A high-altitude wind power new energy supply method and system

Abstract

The application discloses a high-altitude area wind power new energy supply method and system, relates to the technical field of renewable energy management, and comprises the following steps: collecting original meteorological data and power data for preprocessing, and generating a data set; based on the data set, a directed acyclic graph is constructed, a maximum likelihood parameter set is generated, the BIC score of the directed acyclic graph is calculated, preliminary screening is performed, the directed acyclic graph with the maximum BIC score is obtained, secondary fine screening is performed, and a most relevant node set is obtained; based on the most relevant node set, a first-order difference operation is performed, a disturbance vector is generated, the disturbance vector is taken as the input of a DRBN network, the output activation value is outputted, an input matrix is constructed and taken as the input of a bidirectional long short-term memory network, the final hidden state is outputted, linear transformation is performed on the final hidden state through a full connection layer, and a final power prediction value is obtained; and the application improves the supplyability and scheduling accuracy of wind power new energy in a complex highland environment.

Description

A method and system for supplying new wind power energy in high-altitude areas Technical Field

[0001] This invention relates to the field of renewable energy management technology, and in particular to a method and system for supplying new wind power energy in high-altitude areas. Background Technology

[0002] With the increasing global demand for renewable energy, wind power, as an important component of clean energy, is expanding its application scenarios in various complex environments, especially in high-altitude areas. Due to the undulating terrain, thin air, and variable climate, these areas possess higher wind energy resource density and theoretically have greater potential for wind energy development. In recent years, with advancements in wind turbine technology and the evolution of smart grid systems, numerous wind power projects have been deployed in plateau regions exceeding 3,000 meters in altitude, such as the Qinghai-Tibet Plateau and the Western Sichuan Plateau. Simultaneously, technologies such as wind power prediction, flexible scheduling of wind power output, and wind-storage complementary regulation are gradually extending from plains areas to high-altitude scenarios. However, in practical applications, the drastic fluctuations in meteorological conditions, enhanced nonlinearity in data distribution, and higher requirements for equipment operational stability in high-altitude areas have led to insufficient accuracy, high response delays, and difficulties in achieving dynamic multi-source optimization in traditional techniques relying on linear statistical models or shallow neural networks for wind power prediction and supply scheduling.

[0003] Existing technologies generally lack effective screening mechanisms for factors affecting wind power, resulting in high dimensionality of input variables, strong redundancy, and poor generalization ability. Furthermore, the structural parameters of the prediction network often rely on manual experience to set, lacking adaptive structural optimization methods. Especially in high-altitude and complex terrain, the distribution of its temporal characteristics changes drastically, and the fixed window structure is difficult to adapt to dynamic fluctuations. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method and system for supplying new wind power energy in high-altitude areas, which solves the problem that the existing technology lacks an effective screening mechanism for factors affecting wind power, resulting in high dimensionality of input variables, strong redundancy, poor generalization ability, and the structural parameters of the prediction network often rely on manual experience to set, lacking adaptive structural optimization methods. Especially in high-altitude complex terrain, the distribution of its temporal characteristics changes drastically, and the fixed window structure is difficult to adapt to dynamic fluctuations.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for supplying wind power as a new energy source in high-altitude areas, comprising:

[0008] Raw meteorological and power data are collected, preprocessed, and used to generate datasets.

[0009] Based on the dataset, a directed acyclic graph is constructed. After generating the maximum likelihood parameter set, the BIC score of the directed acyclic graph is calculated to perform preliminary screening. The directed acyclic graph with the maximum BIC score is then subjected to secondary fine screening to obtain the set of most relevant nodes.

[0010] Based on the most relevant set of nodes, a first-order difference operation is performed to generate a perturbation vector. The perturbation vector is used as the input of the DRBN network, and the output activation value is used to construct an input matrix as the input of the bidirectional long short-term memory network. The output of the final hidden state is then linearly transformed through a fully connected layer to obtain the final power prediction value.

[0011] After optimizing and feeding back the parameters of the DRBN network and the fully connected layer using a parallel heterogeneous optimization mechanism, the final power prediction value is re-output, and the power difference is calculated to execute the supply scheduling of wind power new energy.

[0012] As a preferred embodiment of the wind power new energy supply method in high-altitude areas described in this invention, the following steps are taken: after constructing a directed acyclic graph and generating a maximum likelihood parameter set, the BIC score of the directed acyclic graph is calculated to perform preliminary screening. Based on the dataset, the equal-frequency binning method is used to label each component with a state and state number. The states include low, medium, and high, and the state numbers include 0, 1, and 2.

[0013] Each component is defined as a node. Based on the conditional dependencies between nodes, directional edges are established between nodes to form a directed acyclic graph. Maximum likelihood estimation is used to perform conditional probability estimation on each node in the directed acyclic graph and generate a set of maximum likelihood parameters.

[0014] Based on the number of states, the number of state combinations of the parent node is calculated. Further, combined with the number of states, the number of independent parameters of each node is calculated and summed to generate the total number of parameters. Using the information criterion, combined with the maximum likelihood parameter set and the dataset, the BIC score of the directed acyclic graph is calculated, and the directed acyclic graph corresponding to the maximum BIC score is extracted.

[0015] As a preferred embodiment of the wind power new energy supply method in high-altitude areas described in this invention, the second fine screening of the directed acyclic graph that obtains the maximum BIC score refers to applying the conditional mutual information method to calculate the conditional mutual information between each node in the graph and the actual wind power output, and combining the mutual information threshold to perform the screening of the most relevant nodes to obtain the set of the most relevant nodes.

