Joint uncertainty assessment method for wind power integration into power system based on time-varying fluctuation

Abstract

The application provides a wind power access power system uncertainty joint evaluation method based on time-varying fluctuation, belongs to the technical field of power system uncertainty evaluation and dispatching auxiliary decision-making, and solves the technical problem that the existing method causes the low coverage and unreasonable interval width of the interval prediction result of the combined output of wind and light due to the rough residual fitting and the insufficient correlation description. The application firstly obtains the edge distribution based on short-term prediction and error fluctuation modeling, then constructs a joint distribution by using R vine Copula, outputs the prediction interval of each station and value under the given confidence, and gives the coverage and interval width indexes, so that the day-ahead or day-in dispatching can be connected, and rolling update and online deployment are realized. The application significantly reduces the interval bandwidth while ensuring the coverage, improves the robustness of the fast changing period, and enhances the safety and economy of new energy consumption.

Description

Technical Field

[0001] This invention relates to the field of power system uncertainty assessment and dispatch auxiliary decision-making technology, and in particular to a joint assessment method for the uncertainty of wind power integration into the power system based on time-varying fluctuations. Background Technology

[0002] Wind and solar power, as core energy sources in new power systems, have seen rapid and continuous growth in installed capacity and power generation, becoming key forces in energy structure transformation. However, wind and solar power generation is constrained by natural conditions, exhibiting significant volatility and randomness. With a high proportion of renewable energy connected to the grid, the uncertainty of its output severely impacts the power balance and frequency stability of the power system, seriously threatening the safe operation of the system and becoming a bottleneck restricting the large-scale development of new energy.

[0003] Against this backdrop, accurately assessing the uncertainty of combined wind and solar power output is a crucial prerequisite for mitigating its impact on system frequency. In particular, wind and solar projects are often distributed in regional clusters, making the overall regional output characteristics more valuable for dispatching than those of individual power stations. Existing assessment methods have significant limitations; traditional correlation coefficient analysis methods struggle to capture local correlation characteristics, while Copula functions lack applicability and accuracy in high-dimensional scenarios. Therefore, developing more accurate and effective high-dimensional uncertainty assessment methods will not only provide technical support for renewable energy consumption but also lay a key foundation for optimizing power system control strategies, possessing significant theoretical and engineering value. Summary of the Invention

[0004] The purpose of this invention is to provide a joint assessment method for uncertainties of wind power integration into the power system based on time-varying fluctuations. By constructing a high-precision unit uncertainty model and a multidimensional correlation model, the method can accurately assess the uncertainty of joint wind power output, provide a reliable basis for power system optimization scheduling and stability control, and effectively mitigate the impact of large-scale wind power integration on the safe operation of the system.

[0005] To better achieve the above-mentioned objectives, this invention is implemented through the following technical solution: a joint assessment method for uncertainties in wind power integration into the power system based on time-varying fluctuations, comprising the following steps:

[0006] S1. For each wind farm and photovoltaic power station unit, based on the historical power output sequence of wind and solar power, the point prediction value is obtained through the Long Short-Term Memory (LSTM) network, and the prediction residual is calculated.

[0007] S2. Input the predicted residuals into the generalized autoregressive conditional heteroscedasticity-difference model GARCH-D to form a long short-term memory-generalized autoregressive conditional heteroscedasticity-difference joint model LSTM-GARCH-D, and obtain the marginal probability distribution function of each unit; perform probability integral transformation on the predicted residuals according to the marginal probability distribution function to obtain the marginal variables located in the interval [0,1], and obtain the inverse normal distribution function of each set of functions;

[0008] S3. The R-vine Copula function model is determined by using the maximum spanning tree MST, and a multidimensional correlation model is constructed to characterize the correlation between the errors of the unit output prediction values.

[0009] S4. Combine the marginal probability distribution functions of each unit obtained in step S2 with the R-vine Copula function model constructed in step S3 to obtain the joint probability density and joint distribution, and output the prediction interval of wind and solar power output of each unit and total power output under a given confidence level.

[0010] Preferably, in step S1, the point prediction values ​​obtained by using a Long Short-Term Memory (LSTM) network based on the historical power output sequence of wind and solar power include the following steps:

[0011] S1-1, Using a one-dimensional vector This represents historical output data. , The step size of the data output for the trained landscape. These represent the wind and solar power output data at each moment;

[0012] S1-2, Transform a one-dimensional vector Divided into training set With test set ,in , , , The length of the training set;

[0013] S1-3, The calculation method of the Long Short-Term Memory (LSTM) network is as follows:

[0014]

[0015] In the formula, for Input values ​​for historical output sequence of moments and scenery. for Output value at all times for The status of the gating unit at all times. , , The results of the state operations for the forget gate, input gate, and output gate. , , These are the bias terms for the forget gate, input gate, and output gate. , , These are the weight matrices for the forget gate, input gate, and output gate, respectively. Let t be the input state of the gate control unit. The weight matrix is ​​the input state of the gating unit. Let δ be its bias term, and let δ be the sigmoid function, ranging from [0,1]. For activation function, This indicates that the elements in the vector are multiplied element by element;

[0016] S1-4, The output of the power set J after passing through the hidden layer is represented as follows: , For the first Predictive data for the output of the wind and light at any given moment.

