Source load correlation frequency modulation signal prediction method for preventing low-frequency load shedding

Abstract

The invention discloses a source load correlation frequency modulation signal prediction method for preventing low-frequency load shedding. Firstly, according to the requirement for preventing low-frequency load shedding, frequency modulation signal statistical variables are extracted to represent the characteristics of frequency modulation signals in a real-time economic dispatching period; secondly, the LSTM neural network is adopted to predict the source load power, and the source load power change at the future moment is obtained; a Copula function is adopted to model historical frequency modulation signal statistical variables and source load power change data, and a correlation model among the historical frequency modulation signal statistical variables, the source load power change data and the historical frequency modulation signal statistical variables is obtained; then, according to probabilitics, obtaining a conditional probability density distribution model of a frequency modulation signal statistical variable based on source load power change; and finally, substituting the predicted source load power change data into the frequency modulation signal statistical variable conditional probability density model to obtain the frequency modulation signal statistical variable probability density at the future moment. According to the method, by capturing the correlation between the frequency modulation signal and the source load power change, direct prediction of the frequency modulation signal is indirectly converted into current more mature and accurate source load power prediction, the prediction accuracy is improved, and the low-frequency load shedding action probability is reduced.

Description

Technical Field

[0001] The invention relates to a method for predicting source-load correlation frequency modulation signals for preventing under-frequency load shedding, and belongs to the technical field of power systems. Background Art

[0002] In modern power systems, with the access of large-scale renewable energy, the stability of power system frequency faces greater challenges. Due to the volatility of renewable energy (such as wind and solar energy), frequency deviation problems have become more frequent and difficult to predict. When the grid frequency drops significantly due to the imbalance between power supply and demand, the system will start the low-frequency load shedding (LS) mechanism to restore the system frequency by automatically cutting off part of the load to prevent grid collapse. Traditional low-frequency load shedding strategies are usually based on preset frequency thresholds and are automatically triggered when the grid frequency falls below a certain critical value. When the load fluctuates violently or the proportion of renewable energy access is high, relying solely on such measures is difficult to adapt to the dynamically changing grid environment, resulting in too frequent load shedding operations, causing economic losses and deteriorating user experience.

[0003] To solve this problem, modern power systems urgently need more intelligent and flexible frequency regulation methods to reduce the triggering frequency of low-frequency load shedding. In this context, accurate prediction of frequency modulation signals has become an important means to reduce low-frequency load shedding actions. Frequency modulation signals can reflect the frequency fluctuation trend in the power grid. By predicting the frequency modulation demand in advance, the system can actively adjust before the frequency drops, thereby avoiding the frequency reaching the triggering point of low-frequency load shedding.

[0004] Most current studies only start from the prediction object itself, while ignoring the correlation between different factors, such as the impact of source load power changes on FM signals.

[0005] The patent document "Method for predicting wind power based on mechanism and data hybrid driven neural network" (patent application number 202410718571.2) proposes a hybrid driven neural network model that effectively combines the mechanism model and the GRU neural network, overcomes the problem of data anomalies, and improves the safety and economy of the power system.

[0006] The patent document "A graph neural network prediction method for power regional load" (patent application number 202110976490.9) proposes a method of using graph neural networks to predict regional power load, improving prediction accuracy.

[0007] The patent document "Photovoltaic power prediction method based on EMA-RElman neural network" (patent application number 202410285011.2) proposes a prediction method that combines EMA and RElman neural networks, using the former to optimize the parameters of the neural network, greatly improving the accuracy of photovoltaic predictions.

[0008] The above patents only start from the prediction object itself, without considering the correlation between different factors, which affects the accuracy of the overall prediction and the economy of subsequent economic scheduling. Summary of the invention

[0009] The purpose of the invention is to provide a source-load correlation frequency modulation signal prediction method for preventing low-frequency load shedding in view of the shortcomings of the above-mentioned background technology, which can be applied to the prediction of frequency modulation signals under high penetration of new energy sources to prevent low-frequency load shedding actions. The present invention first extracts the statistical variables of the frequency modulation signal to characterize the characteristics of the frequency modulation signal in the real-time economic dispatch cycle; secondly, the LSTM neural network is used to predict the source-load power to obtain the source-load power change at future moments; then the Copula function is used to model the historical frequency modulation signal statistical variables and the source-load power change data to obtain the correlation model between the three; then, according to the relevant probability theory, the conditional probability density distribution model of the frequency modulation signal statistical variables based on the source-load power change is obtained; finally, the predicted source-load power change data is substituted into the conditional probability density model of the frequency modulation signal statistical variables to obtain the frequency modulation signal statistical variables at future moments.