[0016] As a preferred embodiment of the wind power new energy supply method for high-altitude areas described in this invention, the following steps are described: performing a first-order difference operation to generate a perturbation vector, using the perturbation vector as the input of a DRBN network, outputting activation values ​​and constructing an input matrix as the input of a bidirectional long short-term memory network, and then performing a linear transformation on the output hidden state through a fully connected layer, which involves extracting the corresponding components based on the most relevant node set and splicing them according to time points to generate an input vector, and then performing a first-order difference operation on each component to form a perturbation vector;

[0017] Using the hidden layers of the DRBN network, a local perturbation mapping is performed on the perturbation vector to obtain the activation value at each time point. Random numbers are generated by a random number generator as the initial window length. The input matrix is ​​constructed using the initial window length and the activation values. The hidden states of the input matrix, including historical and future hidden states, are calculated through the forward and backward LSTM layers of the Bi-LSTM network. The hidden states are concatenated to obtain the final hidden state. A fully connected layer is used to perform a linear transformation on the final hidden state to output the final power prediction value.

[0018] As a preferred embodiment of the wind power new energy supply method for high-altitude areas described in this invention, the method involves: after optimizing and feeding back the parameters of the DRBN network and the fully connected layer using a parallel heterogeneous optimization mechanism, the final power prediction value is re-output. This involves extracting the number of hidden neurons in the hidden layer, combining it with the initial window length, defining it as the individual chromosome required by the genetic algorithm, generating the population using a uniform random generation method for initialization, and then using the mean square error formula to obtain the mean square error value between the final power prediction value and the actual wind power output power. The reciprocal of the mean square error value is then taken as the fitness function value of the individual chromosome for maximization.

[0019] Calculate the fitness function value, and use roulette wheel selection to select stained individuals for crossover to generate new individuals. Perform perturbation mutation on the new individuals using standard Gaussian perturbation. After iterating to the maximum number of times, output the optimal number of hidden neurons and window length.

[0020] The parameters to be optimized are extracted in parallel, including the center vector and variance of the hidden neurons in the hidden layer, the weight matrix and bias term of the fully connected layer, which constitute the individual of the shark optimization algorithm. After initialization by generating a population with random uniform distribution, a standard initial velocity is set for each individual.

[0021] The fitness function of the chromatic individual is used as the fitness function of each shark individual, and the fitness function value of each shark individual is calculated and iteratively updated. During the iteration process, the shark individuals are sorted in descending order according to their fitness function values, and the shark individual with the maximum fitness function value is selected as the optimal shark individual. The velocity of each shark individual is updated by combining the inertial weight decay mechanism of the particle swarm algorithm.

[0022] The system generates random numbers from a standard normal distribution using a random number generator. Combined with the updated velocity, the position of each individual shark is perturbed and updated. After the maximum number of iterations, the optimal individual shark is output, and the optimal center vector, variance, weight matrix, and bias term are extracted.

[0023] The extracted center vector, variance, weight matrix, and bias term, along with the number of hidden neurons and window length, are fed back to the corresponding hidden layers and fully connected layers, and the final power prediction value is obtained again.

[0024] As a preferred embodiment of the wind power new energy supply method in high-altitude areas described in this invention, the calculation of the power difference is performed by the wind power new energy supply scheduling method, which uses subtraction to solve for the difference between the final power prediction value and the actual wind power output power.

[0025] The scheduling threshold is set according to the critical balance point of wind power supply, and the supply scheduling of wind power new energy is carried out in combination with the difference.

[0026] When the difference is greater than the scheduling threshold, it indicates that the current wind turbine has surplus power, and the excess energy can be stored or shifted to other peaks.

[0027] When the difference equals the scheduling threshold, it means that the power of the current wind turbine is precisely matched, and there is no need to perform scheduling to maintain the current power supply or supply-demand balance.

[0028] When the difference is less than the scheduling threshold, it indicates that the current power supply of the wind turbine is insufficient and there is no need to meet the load demand. In this case, the compensation line should be started or the load reduction should be implemented.

[0029] As a preferred embodiment of the wind power new energy supply method in high-altitude areas described in this invention, the step of collecting raw meteorological data and power data and preprocessing them to generate a dataset refers to collecting raw meteorological data and power data of wind farms and wind turbines through multiple sensors.

[0030] The raw meteorological data includes wind speed, wind direction, temperature, relative humidity, solar radiation intensity, precipitation, and atmospheric pressure;

[0031] All data are denoised, interpolated, and standardized. The standardized meteorological data and power data are divided using the sliding window technique. The standardized meteorological data in the same window are concatenated using the feature concatenation operation to obtain the meteorological feature vector. A weighted operation is performed based on the standardized power data to obtain the actual wind power output of each window. The actual wind power output of each window is then combined with the meteorological feature vector to form the dataset for each window.

[0032] Secondly, the present invention provides a wind power renewable energy supply system for high-altitude areas, comprising,

[0033] The data acquisition and generation module is used to collect raw meteorological and power data, preprocess them, and generate datasets.

[0034] The feature optimization module is used to construct a directed acyclic graph. After generating the maximum likelihood parameter set, it calculates the BIC score of the directed acyclic graph to perform preliminary screening. The directed acyclic graph with the maximum BIC score is then subjected to secondary fine screening.

[0035] The serial prediction module is used to perform first-order difference operations, generate perturbation vectors, use perturbation vectors as input to the DRBN network, output activation values ​​and construct an input matrix as input to the bidirectional long short-term memory network, and then output the final hidden state through a fully connected layer for linear transformation.

