based on the health index. Finally, the equivalent actual capacity of the remaining turbines is introduced into the error decoupling model to quantitatively evaluate the prediction errors caused by each stage. The proportion of prediction errors caused by each stage is determined, identifying the sources of wind power errors after considering turbine failures and capacity decay.

[0004] However, existing technologies typically employ fixed-length historical time windows, which are ill-suited to the inherent non-stationarity and volatility of wind power. This leads to the introduction of redundant noise during stable periods and the loss of crucial precursory information during periods of rapid change. Furthermore, existing technologies often treat neighboring wind farm data as independent features or employ static spatial relationship modeling, neglecting the dynamic changes in the intensity of influence between wind farms over time and operational status, resulting in insufficient utilization of spatial correlation information. Moreover, most existing technologies focus on the single objective of point prediction accuracy or require training multiple models separately for probabilistic prediction, lacking an effective mechanism for collaboratively optimizing point prediction accuracy and interval prediction reliability within a unified framework. Additionally, existing technologies often treat prediction models as purely data-driven black boxes, and the output results may violate the physical operational constraints of wind turbines (such as rated upper limits and ramping limitations), affecting the practicality and acceptability of the prediction results.

[0005] Therefore, there is an urgent need to develop a wind power prediction method based on multi-objective optimization to effectively solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a wind power prediction method based on multi-objective optimization to solve the problem of accurate short-term wind power prediction.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A wind power prediction method based on multi-objective optimization includes the following steps:

[0009] S1. Wind power data acquisition and training dataset construction:

[0010] The raw data collection was completed by continuously recording the operating status of the wind farm at fixed time intervals. After obtaining raw sequence data with a long time span, a training sample set for supervised learning was constructed. The dataset was divided into non-overlapping training set, validation set and test set according to time order, which were used for learning model parameters, adjusting hyperparameters and final evaluation of generalization performance, respectively.

[0011] S2. Adaptive sliding window and dynamic fluctuation decomposition of wind power time series data:

[0012] An adaptive sliding window strategy based on local fluctuation intensity perception is adopted to perform dynamic mode decomposition and separate fluctuation components at different time scales from the original data;

[0013] S3. Construct a short-term power prediction model:

[0014] S31. Mining of collaborative features of neighboring wind fields based on spatiotemporal graphs;

[0015] S32. Short-term wind power prediction based on multi-objective optimization;

[0016] S33. Loss function calculation;

[0017] S34. Training of short-term power prediction model;

[0018] S4. Short-term power prediction:

[0019] Once the short-term power prediction model is trained, it can be used to predict short-term power on new and unseen data. In the practical application stage, the latest historical power data sequence of the target wind farm and its neighboring wind farms, as well as the corresponding latest meteorological characteristic data, are first obtained in real time.

[0020] Further, in step S1, the collected raw data includes the actual power output time series data of the wind farm group and multi-source meteorological data closely related to wind power generation; among them, for a small number of missing values ​​caused by equipment failure or communication interruption, linear interpolation or the average of the effective values ​​before and after the time is used to fill in the missing values.

[0021] Furthermore, in step S1, each training sample consists of model input features and corresponding labeled ground truth values. For each moment in the time series, the power data and related meteorological data of its historical period are used as input features, and the labeled content is the future continuous time after that moment, that is, the predicted real power value of the target wind farm within the field of view. The labeling method is the natural alignment of the time series, that is, each input feature sample corresponds to its real power output value at a specific future time step as a supervision signal, and the labeling category is continuous power value.

[0022] Further, step S2 specifically includes the following steps:

[0023] S21. Calculate the local wave intensity index:

[0024] To quantify the intensity of fluctuations in the original power sequence near each sample point, a local fluctuation intensity index is calculated based on the absolute value of power change within the basic observation window and its adaptive weights.

[0025] S22. Determine the adaptive history window length:

[0026] Based on the calculated local fluctuation intensity index, the length of the historical window used as the input to the prediction model is dynamically determined to ensure that more historical information is included when fluctuations are severe in order to capture abrupt change patterns, and the window is shortened when fluctuations are gentle to avoid noise interference.

[0027] S23. Perform dynamic fluctuation decomposition and feature construction:

[0028] The original power sequence of variable length is truncated according to the adaptive historical window length, and then the sequence is dynamically filtered and decomposed to separate the low-frequency trend component and the high-frequency fluctuation component. Then, the decomposed components are concatenated to obtain the decomposed time series feature vector.

[0029] Further, in step S31, the spatial relationships between wind farms are explicitly modeled by constructing a dynamic spatiotemporal graph, and a graph attention network with a gating mechanism is used to aggregate spatial features and mine spatial collaborative features. The specific steps are as follows:

[0030] S311. Constructing a dynamic spatiotemporal graph and calculating the adjacency matrix:

[0031] To model the dynamic spatial dependencies between wind farms, a weighted directed graph is constructed for each sample time. A set of nodes is defined, which includes the target wind farm and its geographically nearest wind farms. The edge set adopts a fully connected approach to indicate the potential influence relationship between nodes. The adjacency matrix is ​​used to quantify the dynamic influence intensity between nodes. Its elements represent the influence weight of one node on another node at a specific time. This weight is calculated by combining geographical distance and instantaneous power correlation.

[0032] S312. Perform spatial feature aggregation based on gated graph attention:

[0033] The decomposed temporal feature vector of each wind farm node at the current time point is used as its initial node feature. An improved graph attention network is used to aggregate spatial features. The improved graph attention network combines a gating mechanism based on the similarity of power change trends with the conventional graph attention network to enhance attention to neighbors with similar evolution patterns when fluctuations are severe, thereby obtaining spatial feature enhancement features.

[0034] Further, in step S32, an end-to-end prediction model is constructed, focusing on multi-point prediction accuracy, interval prediction reliability, and adherence to physical laws. This model deeply integrates and decomposes temporal and spatial collaborative features, and introduces dynamically weighted multi-source meteorological features. A quantile regression framework is used to simultaneously output point and interval predictions for multiple future time steps. Simultaneously, a constraint layer based on physical priors is employed to correct the original predictions. Specifically, this includes the following steps:

[0035] S321. Feature Fusion and Dynamic Meteorological Feature Injection:

[0036] To integrate features from different sources and processing levels, firstly, the decomposed temporal feature vector of the target wind farm at the target time is concatenated with its spatial feature enhancement features to form the basic spatiotemporal fusion feature. Secondly, external multi-source meteorological features are introduced to provide key driving information. Considering that different meteorological factors have different weights on power at different times, an attention-based dynamic weighted encoder is used to encode the meteorological features and inject them into the fusion feature.

