Power grid optimization dispatching method and device based on dynamic selection of wind power prediction model

Abstract

The present invention discloses a power grid optimization scheduling method and device based on dynamic selection of wind power prediction models. The method includes: obtaining a prediction error probability density function of each wind power prediction model based on multiple wind power prediction models; obtaining real-time meteorological data of wind turbines, performing power prediction based on multiple wind power prediction models, obtaining multiple wind power real-time prediction values, and generating wind power real-time prediction scenario sets in combination with corresponding prediction error probability density functions; based on multiple wind power real-time prediction scenario sets, generating real-time control strategies corresponding to the scenario sets through a random unit combination model, and selecting the real-time control strategy with the lowest sum of wind power curtailment and load loss among the multiple real-time control strategies as the optimized control strategy for the wind turbine. Based on the scenario sets generated by multiple prediction models, random unit combinations are performed respectively, and multiple real-time control strategies are generated in combination with the influence of meteorological factors, and finally the strategy with the lowest sum of wind power curtailment and load loss is selected as the optimized control strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid control, and in particular to a power grid optimization dispatching method and device based on dynamic selection of a wind power prediction model. Background Art

[0002] For power systems containing wind power resources, traditional optimization scheduling models usually select a prediction model with the smallest error among multiple wind power prediction models, and use this prediction model for a long time in the random optimization scheduling model, resulting in the scheduling side only being able to "passively" receive one prediction result (the probability distribution of the prediction value and the prediction error). It should be noted that no prediction model can obtain better prediction results in all cases, and this model ignores other prediction results, resulting in the scheduling side losing potential decision-making results that may be more suitable for actual conditions. Summary of the Invention

[0003] The purpose of the embodiment of the present invention is to provide a power grid optimization scheduling method and device based on the dynamic selection of a wind power prediction model. By fully considering the error characteristics of the wind power prediction model and the uncertainty of wind power affected by meteorological factors, a variety of wind power real-time prediction scenario sets are generated, and multiple real-time control strategies are formulated with the help of a random unit combination model. Finally, the strategy with the lowest sum of wind power abandonment and load loss is selected as the optimized control strategy, which effectively improves the utilization efficiency of wind power resources, enhances the ability of the power grid to cope with wind power fluctuations, optimizes power grid scheduling, and helps to achieve safe, stable and efficient operation of the power system.

[0004] To solve the above technical problems, a first aspect of an embodiment of the present invention provides a power grid optimization scheduling method based on dynamic selection of a wind power prediction model, comprising the following steps:

[0005] Based on multiple wind power prediction models, obtaining a prediction error probability density function of each wind power prediction model to obtain several wind power prediction error scenario sets;

[0006] Acquire real-time meteorological data of the wind turbine generator set, perform power forecasts based on the multiple wind power forecast models, obtain a number of wind power real-time forecast values, and generate a number of wind power real-time forecast scenario sets based on the corresponding wind power forecast error scenario sets;

[0007] Based on the several wind power real-time prediction scenario sets, several real-time control strategies corresponding to the wind power real-time prediction scenario sets are generated one by one through a random unit combination model, and the real-time control strategy with the lowest sum of wind power curtailment and load loss among the several real-time control strategies is selected as the optimized control strategy for the wind turbine set.

[0008] Furthermore, the types of real-time meteorological data of the wind turbine generator system include: wind speed, wind direction, temperature and / or air pressure.

[0009] Furthermore, the obtaining of a prediction error probability density function of each wind power prediction model based on a plurality of wind power prediction models includes:

[0010] Obtain historical forecast data and historical actual data of wind turbines;

[0011] Based on the historical forecast data, establishing nonlinear mapping relationships between meteorological data and multiple wind power prediction models;

[0012] Calculating a prediction error of each wind power prediction model based on the historical actual data;

[0013] Based on the prediction error, a prediction error probability density function of each wind power prediction model is obtained respectively by kernel density estimation.

[0014] Furthermore, the wind power prediction model includes: a neural network model, a support vector machine model and a random forest model;

[0015] The corresponding nonlinear mapping relationship of the neural network model is:

[0016]

[0017] The corresponding nonlinear mapping relationship of the support vector machine model is:

[0018]

[0019] K SVM (X i ,X)=exp(-γ||X i -X|| 2 );

[0020] The nonlinear mapping relationship corresponding to the random forest model is:

[0021]

[0022] Among them, f NN (·),f SVM (·) and f RF (·) are the mapping relationships between wind power prediction values ​​and meteorological data X based on neural network model, support vector machine model and random forest model, respectively. l 、W l and are the activation function, weight vector, and bias vector of the lth layer of the neural network, and l=1,2,3,...,L, L is the number of layers of the neural network, W and b NN are the weight matrix and bias matrix of the neural network respectively, N S is the number of support vectors in the support vector machine model, α i is the Lagrange multiplier of the i-th support vector of the support vector machine model, y i is the true label of the i-th support vector of the support vector machine model, K SVM (X i ,X) is the kernel function of the support vector machine model, γ is a positive parameter, X i is the X corresponding to the i-th sample point, b SVM is the bias term of the support vector machine model, N RF is the number of decision trees in the random forest model, is the predicted value of the i-th decision tree for input X, p wind,NN 、p wind,SVM and p wind,RF These are the wind power prediction values ​​obtained based on the neural network model, support vector machine model and random forest model respectively.

[0023] Furthermore, the prediction error probability density function of the neural network model is:

[0024]

[0025] The corresponding prediction error probability density function of the support vector machine model is:

[0026]

[0027] The corresponding prediction error probability density function of the random forest model is:

[0028]

[0029] Among them, K(·) is the kernel density estimation function, h is the bandwidth, N S is the number of samples, Δp wind is the wind power prediction error random variable, X i and are the meteorological factor input vector and historical actual wind power of the i-th historical data sample point, and are the wind power prediction errors of the i-th historical data sample point of the neural network model, support vector machine model and random forest model, and These are the prediction error probability density functions of the neural network model, support vector machine model, and random forest model, respectively.