[0036] The feedback execution module is used to optimize the parameters of the DRBN network and the fully connected layer using a parallel heterogeneous optimization mechanism, then re-output the final power prediction value and calculate the power difference to execute the supply scheduling of wind power new energy.

[0037] The beneficial effects of this invention are as follows: By constructing a directed acyclic graph based on maximum likelihood estimation and combining it with a conditional mutual information filtering mechanism, the most relevant meteorological feature variables for wind power output are effectively extracted, reducing input dimensional redundancy and improving modeling efficiency. Furthermore, by adopting a serial network structure, local mapping of disturbance features and global trend modeling are achieved, enhancing the ability to perceive high-frequency disturbances and long-term dependencies. Secondly, the introduction of genetic algorithms and shark optimization algorithms to construct a heterogeneous parallel optimization mechanism enables this invention to adaptively optimize the network structure and hyperparameter combination, avoiding getting trapped in local optima. Therefore, this invention improves the supplyability and scheduling accuracy of wind power renewable energy in complex plateau environments. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 is a flowchart of the wind power new energy supply method in high-altitude areas in Example 1.

[0040] Figure 2 is a structural diagram of the wind power new energy supply system in high-altitude areas in Example 1.

[0041] Figure 3 is a flowchart of obtaining the most relevant set of nodes in Example 1. Detailed Implementation

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0044] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention.

[0045] Example 1, referring to Figures 1-3, is the first embodiment of the present invention. This embodiment provides a method for supplying wind power renewable energy in high-altitude areas, including the following steps:

[0046] S1. Collect raw meteorological and power data, preprocess them, and generate a dataset;

[0047] Specifically, collecting raw meteorological data and power data for preprocessing to generate datasets refers to collecting raw meteorological data and power data of wind farms and wind turbines through multiple sensors.

[0048] The raw meteorological data includes wind speed, wind direction, temperature, relative humidity, solar radiation intensity, precipitation, and atmospheric pressure;

[0049] The multiple sensors include wind speed, wind direction, temperature, humidity, solar radiometer, precipitation and barometric pressure sensors;

[0050] All data are denoised, interpolated, and standardized. The standardized meteorological data and power data are divided using the sliding window technique. The standardized meteorological data in the same window are then stitched together using the feature stitching operation to obtain the meteorological feature vector.

[0051] We perform a weighted operation based on standardized power data to obtain the actual wind power output of each window, and then combine it with meteorological feature vectors to form the dataset for each window.

[0052] The weighting operation based on power data is performed using the following formula:

[0053]

[0054] In the formula, Indicates a point in time The actual wind power output at that time This indicates the total number of wind turbine units. Indicates the first Typhoon turbine units at the time point Power data at that time.

[0055] By fusing core meteorological factors and turbine power data from high-altitude wind farms using multi-sensor fusion, a sliding window time-series sample was constructed and standardized. Combined with feature splicing and power weighting, unified data labels were generated, significantly improving the temporal expressiveness of the data structure and the driving accuracy of the supply strategy. Mean fusion of multiple turbine power values ​​eliminated local abnormal fluctuations, enhancing the representativeness of the output labels and ensuring greater stability and generalization ability in model training.

[0056] S2. Based on the dataset, construct a directed acyclic graph, generate the maximum likelihood parameter set, calculate the BIC score of the directed acyclic graph to perform preliminary screening, and perform a second fine screening on the directed acyclic graph with the maximum BIC score to obtain the set of most relevant nodes.

[0057] Specifically, after constructing a directed acyclic graph and generating a set of maximum likelihood parameters, the BIC score of the directed acyclic graph is calculated to perform preliminary screening. Based on the dataset, the equal frequency binning method is used to label the state and state number of each component (wind speed, wind direction, temperature, relative humidity, solar radiation intensity, precipitation, atmospheric pressure and output power). The states include low, medium and high, and the state numbers include 0, 1 and 2.

[0058] In the process of performing equal-frequency binning, each component is first arranged in ascending order to obtain an ordered sequence of each component.

[0059] Based on the state, the positions of the segmentation points are calculated, including the first segmentation point and the second segmentation point, using the following formula:

[0060]

[0061]

[0062] In the formula, Indicates the first dividing point. This represents the total number of components within an ordered sequence. This indicates a round-up operation. Indicates the second dividing point;

[0063] Extracting split point values ​​from an ordered sequence, including the value of the first split point. Second split point value The dataset is then assigned a state based on the segmentation point values, using the following formula:

[0064]

[0065] In the formula, Indicates state, Indicates the first The standardized value of each component, where 0 corresponds to the low state, 1 corresponds to the medium state, and 2 corresponds to the high state;

[0066] Each component is defined as a node. Based on the conditional dependencies between nodes, directional edges are established between nodes to form a directed acyclic graph. Maximum likelihood estimation is used to perform conditional probability estimation on each node in the directed acyclic graph and generate a set of maximum likelihood parameters.