[0010] The present invention is achieved through the following technical solutions:

[0011] A method for predicting source-load correlation frequency modulation signals for preventing underfrequency load shedding comprises the following steps:

[0012] Step S1, extracting the statistical variable data of the frequency modulation signal to characterize the real-time economic dispatch cycle, that is, the overall characteristics of the frequency modulation signal within a five-minute time scale;

[0013] Step S2: Based on the historical renewable energy power generation and load power data, the LSTM neural network is used to predict the renewable energy power generation and load power at future moments, and the renewable energy power generation and load power at adjacent moments in the predicted data are respectively subtracted to obtain the source-load power changes at adjacent moments in the future;

[0014] Step S3, establishing a correlation model between the frequency modulation signal statistical variables and the source load power change through the Copula function, and obtaining the conditional probability distribution of the frequency modulation signal statistical variables based on the source load power change;

[0015] Step S4: Substitute the predicted source load power change into the conditional probability distribution of the frequency modulation signal statistical variable based on the source load power change, and obtain the conditional probability distribution of the frequency modulation signal statistical variable based on the source load power change at the future moment.

[0016] The purpose of the present invention can also be further achieved by the following technical measures:

[0017] Preferably, step S1 includes:

[0018] First, the historical data of frequency modulation signals are processed to extract relevant statistical variables within the real-time economic dispatch cycle to characterize the imbalance of distribution network operation during this period. The extracted statistical variables are: cumulative frequency modulation energy Cumulative frequency modulation mileage M n , FM amplitude quantile and the adjustment rate quantile The details are as follows:

[0019] Cumulative FM energy:

[0020]

[0021] In the formula, represents the accumulated energy in the nth economic dispatch cycle; τ represents the time span of a single frequency modulation signal; T represents the time span of a single real-time economic dispatch cycle; j represents the jth frequency modulation signal in an economic dispatch cycle; Indicates the uplink and downlink frequency modulation signal values;

[0022] Cumulative frequency tuning mileage:

[0023]

[0024] Where M n Indicates the accumulated frequency regulation mileage during the nth real-time economic dispatch time; AGC j represents the jth FM signal, AGC j-1 represents the j-1th FM signal;

[0025] FM amplitude quantile:

[0026]

[0027] In the formula, is a judgment function. Its specific meaning is that if the formula in the brackets is true, its value is 1, otherwise it is 0, that is, is the uplink or downlink frequency modulation amplitude quantile; is the number of uplink or downlink signals in the frequency modulation signal sequence in the nth real-time economic dispatch cycle; α MATo select the proportion value corresponding to the quantile of the frequency modulation amplitude;

[0028] Adjustment rate quantile:

[0029]

[0030] In the formula, represents the upward and downward quantiles of the regulation rate; α RR is the proportional value corresponding to the selected adjustment rate quantile; AGC j / j-1 is the adjustment rate of the FM signal from the jth to the j-1th; AGC j-1 / j is the adjustment rate from the j-1th to the jth frequency modulation signal; It represents the vector of the statistical variables of the frequency modulation signal in the nth time period, and its specific expression is as follows:

[0031]

[0032] Preferably, step S2 includes:

[0033] First, the LSTM neural network used for source load power prediction consists of an input layer, a hidden layer, and an output layer, where the number of hidden layers is 1 and the number of neurons in the hidden layer is 200;

[0034] The gate structures of the LSTM neural network are respectively a forget gate, an input gate and an output gate. The gate structure realizes selective forgetting and inheritance of input data, thereby avoiding the gradient explosion problem that may be caused by too much data. The specific workflow is shown as follows:

[0035] The forget gate will selectively discard some state information. The specific principle is expressed as follows:

[0036] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0037] In the formula, f t represents the output of the forget gate at time t; σ represents the sigmoid function; W f represents the weight of the forget gate; [h t-1 ,x t ] represents the output data h of the previous moment t-1 and the output data x at time t t The vector formed by connecting; b f Represents the bias value of the forget gate;

[0038] The input gate is responsible for filtering the input data and determining the cell state information that needs to be updated. The specific principle is expressed as follows:

[0039] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0040]