[0037] S322, Multi-objective prediction head based on quantile regression:

[0038] To obtain point predictions and uncertainty interval predictions simultaneously, the model adopts a quantile regression framework. Specifically, it uses a shared backbone network to process fused features and then connects multiple quantile-specific output layers in parallel. Each output layer corresponds to a target quantile and outputs the power prediction value for several future time steps. At the same time, an additional point prediction branch is defined, whose loss is co-optimized with that of the quantile prediction branch.

[0039] S323. Correction of prediction results based on physical priors:

[0040] The original prediction output may violate the physical operating constraints of the wind turbine. To ensure the physical rationality of the prediction results, a differentiable correction layer is used to constrain the prediction value of each quantile at each time step. The correction function considers both the upper limit of rated power and the maximum allowable rate of power change.

[0041] S324. Model Integration and Output:

[0042] The final model outputs the future. For each time step, the corrected predicted values ​​of all target quantiles and the corrected point predictions are used. For point predictions, the corrected result of the median quantile is taken as the final point prediction value, or the result of the point prediction branch is processed by the same correction function.

[0043] Further, in step S33, a composite loss function is adopted. This composite loss function integrates the accuracy loss of point-oriented prediction and the reliability loss of interval-oriented prediction, and balances them through an adaptive weighting mechanism based on local fluctuation intensity. This effectively drives the joint optimization of the multi-objective prediction model, specifically including the following steps:

[0044] S331. Constructing a composite loss for point prediction and interval prediction:

[0045] The loss function consists of two parts: point prediction loss and quantile regression loss. The point prediction loss directly optimizes the model's ability to predict the most likely value of future power, while the quantile regression loss ensures that the prediction interval can cover the true value with a preset probability by minimizing the quantile loss at multiple quantiles.

[0046] S332, Adaptive weighted loss based on fluctuation intensity:

[0047] The fluctuation characteristics of wind power series are time-varying. The model should have different optimization focuses during the stable period and the period of dramatic change. During the stable period, more attention should be paid to the accuracy of point prediction, while during the period of dramatic change, it is necessary to strengthen the learning of the reliability of the uncertainty interval. Therefore, the weights of the two items in the composite loss should be dynamically adjusted by combining the local fluctuation intensity index.

[0048] Further, in step S34, the training process uses the training dataset constructed in step S1 for iterative optimization. In each training cycle, samples from the training dataset are randomly grouped into mini-batches and input into the short-term power prediction model. The model first performs adaptive sliding window and dynamic fluctuation decomposition in step S2 based on the input historical power data to obtain the decomposed temporal feature vector. Then, step S31 is executed to construct a dynamic spatiotemporal map and perform spatial feature aggregation to obtain spatial feature enhancement features. Next, step S32 is executed to perform multi-scale feature fusion and meteorological feature injection, and through the quantile regression prediction head and physical correction layer, the corrected post-point prediction values ​​and quantile prediction values ​​for multiple future time steps are output. Finally, according to the dynamic weighted loss function defined in step S33, the total loss value between the model's predicted output and the corresponding real power label is calculated.

[0049] Furthermore, an adaptive moment estimation optimizer is employed to backpropagate and update all trainable parameters in the model based on the calculated total loss. After each training cycle, the model performance is evaluated using a validation set, and the validation loss is calculated. Training continues until the validation loss no longer decreases significantly over multiple consecutive training cycles, or until the preset maximum number of training cycles is reached. At this point, the model can be considered to have fully learned the patterns in the data and reached a convergence state, at which point the training process ends. After training is completed, the model parameters with the best validation set performance are saved for subsequent prediction use.

[0050] Further, in step S4, for the moment when prediction is needed, the process is carried out in complete agreement with the training phase: Following step S2, the local fluctuation intensity index for that moment is calculated, the adaptive historical window length is dynamically determined, and the historical power sequence within the window is dynamically decomposed to generate a fixed-dimensional decomposed temporal feature vector; then, following step S31, a dynamic adjacency matrix is ​​calculated based on the power data of each wind farm at the current moment, a spatiotemporal graph is constructed, and the decomposed temporal feature vectors of each node are passed through a trained spatial feature aggregation network based on gated graph attention to obtain the target wind farm's... The spatial feature enhancement feature is then performed. Following step S32, the decomposed temporal feature vector of the target wind farm, the spatial feature enhancement feature, and the multi-source meteorological features after dynamic weighted encoding are fused to form the final fused feature vector. This feature vector is then input into the trained, parameter-fixed prediction model backbone network and quantile output layer to obtain the original point predictions and quantile predictions for multiple future time steps. Finally, these original predictions are processed by the same parameter-fixed physical prior correction layer to output the final corrected point predictions that conform to physical constraints and the power interval predictions at each quantile.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] 1. This invention employs an adaptive sliding window and dynamic filtering decomposition based on local fluctuation intensity perception, which can autonomously adjust the historical window length and decomposition scale according to the inherent fluctuation characteristics of the data, thereby achieving multi-scale accurate characterization of non-stationary power sequences.

[0053] 2. A dynamic spatiotemporal graph integrating the correlation between geographical distance and instantaneous power was constructed, and a graph attention network with trend similarity gating was designed to achieve dynamic and refined modeling of spatial relationships.

[0054] 3. Establish a multi-objective optimization framework that integrates quantile regression and point prediction. Dynamically balance the accuracy of point prediction and the reliability of interval prediction through a loss function that adapts to fluctuation intensity, and simultaneously output deterministic and probabilistic prediction results.

[0055] 4. Design a differentiable physical constraint correction layer to softly embed rated power limits and ramp rate constraints into the model, ensuring that the prediction results conform to the actual operating laws of wind turbine units. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart illustrating the steps of the wind power prediction method based on multi-objective optimization of the present invention.

[0058] Figure 2 A flowchart of adaptive sliding window and dynamic fluctuation decomposition for wind power time series data;

[0059] Figure 3 This is a flowchart of the collaborative feature mining process for neighboring wind fields based on a spatiotemporal graph.