[0067] The conditional probability estimation for each node in the directed acyclic graph is performed using maximum likelihood estimation, as shown in the following formula:

[0068]

[0069] In the formula, Indicates when the first Each component takes the value of And the set of parent nodes Values The conditional probability at that time, Indicates the first The nodes corresponding to each component Indicates the state (i.e. the state of the component, low, medium, high). Indicates the first The set of parent nodes corresponding to each component (which can be generated by counting all the nodes pointed to by the nodes in the directed acyclic graph). Indicates the state of the parent node. Indicates the frequency of occurrence (which can be obtained from a directed acyclic graph using statistical methods);

[0070] The formula for generating the maximum likelihood parameter set is:

[0071]

[0072] In the formula, Denotes the set of maximum likelihood parameters. Represents mathematical logic symbols (i.e., for all, any);

[0073] Based on the number of states (i.e., low, medium, and high), the number of state combinations of the parent node is calculated. Then, combined with the number of states, the number of independent parameters of each node is calculated and summed to generate the total number of parameters. Using the information criterion (BIC) in combination with the maximum likelihood parameter set and the dataset, the BIC score of the directed acyclic graph is calculated, and the directed acyclic graph corresponding to the maximum BIC score is extracted.

[0074] The formula for calculating the number of state combinations of the parent node is:

[0075]

[0076] In the formula, Indicates the first The number of state combinations of all parent nodes of each component. Indicates the first The parent node of each component , Indicates belonging to, Indicates the parent node The number of states (i.e., low, medium, high).

[0077] The formula for summing the number of independent parameters for each node is as follows:

[0078]

[0079]

[0080] In the formula, Indicates the first The number of independent parameters of the node corresponding to each component. Indicates the first The number of states of the nodes corresponding to each component. Indicates the total number of parameters. Indicates the total number of nodes;

[0081] The formula for calculating the BIC score of a directed acyclic graph is as follows:

[0082]

[0083] In the formula, Represents a directed acyclic graph BIC score, The logarithmic function is defined as having the natural logarithm as its base. In a given dataset Directed acyclic graph and the maximum likelihood parameter set Conditional probability under given conditions Represents the dataset The total number of samples in the sample;

[0084] in, It can be obtained by summing the conditional probabilities of all nodes, as shown in the following formula:

[0085]

[0086] In the formula, Represents the time points in the dataset Within a time The observed values ​​(standardized values) of each component node. Represents the time points in the dataset Within a time The log-conditional probability of each component node (can be generated by performing a logarithmic operation on the conditional probability estimate in the form of frequency from the maximum likelihood parameter set).

[0087] By introducing the equal-frequency binning method to discretize the components, the original meteorological and power variables are given a unified state expression, and a directed acyclic graph structure is established to effectively characterize the conditional dependencies between variables. Furthermore, maximum likelihood estimation is used to obtain the conditional probabilities of nodes, forming a complete probability parameter system without relying on manual experience. A BIC model evaluation mechanism is constructed through constraints on the number of parameters and sample complexity, which can screen the optimal network topology from multiple possible structures, improving the model's generalization ability. Secondly, this invention achieves multi-time-point and multi-node joint modeling by expanding the log-likelihood function over time, adapting to the non-stationary characteristics of frequent disturbances in wind power in plateau environments. This not only allows the invention to consider the overall closed loop of variable selection, structural modeling, and complexity control, but also establishes a direct link between the model structure and the final new energy supply strategy, realizing a true "from modeling to regulation" linkage chain.

[0088] Furthermore, the directed acyclic graph with the highest BIC score is subjected to a secondary fine screening. This involves applying the conditional mutual information method to calculate the conditional mutual information between each node in the graph and the actual wind power output. Combined with the mutual information threshold (given the significant differences in wind conditions and climate characteristics among different wind farms, the mutual information threshold is usually set to 0.03 to 0.1 in the prior art. When the threshold is less than 0.03, redundant variables are easily introduced, leading to a complex network structure and an increased risk of overfitting. When the threshold is greater than 0.1, some variables with actual impact but low information content will be eliminated, resulting in the problem of missing key features. Therefore, in this invention, 0.05 can be taken as the default value of the mutual information threshold, so that this invention can exclude invalid variables) the most relevant node screening is performed to obtain the set of most relevant nodes.

[0089] The conditional mutual information method is used to calculate the conditional mutual information between each node in the graph and the actual wind power output, and the formula is as follows:

[0090]

[0091] In the formula, Indicates that in a given number The set of parent nodes of each component node Under the condition of component nodes and Conditional mutual information, Indicates the first The state values ​​of each component node. Indicates the first The state values ​​of each component node. Indicates the first and The probability that the state values ​​of each component node and its parent node appear together. Indicates that in a given Given the values ​​of the parent node's state, the first The state value of each component node is... The probability, This indicates the state of the first node given the state values ​​of its parent node. The state value of each component node is... The probability of;

[0092] , , All of these can be obtained through statistical estimation from the directed acyclic graph corresponding to the maximum BIC score using statistical methods;

[0093] The most relevant nodes are selected by combining mutual information thresholds. The specific steps are as follows: When If the mutual information is greater than or equal to the mutual information threshold, it means that the conditional mutual information between the current node and the actual wind power output meets the screening conditions, and the corresponding node should be retained; otherwise, the current node should be removed.

[0094] By extracting the directed acyclic graph (DAG) corresponding to the maximum BIC score, which possesses optimal causal expressive power and minimal redundancy, unnecessary complex connections are avoided, thus reducing the computational burden of subsequent modeling and improving the overall interpretability of the model. This is particularly effective in high-dimensional wind power data modeling, preventing overfitting and interference from invalid variables. Furthermore, by utilizing conditional mutual information while maintaining the existing parent-child dependency structure, the "conditional contribution" of variables to wind power is further quantified, achieving more semantically accurate variable selection. This mechanism significantly improves the statistical efficiency and predictive ability of the model structure. Moreover, by setting a mutual information threshold, the sensitivity of variable retention can be flexibly controlled to adapt to the differences in climate characteristics of wind farms in different regions. This strategy has good engineering adjustability and is suitable for deployment in personalized modeling scenarios for different wind farms. Secondly, the obtained set of most relevant nodes not only reduces the input dimension and improves the model's generalization ability, but also makes the input features and wind power have a stronger causal relationship through the conditional mutual information filtering mechanism, thus providing a more reliable data foundation for subsequent serial network modeling (such as DRBN and Bi-LSTM).