[0030] Furthermore, the generating of wind power real-time prediction scenario sets respectively based on the corresponding prediction error probability density functions includes:

[0031] Based on the prediction error probability density function, the corresponding wind power prediction error scenario set is generated by Latin hypercube sampling and Cholesky decomposition:

[0032]

[0033] Among them, Ω NN,err ,Ω SVM,err and Ω RF,err are wind power prediction error scenario sets for neural network model, support vector machine model and random forest model, (Δp wind,NN ) i 、(Δp wind,SVM ) i and (Δp wind,RF ) i The i-th wind power prediction error scenario generated by the neural network model, support vector machine model and random forest model, i = 1, 2, 3, ..., N sg , N sg is the number of scenes.

[0034] Based on the wind power real-time prediction value and the wind power prediction error scenario set, a wind power real-time prediction scenario set is obtained:

[0035]

[0036] Among them, Ω NN ,Ω SVM and Ω RF The wind power real-time prediction scenario sets of neural network model, support vector machine model and random forest model are respectively, wind ,NN ,pre , and p wind ,RF ,pre The real-time wind power prediction values ​​of the neural network model, support vector machine model and random forest model are respectively (p wind,NN ) i 、(p wind,SVM ) i and (p wind,RF ) i The i-th wind power prediction scenario generated by the neural network model, support vector machine model and random forest model, i = 1, 2, 3, ..., N sg , N sg is the number of scenes.

[0037] Furthermore, the wind power prediction model includes: a neural network model, a support vector machine model and a random forest model;

[0038] The random unit combination model is:

[0039]

[0040] Among them, f(p wind,output ,x) is the objective function of the random unit combination model, z(p wind,output ,x) is the objective function value, g(p wind,output ,x)≤0 is the constraint condition of the random unit model, including wind power output scenario constraint, power balance constraint and flow constraint, Represents the prediction scenario set of neural network model, support vector machine model and random forest model, p wind,output is the control strategy, and x is the output of other power generation equipment in the power system and the node power angle variable.

[0041] Furthermore, the selection formula of the real-time control strategy is:

[0042]

[0043] Among them, (p wind,output,NN ,x NN )、(p wind,output,SVM ,x SVM ) and (p wind,output,r ,x SVM ) are the real-time control strategies corresponding to the wind power real-time prediction scenario set obtained by power prediction based on the neural network model, support vector machine model and random forest model, and B is the final optimized control strategy. (p wind,output,r ,x r ) is the one with the lowest sum of wind power curtailment and load loss among the real-time control strategies corresponding to the wind power real-time prediction scenario set obtained by prediction based on the neural network model, support vector machine model and random forest model.

[0044] Accordingly, a second aspect of an embodiment of the present invention provides a power grid optimization dispatching system based on dynamic selection of a wind power prediction model, comprising:

[0045] A function construction module, which is used to obtain a prediction error probability density function of each wind power prediction model based on multiple wind power prediction models, and obtain a plurality of wind power prediction error scenario sets;

[0046] a scenario set generation module, which is used to obtain real-time meteorological data of the wind turbine, perform power prediction based on the multiple wind power prediction models, obtain a number of wind power real-time prediction values, and generate a number of wind power real-time prediction scenario sets in combination with the corresponding wind power prediction error scenario sets;

[0047] An optimization scheduling module is used to generate a number of real-time control strategies corresponding to the wind power real-time prediction scenario sets based on the several wind power real-time prediction scenario sets through a random unit combination model, and select the real-time control strategy with the lowest sum of wind power curtailment and load loss among the several real-time control strategies as the optimized control strategy for the wind turbine set.

[0048] Accordingly, a third aspect of an embodiment of the present invention provides an electronic device, comprising: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the above-mentioned grid optimization scheduling method based on dynamic selection of a wind power prediction model.

[0049] Accordingly, a fourth aspect of an embodiment of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-mentioned power grid optimization scheduling method based on dynamic selection of a wind power prediction model.

[0050] The above technical solutions of the embodiments of the present invention have the following beneficial technical effects:

[0051] 1. By utilizing historical forecast data and actual historical data of wind turbines, nonlinear mapping relationships are established between various wind power forecast models (neural network model, support vector machine model, and random forest model) and meteorological data, and the forecast error is calculated to obtain the forecast error probability density function. This allows for accurate grasp of the forecast characteristics and error distribution of each model. When combined with real-time meteorological data for power forecasting and the generation of a real-time forecast scenario set, the real-time wind power forecast scenario set generated through Latin hypercube sampling and Cholesky decomposition, based on these accurate model characteristics and error analysis, is more in line with actual conditions and comprehensively covers the possible range of wind power variations. This provides reliable and comprehensive basic data for subsequent optimized scheduling, effectively improving the accuracy and rationality of wind power forecasting and scenario simulation.

[0052] 2. Based on multiple sets of real-time wind power prediction scenarios, a random unit combination model is used to generate multiple corresponding real-time control strategies. This random unit combination model comprehensively considers multiple key factors such as wind power output scenario constraints, power balance constraints, and flow constraints, ensuring that the generated control strategies can meet the basic requirements of grid operation under different wind power scenarios. In addition, since corresponding control strategies are generated for different prediction model scenario sets, this multi-strategy approach greatly enhances the adaptability of the grid dispatch strategy to wind power uncertainty. Finally, by selecting the real-time control strategy with the lowest sum of wind power curtailment and load loss as the optimized control strategy, the grid dispatch is optimized from the perspective of resource utilization efficiency, effectively reducing the waste of wind power resources and improving the economy and sustainability of the entire power system operation.