[0017] Preferably, the Long Short-Term Memory-Generalized Autoregressive Conditional Heteroskedasticity-Difference Joint Model (LSTM-GARCH-D) described in step S2 is calculated according to the following formula:

[0018]

[0019] In the formula, J represents the wind and solar power output data at each moment. t This represents the predicted data at time t. This represents the residual of the predicted value at time t. The variance of the predicted values ​​at time t is represented by q, p, and m, which are the orders of the generalized autoregressive conditional heteroscedasticity model GARCH, respectively; A, B i C j D k+1 All are parameters obtained from the fitting; m is the order of the difference correction term for the predicted value; Let t be the predicted difference value of the wind and solar power output points.

[0020] Preferably, in step S3, the R-vine Copula function model is determined using the Maximum Spanning Tree (MST), which includes the following steps:

[0021] S3-1 Calculate the N-dimensional historical power output sequence of wind and solar power. The Kendall rank correlation coefficients between the sequences are used as edge weights;

[0022] S3-2. Determine the candidate set of Pair-Copula functions; the candidate set of Pair-Copula functions includes one or more of the following: Frank-Copula, Clayton-Copula, Gaussian Copula, t-Copula, and Gumbel-Copula.

[0023] S3-3. Use the pair-copula candidate function to fit each group of data sequentially, and calculate the corresponding Akaike Information Criterion (AIC) index. The Akaike Information Criterion (AIC) index is calculated according to the following formula:

[0024]

[0025] In the formula, I AIC For the Akaike Information Criterion (AIC) index, m p L represents the number of parameters in the pair-copula model, and L is the maximum likelihood function value.

[0026] S3-4. For each combination, select the Pair-Copula function with the smallest corresponding Akaike Information Criterion (AIC) index as the connection structure of the R-Fuji Copula function in that layer.

[0027] S3-5. Repeat S3-2 to S3-4 to obtain the N-dimensional R-vine Copula function model, and calculate the N-dimensional wind-sun joint probability density function based on the vine Copula function of the N-dimensional R-vine Copula function model.

[0028] Preferably, in steps S3-5, the N-dimensional wind-sun joint probability density function is calculated using the vine Copula function based on the N-dimensional R-vine Copula function model, and the calculation formula is as follows:

[0029]

[0030] In the formula, Here, a and b are the conditional values, representing the sequence numbers of the historical power output of the wind and solar power system. A set representing the historical output sequence of N-dimensional landscape; This represents the joint probability density function of the candidate set of the pair-copula function.

[0031] Preferably, in step S4, the predicted ranges for the wind and solar power output of each unit and the total power output are output. Assuming there are N wind and solar power output units, the power output at a certain time t is respectively... , , , ... Output data and values The historical power output data of the wind and solar power generation units are represented by a set. The historical power output data of wind and solar power generation unit No. 1 is: The steps for predicting the power output range of wind and solar power at time t+1 are as follows:

[0032] S4-1. Use the LSTM algorithm to perform point prediction on N+1 sets of data, including the wind and solar power output units and their sums at time t+1. Then the predicted data of the wind and solar power output units at time t+1 are as follows: , , , ... and ;

[0033] S4-2. For the predicted wind and solar power output data at time t+1 obtained in step S4-1, the residual probability distribution is modeled using the generalized autoregressive conditional heteroscedasticity-difference model GARCH-D. The input data for this step is... , , , ... , and the point prediction data obtained in step S4-1 , , , ... , The probability distribution functions of each wind and solar power output unit and its sum at time t+1 are obtained, and the inverse normal distribution function of each set of functions is calculated as the input function of the R-copula function model in step S3, denoted as follows: , , , ... , ;

[0034] S4-3, will , , , ... , Transform into a probability distribution sequence , , , ... , Following the fitting method of the R-vine Copula function model in step S3, the N+1-dimensional probability density function is obtained by fitting layer by layer. ,Will Substituting these known variables into the N-dimensional wind-sun joint probability density function yields the function. , which is the probability density function of the combined wind and solar residuals at time t+1.

