Regional medium-and-long-term wind power generation probability prediction method

Abstract

The invention discloses a regional medium-and-long-term wind power generation probability prediction method, which comprises the following specific steps: step 1, acquiring medium-and-long-term weather prediction data in a target region, and acquiring week-by-week weather field data from the future 4-26 weeks from a global climate prediction system; and 2, constructing a regional wind power plant group spatial topology network, and constructing a weighted spatial adjacency matrix based on geographic coordinates and installed capacities of all wind power plants in a target region. According to the regional medium-and-long-term wind power generation capacity probability prediction method, large-scale meteorological field data output by a global climate prediction system is fused, and an empirical orthogonal function dominant mode is extracted as a medium-and-long-term driving factor, so that the time scale limitation that a traditional method is limited to short-term weather forecast is broken through; effective tracking of wind resource evolution trends from several weeks to several months in the future is realized, and a weighted spatial topology network based on geographic distance and installed capacity is constructed.

Description

Technical Field

[0001] This invention relates to the field of power systems and their automation technology, specifically to a method for predicting the probability of medium- and long-term wind power generation in a region. Background Technology

[0002] With the large-scale grid connection of renewable energy, the penetration rate of wind power in the power system continues to rise. The strong fluctuations and uncertainties in its output pose a severe challenge to the safe and stable operation of the power grid. As a key technical foundation for supporting power planning, cross-seasonal dispatching arrangements, and the formulation of medium- and long-term contracts in the power market, the probability forecasting of regional medium- and long-term wind power generation needs to comprehensively consider the evolution of meteorological systems, the spatial correlation between regional wind farms, and the time scale extension characteristics from several weeks to several months.

[0003] Referring to patent publication number "CN112217199B", a method for fitting the probability distribution of wind speed in a wind farm is disclosed, including the following steps: Step 1: Measure the wind speed of the wind farm under various conditions, statistically analyze the probability of wind speed, and plot the wind speed probability distribution; Step 2: Approximate the wind speed of the wind farm using the Weibull distribution function and determine the parameters of the Weibull distribution function; Step 3: Calculate the error between the wind speed probability distribution and the Weibull distribution function, denoted as the probability distribution of the landing form; Step 4: Segment the probability distribution of the landing form.

[0004] Step 5: Extract sample data from the probability distribution of the segmented landing patterns.

[0005] As described above, the regional medium- and long-term wind power generation probability prediction aims to construct a probability distribution model with time evolution capabilities by integrating climate models, historical statistical characteristics, and spatial topology to quantify the uncertainty range of wind power output in the coming weeks to months. Existing technologies face multiple bottlenecks in achieving this goal. On the one hand, although some methods introduce Copula theory or typical scenario generation techniques for probability modeling, their inputs rely on short-term numerical weather forecasts, and the time scale is limited to intraday or real-time frequency regulation scenarios, without incorporating medium- and long-term climate signals, resulting in an inability to reflect the systematic shift of wind resources in the coming weeks or more. On the other hand, another type of method based on historical wind speed distribution fitting improves the accuracy of single-point probability modeling, but completely ignores external meteorological driving information and adopts static statistical assumptions, making it difficult to adapt to non-stationary wind conditions caused by climate change. In addition, existing schemes generally focus on single wind farms or isolated nodes, lacking a collaborative modeling mechanism for the spatial correlation and aggregation uncertainty among multiple sites in the region, resulting in significant bias and conservatism in the prediction results at the regional scale. Therefore, there is an urgent need for a probabilistic prediction method that can integrate medium- and long-term meteorological forecast information, dynamically characterize regional spatial correlations, and support temporal evolution, so as to overcome the comprehensive limitations of existing technologies in terms of time scale, driving mechanism, and spatial dimension. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for predicting the probability of medium- and long-term wind power generation in a region. This method solves the comprehensive limitations of existing prediction methods in terms of time scale, driving mechanism, and spatial dimension, and the inability to effectively integrate medium- and long-term meteorological forecast information, dynamically characterize regional spatial correlation, and support temporal evolution probability modeling.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the probability of medium- and long-term wind power generation in a region, comprising the following specific steps:

[0008] Step 1: Obtain medium- and long-term meteorological forecast data for the target area, and obtain weekly meteorological field data for the next 4 to 26 weeks from the global climate prediction system;

[0009] Step 2: Construct a spatial topology network for the regional wind farm cluster. Based on the geographical coordinates and installed capacity of all wind farms in the target area, construct a weighted spatial adjacency matrix.