[0053] 3. By comprehensively applying multiple wind power prediction models such as neural network models, support vector machine models and random forest models, the advantages of different models are fully utilized, such as the powerful learning and fitting capabilities of neural network models, the good performance of support vector machine models on small samples and nonlinear problems, and the resistance to overfitting and high-dimensional data processing of random forest models. This multi-model fusion approach avoids the limitations of a single model and can complement and verify each other in the entire power grid optimization and scheduling process, such as obtaining the prediction error probability density function, generating a set of real-time wind power prediction scenarios, and formulating real-time control strategies, thereby improving the overall performance of the entire power grid optimization and scheduling method, enabling the power grid to operate more stably and efficiently when facing complex and changeable wind power conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of a power grid optimization scheduling method based on dynamic selection of a wind power prediction model provided by an embodiment of the present invention;

[0055] Figure 2 This is a simplified diagram of a 500kV grid provided by an embodiment of the present invention;

[0056] Figure 3 Schematic diagram of wind power prediction values ​​(NN, SVM, RF) and actual values ​​of wind farm A provided by an embodiment of the present invention;

[0057] Figure 4 Schematic diagram of the probability density distribution (NN, SVM, RF) of wind power prediction error of wind farm A provided by an embodiment of the present invention;

[0058] Figure 5 Schematic diagram of wind power prediction error time correlation (left) and wind power scenario set (right) provided by an embodiment of the present invention;

[0059] Figure 6This is a schematic diagram of the total output of other power sources (thermal power, hydropower, and DC) based on the prediction results of NN, SVM, and RF provided by an embodiment of the present invention;

[0060] Figure 7 This is a block diagram of a power grid optimization dispatching device module based on dynamic selection of a wind power prediction model provided by an embodiment of the present invention.

[0061] Reference numerals:

[0062] 1. Function construction module, 2. Scenario set generation module, 3. Optimization scheduling module. DETAILED DESCRIPTION

[0063] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present invention.

[0064] Please refer to Figure 1 A first aspect of an embodiment of the present invention provides a method for optimizing power grid scheduling based on dynamic selection of a wind power prediction model, comprising the following steps:

[0065] Step S100 : Based on a plurality of wind power prediction models, a prediction error probability density function of each wind power prediction model is obtained to obtain a plurality of wind power prediction error scenario sets.

[0066] In step S200, real-time meteorological data of the wind turbine is obtained, and power prediction is performed based on multiple wind power prediction models to obtain several real-time wind power prediction values, and several wind power real-time prediction scenario sets are generated in combination with corresponding wind power prediction error scenario sets.

[0067] Step S300: Based on several wind power real-time prediction scenario sets, a random unit combination model is used to generate several real-time control strategies that correspond one to one with the wind power real-time prediction scenario sets, and the real-time control strategy with the lowest sum of wind power curtailment and load loss among the several real-time control strategies is selected as the optimized control strategy for the wind turbine.

[0068] First, the present invention uses historical data from wind turbines to train multiple prediction models, generating corresponding training models and calculating the prediction error probability density function for each model. This results in several wind power prediction error scenario sets. Subsequently, these trained models are used to predict wind power based on current actual operating data, and a set of real-time wind power prediction scenarios is constructed. Finally, by systematically randomizing the real-time prediction scenario sets generated by different prediction models, the optimal decision-making solution is proactively selected with the goal of minimizing the sum of wind power curtailment and load loss, thereby minimizing the sum of wind power curtailment and load loss in the wind power system.

[0069] This approach no longer selects a prediction method based on a prediction error evaluation metric. Instead, it simultaneously transmits multiple prediction results to the grid system's dispatcher, generating corresponding decision-making results. Ultimately, the decision is proactively selected based on minimizing the sum of wind power curtailment and load loss. This dispatch model breaks the mutually exclusive relationship between prediction models and establishes a mutually cooperative and complementary relationship. It also considers that the goal of decision-making in a random environment is to achieve statistically optimal results.

[0070] In a specific implementation of an embodiment of the present invention, the types of real-time meteorological data of a wind turbine include: wind speed, wind direction, temperature and / or air pressure. Wind speed is the most critical factor affecting wind power generation, and its size directly determines the output power of the wind turbine. Wind speed is positively correlated with output power. Wind direction affects the wind energy capture efficiency of the wind turbine, and the yaw system needs to adjust the direction of the wind rotor in time to face the incoming wind. The layout of the wind farm also needs to consider wind direction to reduce the impact of wake. Changes in temperature indirectly affect power generation by changing air density. As the temperature rises, the air density decreases, and the output power decreases, and vice versa. Changes in air pressure cause changes in air density, thereby indirectly affecting power generation. Air pressure is positively correlated with output power.

[0071] Specifically, the step S100 of obtaining the prediction error probability density function of each wind power prediction model based on multiple wind power prediction models includes:

[0072] Step S110: Acquire historical prediction data and historical actual data of the wind turbine generator system.

[0073] Historical forecast data typically comes from previous predictions generated using wind power forecasting models. These models may be based on different algorithms, such as neural networks and support vector machines. Historical actual data refers to the actual power output generated by wind turbines during past operation. This data can be collected through the wind farm's monitoring system, which records the actual power generated by wind turbines at different points in time, as well as relevant information such as prevailing weather conditions. The data collection period must be long enough to cover a wide range of weather conditions and wind power variations. For example, data collection may require months to several years to fully reflect seasonal variations and wind power variations under different weather patterns.

[0074] Data accuracy and completeness are crucial for subsequent steps. Significant deviations from historical forecast data may be due to improper model parameter settings or incorrect input data. Incomplete historical data, such as missing data from critical periods (such as those during extreme weather conditions), can hinder understanding of actual wind power generation trends. Therefore, after data collection, preprocessing, such as data cleaning, is necessary to remove outliers and erroneous data.