[0035] Furthermore, to better achieve the aforementioned objectives, this invention also proposes a wind power system for joint assessment of uncertainties in wind power integration into the power system based on time-varying fluctuations. The system includes:

[0036] The point prediction module is used to obtain point prediction values ​​for each wind farm and photovoltaic power station unit based on the historical power output sequence of wind and solar power, and to calculate the prediction residuals.

[0037] The edge modeling module is used to input the prediction residuals into GARCH-D to form LSTM-GARCH-D and obtain the edge probability distribution function of each unit; according to the edge probability distribution function, the prediction residuals are subjected to probability integral transformation to obtain the edge variables located in the interval [0,1], and the inverse normal distribution function of each set of functions is obtained.

[0038] The vine structure construction module is used to determine the R-vine Copula function model using the maximum spanning tree MST, construct a multidimensional correlation model, and characterize the correlation between the errors of the unit output prediction values.

[0039] The joint evaluation and output module is used to combine the edge distribution and the vine structure to obtain the joint probability density and joint distribution, and output the prediction range of the output of each unit and the total output at a given confidence level.

[0040] Furthermore, in order to better achieve the above-mentioned objectives, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed, it implements the steps of the method of the present invention.

[0041] Finally, in order to better achieve the above-mentioned objectives, the present invention proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] (1) This invention uses the Long Short-Term Memory-Generalized Autoregressive Conditional Heteroscedasticity-Difference Joint Model LSTM-GARCH-D model for residual fitting. The Long Short-Term Memory Network LSTM model can capture the long and short-term time-series dependence of the residuals, and the GARCH-D model accurately describes the time-varying fluctuation characteristics through the difference correction term. The synergistic effect of the two significantly improves the accuracy of the single-unit uncertainty model, reduces the cumulative error in high-dimensional joint evaluation from the source, and provides a high-quality data foundation for correlation analysis.

[0044] (2) This invention introduces the R-vine Copula function for correlation analysis. By combining the hierarchical vine structure with the Pair-Copula function, it breaks through the limitations of traditional methods in characterizing multidimensional correlation. It can capture the overall correlation between wind and light units and accurately reflect the local specific correlation, while maintaining high modeling accuracy in high-dimensional scenarios.

[0045] (3) This invention constructs a collaborative framework of “accurate residual fitting and fine correlation characterization”. The LSTM-GARCH-D model provides a reliable single-unit uncertainty boundary, and the R-Teng Copula function realizes the reasonable aggregation of unit uncertainty, which significantly improves the coverage and accuracy of the joint output interval prediction, and provides a more accurate decision basis for the power system to formulate new energy consumption strategies and stability control schemes. Attached Figure Description

[0046] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0047] Figure 1 This is a structural block diagram of the method of the present invention.

[0048] Figure 2 This is a diagram showing the power output data of the wind and solar turbine units according to an embodiment of the present invention.

[0049] Figure 3 This is a graph showing the prediction results of combined wind and solar power output based on LSTM in an embodiment of the present invention.

[0050] Figure 4 This is a graph showing the error results of the LSTM-based joint wind power prediction in an embodiment of the present invention.

[0051] Figure 5 This is a wind-solar joint uncertainty assessment diagram based on R-vine Copula correlation analysis according to an embodiment of the present invention.

[0052] Figure 6 This is a graph illustrating the uncertainty assessment of wind-solar joint operation based on the GARCH model, according to an embodiment of the present invention.

[0053] Figure 7This is a diagram showing the interval coverage index according to an embodiment of the present invention.

[0054] Figure 8 This is a diagram showing the interval bandwidth index of an embodiment of the present invention.

[0055] Figure 9 This is a comprehensive evaluation index chart of the interval bandwidth index in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0057] Example 1: This example demonstrates a joint assessment method for the uncertainty of wind power integration into the power system based on time-varying fluctuations. It applies six sets of output data from three photovoltaic power plants and three wind farms in a certain region in January 2023. The wind and solar turbine output data are as follows: Figure 2 As shown. Reference Figure 1 As shown, the specific steps in this embodiment include:

[0058] (1) Uncertainty assessment based on LSTM-GARCH-D residual fitting:

[0059] ① First, the power output of the wind and solar units is predicted using an LSTM model:

[0060] The LSTM algorithm is trained using hidden layers composed of numerous cascaded cell structures; the number of cascaded cell structures corresponds to the number of hidden layers. Historical wind and solar power output data is represented by a one-dimensional vector. express, , The step length of the wind power output data during training. These represent the wind and solar power output data at each moment, and these data will be divided into training sets. With test set , , , , This represents the length of the training set. Assume that in... At that moment, the input to the LSTM cell structure is: Historical output sequence of moment and scenery, input value , Output value at time Gating unit status The output of the LSTM cell structure is: the output value at time t. Time-based gating unit status The final output of the LSTM cell structure is determined by the output gate state and the gate control unit state, and the calculation method is as follows:

[0061]

[0062] In the formula, , , The results of the state operations for the forget gate, input gate, and output gate; , , These are the bias terms for the three types of gates; , , These are the weight matrices for the three types of gates; Let t be the input state of the gate control unit; δ is the weight matrix for the input state of the gated unit; δ is the sigmoid function, ranging from [0,1]. Its bias term; For activation functions; This indicates that the elements in the vector are multiplied bitwise. , , , , , These are common parameters for all LSTM cell structures in the same hidden layer, representing the logic layer in a neural network that requires extensive training. When training the hidden layer, the initial gate state parameters, the number of hidden layers, and the training set are first set. and test set Then, the output value of each neuron is calculated forward, which is... , , , , The value of is calculated, then the error of the output value of each neuron is calculated in reverse, and finally the weight matrix coefficients and bias terms are calculated based on the error. By increasing the number of iterations, the error of the output value is reduced, and the state parameters of each operational gate in the cell structure are obtained. The output of the output set J after passing through the hidden layer can be expressed as:

[0063]

[0064] In the formula, For the first Predictive data for the output of the wind and light at any given moment.

[0065] ② Based on the prediction of wind and solar power output, the probability distribution function of future wind and solar power output data is obtained using the GARCH-D uncertainty modeling method. Traditional GARCH models are often used for volatility analysis of time series data in fields such as economics, and their expression is as follows:

[0066]

[0067] In the formula This represents the residual of the predicted value at time t. It is the variance of the predicted value at time t. , Let A and B be the order of the GARCH model. i C j All parameters are fitted, and the residuals of the predicted values ​​at this time follow a normal distribution. However, in the traditional GARCH model, the variance of the predicted value at the current time step is only related to the variance and residuals of the predicted values ​​at one or more previous time steps, but ignores the changing trend of the wind and solar power output prediction sequence calculated in the previous section, which may cause the residuals of the predicted values ​​fitted by the GARCH model to deviate from the actual situation. To this end, this invention introduces the difference sequence of wind and solar power output point predictions as a correction term into the traditional GARCH model, and then uses the modified GARCH-D (GARCH-Difference) model to fit the residuals, the expression of which is:

[0068]

[0069] In the formula, Let D be the predicted difference value of the wind and solar power output point at time t. k+1 Let be the parameters to be fitted, and m be the order of the difference correction term for the predicted values. The formula introduces a difference correction term to reflect the changing trend of the wind and solar power output prediction sequence, thereby improving the fitting accuracy of the residuals. The corrected GARCH-D model can then be combined with the LSTM model to form the LSTM-GARCH-D model, thus constructing the probability distribution function.

[0070]

[0071] In the formula, Let t be the predicted scenic spot value based on the LSTM model. The residual term is the input sequence of the R-copula function.

[0072] (2) Uncertainty assessment of wind-solar joint operation based on R-vine Copula correlation analysis

[0073] ①Vine Copula

[0074] The output characteristics of power generation units, primarily wind and solar power, exhibit a degree of randomness and correlation. The uncertainty caused by the combined output prediction error is influenced by various factors, resulting in time-varying characteristics and a certain degree of negative correlation (i.e., complementarity). When wind and solar power are simultaneously connected to the grid, the prediction error of the combined output shows a gradual decreasing trend. To improve the accuracy and precision of the combined wind and solar output range prediction, it is necessary to fully explore the correlation between their respective outputs and reduce the accuracy reduction caused by the complementarity of the combined output.

[0075] Therefore, based on the evaluation of the uncertainty in wind and solar unit output prediction using the LSTM-GARCH-D method, a multivariate Copula model is used to describe the correlation problem among high-dimensional variables of combined wind and solar power output. The multivariate Copula model consists of multiple basic Copula functions joined pairwise (i.e., Pair-Copula functions) for an N-dimensional random vector of wind or solar power output. Its probability density function can be expressed by the conditional probability density function as:

[0076]

[0077] In the formula, Represents the probability density function. Indicates in Under the conditions of occurrence The probability of occurrence. In Copula theory, the joint probability density function of an N-dimensional random vector X can be expressed by the Copula function and the marginal distribution density function:

[0078]

[0079] In the formula, express The Copula function is composed of Let x1 be the probability distribution function.