[0041] In the formula, i t Represents the output of the sigmoid module; represents the output of the tanh module; W c and W i represents the weight of the input gate; b i and b c represents the bias of the input gate; tanh represents the tanh function;

[0042] The neuron cell state first passes through the forget gate to selectively forget some information, and then passes through the input gate to update some cell state information, and finally obtains the current cell state C. t , which is specifically expressed as follows:

[0043]

[0044] In the formula, C t is the neuron cell state at time t; C t-1 represents the neuron cell state at time t-1;

[0045] The output gate determines the output cell state information, and its specific principle is expressed as follows:

[0046] O t =σ(W o ·[h t-1 ,x t ]+b o )

[0047] h t =O t tanh(C t )

[0048] In the formula, O t Represents the output of the sigmoid module; h t represents the output of the neuron at time t; W o represents the weight of the output gate; b o Represents the bias of the output gate; W o Represents the weight of the output gate;

[0049] Finally, the constructed LSTM neural network is used to predict the source load power change data at future moments.

[0050] Preferably, step S3 specifically includes:

[0051] First, the Copula function is used to model the correlation between the FM signal statistical variables and the source load power changes. The specific expression is as follows:

[0052]

[0053] In the formula, F(ξ AGC ,ΔP RE ,ΔP LD ) represents the joint probability distribution function of the frequency modulation signal statistical variable and the source load power change; is the Copula function; and They are the marginal distributions of frequency modulation signal statistical variables, renewable energy power generation changes, and load power changes;

[0054] Then, the Bayesian Information Criterion (BIC) was used as the model fit goodness evaluation index to select the optimal Copula function type, which is specifically expressed as follows:

[0055] BIC = kln(n)-2ln(L)

[0056] In the formula, k represents the number of model parameters; n represents the number of samples; L is the likelihood function;

[0057] Based on the established joint probability distribution of the frequency modulation signal statistical variables and the source-load power change, the joint probability density function of the frequency modulation signal statistical variables and the source-load power change is derived. Therefore, the joint probability density of the frequency modulation signal statistical variables and the new energy generation and load power change is expressed as follows:

[0058]

[0059] In the formula, f(ξ AGC ,ΔP RE ,ΔP LD ) represents the joint probability density of the frequency modulation signal statistical variables and the changes in renewable energy generation and load power; Represents the three-dimensional probability density distribution of the Copula function; and They represent the marginal probability density functions of the frequency modulation signal statistical variables, the renewable energy power generation changes, and the load power changes respectively;

[0060] The conditional probability density of the frequency modulation signal statistical variable based on the source-load power change is obtained by dividing the joint probability density by the two-dimensional joint probability density of the renewable energy power change and the load power change, which is expressed as follows:

[0061]

[0062] In the formula, f(ξ AGC |(ΔP RE ,ΔP LD )) represents the conditional probability density of the FM signal statistical variable based on the source load power change; f(ΔP RE ,ΔP LD ) represents the two-dimensional joint probability density of the change in renewable energy power generation and load power; is the Copula density function corresponding to the two-dimensional joint probability density.

[0063] Preferably, step S4 includes:

[0064] Substitute the source load power change data predicted by the LSTM neural network into the conditional probability density function of the FM signal statistical variable based on the source load power change to obtain the required FM signal statistical variable prediction value, which is expressed as follows:

[0065]

[0066] In the formula, It represents the conditional probability density of the future time-frequency signal statistical variables based on the source load power change obtained by using the predicted source load power change data; It represents the joint probability density of the frequency modulation signal statistical variables and the changes in renewable energy power generation and load power at future moments obtained by using the predicted source-load power change data; It represents the two-dimensional joint probability density of the future renewable energy power generation change and load power change obtained by using the predicted source-load power change data.

[0067] Compared with the prior art, the beneficial effects of the present invention are: the present invention constructs a method for predicting statistical variables of frequency modulation signals taking into account the source-load correlation, successfully captures the correlation between frequency modulation signals and source-load power changes, and converts direct prediction of frequency modulation signals into a more mature and reliable source-load power prediction, which not only reduces the difficulty of predicting frequency modulation signals, but also improves the accuracy and reliability of predictions, and reduces subsequent economic dispatch operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic diagram of the implementation process of the present invention;

[0069] Figure 2A flow chart for predicting statistical variables of FM signals considering source-load correlation;

[0070] Figure 3 Neural network prediction diagram for wind power generation;