[0075] Step S120 : establishing nonlinear mapping relationships between meteorological data and a plurality of wind power prediction models based on historical prediction data.

[0076] Neural network models possess powerful nonlinear fitting capabilities. When establishing mapping relationships, the neural network's neuron structure can automatically adjust weights and thresholds by learning patterns from historical forecast data to find the complex nonlinear relationships between meteorological data (such as wind speed, wind direction, temperature, and air pressure) and wind power forecasts. For example, in a multi-layer perceptron (MLP) neural network, meteorological data is received at the input layer, processed by neurons in the hidden layer, and ultimately, the wind power forecast is obtained at the output layer.

[0077] The support vector machine model works by finding an optimal hyperplane to divide the input space (meteorological data space) and the output space (wind power prediction space). This transforms nonlinear problems into linear problems. Using a kernel function, the meteorological data is mapped into a high-dimensional feature space, establishing a linear relationship within this high-dimensional space. Common kernel functions include the Gaussian kernel function, which effectively handles the nonlinear mapping between meteorological data and wind power.

[0078] The random forest model is a decision tree-based ensemble learning method that builds multiple decision trees to learn the relationship between meteorological data and wind power. Each decision tree splits based on different characteristics of the meteorological data, and the results of multiple decision trees are combined to produce a predicted wind power value. This model is capable of processing high-dimensional meteorological data and is highly resistant to overfitting.

[0079] Establishing this mapping relationship can better reflect the complex impact of meteorological factors on wind power. The relationship between meteorological data and wind power is not a simple linear one. For example, wind speed and wind power may have a nearly quadratic relationship within a certain range, but in other cases, the relationship becomes more complex due to the combined influence of factors such as wind direction and temperature. These nonlinear mapping relationships allow for more accurate consideration of the combined effects of meteorological factors in power forecasting.

[0080] Step S130 : calculating the prediction error of each wind power prediction model based on historical actual data.

[0081] By calculating the prediction error, we can intuitively understand the accuracy of each wind power forecast model. Different models may exhibit different error performance under different meteorological conditions and power output ranges. For example, some models may produce larger errors under conditions of large wind speed fluctuations, while the errors may be smaller under relatively stable wind speeds. Analyzing these errors can help identify the strengths and weaknesses of the models, providing a basis for subsequent improvement and selection.

[0082] Step S140 : Based on the prediction error, a prediction error probability density function of each wind power prediction model is obtained by kernel density estimation.

[0083] Kernel density estimation is a non-parametric estimation method used to estimate the probability density function of a random variable. For the prediction error of the wind power prediction model, a random variable, the kernel density estimation method is to regard each prediction error data point as the center of a probability distribution, and then use a kernel function (such as a Gaussian kernel function) to estimate the probability density around this point. After obtaining the prediction error probability density function, we can fully understand the distribution of the prediction error. For example, the probability density function can be used to determine the probability of the prediction error occurring within a certain interval, which is very important for evaluating the reliability of the wind power prediction model. If the prediction error probability density function of a model has a high peak near zero and the distribution is relatively concentrated, it means that the prediction error of the model is small and relatively stable; conversely, if the probability density function distribution is relatively dispersed, it means that the prediction error of the model fluctuates greatly and the reliability is low.

[0084] Correspondingly, the nonlinear mapping relationship of the neural network model is:

[0085]

[0086] Neural network models use complex neuron structures to construct mapping relationships, enabling them to fit extremely complex nonlinear functions. In wind power forecasting, the relationship between meteorological data (including factors such as wind speed, wind direction, temperature, and air pressure) and wind power is often nonlinear. For example, the impact of wind speed on power is not a simple linear relationship. Its contribution to power varies across different wind speed ranges and is also influenced by the interactions of other meteorological factors. Neural networks can automatically learn these complex relationship patterns using large amounts of training data.

[0087] During training, the neural network continuously adjusts the connection weights between neurons using the error backpropagation algorithm, adaptively optimizing the mapping relationship to better approximate the relationship between real-world weather data and wind power. For example, when new weather data and corresponding wind power data are added to the training set, the neural network can learn these new patterns by updating the weights, thereby improving the accuracy of the prediction.

[0088] Meteorological data is often high-dimensional, and individual features may interact with each other. Neural networks can effectively handle this situation. Using neurons in the hidden layer, they extract and combine features from high-dimensional data, capturing the interactions between different meteorological features and generating more accurate wind power forecasts. For example, different combinations of wind direction and wind speed may have different effects on wind power. Neural networks can learn these complex relationships and reflect them in the mapping relationship.

[0089] Correspondingly, the nonlinear mapping relationship of the support vector machine model is:

[0090]

[0091] K SVM (X i ,X)=exp(-γ||X i -X|| 2 );

[0092] By mapping meteorological data into a high-dimensional feature space using a kernel function, the relationship between meteorological data and wind power, which was originally nonlinear in a low-dimensional space, can be transformed into a linearly separable one in this high-dimensional space. This transformation capability enables support vector machines to effectively handle complex nonlinear relationships. For example, when the relationship between meteorological data and wind power exhibits a complex curve, using an appropriate kernel function (such as a Gaussian kernel function) allows the support vector machine to find a hyperplane in the high-dimensional space to divide the data, thereby achieving accurate predictions.

[0093] The support vector machine model focuses on maximizing the classification margin when searching for the optimal partitioning hyperplane, resulting in excellent generalization performance. In wind power forecasting, it can accurately predict wind power when new meteorological data is input, based on the learned mapping relationships. Even when faced with meteorological conditions that slightly deviate from the training data distribution, the support vector machine, thanks to its generalization capabilities, can provide reasonable predictions, reducing overfitting.