[0080] The joint probability density of an N-dimensional random vector X is expressed by a conditional probability density function:

[0081]

[0082] In the formula, For conditional Pair-Copula functions, it means and Under the conditions of occurrence and The copula function is composed of these components. The expression for the conditional distribution function is then:

[0083]

[0084] This shows that a multidimensional Copula model can be obtained by repeatedly calculating multiple marginal distribution functions using the above formula through Pair-Copula functions, and the connection topology between adjacent levels is also described using Pair-Copula functions. Due to the complexity and variability of the topology during the connection process, there are typically N! connection methods for an N-dimensional random vector. To more clearly describe this connection topology, the R-vine structure is used to express the correlation between different Pair-Copulas in the multidimensional Copula model. The N-dimensional wind-sun joint probability density function based on the vine Copula function is as follows:

[0085]

[0086] In the formula, Here, a and b are the condition values, representing the sequence numbers of the scenic landscape. A collection representing N-dimensional landscape data; Let represent the joint probability density function of all Pair-Copulas. The joint probability density corresponding to N-dimensional wind and light data is obtained by multiplying each marginal distribution by the joint probability density of all Pair-Copulas.

[0087] ② Construction of a joint output uncertainty model based on R.Teng Copula

[0088] The R-vine Copula model has a multi-level tree structure. When the dimension N of the wind and solar power output random vector is small, the number of tree structure combinations is limited, so the optimal model structure can be selected by traversing all possibilities. However, generally speaking, the dimension of the wind and solar power output random vector is high, and the number of possible tree structure types in the corresponding R-vine Copula model will increase rapidly. For an R-vine structure composed of N-dimensional variables, the maximum number of possible tree structure types is... There are several types. Traversing all combinations at this point could significantly increase computational complexity; therefore, a suitable algorithm is needed to generate the optimal R-vine Copula model structure. This paper uses the Maximum Spanning Tree (MST) method to fit the R-vine tree structure layer by layer. The generation steps are as follows:

[0089] 1) Calculate the N-dimensional historical power output sequence of wind and solar power. The Kendall rank correlation coefficients between the sequences are used as edge weights;

[0090] 2) Select the Frank-Copula function, Clayton-Copula function, normal Copula function, t-Copula function, and Gumbel-Copula function as alternative Pair-Copula functions for different data units;

[0091] 3) Use Frank, Clayton, and other Copula functions to fit each group of data sequentially, and calculate the corresponding Akaike's Information Criterion (AIC) evaluation index values ​​for all combinations. The expression for the AIC index is: , where I AIC This represents the AIC index value, m p The parameter represents the number of parameters in the Coupla model. Different Coupla functions have different numbers of parameters: Clayton-Copula, Gauss-Copula, and Gumble-Copula all have 3 parameters, while t-Copula has four. L represents the maximum likelihood function value, indicating the goodness of fit of the Coupla model.

[0092] 4) For each combination, select the Pair-Copula function with the smallest corresponding AIC as the connection structure of the R-copula function in that layer;

[0093] 5) At this point, each “point” in the first-level tree structure and the “edge” it forms have been obtained. The input data for the next level can be calculated. The total dimension of the input data at this point is one less than that of the first level. Repeat steps 2 to 4 to obtain the tree structure from the second to the Nth level until the final N-dimensional R-vine Copula function model is obtained.

[0094] ③ Uncertainty Assessment Steps for Combined Wind and Solar Power Output

[0095] Assume there are N wind and solar power units, and their power outputs at a certain time t are respectively , , , ... Output data and values The historical power output data of the wind and solar power units can then be represented by a set, such as the historical power output data of wind and solar power unit No. 1. The steps for predicting the combined wind and solar power output range at time t+1 are as follows:

[0096] 1) Use the LSTM algorithm to perform point prediction on N+1 sets of data, including the wind and solar power output units and their sums at time t+1, such as... Then the predicted power output data for wind and solar power at time t+1 are as follows: , , , ... and ;

[0097] 2) For the obtained wind and solar power output prediction data at time t+1, GARCH-D is used to model the residual probability distribution. The input data for this step is... , , , ... , And the point prediction data obtained in the previous step , , , ... , Finally, the probability distribution functions of each wind and solar power output unit and its sum at time t+1 can be obtained, and the inverse normal distribution function of each set of functions can be obtained as the input function of the next step, the R-copula function model, which can be denoted as follows: , , , ... , ;

[0098] 3) , , , ... , Transform into a probability distribution sequence , , , ... , Following the fitting method of the R-Fold Copula function, the probability density function of N+1 dimensions is obtained by fitting layer by layer. At this time Substituting these variables as known variables into the multidimensional joint probability density function formula, we obtain the function. , which is the probability density function of the joint residual of wind power at time t+1.