[0010] Step 3: Establish a dynamic spatial correlation modeling module, and input the meteorological dominant mode time coefficients extracted in Step 1 as external driving factors into the graph attention network;

[0011] Step 4: Design the medium- and long-term probabilistic prediction backbone model, adopting a hybrid architecture of temporal convolutional network and long short-term memory network. The temporal convolutional network is used to extract meteorological driving factors and local temporal patterns in historical power output sequences. The convolutional kernel width is set to 5, and the expansion factor increases to 32 in powers of 2. The long short-term memory network is used to capture long-term dependencies across weekly scales. The hidden layer dimension is set to 128. The model input includes the meteorological modal coefficients from Step 1, the dynamic spatial weighting matrix output from Step 3, and the weekly average power output data of each wind farm over the past 52 weeks.

[0012] Step 5: Implement probabilistic output modeling, connect the output of the backbone model in Step 4 to the quantile regression layer, and finally output the probability distribution function of the total wind power output in the future week, supporting the extraction of prediction intervals at any confidence level;

[0013] Step Six: Perform regional aggregation and uncertainty propagation calculations. After obtaining the individual probability prediction results for each wind farm, the spatial correlation structure is based on the dynamic learning in Step Three.

[0014] Preferably, the meteorological field data includes sea level pressure, 500 hPa geopotential height, near-surface wind speed and direction, temperature and humidity variables. The meteorological field data is downscaled in time and space, and its spatial resolution is resampled to a 10 km grid using bilinear interpolation. The time coefficient sequences of the first six dominant modes are extracted by empirical orthogonal function decomposition as a low-dimensional representation of large-scale circulation anomalies. The meteorological field data provided by the global climate prediction system in step one is updated weekly, providing prediction results for 50 ensemble members each time. The empirical orthogonal function decomposition is based on historical meteorological data from the same period over the past 30 years to construct the background field.

[0015] Preferably, in step two, the adjacency weight is calculated using a Gaussian decay function with a decay coefficient that is the reciprocal of 0.05 km. The spatial distance threshold is set to 150 km, and the weight between wind farms exceeding this threshold is set to 0, forming a sparsely connected regional wind power spatial coupling structure. In step two, the terrain correction coefficient is calculated based on the digital elevation model data of the grid where the wind farm is located. A linear fitting method is used to establish a regression relationship between altitude and annual average wind speed, with the regression slope ranging from 1.2 m / s to 1.8 m / s for every kilometer of rising wind speed.

[0016] Preferably, the attention network is based on the spatial topology network constructed in step two as a graph structure. The node features are the historical power output standardized sequence and terrain correction coefficient of each wind farm. Through a multi-head attention mechanism, the network dynamically learns the changes in the correlation strength between nodes under different meteorological modes and outputs a weighted adjacency matrix that evolves over time, reflecting the non-stationary characteristics of the correlation of regional wind power output.

[0017] Preferably, the feature sequence output by the temporal convolutional network in step four is used as the input sequence of the long short-term memory network. The hidden state of the last time step of the long short-term memory network is concatenated with the cell state and used as the comprehensive temporal representation of the backbone model. The dimension of this representation vector is 256.

[0018] Preferably, the quantile regression layer in step five contains 99 parallel output channels, each corresponding to an equally spaced quantile from 0.5% to 99.5%, and is trained end-to-end using a smooth quantile loss function.

[0019] Preferably, in step five, the quantile regression layer uses an adaptive learning rate optimization algorithm during the training phase. The initial learning rate is set to 0.001, decays by 10% every 10 training cycles, the batch size is set to 64, and the training data is divided into a 70% training set, a 15% validation set, and a 15% test set. The validation set is used for an early stopping mechanism, and the maximum number of training cycles is set to 500. Training is terminated when the validation loss does not decrease for 20 consecutive cycles.

[0020] Preferably, step six uses an improved Gaussian Copula model to model the joint probability distribution of the power output of each station. The rank correlation coefficient matrix of the Copula function is obtained by normalizing and mapping the dynamic weight matrix output by the graph attention network. Then, the probability density curve of the total regional power output is generated by Monte Carlo sampling to complete the propagation and aggregation of uncertainty in the spatial dimension.