[0094] Support vector machines construct mapping relationships by selecting support vectors (sample points located on the classification boundary). These support vectors play a key role in determining the classification hyperplane. When dealing with wind power forecasting problems, they can focus on the sample points that are most important for dividing the relationship between meteorological data and wind power, thereby more effectively utilizing limited training sample information. This advantage is particularly evident when the sample data is small.

[0095] Correspondingly, the nonlinear mapping relationship of the random forest model is:

[0096]

[0097] The mapping relationship in the random forest model is achieved by integrating multiple decision trees. This reduces the risk of overfitting in a single decision tree and improves the accuracy and stability of predictions. In wind power forecasting, each decision tree can be viewed as a different way of partitioning the relationship between meteorological data and wind power. By combining the results of multiple decision trees, a more robust prediction can be obtained. For example, if a decision tree is biased by abnormal meteorological data, other decision trees can provide a counterbalance and correction.

[0098] Random forests can handle the complex distribution and noise of meteorological data. Because wind power is affected by multiple meteorological factors, its data distribution can be complex and subject to noise such as measurement errors. Random forests adapt to this complex data environment by using random sampling and feature selection across multiple decision trees. Each decision tree is trained on a different subset of data and feature subsets, capturing distinct data patterns and feature relationships, effectively addressing noise in the data and improving the robustness of wind power forecasts.

[0099] Random forest models offer a certain degree of interpretability. During the mapping process, the relative importance of individual meteorological data features for wind power forecasting can be assessed by analyzing the frequency and importance of features within each decision tree. For example, the importance of factors such as wind speed, wind direction, temperature, and air pressure in predicting wind power can be determined. This is crucial for understanding the impact of meteorological factors and further optimizing prediction models.

[0100] Among them, f NN(·),f SVM (·) and f RF (·) are the mapping relationships between wind power prediction values ​​and meteorological data X based on neural network model, support vector machine model and random forest model, respectively. l 、W l and are the activation function, weight vector, and bias vector of the lth layer of the neural network, and l=1,2,3,...,L, L is the number of layers of the neural network, W and b NN are the weight matrix and bias matrix of the neural network respectively, N S is the number of support vectors in the support vector machine model, α i is the Lagrange multiplier of the i-th support vector of the support vector machine model, y i is the true label of the i-th support vector of the support vector machine model, K SVM (X i ,X) is the kernel function of the support vector machine model, γ is a positive parameter, X i is the X corresponding to the i-th sample point, b SVM is the bias term of the support vector machine model, N RF is the number of decision trees in the random forest model, is the predicted value of the i-th decision tree for input X, p wind,NN 、p wind,SVM and p wind,RF These are the wind power prediction values ​​obtained based on the neural network model, support vector machine model and random forest model respectively.

[0101] Furthermore, the corresponding prediction error probability density function of the neural network model is:

[0102]

[0103] The corresponding prediction error probability density function of the support vector machine model is:

[0104]

[0105] The corresponding prediction error probability density function of the random forest model is:

[0106]

[0107] Among them, K(·) is the kernel density estimation function, h is the bandwidth, N S is the number of samples, Δp wind is the wind power prediction error random variable, X i and are the meteorological factor input vector and historical actual wind power of the i-th historical data sample point, and are the wind power prediction errors of the i-th historical data sample point of the neural network model, support vector machine model and random forest model, and These are the prediction error probability density functions of the neural network model, support vector machine model, and random forest model, respectively.

[0108] Specifically, the step S200 of combining the corresponding wind power prediction error scenario sets to generate wind power real-time prediction scenario sets includes:

[0109] Step S201: Based on the prediction error probability density function, Latin hypercube sampling and Cholesky decomposition are used to generate a corresponding wind power prediction error scenario set:

[0110]

[0111] Among them, Ω NN,err ,Ω SVM,err and Ω RF,err are wind power prediction error scenario sets for neural network model, support vector machine model and random forest model, (Δp wind,NN ) i 、(Δp wind,SVM ) i and (Δp wind,RF ) i The i-th wind power prediction error scenario generated by the neural network model, support vector machine model and random forest model, i = 1, 2, 3, ..., N sg , N sg is the number of scenes.

[0112] Latin hypercube sampling is a stratified sampling technique. In the context of wind power forecasting scenarios, the range of each input variable (here, the wind power forecast value and its forecast error-related variables) is divided into non-overlapping intervals of equal probability. For example, if the wind power forecast error ranges from -100 kW to 100 kW, assuming 10 intervals, each interval is 20 kW wide, and each interval has an equal probability of being sampled. A sample point is then randomly drawn from each interval. Unlike simple random sampling, Latin hypercube sampling ensures a more even distribution of samples across the entire range, better covering the full range of possible variable values. In wind power forecasting, this evenly distributed sample can more comprehensively represent all possible combinations of wind power forecast values ​​and errors. These sample points can serve as the basis for constructing wind power forecast scenarios. For example, by extracting different wind power forecast values ​​and corresponding forecast error sample points, multiple different possible wind power states can be generated, providing a diverse data source for constructing scenario sets.

[0113] These sample points are transformed using the lower triangular matrix obtained through Cholesky decomposition. This process injects correlation information into the sample points, resulting in a set of scenarios that not only encompasses the various possible values ​​of the real-time wind power forecast and forecast error, but also considers the relationship between the two. The resulting real-time wind power forecast scenario set is a collection of multiple scenarios, each representing a possible real-time change in wind power. These scenarios can be used to simulate the impact of wind power on power system operation under different circumstances, such as providing a basis for developing different real-time control strategies in random unit combination models.