[0060] Figure 4 This is a flowchart of short-term wind power prediction based on multi-objective optimization. Detailed Implementation

[0061] The present invention will be further described below with reference to embodiments:

[0062] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0063] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0064] like Figure 1 As shown, this invention proposes a wind power prediction method based on multi-objective optimization, comprising the following steps:

[0065] S1. Wind power data acquisition and training dataset construction:

[0066] The dataset required for model training and testing is collected. The data source is the monitoring and data acquisition system deployed in the wind farm, which continuously records the operating status of the wind farm at fixed time intervals, such as 15 minutes.

[0067] The collected raw data mainly includes two categories: 1. Time series data of actual power output of wind farm clusters, covering the target wind farm and several geographically adjacent wind farms; 2. Multi-source meteorological data closely related to wind power generation, such as wind speed, wind direction, ambient temperature and air pressure, which can come from the meteorological tower monitoring equipment on the wind farm itself or external numerical weather forecast products.

[0068] The data acquisition process must ensure the continuity and integrity of the data. For a small number of missing values ​​caused by equipment failure or communication interruption, linear interpolation or the average of the valid values ​​before and after the time interval should be used to fill in the missing values.

[0069] After obtaining the raw sequence data over a long period, a training sample set needs to be constructed for supervised learning. Each training sample consists of model input features and corresponding labeled ground truth values. Specifically, for each moment in the time series, the power data and related meteorological data from its historical period are used as input features, and the labeled content is the future continuous time period after that moment, i.e., the predicted true power value of the target wind farm within the field of view.

[0070] The annotation method is the natural alignment of the time series, that is, each input feature sample corresponds to its true power output value at a specific future time step as a supervision signal, and the annotation category is continuous power value, usually in megawatts.

[0071] The final dataset will be divided into non-overlapping training, validation, and test sets in chronological order, which will be used for learning model parameters, tuning hyperparameters, and final evaluation of generalization performance, respectively.

[0072] S2. Adaptive sliding window and dynamic fluctuation decomposition of wind power time series data:

[0073] The actual power output data of wind farms is time series data with strong non-stationarity, high volatility and significant intermittency. Its fluctuation pattern changes rapidly with the weather system. Conventional techniques usually use a fixed-length time window to extract historical data as model input. Static windows cannot adapt to the changes in the inherent fluctuation intensity of the data. In the stable period, it may contain redundant information, and in the period of drastic change, it may lose key precursor features, making it difficult for the model to accurately capture short-term power changes in the future.

[0074] like Figure 2As shown, this invention employs an adaptive sliding window strategy based on local fluctuation intensity perception, and performs dynamic mode decomposition on this basis to separate fluctuation components at different time scales from the original data. The specific steps are as follows:

[0075] S21. Calculate the local wave intensity index:

[0076] To quantify the intensity of fluctuations in the original power sequence near each sample point, a local fluctuation intensity index is calculated based on the absolute value of power changes within the baseline observation window and its adaptive weights, and is expressed as:

[0077] ;

[0078] In the formula, Indicates the first The local fluctuation intensity index at a given time point is used to quantify the severity of power changes near that point; the larger the value, the more severe the fluctuation.

[0079] Integer time index representing the current target sample point;

[0080] Indicates the wind farm in the 1st The actual power output value at each time point Indicates the wind farm in the 1st The actual power output value at each point in time, in megawatts, is directly obtained from the wind farm monitoring and data acquisition system.

[0081] This represents the length of the baseline observation window, which is the length of the baseline observation window used when calculating the intensity of fluctuations. It is used to define the range of historical data examined when calculating the intensity of local fluctuations. An example value is 24, which corresponds to a 6-hour historical data window.

[0082] This represents an adaptive weighting function, used to assign different importance to historical data points; its value varies with the historical data points. With the current point The power decreases as the distance increases, meaning that the contribution of historical power changes closer to the current point to the current fluctuation intensity is greater. This is calculated as follows: ;

[0083] Indicates difference from Integer time index;

[0084] This represents the decay parameter of the adaptive weight function, which controls the rate at which the weight decays as the distance between the historical point and the current point increases. An example value is 0.1.

[0085] S22. Determine the adaptive history window length:

[0086] Based on the calculated local fluctuation intensity index, the length of the historical window used as the input to the prediction model is dynamically determined. This ensures that more historical information is included to capture abrupt change patterns during periods of intense fluctuation, while the window is shortened during periods of calm to avoid noise interference. This is expressed as follows:

[0087] ;

[0088] In the formula, Indicates for the first The length of the adaptive historical window used for each target prediction time, in time steps, is the length of the historical data used as input to the prediction model.

[0089] This represents the base window length, which serves as the benchmark for window adjustments. An example value is 16.

[0090] This represents the upper limit of the window length adjustment range, constraining the maximum change in the adaptive historical window length relative to the base window. An example value is 8.

[0091] This represents the hyperbolic tangent activation function, which maps the input to the interval [0, 1]. Inside;

[0092] This represents the scaling factor, used to control the sensitivity of the local fluctuation intensity index to the influence of the window length. An example value is 2.0.

[0093] This represents the fluctuation intensity threshold, which serves as a benchmark for determining the level of power fluctuations. An example value is 0.5.

[0094] This indicates a function to round down, ensuring that the window length is an integer.

[0095] S23. Perform dynamic fluctuation decomposition and feature construction:

[0096] The original power sequence of variable length is truncated according to the adaptive historical window length. Then, the sequence is dynamically filtered and decomposed to separate the low-frequency trend component and the high-frequency fluctuation component. Finally, the decomposed components are concatenated to obtain the decomposed time series feature vector, represented as:

[0097] ;

[0098] In the formula, Indicates the first The decomposed temporal feature vectors at each time point have a length of . , characterizing the feature representation of the original power sequence after dynamic fluctuation decomposition;

[0099] Represents the original power sequence The low-frequency trend component obtained after low-pass filtering characterizes the slow changing trend of the power data, reflecting the long-term evolution pattern, and has a length of [missing information]. The method of obtaining is expressed as ;

[0100] Represents the original power sequence The high-frequency fluctuation component obtained after high-pass filtering contains fast fluctuations and noise components in the power sequence, and its length is [missing information]. The method of obtaining is expressed as ;

[0101] This represents the column vector concatenation operation;

[0102] Indicates the kernel size as The one-dimensional average convolution operation acts as a low-pass filter by calculating the average value of the elements in the kernel, and fills in zeros to make the output length consistent with the input length.