[0021] Preferably, in step six, the improved Gaussian Copula model introduces a time decay factor when calculating the rank correlation coefficient matrix, the contribution weight of the recent dynamic spatial weight is 1.5 times that of the long-term weight, and historical correlation information is fused using an exponentially weighted moving average method.

[0022] Preferably, the method further includes a post-processing calibration step for the prediction results, using nonlinear least squares method to correct the bias of the predicted probability distribution, with the predicted quantile coverage rate during the historical backtesting period as the objective function, so that the error between the actual coverage rate and the nominal coverage rate of each quantile is less than 5 percentage points.

[0023] This invention provides a method for predicting the probability of medium- to long-term wind power generation in a region. Compared with existing technologies, it has the following advantages:

[0024] 1. The medium- and long-term wind power generation probability prediction method in this region, by integrating large-scale meteorological field data output from the global climate prediction system, extracts the dominant mode of empirical orthogonal function as a medium- and long-term driving factor, breaks through the time scale limitation of traditional methods that are limited to short-term weather forecasts, and realizes the effective tracking of wind resource evolution trends over the next few weeks to months. By constructing a weighted spatial topology network based on geographical distance and installed capacity, and introducing a graph attention mechanism to dynamically respond to changes in external meteorological modes, the method achieves time-varying modeling of the spatial correlation of regional wind farm clusters for the first time, overcoming the prediction bias caused by the static correlation assumption.

[0025] 2. The medium- and long-term wind power generation probability prediction method in this region adopts a hybrid architecture of temporal convolutional network and long short-term memory network, taking into account both local meteorological fluctuation characteristics and long-term seasonal memory, which improves the model's ability to represent complex time patterns. It directly outputs the full probability distribution through the quantile regression layer and combines it with the improved Gaussian Copula model to aggregate uncertainty space, forming a probability transmission chain from meteorological drive to regional power output. The prediction results have physical consistency and statistical reliability.

[0026] 3. The regional medium- and long-term wind power generation probability prediction method significantly improves the accuracy, dynamic adaptability, and probability calibration of regional medium- and long-term wind power generation prediction, providing a high-confidence decision-making basis for cross-seasonal power system dispatch, power source planning, and medium- and long-term electricity market transactions. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0028] Figure 2 This is a flowchart illustrating the logical flow of the long-term probability prediction backbone model in this invention.

[0029] Figure 3 This is a schematic diagram of the multi-level interaction relationship and data flow of the region aggregation and uncertainty propagation calculation of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1-3 The medium- and long-term wind power generation probability prediction method in this region offers two technical solutions:

[0032] The first implementation method includes the following specific steps:

[0033] Step 1: Obtain medium- and long-term meteorological forecast data for the target area. Obtain weekly meteorological field data for the next 4 to 26 weeks from the global climate prediction system, including sea level pressure, 500 hPa geopotential height, near-surface wind speed and direction, temperature and humidity variables. Perform spatiotemporal downscaling on the meteorological field data, and use bilinear interpolation to resample its spatial resolution to a 10 km grid. Extract the time coefficient sequences of the first 6 dominant modes through empirical orthogonal function decomposition as a low-dimensional representation of large-scale circulation anomalies.

[0034] In step one, the meteorological field data provided by the global climate prediction system is updated weekly, with each update providing prediction results for 50 ensemble members. The empirical orthogonal function decomposition constructs the background field based on historical meteorological data from the same period over the past 30 years.

[0035] Step 2: Construct a spatial topology network for the regional wind farm cluster. Based on the geographical coordinates and installed capacity of all wind farms in the target area, construct a weighted spatial adjacency matrix. The adjacency weights are calculated using a Gaussian decay function with a decay coefficient that is the reciprocal of 0.05 km. The spatial distance threshold is set to 150 km. The weights between farms exceeding this threshold are set to 0, forming a sparsely connected regional wind power spatial coupling structure.

[0036] In step two, the terrain correction coefficient is calculated based on the digital elevation model data of the grid where the wind farm is located. A linear fitting method is used to establish the regression relationship between altitude and annual average wind speed. The regression slope ranges from 1.2 m / s to 1.8 m / s for every kilometer of rising wind speed.