[0099] To verify the effectiveness of the method of this invention, six sets of data from three photovoltaic power plants and three wind farms in a certain region in January 2023 were used for simulation analysis. The sampling density of each set of data was 144 points / day, with a total of 288 sampling points. For each wind and solar power plant, 144 points were taken as the rolling bandwidth for rolling prediction, predicting the output values ​​at the next 1, 3, and 6 time points (i.e., 10 minutes, 30 minutes, and 1 hour). The data from the last 144 sampling points were used as control test data to verify the prediction effect of the model. The number of hidden layer neurons in the LSTM-GARCH-D algorithm was set to 200, the number of iterations was 500, the gradient threshold was 1, the initial learning rate was 0.01, and the learning rate was multiplied by a multiplication factor of 0.5 after every 100 iterations.

[0100] Based on the settings of the basic example, the LSTM algorithm is first used to predict the magnitude of the combined wind and solar power output. The prediction results are as follows: Figure 3 As shown, the LSTM algorithm has a good fit for the ultra-short-term wind and solar power output prediction results. For example, the errors between the 10-minute and 30-minute prediction results and the actual values ​​are not significant. However, when the time span is long, the prediction accuracy drops significantly. The 1-hour prediction result around 8:00 has an error of nearly 50% compared with the actual value. This may be because the period before that time is the rising period of new energy power output. The LSTM algorithm has a large "inertia" in its time series prediction results, which causes the prediction results to deviate significantly from the actual values.

[0101] Figure 4 The results of the LSTM wind and solar joint prediction error are presented. By analyzing the error magnitude at each time step, the magnitude of the LSTM prediction error can be seen more intuitively. The larger the prediction time step, the larger the possible error will be. Moreover, the time step will cause the error to expand in a non-linear form. For example, the prediction error of 10 minutes around 10:00 is only 200kW, while the prediction error of 30 minutes is as high as 700kW, and the prediction error of 60 minutes is as high as nearly 1200kW. At the same time, there is a correlation between the errors of different prediction time steps at the same time. Not only are the error magnitudes of different prediction time steps at most times the same positive or negative, but the trend of change is also roughly the same.

[0102] Since the output of the R-Teng Copula function is a probability density function, it is difficult to intuitively display the evaluation results for multi-time-point combined wind and solar power output sequences. Therefore, it is further processed and represented as bandwidth intervals at different confidence levels. With a prediction step size of 60 minutes, based on the prediction results of the LSTM-GARCH-D model, the method of this invention (LSTM-GARCH-D+R-Teng Copula) is compared with the traditional GARCH model to evaluate the interval prediction performance at 80% and 90% confidence levels. The uncertainty evaluation results of different methods are as follows: Figure 5 (Wind-solar joint uncertainty assessment based on R-vine Copula correlation analysis) and Figure 6 (Uncertainty assessment of joint wind and solar power based on GARCH model) is shown below, with comparisons:

[0103] ① Interval Coverage: For output fluctuations around 10:00, the method of this invention can cover the maximum error at 80% confidence level, while the GARCH model requires 90% confidence level to cover it, indicating that the present invention captures the actual fluctuations more accurately.

[0104] ② Interval bandwidth: The interval bandwidth of the GARCH model is significantly wider, especially during the period of high output from afternoon to evening, when the bandwidth expands sharply with the increase in output, deviating from the actual error fluctuation trend; the method of this invention, due to the consideration of wind and solar correlation, has a more compact interval bandwidth, which is consistent with the actual error change trend and has less redundant information.

[0105] Therefore, it can be seen that the method of the present invention, by introducing R-vine Copula to characterize correlation, significantly reduces the interval bandwidth while ensuring coverage, and the evaluation results are closer to reality.

[0106] Example 2: This example uses the wind and solar power output data from Example 1 to assess uncertainty using the conventional LSTM-GARCH method, the LSTM-GARCH-D method, and the proposed method, and compares the results. Ten confidence levels were set: {50%, 55%, 60%, ..., 95%}. Uncertainty assessment results for the 1st, 3rd, and 6th step intervals (i.e., 10 minutes, 30 minutes, and 1 hour later) were obtained using different methods at different confidence levels. The uncertainty assessment results were then compared using corresponding interval coverage, interval average bandwidth, and comprehensive evaluation indicators. The example data is the same as in Example 1, using the same wind and solar power output data to evaluate the uncertainty indicators predicted at 10 minutes (step 1), 30 minutes (step 3), and 60 minutes (step 6). The evaluation indicators are interval coverage, interval average bandwidth, and comprehensive evaluation indicators. The results and analysis are as follows:

[0107] ①Interval coverage: such as Figure 7 As shown, the method of this invention has the highest coverage at all confidence levels and prediction step sizes, and it approaches 1 as the confidence level increases. For example, the coverage of the 6-step prediction at a 90% confidence level reaches 95%, which is significantly higher than LSTM-GARCH-D (82%) without correlation analysis and conventional LSTM-GARCH (70%), indicating that its reliability is superior.