[0103] Indicates from the first From the point in time to the The original power sequence at each time point has a length of [number]. ,Right now ;

[0104] Indicates the wind farm in the 1st The actual power output value at each time point Indicates the wind farm in the 1st The actual power output value at each time point Indicates the wind farm in the 1st The actual power output value at each point in time, in megawatts, is directly obtained from the wind farm monitoring and data acquisition system.

[0105] This represents the dynamically adjusted kernel size used in one-dimensional average convolution operations, ensuring that the filter scale is proportional to the data window length. The calculation method is expressed as follows: ;

[0106] This represents the floor function;

[0107] This represents the scaling factor that controls the strength of the low-pass filter. An example value is 0.25, which is used to dynamically determine the kernel size.

[0108] In practical implementation, the low-frequency trend component is obtained by performing a one-dimensional average convolution operation on the original sequence. This component is used to characterize the slow changing trend of the power data. The convolution kernel size is dynamically adjusted according to the window length, specifically by using a kernel of length [missing information]. The convolution kernel for the sequence It is implemented by performing a convolution operation with a stride of 1 and padding both ends with zero values ​​to keep the output length unchanged.

[0109] Furthermore, the time-series feature vectors will be decomposed. Unified adjustment to fixed dimensions ,like Then for Perform a center cutoff, retaining the middle section. If an element, Then in Fill both ends with zero values ​​to make its length reach In the formula, This indicates the preset fixed input feature dimension, with an example value of 64.

[0110] S3. Construct a short-term power prediction model:

[0111] S31. Collaborative feature mining of neighboring wind fields based on spatiotemporal graph construction:

[0112] The power output of a single wind farm depends not only on its own historical state, but also on the spatial correlation of meteorological conditions of neighboring wind farms, such as wind speed, wind direction and wake effect. As a typical spatiotemporal prediction problem, conventional techniques often only use the historical power sequence of the local farm, ignoring the spatial correlation information, or simply use the power of neighboring wind farms as an additional independent time series feature input, making it difficult to model its spatial topological relationship. This results in the model being unable to effectively use spatial dependence to correct local prediction errors.

[0113] like Figure 3 As shown, this invention explicitly models the spatial relationships between wind farms by constructing a dynamic spatiotemporal graph, and uses a graph attention network with a gating mechanism to aggregate spatial features, thereby mining valuable spatial collaborative features. The specific steps are as follows:

[0114] S311. Constructing a dynamic spatiotemporal graph and calculating the adjacency matrix:

[0115] To model the dynamic spatial dependencies between wind farms, a weighted directed graph is constructed for each sample time. Specifically, a node set is defined, containing the target wind farm and its geographically nearest neighboring wind farms. The edge set uses a fully connected approach to represent potential influence relationships between nodes. The adjacency matrix is ​​used to quantify the dynamic influence strength between nodes; its elements represent the influence weight of one node on another at a specific time. This weight is calculated by combining geographical distance and instantaneous power correlation, and is expressed as:

[0116] ;

[0117] In the formula, Representing a dynamic spacetime graph adjacency matrix The Middle line, number The elements of a column represent nodes. For nodes In the The dynamic impact weight at each moment, with a larger value indicating a stronger impact;

[0118] Indicates the first A dynamic spatiotemporal graph constructed at each time step is used to model the spatial relationships between wind farms, where nodes represent wind farms and edge weights are determined by the adjacency matrix. definition;

[0119] Represents the natural exponential function;

[0120] Indicates wind farm With wind farm The standardized geographical distance between them, which has been normalized based on the maximum distance between all wind farm pairs, so that its value is within... interval;

[0121] This represents the distance decay coefficient, used to adjust the rate at which the influence of geographical distance on the weight decays. An example value is 2.0.

[0122] Indicates that in the first The wind farm calculated within a short time window centered on a given moment With wind farm The Pearson correlation coefficient of the output power time series is used to characterize the instantaneous synchronicity of the power changes between the two.

[0123] This represents the maximum value function, used to ensure that only positive correlations are retained and negative correlations are ignored, meaning that negative correlations are considered not to constitute an effective synergistic effect;

[0124] This represents the total number of nodes in the dynamic spatiotemporal graph, including the target wind farm and its... A geographically adjacent wind farm, namely ,in This is a preset neighbor hyperparameter, with an example value of 5;

[0125] Indicates the index of a node in the first dynamic spatiotemporal graph;

[0126] This represents the index of a node in the second dynamic spatiotemporal graph.

[0127] In practical implementation, It is a wind farm and In the first The Pearson correlation coefficient of the power series within a short time window centered on a specific time point is obtained by calculating the Pearson correlation coefficient using the power value series of two wind farms within this window. The short time window refers to... Centered on, with a length of The time step window is used to capture instantaneous correlations. This represents the preset time window radius parameter. An example value of 2 indicates that the window will be examined for a total of 5 time steps.

[0128] S312. Perform spatial feature aggregation based on gated graph attention:

[0129] The decomposed temporal feature vector of each wind farm node at the current time point is used as its initial node feature. An improved graph attention network is used for spatial feature aggregation. The improved graph attention network, based on the conventional graph attention network, combines a gating mechanism based on the similarity of power change trends to enhance attention to neighbors with similar evolution patterns during periods of drastic fluctuations, thereby obtaining spatial feature enhancement features, represented as:

[0130] ;

[0131] In the formula, Represents a node Spatial feature enhancement features are nodes The spatial feature enhancement feature obtained after spatial feature aggregation based on gating graph attention has a dimension of [dimensionality missing]. ,node Represents the first in the dynamic spacetime graph Each wind farm node, at each time point For each sample, the decomposed temporal feature vectors of all nodes are represented.

[0132] This represents the exponential linear unit activation function, which introduces non-linearity into neural networks to help the model learn complex patterns;

[0133] This represents the final attention weight after gating and modulation, which combines feature similarity and trend similarity. The calculation method is as follows:

[0134] ;

[0135] Represents a node with neighboring nodes The original attention coefficients, calculated by the standard graph attention mechanism, are used to measure the neighbors. For the target node The importance is expressed as follows:

[0136] ;

[0137] This represents a gating value based on trend similarity, used to modulate the original attention weights. The closer the value is to 1, the stronger the node's performance. and The more similar the trends, the more likely the neighbor information should be retained. The calculation method is expressed as follows: ;

[0138] The weight matrix, representing the linear transformation of node features, is a trainable parameter with dimension O(n). ;

[0139] The weight vector used to calculate the attention score is a trainable parameter with dimension . , express transpose;

[0140] Represents a node In the The decomposed temporal feature vectors at each time point have a dimension of . ;

[0141] Represents a node In the The decomposed temporal feature vectors at each time point have a dimension of . ;

[0142] The output dimension of node features after spatial feature aggregation is represented by a hyperparameter that is adjustable and used to control the dimension of feature mapping.