[0095] S3. Based on the most relevant set of nodes, perform a first-order difference operation to generate a perturbation vector. Use the perturbation vector as the input of the DRBN network, output the activation value, and construct the input matrix as the input of the bidirectional long short-term memory network. Then, output the final hidden state and perform a linear transformation through a fully connected layer to obtain the final power prediction value.

[0096] Specifically, a first-order difference operation is performed to generate a perturbation vector. The perturbation vector is used as the input to the DRBN network. The output activation value is used to construct an input matrix as the input to the bidirectional long short-term memory network. The output of the final hidden state is then linearly transformed by a fully connected layer. Based on the set of most relevant nodes, the corresponding components are extracted and concatenated according to time points to generate the input vector. Then, a first-order difference operation is performed on each component to form the perturbation vector.

[0097] Using the hidden layers of the DRBN network, a local perturbation mapping is performed on the perturbation vector to obtain the activation value at each time point, as shown in the formula:

[0098]

[0099] In the formula, Indicates the first A hidden neuron at time point The activation value at that time, This represents the natural exponential function. Describing the L2 norm, Indicates a point in time The perturbation vector at time, Indicates the first The center vector of each hidden neuron It is a vector with the same dimension as the perturbation vector. It can represent the "center position" of the current hidden neuron in the high-dimensional input space and is used to measure the distance between the current perturbation vector and the current hidden neuron, thereby determining the response strength of the hidden neuron. For example, the closer the distance, the higher the activation value. Indicates the first The variance of a hidden neuron is a response range that controls the Gaussian kernel, i.e., the sensitivity of the hidden neuron to the perturbation vector. It determines the "perception range" of the current hidden neuron to the perturbation vector. For example, The larger the value, the lower the sensitivity of the hidden neuron to distance, and the wider its response range. The smaller the value, the narrower the response range, thus making the hidden neurons more selective;

[0100] pass and The configuration gives the DRBN network interpretable "local awareness" capabilities, and through subsequent optimization, it enables the DRBN network to self-adjust its structure.

[0101] Random numbers are generated using a random number generator as the initial window length. The input matrix is ​​constructed using the initial window length and activation values. The hidden states of the input matrix, including historical and future hidden states, are calculated through the forward and backward LSTM layers of a Bi-LSTM network (Bi-LSTM uses mean squared error as the loss function and uses the Adam optimizer to iteratively update the model. During the iteration process, the iteration stops when the decrease in the loss function no longer decreases significantly, and the trained Bi-LSTM network is output). The hidden states are concatenated to obtain the final hidden state. A fully connected layer is used to perform a linear transformation on the final hidden state to output the final power prediction value.

[0102] The hidden state of the input matrix is ​​calculated using the following formula:

[0103]

[0104]

[0105] In the formula, Indicates a point in time The historical hidden state output to the LSTM layer before the time step. The standard network function representing the feedforward LSTM layer, Indicates the start time point is and the termination time point is The input matrix, Indicates a point in time The future hidden state output to the LSTM layer after time. The standard network function representing the backward LSTM layer;

[0106] The final hidden state is linearly transformed using a fully connected layer, as shown in the formula:

[0107]

[0108] In the formula, Indicates a point in time The final power forecast value of the wind turbine at that time. Represents a fully connected layer The weight matrix, Indicates a point in time The final hidden state at that time Represents a fully connected layer The bias term;

[0109] The number of hidden neurons in the hidden layer, parameters , The weight matrix and bias terms of fully connected layers can both be initialized using a random number generator. For example, the number of hidden neurons can be a random integer, and the parameters... and The hidden layer dimensions of the Bi-LSTM network can be randomly and uniformly generated, and the weight matrix and bias terms of the fully connected layer can be initialized by Xavier and set by constants.

[0110] By applying first-order differencing to multi-source meteorological and environmental variables, a perturbation vector is obtained, enabling the model to keenly capture rapidly changing dynamic features. This is particularly suitable for modeling the sudden changes in wind power output under variable climates in high-altitude regions. The hidden layers of the DRBN are activated by responding to the perturbation vector with Gaussian kernels, allowing the network to form local feature clusters under high-frequency perturbation conditions, improving the fitting accuracy of the nonlinear mapping relationship between input perturbations and wind power output. The introduction of a random window length avoids the mismatch between fixed window structures and data feature scales, giving the model adaptive time-series awareness and improving prediction robustness under different wind conditions. Furthermore, the Bi-LSTM network combines historical and future states, effectively enhancing the model's ability to learn from long-term dependencies and adapting to the periodic trends of wind energy changes with climate rhythms in high-altitude environments. Finally, the hidden states are linearly transformed through fully connected layers to form the output prediction value. This structure retains the ability to handle nonlinearities while providing physical traceability of the output results, offering a clear basis for formulating new energy supply strategies.

[0111] S4. After optimizing and feeding back the parameters of the DRBN network and the fully connected layer using a parallel heterogeneous optimization mechanism, the final power prediction value is re-output, and the power difference is calculated to execute the supply scheduling of wind power new energy.