[0114] Step S202: obtaining a wind power real-time prediction scenario set based on the wind power real-time prediction value and the wind power prediction error scenario set:

[0115]

[0116] Among them, Ω NN ,Ω SVM and Ω RF The wind power real-time prediction scenario sets of neural network model, support vector machine model and random forest model are respectively, wind ,NN ,pre , and p wind ,RF ,pre The real-time wind power prediction values ​​of the neural network model, support vector machine model and random forest model are respectively (p wind,NN ) i 、(p wind,SVM ) iand (p wind,RF ) i The i-th wind power prediction scenario generated by the neural network model, support vector machine model and random forest model, i = 1, 2, 3, ..., N sg , N sg is the number of scenes.

[0117] In addition, when the wind power prediction model includes: neural network model, support vector machine model and random forest model, the corresponding random unit combination model is:

[0118]

[0119] Among them, f(p wind,output ,x) is the objective function of the random unit combination model, z(p wind,output ,x) is the objective function value, g(p wind,output ,x)≤0 is the constraint condition of the random unit model, including wind power output scenario constraint, power balance constraint and flow constraint, Represents the prediction scenario set of neural network model, support vector machine model and random forest model, p wind,output is the control strategy, and x is the output of other power generation equipment in the power system and the node power angle variable.

[0120] The control strategy is a key variable in optimization. By adjusting the control strategy, we can determine which units should be started and shut down, as well as their respective power generation capacities. Taking into account the uncertainty of wind power, the control strategy needs to be flexibly adjusted based on different wind power forecast scenarios to achieve the optimal objective function. For example, when the wind power forecast scenario indicates high wind power, the output of other power generation equipment can be appropriately reduced to increase wind power absorption.

[0121] In addition, wind power output scenario constraints are key constraints to consider the uncertainty of wind power. Since wind power will vary according to factors such as different meteorological conditions, Ω NN |Ω SVM |Ω RFThese forecast scenario sets contain different wind power scenarios. Wind power output scenario constraints require that, in each scenario, wind power output conform to actual physical laws and the range of the forecast model. For example, wind power output cannot be negative, and its output must remain within a reasonable range under varying wind speed and other meteorological conditions. Power balance constraints are fundamental constraints on power system operation. In the stochastic unit commitment model, since multiple wind power forecast scenarios are considered, power balance constraints must be met in every scenario. This requires that when formulating unit commitment and control strategies, the output of other generating equipment should be appropriately arranged based on the variations in wind power under different scenarios. Power flow constraints primarily consider the physical characteristics of the power grid, such as line capacity limitations and node voltage limits. In the stochastic unit commitment model, power flow constraints ensure safe and stable operation of the power grid under different wind power scenarios. For example, when wind power suddenly increases or decreases, the power flow distribution in the power grid changes. Reasonable unit commitment and control strategies are required to ensure that the power flow remains within the permitted range and avoid problems such as line overload.

[0122] Furthermore, the selection formula of the real-time control strategy is:

[0123]

[0124] Among them, (p wind,output,NN ,x NN )、(p wind,output,SVM ,x SVM ) and (p wind,output,r ,x SVM ) are the real-time control strategies corresponding to the wind power real-time prediction scenario set obtained by power prediction based on the neural network model, support vector machine model and random forest model, and B is the final optimized control strategy. (p wind,output,r ,x r ) is the one with the lowest sum of wind power curtailment and load loss among the real-time control strategies corresponding to the wind power real-time prediction scenario set obtained by prediction based on the neural network model, support vector machine model and random forest model.

[0125] The following is a practical application scenario to further illustrate the specific process of the present invention:

[0126] Based on a DC receiving-end power grid, the 500kV grid diagram of the provincial grid is as follows: Figure 2 As shown in the figure, the power generation units participating in the optimized dispatch of the provincial grid include 37 thermal power units, 10 cascade hydropower stations, and 12 wind farms. The grid structure studied is 220kV and above, and after equivalent simplification, it contains a total of 489 nodes.

[0127] The following is an example of a certain day: The prediction results of wind farm A on a certain day are as follows: Figure 3, where NN, SVM, and RF represent neural network, support vector machine, and random forest, respectively.

[0128] Figure 4 The prediction error probability distribution of wind farm A under various prediction methods is shown. It can be seen that the prediction error distribution of each prediction method presents different "skewness", and the gap between the distributions is large, indicating that different prediction models with the same prediction level have significantly different "overestimation" or "underestimation" prediction "preferences".

[0129] Based on the prediction results obtained by the above three wind power prediction methods, the time correlation between the prediction errors is further analyzed, and based on Latin hypercube sampling and Cholesky decomposition, the wind power prediction error scenario sets of each prediction model are generated respectively in combination with the corresponding prediction error probability density function. Based on the wind power real-time prediction value of each prediction model and the corresponding wind power prediction error scenario set, the wind power real-time prediction scenario set of each prediction model is obtained, such as Figure 5 shown. Figure 5 It can be seen that the temporal correlation of the prediction errors for each prediction method shows a decreasing trend as the time interval increases; locally, there are slight differences in the correlation of the prediction errors produced by different prediction methods. Based on the predicted wind power values ​​and wind power prediction errors, a set of wind power scenarios can be generated. Because the scenario set is closely related to the probability distribution of the wind power prediction values ​​and prediction errors, there are obvious differences in the wind power scenarios obtained by different prediction methods. However, overall, the scenario set obtained based on any prediction method can basically cover the fluctuation range of the actual wind power value.

[0130] In the stochastic optimization scheduling model considering the uncertainty of wind power, only wind power output has uncertainty. Other power sources including thermal power, hydropower, DC, etc. do not consider uncertainty and can all be output as planned. Figure 6 The total output of other power sources, as predicted by the NN, SVM, and RF prediction methods, is shown. The resulting dispatch plans differ significantly from one another. At 2:00 PM, the difference in total output between the SVM-based and RF-based predictions reached 621.37 MW, the largest difference between the three predictions.

[0131] Based on the 100-day scheduling plan and actual observations, the following "actively" selects the prediction model with the lowest expected system operating cost among the three prediction models of neural network, support vector machine, and random forest for scheduling (dynamic scheduling mode).