[0143] This represents a vector concatenation operation;

[0144] This represents the activation function of a linear rectifier unit with leakage.

[0145] Represents a node The set of neighboring nodes, containing all nodes in the dynamic spatiotemporal graph related to the node. Connected nodes;

[0146] This represents the Sigmoid activation function, which maps the input to... interval;

[0147] This represents the weight vector used to calculate the gate value, and it is a trainable parameter.

[0148] This represents the bias scalar used to calculate the gate value, and it is a trainable parameter.

[0149] express The difference vectors at the most recent time steps are used to characterize the short-term power variation trend, and are calculated as follows: ;

[0150] This indicates the step size for trend calculation; an example value is 2.

[0151] Represents a node In the The decomposed temporal feature vectors at each time point have a dimension of . ;

[0152] express The difference vector at the most recent time steps.

[0153] It should be noted that nodes Represents the first in the dynamic spacetime graph Each wind farm node, at each time point For a given sample, the sample contains the decomposed temporal feature vectors of all nodes. It is a node In time Enhanced features resulting from spatial feature aggregation.

[0154] S32. Short-term wind power forecasting based on multi-objective optimization:

[0155] Short-term wind power forecasting not only requires accurate point prediction values, but also quantifies the uncertainty of the forecast to support the grid's risk perception and optimized scheduling. Conventional techniques usually build a single model to minimize point prediction errors, ignoring the dual requirements that the prediction results should simultaneously meet the requirements of deterministic accuracy and reliability range. Furthermore, the model usually directly inputs the processed spatiotemporal features into the standard regression network, failing to effectively integrate external meteorological factors and lacking explicit modeling of the physical constraints of wind power, which may lead to the model output violating physical laws.

[0156] like Figure 4 As shown, this invention constructs an end-to-end prediction model that addresses the accuracy of multi-point predictions, the reliability of interval predictions, and adherence to physical laws. It deeply integrates and decomposes temporal and spatial collaborative features, and introduces dynamically weighted multi-source meteorological features. Through a quantile regression framework, it simultaneously outputs point and interval predictions for multiple future time steps. Furthermore, it employs a constraint layer based on physical priors to correct the original predictions. The specific steps are as follows:

[0157] S321. Feature Fusion and Dynamic Meteorological Feature Injection:

[0158] To fuse features from different sources and processing levels, firstly, the decomposed temporal feature vector of the target wind farm at the target time is concatenated with its spatial feature enhancement features to form the basic spatiotemporal fusion feature. Secondly, external multi-source meteorological features are introduced to provide key driving information. Considering that different meteorological factors have different weights on power at different times, an attention-based dynamic weighted encoder is used to encode the meteorological features and inject them into the fusion feature, represented as:

[0159] ;

[0160] In the formula, Indicates the node used for the target wind farm In the The final fused feature vector predicted at each time step has a dimension of [dimensionality]. By comprehensively considering the decomposition time-series characteristics, spatial coordination characteristics, and dynamically weighted multi-source meteorological characteristics of the target wind farm, the time dependence, spatial correlation, and external driving factors affecting power prediction can be fully characterized.

[0161] This indicates the total number of meteorological feature types considered, such as wind speed, wind direction, temperature, and air pressure. An example value is 4.

[0162] Indicates the first The moment for the first The dynamic attention weights for meteorological features are used to quantify the importance of these meteorological factors to the power forecast at the current moment. The calculation method is expressed as follows:

[0163] ;

[0164] Let represent the linear transformation weight matrix used to map the original meteorological features to the coding space, and be trainable parameters with dimension . ;

[0165] Indicates the first The first time obtained from observation or forecast at the nth moment Meteorological feature vector, dimension is Presets can be obtained from external sources, such as meteorological monitoring systems or numerical weather prediction.

[0166] The query vector is obtained by linear transformation of the basic spatiotemporal fusion features, i.e. The correlation between the key vectors of meteorological features and the dynamic attention weights is calculated to determine the dynamic attention weights, enabling the model to adaptively adjust the contribution of different meteorological features to the prediction based on the current spatiotemporal state.

[0167] express transpose;

[0168] This represents the weight matrix used to generate the query vector; it is a trainable parameter with dimensions of . ;

[0169] Indicates the first The key vector of meteorological features is a trainable parameter vector used to correlate with the query vector. Calculate the correlation to determine attention weights;

[0170] The dimension representing the meteorological feature encoding is an adjustable hyperparameter;

[0171] This represents the input dimension of each type of original meteorological feature vector;

[0172] This represents the weight matrix used to generate the query vector, and it consists of trainable parameters.

[0173] S322, Multi-objective prediction head based on quantile regression:

[0174] To simultaneously obtain point predictions and uncertainty interval predictions, the model employs a quantile regression framework. Specifically, it uses a shared backbone network to process and fuse features, then connects in parallel to multiple quantile-specific output layers. Each output layer corresponds to a target quantile, outputting power prediction values ​​for several future time steps. Simultaneously, an additional point prediction branch is defined, whose loss is co-optimized with that of the quantile prediction branch, expressed as:

[0175] ;

[0176] ;

[0177] In the formula, The model represents the nodes of the wind farm. In the future Time, corresponding to quantile The predicted power value is in megawatts;

[0178] Indicated for quantiles The designed multilayer perceptron output layer has trainable parameters as follows: This is used to map the features extracted by the shared backbone network to the power prediction values ​​of the corresponding quantiles, thereby achieving quantile regression;

[0179] This represents a shared feature extraction backbone network, whose trainable parameters are: All quantiles share this backbone, which is used to extract abstract features from the fused features. These features are used by all quantile output layers to improve model efficiency and consistency of feature representation.

[0180] This represents the target quantile, which belongs to a predefined set of quantiles. ,For example ,in, This is the median prediction;

[0181] Indices representing quantiles ;

[0182] Represents the set of quantiles The size of , with an example value of 5;

[0183] Indicates the prediction time step index, for example, ;

[0184] This indicates the predicted field of view length, which is the total number of time steps that need to be predicted in the future. An example value is 6.