[0112] Specifically, after optimizing and feeding back the parameters of the DRBN network and the fully connected layer using a parallel heterogeneous optimization mechanism, the final power prediction value is re-output. This involves extracting the number of hidden neurons in the hidden layer, combining it with the initial window length, and defining them as the individual chromosomes required by the genetic algorithm. After initializing the population using a uniform random generation method, the mean square error formula is used to obtain the mean square error values ​​of the final power prediction value and the actual wind power output power. The reciprocal of the mean square error value is then taken as the fitness function value of the individual chromosomes and maximized.

[0113] The reciprocal of the mean squared error value is used as the fitness function value of the chromosomal individual for maximization, and the formula is:

[0114]

[0115] In the formula, Represents chromosomal individuals The fitness function value, Indicates the total number of sliding windows;

[0116] Calculate the fitness function value, and use roulette wheel selection to select stained individuals for crossover to generate new individuals. Perform perturbation mutation on the new individuals using standard Gaussian perturbation. After iterating to the maximum number of times, output the optimal number of hidden neurons and window length.

[0117] The parameters to be optimized are extracted in parallel, including the center vector and variance of the hidden neurons in the hidden layer, the weight matrix and bias term of the fully connected layer, which constitute the individual of the shark optimization algorithm. After initialization by generating a population with random uniform distribution, a standard initial velocity is set for each individual (for example, taking the standard initial value "0").

[0118] The fitness function of the chromatic individuals is used as the fitness function for each shark individual. The fitness function value of each shark individual is calculated and iteratively updated. During the iteration process, the shark individuals are sorted in descending order according to their fitness function values, and the shark individual with the highest fitness function value is selected as the optimal shark individual. The velocity of each shark individual is updated using the inertia weight decay mechanism of the particle swarm optimization algorithm, as shown in the following formula:

[0119]

[0120] In the formula, Indicates the first The first individual shark in the... Speed ​​during round iteration Indicates the first Inertia weights during round iterations, Indicates the first The first individual shark in the... The initial velocity during round iteration, Represents the learning factor. This represents a uniformly distributed random number (generated by a random number generator). Indicates the first The position of the optimal shark individual during round iteration. Indicates the first The first individual shark in the... Position during round iteration;

[0121] Learning factor In the prior art, the standard value is usually 1.5 to 1.8. However, in order to avoid local optima and divergence in the present invention, 1.5 can be taken as the default value of the present invention.

[0122] The inertia weight is obtained as follows:

[0123]

[0124] In the formula, This represents the maximum inertia weight. This represents the minimum inertia weight. Indicates the maximum number of iterations;

[0125] The maximum and minimum inertia weights in this invention can be set empirically. However, in existing technologies, if the inertia weight is kept high (e.g., close to 1.0), the population is prone to excessive jumps, oscillations, and difficulty in convergence in high-dimensional space. On the other hand, if the inertia weight is too small in the initial stage (e.g., less than 0.3), the search is prone to getting stuck in local extrema and lacks exploratory power. Therefore, in this invention, the maximum inertia weight can be set to 0.9 and the minimum inertia weight to 0.4. The setting of 0.9 not only gives this invention a greater jumping ability but also allows for coarse search. Furthermore, the setting of 0.4 can reduce search oscillations, allowing the algorithm to converge to a stable region.

[0126] Random numbers following a standard normal distribution are generated using a random number generator. These numbers, combined with the updated velocity, are used to perturb and update the position of each individual shark. After the maximum number of iterations, the optimal individual shark is output, and the optimal center vector, variance, weight matrix, and bias term are extracted. The formula is as follows:

[0127]

[0128] In the formula, Indicates the first The first individual shark in the... Update position during round iteration This represents the disturbance coefficient. Random numbers representing a standard normal distribution;

[0129] Disturbance coefficient The setting of the value is related to the stability and search depth of the present invention. In the prior art, based on empirical knowledge and related perturbation mechanisms, the value range is usually set to 0.01 to 0.1. If the perturbation coefficient is less than 0.01, the perturbation is too weak, causing the algorithm to be unable to escape local optima. However, if the perturbation coefficient is greater than 0.1, the perturbation will be too large, making the algorithm unstable. Therefore, in the present invention, the perturbation coefficient can be set to 0.05 as the default value, thereby ensuring that the search process of the present invention has diversity and effectively escapes local extrema.

[0130] The extracted center vector, variance, weight matrix, and bias term, along with the number of hidden neurons and window length, are fed back to the corresponding hidden layers and fully connected layers, and the final power prediction value is obtained again.

[0131] By employing a parallel heterogeneous optimization mechanism, the model's structural parameters and weight hyperparameters are jointly optimized, thereby achieving high-precision modeling of wind power and dynamic new energy supply decision support under complex meteorological conditions and nonlinear disturbances. Furthermore, this parallel heterogeneous optimization mechanism effectively overcomes the shortcomings of traditional neural network structures that rely on empirical settings, automatically adjusting neuron size and time windows to improve the model's adaptability to wind field time-series patterns in high-altitude areas. The genetic and shark optimization algorithms each undertake the responsibilities of structural optimization and parameter optimization, ensuring spatial coverage while maintaining algorithm convergence and avoiding getting trapped in local optima. Secondly, this parallel heterogeneous optimization mechanism, through a unified fitness function definition, forms a unified evaluation standard for heterogeneous model optimization, achieving a two-way feedback loop between model performance and actual supply effect. Thirdly, the genetic and shark optimization algorithms are used in optimization tasks in two "parameter domains," and the results are fed back to the same prediction network through a unified fitness function, forming a fused, interconnected optimization loop. This design enables this invention to form a complete logical link: structural optimization → parameter refinement → supply prediction → accuracy feedback.