[0108] ② Average bandwidth of the interval: such as Figure 8As shown, the bandwidth of all methods increases with the increase of the prediction step size, but the bandwidth of the method of this invention is always the smallest. Taking 6-step prediction and 90% confidence level as an example, the bandwidth of this invention is 0.08, which is only 22.9% of that of conventional LSTM-GARCH (0.35), indicating that its evaluation results are more reliable.

[0109] ③ Comprehensive evaluation indicators: such as Figure 9 As shown, the comprehensive index of the method of this invention significantly outperforms the method in each prediction step, and the increase is the largest with the increase of confidence level. In 6-step prediction at 90% confidence level, the comprehensive index reaches 0.92, which is much higher than LSTM-GARCH-D (0.75) without correlation analysis and conventional LSTM-GARCH (0.58), demonstrating the superiority of the "residual fitting + correlation analysis" synergistic framework.

[0110] Therefore, the method of the present invention is superior to traditional methods in terms of coverage, bandwidth control and overall performance, especially in long step size prediction, which verifies its effectiveness and superiority.

[0111] Example 3: This example proposes a joint assessment system for uncertainties in wind power integration into the power system based on time-varying fluctuations, including:

[0112] The point prediction module is used to generate the point prediction value for the next time step based on the historical output sequence using LSTM, and output the prediction residual.

[0113] The edge modeling module is used to input the prediction residuals into GARCH-D to form LSTM-GARCH-D and obtain the edge probability distribution function of each unit; the prediction residuals are subjected to probability integral transformation according to the edge probability distribution function to obtain the edge variables located in the interval [0,1], and the inverse normal distribution function of each set of functions is obtained.

[0114] The vine structure construction module is used to determine the R-vine Copula function model using the maximum spanning tree MST, construct a multidimensional correlation model, and characterize the correlation between the errors of the unit output prediction values.

[0115] The joint evaluation and output module is used to combine the edge distribution and the vine structure to obtain the joint probability density and joint distribution, and output the prediction range of the output of each unit and the total output at a given confidence level.

[0116] Example 4: This example proposes an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method steps of the present invention.

[0117] Example 5: This example proposes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the method of the present invention, which will not be described in detail here.

[0118] It should be noted that the processing flow of Examples 3-5 corresponds to the specific steps of the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the methods provided in Examples 1-2 of the present invention.

[0119] The specific implementation schemes described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific implementation schemes of the present invention and are not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.

Claims

1. A joint assessment method for uncertainties in wind power integration into the power system based on time-varying fluctuations, characterized in that, include: S1. For each wind farm and photovoltaic power station unit, based on the historical power output sequence of wind and solar power, the point prediction value is obtained through the Long Short-Term Memory (LSTM) network, and the prediction residual is calculated. S2. Input the predicted residuals into the generalized autoregressive conditional heteroscedasticity-difference model GARCH-D to construct the long short-term memory-generalized autoregressive conditional heteroscedasticity-difference joint model LSTM-GARCH-D, and obtain the marginal probability distribution function of each unit: ； In the formula, These represent the wind and solar power output data at each moment. J t express t Predicted data for each moment, express t The predicted residual at time 10:

00. express t The variance of the predicted values ​​at time points; q , p , m These are the orders of the generalized autoregressive conditional heteroscedasticity model GARCH, respectively. A , B i , C j , D k+1 All parameters are obtained through fitting; m The order of the difference correction term for the predicted value; for t Predictive difference of the output point of the scenery at any time; The prediction residuals are subjected to probability integral transformation based on the marginal probability distribution function to obtain marginal variables located in the interval [0,1], and the inverse normal distribution function of each set of functions is obtained. S3. The R-vine Copula function model is determined by using the maximum spanning tree MST, and a multidimensional correlation model is constructed to characterize the correlation between the errors of the unit output prediction values. S4. Combine the marginal probability distribution functions of each unit obtained in step S2 with the R-vine Copula function model constructed in step S3 to obtain the joint probability density and joint distribution, and output the prediction interval of wind and solar power output of each unit and total power output under a given confidence level.