[0071] Figure 4 It is the conditional probability density distribution diagram of the upper amplitude quantile of the FM signal. DETAILED DESCRIPTION

[0072] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0073] like Figure 1 , 2 As shown, the present invention provides a method for predicting source-load correlation frequency modulation signals for preventing underfrequency load shedding, comprising the following steps:

[0074] Step S1, extracting the statistical variable data of the frequency modulation signal to characterize the overall characteristics of the frequency modulation signal within the real-time economic dispatch cycle, i.e., within a five-minute time scale;

[0075] First, the historical data of frequency modulation signals are processed to extract relevant statistical variables within the real-time economic dispatch cycle to characterize the imbalance of distribution network operation during this period. The extracted statistical variables are: cumulative frequency modulation energy Cumulative frequency modulation mileage M n , FM amplitude quantile and the adjustment rate quantile

[0076] Cumulative FM energy:

[0077]

[0078] In the formula, represents the accumulated energy in the nth economic dispatch cycle; τ represents the time span of a single frequency modulation signal; T represents the time span of a single real-time economic dispatch cycle; j represents the jth frequency modulation signal in an economic dispatch cycle; Indicates the uplink and downlink FM signal values.

[0079] Cumulative frequency tuning mileage:

[0080]

[0081] Where M n Indicates the accumulated frequency regulation mileage during the nth real-time economic dispatch time; AGC j represents the jth FM signal, AGC j-1 represents the j-1th FM signal.

[0082] FM amplitude quantile:

[0083]

[0084] In the formula, is a judgment function. Its specific meaning is that if the formula in the brackets is true, its value is 1, otherwise it is 0, that is, is the uplink or downlink frequency modulation amplitude quantile; is the number of uplink or downlink signals in the frequency modulation signal sequence in the nth real-time economic dispatch cycle; α MA Select the proportion value corresponding to the FM amplitude quantile.

[0085] Adjustment rate quantile:

[0086]

[0087] In the formula, Indicates the upward and downward quantiles of the regulation rate; α RR is the proportional value corresponding to the selected adjustment rate quantile; AGC j / j-1 is the adjustment rate of the FM signal from the jth to the j-1th; AGC j-1 / j is the modulation rate from the j-1th to the jth frequency modulation signal.

[0088] To facilitate subsequent calculations, we use It represents the vector of the statistical variables of the frequency modulation signal in the nth time period, and its specific expression is as follows:

[0089]

[0090] Step S2: Based on the historical renewable energy power generation and load power data, the LSTM neural network is used to predict the renewable energy power generation and load power at future moments, and the renewable energy power generation and load power at adjacent moments in the predicted data are respectively subtracted to obtain the source-load power changes at adjacent moments in the future;

[0091] As an implementation of the embodiment of the present invention, step S2 includes:

[0092] First, the LSTM neural network used for source load power prediction consists of an input layer, a hidden layer, and an output layer, where the number of hidden layers is 1 and the number of hidden layer neurons is 200.

[0093] Furthermore, the LSTM neural network has a unique gate structure, which includes the forget gate, input gate, and output gate. This gate structure enables selective forgetting and inheritance of input data, thereby avoiding the gradient explosion problem that may be caused by too much data. The specific workflow is as follows:

[0094] The forget gate will selectively discard some state information. The specific principle is expressed as follows:

[0095] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0096] In the formula, f t represents the output of the forget gate at time t; σ represents the sigmoid function; W f represents the weight of the forget gate; [h t-1 ,x t ] represents the output data h of the previous moment t-1 and the output data x at time t t The vector formed by connecting; b f Indicates the bias value of the forget gate.

[0097] The input gate is responsible for filtering the input data and determining the cell state information that needs to be updated. The specific principle is expressed as follows:

[0098] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0099]

[0100] In the formula, i t Represents the output of the sigmoid module; represents the output of the tanh module; W c and W i represents the weight of the input gate; b i and b c Represents the bias of the input gate; tanh represents the tanh function.

[0101] The neuron cell state first passes through the forget gate to selectively forget some information, and then passes through the input gate to update some cell state information, and finally obtains the current cell state C. t The specific expression is as follows:

[0102]

[0103] In the formula, C t is the neuron cell state at time t; C t-1 Represents the state of the neuron cell at time t-1.