[0132] Furthermore, the final output parameters are back-injected into the DRBN and Bi-LSTM network structures, driving the entire wind power prediction module to be reconstructed into a self-evolving multi-stage optimization system, enabling the model to dynamically adjust the prediction structure and supply strategy according to different high-altitude meteorological inputs.

[0133] Furthermore, the power difference calculation method for wind power new energy supply dispatching uses subtraction to solve for the difference between the final power prediction value and the actual wind power output power.

[0134] The scheduling threshold is set according to the critical balance point of wind power supply (for example, the value of the critical balance point is usually "0" in the standard value, so "0" can be set as the scheduling threshold of the present invention), and the supply scheduling of wind power new energy is carried out in combination with the difference.

[0135] When the difference is greater than the scheduling threshold, it indicates that the current wind turbine has surplus power, and the excess energy can be stored or shifted to other peaks.

[0136] When the difference equals the scheduling threshold, it means that the power of the current wind turbine is precisely matched, and there is no need to perform scheduling to maintain the current power supply or supply-demand balance.

[0137] When the difference is less than the scheduling threshold, it indicates that the current power supply of the wind turbine is insufficient and there is no need to meet the load demand. In this case, the compensation line should be started or the load reduction should be implemented.

[0138] The difference between the predicted wind power and the actual load power is obtained through subtraction. Combined with a preset scheduling threshold, the system is determined to be in a state of power surplus, precise matching, or insufficient supply. Based on the determination, the system executes supply actions such as energy storage charging, maintaining the status quo, activating backup power, and load reduction, thereby achieving precise matching between wind power output and actual electricity demand. This invention not only constructs a closed-loop control path between wind power prediction results and new energy supply strategies but also significantly improves the system's rapid response capability and scheduling accuracy to power fluctuations.

[0139] This embodiment also provides a wind power renewable energy supply system for high-altitude areas, including:

[0140] The data acquisition and generation module is used to collect raw meteorological and power data, preprocess them, and generate datasets.

[0141] The feature optimization module is used to construct a directed acyclic graph. After generating the maximum likelihood parameter set, it calculates the BIC score of the directed acyclic graph to perform preliminary screening. The directed acyclic graph with the maximum BIC score is then subjected to secondary fine screening.

[0142] The serial prediction module is used to perform first-order difference operations, generate perturbation vectors, use perturbation vectors as input to the DRBN network, output activation values ​​and construct an input matrix as input to the bidirectional long short-term memory network, and then output the final hidden state through a fully connected layer for linear transformation.

[0143] The feedback execution module is used to optimize the parameters of the DRBN network and the fully connected layer using a parallel heterogeneous optimization mechanism, then re-output the final power prediction value and calculate the power difference to execute the supply scheduling of wind power new energy.

[0144] This embodiment also provides a computer device applicable to the wind power renewable energy supply method in high-altitude areas, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the wind power renewable energy supply method in high-altitude areas as proposed in the above embodiment.

[0145] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0146] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for supplying wind power renewable energy in high-altitude areas as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0147] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for supplying new wind power energy in high-altitude areas, characterized in that: include, Raw meteorological and power data are collected, preprocessed, and used to generate datasets. Based on the dataset, a directed acyclic graph (DAG) is constructed. After generating the maximum likelihood parameter set, the BIC score of the DAG is calculated to perform preliminary screening. The DAG with the maximum BIC score is then subjected to secondary fine screening to obtain the most relevant node set. Based on the most relevant node set, a first-order difference operation is performed to generate a perturbation vector. The perturbation vector is used as the input of the DRBN network, and the output activation value is used to construct an input matrix as the input of the bidirectional long short-term memory network. The output final hidden state is then linearly transformed through a fully connected layer to obtain the final power prediction value. After optimizing and feeding back the parameters of the DRBN network and the fully connected layer using a parallel heterogeneous optimization mechanism, the final power prediction value is re-output, and the power difference is calculated to execute the supply scheduling of wind power new energy.

2. The method for supplying new wind power energy in high-altitude areas as described in claim 1, characterized in that: The process involves constructing a directed acyclic graph (DAG), generating a maximum likelihood parameter set, and then calculating the BIC score of the DAG to perform preliminary screening. Based on the dataset, equal-frequency binning is used to label each component with a state and state number. States include low, medium, and high, and state numbers include 0, 1, and 2. Each component is defined as a node, and connection edges are established between nodes according to their conditional dependencies, forming the DAG. Maximum likelihood estimation is used to perform conditional probability estimation on each node in the DAG, generating a maximum likelihood parameter set. Based on the number of states, the number of state combinations for the parent node is calculated. Further, combining the number of states, the number of independent parameters for each node is calculated and summed to generate the total number of parameters. Using information criteria, combined with the maximum likelihood parameter set and the dataset, the BIC score of the DAG is calculated, and the DAG corresponding to the maximum BIC score is extracted.

3. The method for supplying new wind power energy in high-altitude areas as described in claim 2, characterized in that: The process of performing secondary fine screening on the directed acyclic graph that obtains the maximum BIC score refers to applying the conditional mutual information method to calculate the conditional mutual information between each node in the graph and the actual wind power output, and then combining the mutual information threshold to perform the screening of the most relevant nodes to obtain the set of the most relevant nodes.