[0037] Step 3: Establish a dynamic spatial correlation modeling module. The meteorological dominant mode time coefficient extracted in Step 1 is used as an external driving factor and input into the graph attention network. The graph attention network is based on the spatial topology network constructed in Step 2. The node features are the historical power output standardized sequence and terrain correction coefficient of each wind farm. Through the multi-head attention mechanism, the correlation strength changes between nodes under different meteorological modes are dynamically learned, and the weighted adjacency matrix that evolves over time is output to reflect the non-stationary characteristics of regional wind power output correlation.

[0038] In step three, the graph attention network contains three layers of graph attention units, with four attention heads in each layer. The attention score is calculated by linearly transforming the node features using a learnable weight matrix. Residual connections and layer normalization mechanisms are introduced, and the attention weights are output after Softmax normalization to ensure that the sum of the spatial correlation strengths in each step is 1.

[0039] By integrating large-scale meteorological field data output from the global climate prediction system, and extracting the dominant mode of empirical orthogonal functions as a medium- to long-term driving factor, this method breaks through the time scale limitation of traditional methods that are limited to short-term weather forecasts. It enables effective tracking of the evolution trend of wind resources over the next few weeks to months. By constructing a weighted spatial topology network based on geographical distance and installed capacity, and introducing a graph attention mechanism to dynamically respond to changes in external meteorological modes, it achieves time-varying modeling of the spatial correlation of regional wind farm clusters for the first time, overcoming the prediction bias caused by the static correlation assumption.

[0040] The second implementation method differs from the first implementation method in the following steps: Step 4: Design a medium- and long-term probabilistic prediction backbone model, which adopts a hybrid architecture of temporal convolutional network and long short-term memory network. The temporal convolutional network is used to extract meteorological driving factors and local temporal patterns in historical power output sequences. The convolutional kernel width is set to 5, and the expansion factor increases to 32 in powers of 2. The long short-term memory network is used to capture long-term dependencies across weekly scales. The hidden layer dimension is set to 128. The model input includes the meteorological modal coefficients from Step 1, the dynamic spatial weighting matrix output from Step 3, and the weekly average power output data of each wind farm over the past 52 weeks.

[0041] In step four, the feature sequence output by the temporal convolutional network is used as the input sequence of the long short-term memory network. The hidden state of the last time step of the long short-term memory network is concatenated with the cell state and used as the comprehensive temporal representation of the backbone model. The dimension of this representation vector is 256.

[0042] Step 5: Implement probabilistic output modeling by connecting the output of the backbone model from Step 4 to the quantile regression layer. The quantile regression layer contains 99 parallel output channels, corresponding to equally spaced quantiles from 0.5% to 99.5% (i.e., For each quantile level τ, the quantile regression layer maps the 256-dimensional features to a scalar predicted value through a separate fully connected layer. The τ quantile, representing the wind farm's output in future week T, is trained end-to-end using a smooth quantile loss function, ultimately outputting the probability distribution function of the total wind power output for each future week, supporting the extraction of prediction intervals at any confidence level.

[0043] In step five, the quantile regression layer uses an adaptive learning rate optimization algorithm during the training phase. The initial learning rate is set to 0.001, decays by 10% every 10 training epochs, the batch size is set to 64, and the training data is divided into a 70% training set, a 15% validation set, and a 15% test set. The validation set is used for the early stopping mechanism, and the maximum number of training epochs is set to 500. Training is terminated when the validation loss does not decrease for 20 consecutive epochs.

[0044] Step Six: Perform regional aggregation and uncertainty propagation calculations. After obtaining the individual probability prediction results of each wind farm, based on the spatial correlation structure learned in Step Three, an improved Gaussian Copula model is used to model the joint probability distribution of the power output of each station. The rank correlation coefficient matrix of the Copula function is obtained by normalizing and mapping the dynamic weight matrix output by the graph attention network. Then, the probability density curve of the total regional power output is generated by Monte Carlo sampling to complete the propagation and aggregation of uncertainty in the spatial dimension.