[0185] Indicates the point prediction branch for wind farm nodes In the future The predicted power point value at time t, in megawatts;

[0186] The output layer of the multilayer perceptron representing the point prediction branch has the following trainable parameters: It is used to directly output point prediction values ​​from features extracted from the shared backbone network, and to coordinate with the quantile regression branch to improve the accuracy of point prediction.

[0187] S323. Correction of prediction results based on physical priors:

[0188] The original forecast output may violate the physical operating constraints of the wind turbine. To ensure the physical rationality of the forecast results, a differentiable correction layer is used to constrain the predicted value of each quantile at each time step. The correction function considers both the rated power upper limit and the maximum allowable power change rate, and is expressed as:

[0189] ;

[0190] In the formula, This indicates the wind farm nodes after physical constraint correction. In the future Time, quantile The final power prediction value is given below, in megawatts;

[0191] This represents the clipping function, which takes the input... Limited to the range Inside, defined as ;

[0192] Represents wind farm nodes The installed rated power is a known constant, and the unit is megawatts;

[0193] This represents the Sigmoid activation function, which maps the input to... interval;

[0194] Represents the hyperbolic tangent function, mapping the input to... interval;

[0195] The positive scalar parameter representing the smoothness of the saturation region is a trainable parameter.

[0196] The positive scalar parameter representing the transition width of the saturation region is a trainable parameter.

[0197] Indicates the rate of change of power The constraint function, used to softly limit the grade ratio, is calculated as follows: ;

[0198] This represents the absolute value of the maximum allowable power change per unit time step, i.e., the ramp rate limit, and is expressed in megawatts per time step.

[0199] This represents the change in the original predicted power between adjacent time points, i.e. ;

[0200] This represents the corresponding quantile. At the previous moment The original power prediction value, for In this case, the value uses the most recent actual observation. Replacement.

[0201] S324. Model Integration and Output:

[0202] The final model outputs the future. The corrected predicted values ​​of all target quantiles and the corrected point predictions at each time step, specifically...

[0203] For point prediction, the corrected result of the median quantile can be used as the final point prediction value, or the result of the point prediction branch's output processed by the same correction function can be used.

[0204] S33. Loss Function Calculation:

[0205] Short-term wind power forecasting models need to simultaneously optimize forecast accuracy and the reliability of uncertainty estimation. Conventional techniques usually use a single loss function, such as mean square error, which cannot guide the model to simultaneously achieve the dual objectives of point forecasting and interval forecasting. In addition, the model forecasting needs to follow the physical variation law of wind power.

[0206] This invention employs a composite loss function, which integrates the accuracy loss of point-oriented prediction and the reliability loss of interval-oriented prediction, and balances them through an adaptive weighting mechanism based on local fluctuation intensity. This effectively drives the joint optimization of multi-objective prediction models. The specific steps are as follows:

[0207] S331. Constructing a composite loss for point prediction and interval prediction:

[0208] The loss function consists of two parts: point prediction loss and quantile regression loss. The point prediction loss directly optimizes the model's ability to predict the most likely value of future power, while the quantile regression loss ensures that the prediction interval can cover the true value with a preset probability by minimizing the quantile loss at multiple quantiles. It is expressed as:

[0209] ;

[0210] In the formula, Indicates the first The basic composite loss of each training sample is used to measure the overall performance of the model in point prediction and interval prediction, driving the model to optimize both prediction accuracy and uncertainty estimation.

[0211] This indicates the length of the predicted field of view, which is the total number of time steps that need to be predicted in the future.

[0212] This represents the prediction time step index, traversing from 1 to... The future time step;

[0213] This indicates the wind farm nodes after physical correction. In the future The median quantile prediction at time 1, which is also used as the final point prediction output, is expressed in megawatts.

[0214] Represents wind farm nodes In the future The actual power value at any given moment, in megawatts;

[0215] The quantile loss function is simplified as follows: ,but .

[0216] S332, Adaptive weighted loss based on fluctuation intensity:

[0217] The fluctuation characteristics of wind power sequences are time-varying. The model should have different optimization focuses during stable and highly variable periods. During stable periods, more attention should be paid to point prediction accuracy, while during highly variable periods, the learning of reliability within the uncertainty interval needs to be strengthened. Therefore, by dynamically adjusting the weights of the two terms in the composite loss function using the local fluctuation intensity index, the final dynamic loss function is expressed as:

[0218] ;

[0219] In the formula, Indicates the first The final dynamic loss of each training sample is the model training objective, which is to minimize the sum of this loss for all samples.

[0220] Indicator of local fluctuation intensity The maximum value over the entire training set is used to... Perform normalization;

[0221] It should be noted that, The dynamic weights used to construct the point prediction loss term The dynamic weights used to construct the quantile loss term, when When the value is low, i.e., in a stable period, the point prediction loss has a high weight, and the optimization focus is biased towards accuracy. When the quantile loss is high, indicating a period of dramatic change, optimization focuses on the reliability of interval prediction.

[0222] S34. Training of short-term power prediction model:

[0223] After completing the model structure design and loss function definition, the model enters the training phase to learn the optimal trainable parameters. The training process uses the training dataset constructed in step S1 for iterative optimization. Specifically, in each training epoch, samples from the training dataset are randomly grouped into mini-batches and input into the short-term power prediction model. The model first performs adaptive sliding windowing and dynamic fluctuation decomposition in step S2 based on the input historical power data to obtain decomposed temporal feature vectors; then, step S31 is executed to construct a dynamic spatiotemporal map and perform spatial feature aggregation to obtain spatial feature enhancement features; then, step S32 is executed to perform multi-scale feature fusion and meteorological feature injection, and through a quantile regression prediction head and a physical correction layer, outputs corrected point prediction values ​​and quantile prediction values ​​for multiple future time steps. Finally, based on the dynamically weighted loss function defined in step S33, the total loss value between the model's predicted output and the corresponding real power label is calculated.

[0224] A gradient-based optimization algorithm, such as the adaptive moment estimator optimizer, is used to backpropagate and update all trainable parameters in the model based on the calculated total loss. A validation set is used to monitor the training process and prevent overfitting.