4. The method for supplying new wind power energy in high-altitude areas as described in claim 3, characterized in that: The process involves performing a first-order difference operation to generate a perturbation vector. This perturbation vector is then used as input to the DRBN network. After outputting activation values ​​and constructing an input matrix, which is then used as input to the bidirectional long short-term memory network, the final hidden state is linearly transformed through a fully connected layer. This process involves extracting the corresponding components based on the most relevant node set and concatenating them at time points to generate an input vector. A first-order difference operation is then performed on each component to form the perturbation vector. The hidden layers of the DRBN network are used to perform local perturbation mapping on the perturbation vector to obtain the activation value at each time point. Random numbers are generated using a random number generator as the initial window length. The initial window length is combined with the activation values ​​to construct the input matrix. The hidden state of the input matrix, including historical and future hidden states, is calculated through the forward and backward LSTM layers of the Bi-LSTM network. These hidden states are concatenated to obtain the final hidden state. A fully connected layer is then used to perform a linear transformation on the final hidden state to output the final power prediction value.

5. The method for supplying new wind power energy in high-altitude areas as described in claim 4, characterized in that: After optimizing the parameters of the DRBN network and fully connected layers using a parallel heterogeneous optimization mechanism, the final power prediction value is re-output. This involves extracting the number of hidden neurons in the hidden layers and defining them as the individual chromosomes required for the genetic algorithm, based on the initial window length. A population is generated using a uniform random generation method for initialization. The mean square error formula is used to obtain the mean square error values ​​of the final power prediction value and the actual wind power output power. The reciprocal of the mean square error value is taken as the fitness function value of the individual chromosome and maximized. The fitness function value is calculated, and a roulette wheel selection method is used to select individual chromosomes for crossover to generate new individuals. Standard Gaussian perturbation is used to perform perturbation mutation on the new individuals. After iterating to the maximum number of iterations, the optimal number of hidden neurons and window length are output. In parallel, the parameters to be optimized, including the center vector and variance of the hidden neurons in the hidden layers, the weight matrix of the fully connected layers, and the bias term, are extracted to form individuals for the shark optimization algorithm. These are then randomly and uniformly distributed. After initializing the population, a standard initial velocity is set for each individual. The fitness function of the chromatic individuals is used as the fitness function for each shark individual, and the fitness function value of each shark individual is calculated and iteratively updated. During the iteration process, the shark individuals are sorted in descending order according to their fitness function values, and the shark individual with the maximum fitness function value is selected as the optimal shark individual. The velocity of each shark individual is updated using the inertial weight decay mechanism of the particle swarm optimization algorithm. A random number generator generates a standard normal distribution of random numbers, and the position of each shark individual is perturbed and updated in combination with the updated velocity. When the maximum number of iterations is reached, the optimal shark individual is output, and the optimal center vector, variance, weight matrix, and bias term are extracted. The extracted center vector, variance, weight matrix, and bias term, along with the number of hidden neurons and the window length, are fed back to the corresponding hidden layers and fully connected layers, and the final power prediction value is obtained again.

6. The method for supplying new wind power energy in high-altitude areas as described in claim 5, characterized in that: The calculation of the power difference for wind power new energy supply scheduling refers to using subtraction to solve for the difference between the final power prediction value and the actual wind power output power; setting a scheduling threshold based on the critical balance point of wind power supply, and combining the difference to execute wind power new energy supply scheduling. When the difference is greater than the scheduling threshold, it indicates that the current wind turbine has surplus power, and the excess energy can be stored or shifted to other peaks. When the difference equals the scheduling threshold, it means that the power of the current wind turbine is precisely matched, and there is no need to perform scheduling to maintain the current power supply or supply-demand balance. When the difference is less than the scheduling threshold, it indicates that the current power supply of the wind turbine is insufficient and there is no need to meet the load demand. In this case, the compensation line should be started or the load reduction should be implemented.

7. The method for supplying new wind power energy in high-altitude areas as described in claim 6, characterized in that: The process of collecting raw meteorological and power data and preprocessing it to generate a dataset refers to collecting raw meteorological and power data of wind farms and wind turbines through multiple sensors. The raw meteorological data includes wind speed, wind direction, temperature, relative humidity, solar radiation intensity, precipitation, and atmospheric pressure. All data are denoised, interpolated, and standardized. The standardized meteorological and power data are divided using a sliding window technique. The standardized meteorological data in the same window are concatenated using a feature concatenation operation to obtain a meteorological feature vector. A weighted operation is performed based on the standardized power data to obtain the actual wind power output power of each window. This data is then combined with the meteorological feature vector to form the dataset for each window.

8. A wind power renewable energy supply system for high-altitude areas, based on the wind power renewable energy supply method for high-altitude areas as described in any one of claims 1 to 7, characterized in that: This includes a data acquisition and generation module, which collects raw meteorological and power data, preprocesses them, and generates a dataset. The feature optimization module is used to construct a directed acyclic graph. After generating the maximum likelihood parameter set, it calculates the BIC score of the directed acyclic graph to perform preliminary screening. The directed acyclic graph with the maximum BIC score is then subjected to secondary fine screening. The serial prediction module is used to perform first-order difference operations, generate perturbation vectors, use perturbation vectors as input to the DRBN network, output activation values ​​and construct an input matrix as input to the bidirectional long short-term memory network, and then output the final hidden state through a fully connected layer for linear transformation. The feedback execution module is used to optimize and feedback the parameters of the DRBN network and the fully connected layer using a parallel heterogeneous optimization mechanism, then re-output the final power prediction value and calculate the power difference to execute the supply scheduling of wind power new energy.

Images