2. The joint assessment method for uncertainties of wind power integration into the power system based on time-varying fluctuations as described in claim 1, characterized in that, In step S1, based on the historical power output sequence of wind and solar power, point prediction values ​​are obtained through a Long Short-Term Memory (LSTM) network, including the following steps: S1-1, Using a one-dimensional vector This represents historical output data. , The step size of the data output for the trained landscape. These represent the wind and solar power output data at each moment; S1-2, Transform a one-dimensional vector Divided into training set With test set ,in , , , The length of the training set; S1-3, The calculation method of the Long Short-Term Memory (LSTM) network is as follows: ； In the formula, for Input values ​​for historical output sequence of moments and scenery. for t Output value at all times for Output value at all times for The status of the gating unit at all times. , , The results of the state operations for the forget gate, input gate, and output gate. , , These are the bias terms for the forget gate, input gate, and output gate. , , These are the weight matrices for the forget gate, input gate, and output gate, respectively. for t The input state of the constant-time gating unit, The weight matrix is ​​the input state of the gating unit. For its bias term, δ It is the sigmoid function, with a range of [0,1]. For activation function, This indicates that the elements in the vector are multiplied element by element; S1-4, gathering efforts J The output after the hidden layer is represented as: , For the first Predictive data for the output of the wind and light at any given moment.

3. The joint assessment method for uncertainties of wind power integration into the power system based on time-varying fluctuations as described in claim 1, characterized in that, In step S3, the R-vine Copula function model is determined using the Maximum Spanning Tree (MST), which includes the following steps: S3-1 Calculate the N-dimensional historical power output sequence of wind and solar power. The Kendall rank correlation coefficients between the sequences are used as edge weights. A collection representing N-dimensional landscape data; S3-2. Determine the candidate set of Pair-Copula functions; the candidate set of Pair-Copula functions includes one or more of the following: Frank-Copula, Clayton-Copula, GaussianCopula, t-Copula, and Gumbel-Copula. S3-3. Use the pair-copula candidate function to fit each group of data sequentially, and calculate the corresponding Akaike Information Criterion (AIC) index. The Akaike Information Criterion (AIC) index is calculated according to the following formula: ； In the formula, I AIC The AIC (Advanced Information Criteria) index is used for information management in Akaike. m p This represents the number of parameters in the pair-copula model. L The value of the maximum likelihood function; S3-4. For each combination, select the Pair-Copula function with the smallest corresponding Akaike Information Criterion (AIC) index as the connection structure of the R-Fuji Copula function. S3-5. Repeat S3-2 to S3-4 to obtain the N-dimensional R-vine Copula function model, and calculate the N-dimensional wind-sun joint probability density function based on the vine Copula function of the N-dimensional R-vine Copula function model.

4. The joint assessment method for uncertainties of wind power integration into the power system based on time-varying fluctuations according to claim 3, characterized in that, In steps S3-5, the N-dimensional wind-sun joint probability density function is calculated using the vine Copula function based on the N-dimensional R-vine Copula function model. The calculation formula is as follows: ； In the formula, For conditional values, a , b Numbering of the sequence of contributions to the scenic history. A set representing the historical output sequence of N-dimensional landscape; This represents the joint probability density function of the candidate set of the pair-copula function.

5. The joint assessment method for uncertainties of wind power integration into the power system based on time-varying fluctuations according to claim 4, characterized in that, In step S4, the predicted ranges for the wind and solar power output of each unit and the total power output are output, assuming there are... N Each scenic power unit, at a certain moment t The output power is respectively , , , ... Output data and values The historical power output data of the wind and solar power generation units are represented by a set. The historical power output data of wind and solar power generation unit No. 1 is: ,right t+ The steps for predicting the wind and solar power output range at time 1 are as follows: S4-1, Using the LSTM algorithm to... t+1 The total output of the wind and solar power units at any given time and their sum N +1 set of data for point prediction ,but t The predicted data for the solar and wind power output units at time +1 are as follows: , , , ... and ; S4-2, Regarding the result obtained in step S4-1 t The predicted wind and solar power output data at time +1 were used to model the residual probability distribution using the generalized autoregressive conditional heteroscedasticity-difference model (GARCH-D). The input data for this step was... , , , ... , and the point prediction data obtained in step S4-1 , , , ... , The sum of the output values ​​of each wind and solar power unit is obtained. t The probability distribution function at time +1 is determined, and the inverse normal distribution function of each set of functions is obtained, which is used as the input function of the R-copula function model in step S3, denoted as follows: , , , ... , ; S4-3, will , , , ... , Transform into a probability distribution sequence , , , ... , Following the fitting method for the R-copula function model in step S3, the model is fitted layer by layer to obtain... N +1-dimensional probability density function ,Will Substituting these known variables into the N-dimensional wind-sun joint probability density function yields the function. That is t The probability density function of the combined wind and light residual at time +1.

Images