[0225] After each training epoch, the model performance is evaluated using a validation set, and the validation loss is calculated. Training continues until the validation loss no longer decreases significantly over multiple consecutive training epochs, or until the preset maximum number of training epochs is reached. At this point, the model is considered to have fully learned the patterns in the data and reached convergence, and the training process ends. After training is complete, the model parameters with the best validation set performance are saved for subsequent predictions.

[0226] S4. Short-term power prediction:

[0227] Once the short-term power prediction model is trained, it can be used to predict short-term power on new, unseen data. In practical applications, the latest historical power data sequences of the target wind farm and its neighboring wind farms, along with the corresponding latest meteorological data, are first acquired in real time.

[0228] For the current moment requiring prediction, the process follows the same procedure as the training phase: Following step S2, the local fluctuation intensity index for that moment is calculated, the adaptive historical window length is dynamically determined, and the historical power sequence within the window is dynamically decomposed to generate a fixed-dimensional decomposed temporal feature vector. Then, following step S31, a dynamic adjacency matrix is ​​calculated based on the power data of each wind farm at the current moment, a spatiotemporal graph is constructed, and the decomposed temporal feature vectors of each node are passed through a trained spatial feature aggregation network based on gated graph attention to obtain the spatial feature enhancement features of the target wind farm. Then, following step S32, the decomposed temporal feature vector of the target wind farm, the spatial feature enhancement features, and the dynamically weighted encoded multi-source meteorological features are fused to form the final fused feature vector. This feature vector is input into the trained, parameter-fixed prediction model backbone network and quantile output layer to obtain the original point predictions and quantile predictions for multiple future time steps. Finally, these original predictions are processed by a physical prior correction layer with the same fixed parameters, outputting the final corrected point predictions that conform to physical constraints and the power interval predictions at each quantile. These forecasts can be directly provided to the power grid dispatching system for subsequent dispatching decisions and risk assessments.

[0229] This invention employs an adaptive sliding window and dynamic mode decomposition mechanism based on local fluctuation intensity perception. It dynamically adjusts the historical input window length according to the real-time fluctuation characteristics of the power sequence and performs targeted decomposition, effectively improving the model's ability to capture different patterns during stationary and rapidly changing periods in non-stationary sequences. This invention uses a dynamic spatiotemporal graph that integrates geographical distance and instantaneous power correlation, and introduces a gated graph attention mechanism based on trend similarity to achieve explicit modeling and adaptive feature aggregation of time-varying spatial dependencies between wind farms. This invention adopts an end-to-end model for multi-objective prediction, simultaneously outputting point and interval predictions through a quantile regression framework, and innovatively incorporates local fluctuation intensity as a weighting factor into the loss function, achieving a dynamic balance between prediction accuracy and uncertainty reliability. This invention uses a differentiable correction layer based on physical priors, embedding rated power and ramp rate constraints into the network in a smooth and differentiable manner. This ensures the physical rationality of the prediction results without interrupting gradient propagation, achieving a deep integration of data-driven and physical knowledge.

[0230] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A wind power prediction method based on multi-objective optimization, characterized in that, Includes the following steps: S1. Wind power data acquisition and training dataset construction: The raw data collection was completed by continuously recording the operating status of the wind farm at fixed time intervals. After obtaining raw sequence data with a long time span, a training sample set for supervised learning was constructed. The dataset was divided into non-overlapping training set, validation set and test set according to time order, which were used for learning model parameters, adjusting hyperparameters and final evaluation of generalization performance, respectively. S2. Adaptive sliding window and dynamic fluctuation decomposition of wind power time series data: An adaptive sliding window strategy based on local fluctuation intensity perception is adopted to perform dynamic mode decomposition and separate fluctuation components at different time scales from the original data; S3. Construct a short-term power prediction model: S31. Mining of collaborative features of neighboring wind fields based on spatiotemporal graphs; S32. Short-term wind power prediction based on multi-objective optimization; S33. Loss function calculation; S34. Training of short-term power prediction model; S4. Short-term power prediction: Once the short-term power prediction model is trained, it can be used to predict short-term power on new and unseen data. In the practical application stage, the latest historical power data sequence of the target wind farm and its neighboring wind farms, as well as the corresponding latest meteorological characteristic data, are first obtained in real time. Step S32 involves constructing an end-to-end prediction model that considers multi-point prediction accuracy, interval prediction reliability, and adherence to physical laws. This model deeply integrates and decomposes temporal and spatial collaborative features, and introduces dynamically weighted multi-source meteorological features. A quantile regression framework is used to simultaneously output point and interval predictions for multiple future time steps. A constraint layer based on physical priors is employed to correct the original predictions. Specifically, this includes the following steps: S321. Feature Fusion and Dynamic Meteorological Feature Injection: To integrate features from different sources and processing levels, firstly, the decomposed temporal feature vector of the target wind farm at the target time is concatenated with its spatial feature enhancement features to form the basic spatiotemporal fusion feature. Secondly, external multi-source meteorological features are introduced to provide key driving information. Considering that different meteorological factors have different weights on power at different times, an attention-based dynamic weighted encoder is used to encode the meteorological features and inject them into the fusion feature. S322, Multi-objective prediction head based on quantile regression: To obtain point predictions and uncertainty interval predictions simultaneously, the model adopts a quantile regression framework. Specifically, it uses a shared backbone network to process fused features and then connects multiple quantile-specific output layers in parallel. Each output layer corresponds to a target quantile and outputs the power prediction value for several future time steps. At the same time, an additional point prediction branch is defined, whose loss is co-optimized with that of the quantile prediction branch. S323. Correction of prediction results based on physical priors: The original prediction output may violate the physical operating constraints of the wind turbine. To ensure the physical rationality of the prediction results, a differentiable correction layer is used to constrain the prediction value of each quantile at each time step. The correction function considers both the upper limit of rated power and the maximum allowable rate of power change. S324. Model Integration and Output: The final model outputs the future. For each time step, the corrected predicted values ​​of all target quantiles and the corrected values ​​of point predictions are used. For point predictions, the corrected result of the median quantile is taken as the final point prediction value, or the result of the output of the point prediction branch is processed by the same correction function. Step S33 involves employing a composite loss function. This function integrates the accuracy loss from point-oriented prediction with the reliability loss from interval-oriented prediction, and balances them through an adaptive weighting mechanism based on local fluctuation intensity. This effectively drives the joint optimization of multi-objective prediction models, specifically including the following steps: S331. Constructing a composite loss for point prediction and interval prediction: The loss function consists of two parts: point prediction loss and quantile regression loss. The point prediction loss directly optimizes the model's ability to predict the most likely value of future power, while the quantile regression loss ensures that the prediction interval can cover the true value with a preset probability by minimizing the quantile loss at multiple quantiles. S332, Adaptive weighted loss based on fluctuation intensity: The fluctuation characteristics of wind power series are time-varying. The model should have different optimization focuses during the stable period and the period of dramatic change. During the stable period, more attention should be paid to the accuracy of point prediction, while during the period of dramatic change, it is necessary to strengthen the learning of the reliability of the uncertainty interval. Therefore, the weights of the two items in the composite loss should be dynamically adjusted by combining the local fluctuation intensity index.