[0045] Specifically, the quantile prediction results for each wind farm are first obtained: Through its corresponding cumulative distribution function Convert to uniformly distributed pseudo-observations Then, the rank correlation coefficient matrix R of the Copula function is constructed. The rank correlation coefficient matrix of the Copula function is obtained by normalizing and mapping the dynamic weight matrix A(t) output by the graph attention network.

[0046] The specific mapping process is as follows: First, row normalize A(t) so that the sum of each row is 1 to obtain the transition probability matrix P(t). Then, calculate the covariance matrix Cov(P(t)) of P(t). Finally, convert the covariance matrix into the rank correlation coefficient matrix R through Fisher z transformation.

[0047] In step six, the improved Gaussian Copula model introduces a time decay factor when calculating the rank correlation coefficient matrix. The contribution weight of the recent dynamic spatial weight is 1.5 times that of the long-term weight. An exponentially weighted moving average method is used to fuse historical correlation information. ,in =1 / 1.5, K is the historical window length, set to 4 weeks. This ensures that the joint distribution modeling reflects both the current meteorological state and preserves long-term statistical regularities. Finally, the probability density curve of the total regional power output is generated through Monte Carlo sampling: a large number of sample vectors U=(u1,...,u) are drawn from the Gaussian Copula. N Then, through the inverse cumulative distribution function of each wind farm... Convert U back to the actual output space Q=(q1,...,q N The total regional output sample is obtained by summing Q. After repeating this process 10,000 times, the probability density curve of the total regional output can be plotted, thus completing the propagation and aggregation of uncertainty in the spatial dimension.

[0048] In step six, the improved Gaussian Copula model introduces a time decay factor when calculating the rank correlation coefficient matrix. The contribution weight of the recent dynamic spatial weight is 1.5 times that of the long-term weight, and the historical correlation information is fused using an exponentially weighted moving average method.

[0049] It also includes a post-processing calibration step for the prediction results, using nonlinear least squares method to correct the bias of the predicted probability distribution, with the predicted quantile coverage rate during the historical backtesting period as the objective function, so that the error between the actual coverage rate and the nominal coverage rate of each quantile is less than 5 percentage points.

[0050] A hybrid architecture combining temporal convolutional networks and long short-term memory networks is adopted to take into account both local meteorological fluctuations and long-term seasonal memory, thereby improving the model's ability to represent complex temporal patterns. The model directly outputs the full probability distribution through quantile regression layers and combines an improved Gaussian Copula model to aggregate uncertainty space, forming a probability transmission chain from meteorological drivers to regional power output. The prediction results have physical consistency and statistical reliability, significantly improving the accuracy, dynamic adaptability, and probability calibration of medium- and long-term wind power generation forecasts in the region. This provides a high-confidence decision-making basis for cross-seasonal power system dispatch, power planning, and medium- and long-term electricity market transactions.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the probability of medium- and long-term wind power generation in a region, characterized in that: The specific steps include the following: Step 1: Obtain medium- and long-term meteorological forecast data for the target area, and obtain weekly meteorological field data for the next 4 to 26 weeks from the global climate prediction system; Step 2: Construct a spatial topology network for the regional wind farm cluster. Based on the geographical coordinates and installed capacity of all wind farms in the target area, construct a weighted spatial adjacency matrix. Step 3: Establish a dynamic spatial correlation modeling module, and input the meteorological dominant mode time coefficients extracted in Step 1 as external driving factors into the graph attention network; Step 4: Design the medium- and long-term probabilistic prediction backbone model, adopting a hybrid architecture of temporal convolutional network and long short-term memory network. The temporal convolutional network is used to extract meteorological driving factors and local temporal patterns in historical power output sequences. The convolutional kernel width is set to 5, and the expansion factor increases to 32 in powers of 2. The long short-term memory network is used to capture long-term dependencies across weekly scales. The hidden layer dimension is set to 128. The model input includes the meteorological modal coefficients from Step 1, the dynamic spatial weighting matrix output from Step 3, and the weekly average power output data of each wind farm over the past 52 weeks. Step 5: Implement probabilistic output modeling, connect the output of the backbone model in Step 4 to the quantile regression layer, and finally output the probability distribution function of the total wind power output in the future week, supporting the extraction of prediction intervals at any confidence level; Step Six: Perform regional aggregation and uncertainty propagation calculations. After obtaining the individual probability prediction results for each wind farm, the spatial correlation structure is based on the dynamic learning in Step Three.