[0132] Table 1 shows the number of times each prediction method was selected. It can be seen that the predictions and decisions based on the SVM were selected the most frequently, being selected on 51 days within the 100-day calculation period. This means that the decisions based on the SVM prediction method resulted in the lowest total system curtailment and load loss during these 51 days.

[0133] Table 1 The number of times each forecast method was selected (100 days)

[0134]

[0135] Accordingly, please refer to Figure 7 A second aspect of an embodiment of the present invention provides a power grid optimization dispatching system based on dynamic selection of a wind power prediction model, comprising:

[0136] Function construction module 1, which is used to obtain the prediction error probability density function of each wind power prediction model based on multiple wind power prediction models, and obtain several wind power prediction error scenario sets;

[0137] Scenario set generation module 2 is used to obtain real-time meteorological data of wind turbines, perform power prediction based on multiple wind power prediction models, obtain several real-time wind power prediction values, and generate several real-time wind power prediction scenario sets based on several corresponding wind power prediction error scenario sets;

[0138] The optimization scheduling module 3 is used to generate a number of real-time control strategies corresponding to the wind power real-time prediction scenario sets based on a number of wind power real-time prediction scenario sets through a random unit combination model, and select the real-time control strategy with the lowest sum of wind power abandonment and load loss among the several real-time control strategies as the optimized control strategy for the wind turbine set.

[0139] Accordingly, a third aspect of an embodiment of the present invention provides an electronic device comprising: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the above-mentioned grid optimization scheduling method based on dynamic selection of a wind power prediction model.

[0140] Accordingly, a fourth aspect of an embodiment of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-mentioned power grid optimization scheduling method based on dynamic selection of a wind power prediction model.

[0141] The embodiment of the present invention aims to protect a method and device for optimizing and dispatching a power grid based on the dynamic selection of a wind power prediction model, comprising the following steps: based on multiple wind power prediction models, obtaining the prediction error probability density function of each wind power prediction model, and obtaining several wind power prediction error scenario sets; obtaining real-time meteorological data of the wind turbine, and performing power prediction based on multiple wind power prediction models to obtain several wind power real-time prediction values, and combining the corresponding wind power prediction error scenario sets to generate several wind power real-time prediction scenario sets; based on several wind power real-time prediction scenario sets, generating several real-time control strategies corresponding to the wind power real-time prediction scenario sets through a random unit combination model, and selecting the real-time control strategy with the lowest sum of wind power abandonment and load loss among the several real-time control strategies as the optimized control strategy for the wind turbine. The above technical solution has the following effects:

[0142] 1. By utilizing historical forecast data and actual historical data of wind turbines, nonlinear mapping relationships are established between various wind power forecast models (neural network model, support vector machine model, and random forest model) and meteorological data, and the forecast error is calculated to obtain the forecast error probability density function. This allows for accurate grasp of the forecast characteristics and error distribution of each model. When combined with real-time meteorological data for power forecasting and the generation of a real-time forecast scenario set, the real-time wind power forecast scenario set generated through Latin hypercube sampling and Cholesky decomposition, based on these accurate model characteristics and error analysis, is more in line with actual conditions and comprehensively covers the possible range of wind power variations. This provides reliable and comprehensive basic data for subsequent optimized scheduling, effectively improving the accuracy and rationality of wind power forecasting and scenario simulation.

[0143] 2. Based on multiple sets of real-time wind power prediction scenarios, a random unit combination model is used to generate multiple corresponding real-time control strategies. This random unit combination model comprehensively considers multiple key factors such as wind power output scenario constraints, power balance constraints, and flow constraints, ensuring that the generated control strategies can meet the basic requirements of grid operation under different wind power scenarios. In addition, since corresponding control strategies are generated for different prediction model scenario sets, this multi-strategy approach greatly enhances the adaptability of the grid dispatch strategy to wind power uncertainty. Finally, by selecting the real-time control strategy with the lowest sum of wind power curtailment and load loss as the optimized control strategy, the grid dispatch is optimized from the perspective of resource utilization efficiency, effectively reducing the waste of wind power resources and improving the economy and sustainability of the entire power system operation.

[0144] 3. By comprehensively applying multiple wind power prediction models such as neural network models, support vector machine models and random forest models, the advantages of different models are fully utilized, such as the powerful learning and fitting capabilities of neural network models, the good performance of support vector machine models on small samples and nonlinear problems, and the resistance to overfitting and high-dimensional data processing of random forest models. This multi-model fusion approach avoids the limitations of a single model and can complement and verify each other in the entire power grid optimization and scheduling process, such as obtaining the prediction error probability density function, generating a set of real-time wind power prediction scenarios, and formulating real-time control strategies, thereby improving the overall performance of the entire power grid optimization and scheduling method, enabling the power grid to operate more stably and efficiently when facing complex and changeable wind power conditions.