2. The wind power prediction method based on multi-objective optimization according to claim 1, characterized in that: Step S1: The collected raw data includes the actual power output time series data of the wind farm group and multi-source meteorological data closely related to wind power generation; among them, for a small number of missing values ​​caused by equipment failure or communication interruption, linear interpolation or the mean of the effective values ​​before and after the time is used to fill in the missing values.

3. The wind power prediction method based on multi-objective optimization according to claim 2, characterized in that: Step S1: Each training sample consists of model input features and corresponding labeled ground truth values. For each moment in the time series, the power data and related meteorological data of its historical period are used as input features, and the labeled content is the future continuous time after that moment, that is, the predicted real power value of the target wind farm within the field of view. The labeling method is the natural alignment of the time series, that is, each input feature sample corresponds to its real power output value at a specific future time step as a supervision signal, and the labeling category is continuous power value.

4. The wind power prediction method based on multi-objective optimization according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Calculate the local wave intensity index: To quantify the intensity of fluctuations in the original power sequence near each sample point, a local fluctuation intensity index is calculated based on the absolute value of power change within the basic observation window and its adaptive weights. S22. Determine the adaptive history window length: Based on the calculated local fluctuation intensity index, the length of the historical window used as the input to the prediction model is dynamically determined to ensure that more historical information is included when fluctuations are severe in order to capture abrupt change patterns, and the window is shortened when fluctuations are gentle to avoid noise interference. S23. Perform dynamic fluctuation decomposition and feature construction: The original power sequence of variable length is truncated according to the adaptive historical window length, and then the sequence is dynamically filtered and decomposed to separate the low-frequency trend component and the high-frequency fluctuation component. Then, the decomposed components are concatenated to obtain the decomposed time series feature vector.

5. The wind power prediction method based on multi-objective optimization according to claim 1, characterized in that, Step S31 involves explicitly modeling the spatial relationships between wind farms by constructing a dynamic spatiotemporal graph, and then using a graph attention network with a gating mechanism to aggregate spatial features and mine spatial collaborative features. The specific steps are as follows: S311. Constructing a dynamic spatiotemporal graph and calculating the adjacency matrix: To model the dynamic spatial dependencies between wind farms, a weighted directed graph is constructed for each sample time. A set of nodes is defined, which includes the target wind farm and its geographically nearest wind farms. The edge set adopts a fully connected approach to indicate the potential influence relationship between nodes. The adjacency matrix is ​​used to quantify the dynamic influence intensity between nodes. Its elements represent the influence weight of one node on another node at a specific time. This weight is calculated by combining geographical distance and instantaneous power correlation. S312. Perform spatial feature aggregation based on gated graph attention: The decomposed temporal feature vector of each wind farm node at the current time point is used as its initial node feature. An improved graph attention network is used for spatial feature aggregation. Based on the conventional graph attention network, the improved graph attention network combines a gating mechanism based on the similarity of power change trends to enhance attention to neighbors with similar evolution patterns when fluctuations are drastic, thereby obtaining spatial feature enhancement features.

6. The wind power prediction method based on multi-objective optimization according to claim 1, characterized in that, Step S34: The training process uses the training dataset constructed in step S1 for iterative optimization. In each training cycle, the samples in the training dataset are randomly ordered into mini-batches and input into the short-term power prediction model. The model first performs adaptive sliding window and dynamic fluctuation decomposition in step S2 based on the input historical power data to obtain the decomposed temporal feature vector. Then, step S31 is executed to construct a dynamic spatiotemporal graph and perform spatial feature aggregation to obtain spatial feature enhancement features. Then, step S32 is executed to perform multi-scale feature fusion and meteorological feature injection. Through the quantile regression prediction head and physical correction layer, the corrected point prediction value and quantile prediction value for multiple future time steps are output. Then, according to the dynamic weighted loss function defined in step S33, the total loss value between the model prediction output and the actual power label at the corresponding time is calculated.

7. The wind power prediction method based on multi-objective optimization according to claim 6, characterized in that: An adaptive moment estimation optimizer is used to backpropagate and update all trainable parameters in the model based on the calculated total loss; after each training cycle, the model performance is evaluated using a validation set and the validation loss is calculated. Training will continue until it is verified that the loss no longer decreases significantly over multiple consecutive training epochs, or until the preset maximum number of training epochs is reached. At this point, it can be considered that the model has fully learned the patterns in the data, reached a convergence state, and the training process ends. After training is complete, save the model parameters that have the best performance on the validation set for use in subsequent predictions.

8. The wind power prediction method based on multi-objective optimization according to claim 1, characterized in that, Step S4: For the moment when prediction is needed, process it in the same way as the training phase: According to step S2, calculate the local fluctuation intensity index at that moment, dynamically determine the adaptive historical window length, perform dynamic fluctuation decomposition on the historical power sequence within the window, and generate a fixed-dimensional decomposed temporal feature vector. Following step S31, the dynamic adjacency matrix is ​​calculated based on the power data of each wind farm at the current moment, a spatiotemporal graph is constructed, and the decomposed temporal feature vectors of each node are processed through a trained spatial feature aggregation network based on gated graph attention to obtain the spatial feature enhancement features of the target wind farm. Then, according to step S32, the decomposed temporal feature vector, spatial feature enhancement features, and multi-source meteorological features after dynamic weighted encoding of the target wind farm are fused to form the final fused feature vector. This feature vector is then input into the trained, parameter-fixed prediction model backbone network and quantile output layer to obtain the original point predictions and quantile predictions for multiple future time steps. Finally, these original predictions are processed by the same parameter-fixed physical prior correction layer to output the final corrected point predictions that meet physical constraints and the power interval predictions at each quantile.

Images