2. The method for predicting the probability of regional medium- and long-term wind power generation according to claim 1, characterized in that: The meteorological field data includes sea level pressure, 500 hPa geopotential height, near-surface wind speed and direction, temperature, and humidity variables. The meteorological field data is downscaled in time and space, and its spatial resolution is resampled to a 10 km grid using bilinear interpolation. The time coefficient sequences of the first six dominant modes are extracted by empirical orthogonal function decomposition as a low-dimensional representation of large-scale circulation anomalies. The meteorological field data provided by the global climate prediction system in step one is updated weekly, providing prediction results for 50 ensemble members each time. The empirical orthogonal function decomposition is based on historical meteorological data from the same period over the past 30 years to construct the background field.

3. The method for predicting the probability of regional medium- and long-term wind power generation according to claim 1, characterized in that: In step two, the adjacency weights are calculated using a Gaussian decay function with a decay coefficient that is the reciprocal of 0.05 km. The spatial distance threshold is set to 150 km, and the weights between wind farms exceeding this threshold are set to 0, forming a sparsely connected regional wind power spatial coupling structure. In step two, the terrain correction coefficient is calculated based on the digital elevation model data of the grid where the wind farm is located. A linear fitting method is used to establish a regression relationship between altitude and annual average wind speed, with the regression slope ranging from 1.2 m / s to 1.8 m / s for every kilometer of rising wind speed.

4. The method for predicting the probability of regional medium- and long-term wind power generation according to claim 1, characterized in that: The attention network is based on the spatial topology network constructed in step two. The node features are the historical power output standardized sequence and terrain correction coefficient of each wind farm. Through a multi-head attention mechanism, it dynamically learns the changes in the correlation strength between nodes under different meteorological modes and outputs a weighted adjacency matrix that evolves over time, reflecting the non-stationary characteristics of the correlation of regional wind power output.

5. The method for predicting the probability of regional medium- and long-term wind power generation according to claim 1, characterized in that: In step four, the feature sequence output by the temporal convolutional network is used as the input sequence of the long short-term memory network. The hidden state of the last time step of the long short-term memory network is concatenated with the cell state and used as the comprehensive temporal representation of the backbone model. The dimension of this representation vector is 256.

6. The method for predicting the probability of regional medium- and long-term wind power generation according to claim 1, characterized in that: In step five, the quantile regression layer contains 99 parallel output channels, each corresponding to an equally spaced quantile from 0.5% to 99.5%, and is trained end-to-end using a smooth quantile loss function.

7. The method for predicting the probability of regional medium- and long-term wind power generation according to claim 1, characterized in that: In step five, the quantile regression layer uses an adaptive learning rate optimization algorithm during the training phase. The initial learning rate is set to 0.001, decays by 10% every 10 training cycles, the batch size is set to 64, and the training data is divided into a 70% training set, a 15% validation set, and a 15% test set. The validation set is used for the early stopping mechanism, and the maximum number of training cycles is set to 500. Training is terminated when the validation loss does not decrease for 20 consecutive cycles.

8. The method for predicting the probability of regional medium- and long-term wind power generation according to claim 1, characterized in that: Step six employs an improved Gaussian Copula model to model the joint probability distribution of power output from each station. The rank correlation coefficient matrix of the Copula function is obtained by normalizing and mapping the dynamic weight matrix output by the graph attention network. Then, the probability density curve of the total regional power output is generated through Monte Carlo sampling, thus completing the propagation and aggregation of uncertainty in the spatial dimension.

9. The method for predicting the probability of medium- and long-term wind power generation in a region according to claim 1, characterized in that: The improved Gaussian Copula model in step six introduces a time decay factor when calculating the rank correlation coefficient matrix. The contribution weight of the recent dynamic spatial weight is 1.5 times that of the long-term weight, and historical correlation information is fused using an exponentially weighted moving average method.

10. The method for predicting the probability of regional medium- and long-term wind power generation according to claim 1, characterized in that: It also includes a post-processing calibration step for the prediction results, using nonlinear least squares method to correct the bias of the predicted probability distribution, with the predicted quantile coverage rate during the historical backtesting period as the objective function, so that the error between the actual coverage rate and the nominal coverage rate of each quantile is less than 5 percentage points.

Images