[0145] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0146] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0147] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0148] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A power grid optimization scheduling method based on dynamic selection of wind power prediction model, characterized in that: The steps include: Based on multiple wind power prediction models, obtaining a prediction error probability density function of each wind power prediction model to obtain several wind power prediction error scenario sets; Acquire real-time meteorological data of the wind turbine generator set, perform power forecasts based on the multiple wind power forecast models, obtain a number of wind power real-time forecast values, and generate a number of wind power real-time forecast scenario sets based on the corresponding wind power forecast error scenario sets; Based on the plurality of wind power real-time prediction scenario sets, generating a plurality of real-time control strategies corresponding to the wind power real-time prediction scenario sets through a random unit combination model, and selecting the real-time control strategy with the lowest sum of wind power curtailment and load loss among the plurality of real-time control strategies as the optimized control strategy for the wind turbine set; The wind power prediction model includes: a neural network model, a support vector machine model and a random forest model; The corresponding nonlinear mapping relationship of the neural network model is: ； ； The corresponding nonlinear mapping relationship of the support vector machine model is: ； ； The nonlinear mapping relationship corresponding to the random forest model is: ； Among them, f NN (·),f SVM (·) and f RF (·) are the mapping relationships between wind power prediction values ​​and meteorological data X based on neural network model, support vector machine model and random forest model, respectively. 、 and are the activation function, weight vector and bias vector of the neural network layer l, respectively, and , is the number of layers of the neural network, and are the weight matrix and bias matrix of the neural network respectively, is the number of support vectors of the support vector machine model, is the Lagrange multiplier of the i-th support vector of the support vector machine model, is the true label of the i-th support vector of the support vector machine model, is the kernel function of the support vector machine model, is a positive parameter, is the X corresponding to the i-th sample point, is the bias term of the support vector machine model, is the number of decision trees in the random forest model, is the predicted value of the i-th decision tree for input X, 、 and These are the wind power prediction values ​​obtained based on the neural network model, support vector machine model, and random forest model respectively; The corresponding prediction error probability density function of the neural network model is: ； ； The corresponding prediction error probability density function of the support vector machine model is: ； ； The corresponding prediction error probability density function of the random forest model is: ； ； Where K(·) is the kernel density estimation function, h is the bandwidth, is the number of samples, is the wind power prediction error random variable, and are the meteorological factor input vector and historical actual wind power of the i-th historical data sample point, 、 and are the wind power prediction errors of the i-th historical data sample point of the neural network model, support vector machine model and random forest model, 、 and These are the prediction error probability density functions of the neural network model, support vector machine model, and random forest model, respectively.

2. The power grid optimization scheduling method based on dynamic selection of wind power prediction model according to claim 1 is characterized in that: The types of real-time meteorological data of the wind turbine generator system include: wind speed, wind direction, temperature and / or air pressure.

3. The power grid optimization scheduling method based on dynamic selection of wind power prediction model according to claim 1 is characterized in that: The obtaining, based on the plurality of wind power prediction models, a prediction error probability density function of each wind power prediction model comprises: Obtain historical forecast data and historical actual data of wind turbines; Based on the historical forecast data, establishing nonlinear mapping relationships between meteorological data and multiple wind power prediction models; Calculating a prediction error of each wind power prediction model based on the historical actual data; Based on the prediction error, a prediction error probability density function of each wind power prediction model is obtained respectively by kernel density estimation.

4. The power grid optimization scheduling method based on dynamic selection of wind power prediction model according to any one of claims 1 to 3, characterized in that: The generating of wind power real-time prediction scenario sets respectively by combining the corresponding plurality of wind power prediction error scenario sets comprises: Based on the prediction error probability density function, the corresponding wind power prediction error scenario set is generated by Latin hypercube sampling and Cholesky decomposition: ； ； ； in, 、 and are the wind power prediction error scenario sets of the neural network model, support vector machine model and random forest model, 、 and are the i-th wind power prediction error scenarios generated by the neural network model, support vector machine model, and random forest model, respectively. , is the number of scenes; Based on the wind power real-time prediction value and the wind power prediction error scenario set, a wind power real-time prediction scenario set is obtained: ； ； ； in, 、 and These are the wind power real-time prediction scenario sets of the neural network model, support vector machine model and random forest model, 、 and are the real-time wind power prediction values ​​of the neural network model, support vector machine model and random forest model, 、 and are the i-th wind power prediction scenarios generated by the neural network model, support vector machine model and random forest model respectively, , is the number of scenes.

5. The power grid optimization scheduling method based on dynamic selection of wind power prediction model according to any one of claims 1 to 3, characterized in that: The wind power prediction model includes: a neural network model, a support vector machine model and a random forest model; The random unit combination model is: ； ； in, is the objective function of the random unit combination model, is the objective function value, are the constraints of the random unit model, including wind power output scenario constraints, power balance constraints and power flow constraints. Represents the prediction scenario set of the neural network model, support vector machine model and random forest model, For control strategy, is the output of other power generation equipment in the power system and the node power angle variable.

6. The power grid optimization scheduling method based on dynamic selection of wind power prediction model according to claim 5 is characterized in that: The selection formula of the real-time control strategy is: ； in, 、 and The real-time control strategies corresponding to the wind power real-time prediction scenario set obtained by power prediction based on neural network model, support vector machine model and random forest model are respectively. is the final optimized control strategy, The one with the lowest sum of wind power curtailment and load loss among the real-time control strategies corresponding to the wind power real-time prediction scenario set obtained by prediction based on the neural network model, support vector machine model and random forest model.

7. A power grid optimization dispatching system based on dynamic selection of wind power prediction model, characterized in that: The method for optimizing the power grid scheduling based on the dynamic selection of the wind power prediction model according to any one of claims 1 to 6 is used to optimize the power grid scheduling, including: A function construction module, which is used to obtain a prediction error probability density function of each wind power prediction model based on multiple wind power prediction models, and obtain a plurality of wind power prediction error scenario sets; a scenario set generation module, which is used to obtain real-time meteorological data of the wind turbine, perform power prediction based on the multiple wind power prediction models, obtain a number of wind power real-time prediction values, and generate a number of wind power real-time prediction scenario sets in combination with the corresponding wind power prediction error scenario sets; An optimization scheduling module is used to generate a number of real-time control strategies corresponding to the wind power real-time prediction scenario sets based on the several wind power real-time prediction scenario sets through a random unit combination model, and select the real-time control strategy with the lowest sum of wind power curtailment and load loss among the several real-time control strategies as the optimized control strategy for the wind turbine set.

8. An electronic device, characterized in that: include: at least one processor; And a memory connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the power grid optimization scheduling method based on dynamic selection of wind power prediction model as described in any one of claims 1-6.

Images