[0104] The output gate mainly determines the output cell state information, and its specific principle is expressed as follows:

[0105] O t =σ(W o ·[h t-1 ,x t ]+b o )

[0106] h t =O t tanh(C t )

[0107] In the formula, O t Represents the output of the sigmoid module; h t represents the output of the neuron at time t; W o represents the weight of the output gate; b o Represents the bias of the output gate.

[0108] Finally, the constructed LSTM neural network is used to predict the renewable energy power generation and load power data in the future. Taking into account the accuracy and speed of the prediction model, the parameters of the neural network model are set as follows after trial and error and experimental verification: the number of hidden layers is 1, the number of neurons is set to 200; the optimization algorithm uses the Adam algorithm to prevent overfitting during the model operation; the learning rate is set to 0.01; the number of iterations is 150. The wind power generation data from January 1 to January 30 in a certain place is selected as the training set data sample, with a time interval of 5 minutes, and the wind power generation data on January 31 is used as the test set for prediction verification. The data sample set is brought into the LSTM neural network prediction model, and the training results are obtained as follows Figure 3 As shown in the figure, the accuracy of prediction using LSTM neural network is verified, and the source-load power change is obtained by subtracting the power generation power and load power at adjacent moments.

[0109] Step S3, establishing a correlation model between the frequency modulation signal statistical variables and the source load power change through the Copula function, and obtaining the conditional probability distribution of the frequency modulation signal statistical variables based on the source load power change based on the probability theory;

[0110] As an implementation of the embodiment of the present invention, step S3 includes:

[0111] First, the Copula function is used to model the correlation between the FM signal statistical variables and the source load power changes. The specific expression is as follows:

[0112]

[0113] In the formula, F(ξ AGC ,ΔP RE ,ΔP LD) represents the joint probability distribution function of the frequency modulation signal statistical variable and the source load power change; is the Copula function; and They are the marginal distributions of frequency modulation signal statistical variables, renewable energy power generation changes, and load power changes.

[0114] Then, the Bayesian Information Criterion (BIC) was used as the model fit goodness evaluation index to select the optimal Copula function type, which is specifically expressed as follows:

[0115] BIC = kln(n)-2ln(L)

[0116] In the formula, k represents the number of model parameters; n represents the number of samples; and L is the likelihood function.

[0117] Further, based on the established joint probability distribution of the frequency modulation signal statistical variables and the source-load power change, the joint probability density function of the frequency modulation signal statistical variables and the source-load power change can be derived. Therefore, the joint probability density of the frequency modulation signal statistical variables and the new energy generation and load power change is expressed as follows:

[0118]

[0119] In the formula, f(ξ AGC ,ΔP RE ,ΔP LD ) represents the joint probability density of the frequency modulation signal statistical variables and the changes in renewable energy generation and load power; Represents the three-dimensional probability density distribution of the Copula function; and They represent the marginal probability density functions of the frequency modulation signal statistical variables, the renewable energy power generation changes and the load power changes respectively.

[0120] Finally, according to the relevant theories of probability theory, the conditional probability density of the frequency modulation signal statistical variable based on the source and load power changes can be obtained by dividing the joint probability density by the two-dimensional joint probability density of the renewable energy power change and the load power change, which can be expressed as follows:

[0121]

[0122] In the formula, f(ξ AGC |(ΔP RE ,ΔP LD )) represents the conditional probability density of the FM signal statistical variable based on the source load power change; f(ΔP RE ,ΔP LD) represents the two-dimensional joint probability density of the change in renewable energy power generation and load power; is the Copula density function corresponding to the two-dimensional joint probability density.

[0123] Step S4, substitute the predicted source load power change into the conditional probability distribution of the frequency modulation signal statistical variable based on the source load power change, and obtain the conditional probability distribution of the frequency modulation signal statistical variable based on the source load power change at the future moment. As an implementation method of an embodiment of the present invention, step S4 includes:

[0124] Substitute the source load power change data predicted by the LSTM neural network into the conditional probability density function of the FM signal statistical variable based on the source load power change to obtain the required FM signal statistical variable prediction value, which is expressed as follows:

[0125]

[0126] In the formula, It represents the conditional probability density of the future time-frequency signal statistical variables based on the source load power change obtained by using the predicted source load power change data; It represents the joint probability density of the frequency modulation signal statistical variables and the changes in renewable energy power generation and load power at future moments obtained by using the predicted source-load power change data; It represents the two-dimensional joint probability density of the future renewable energy power generation change and load power change obtained by using the predicted source load power change data. Taking the upper quantile of the frequency modulation amplitude in the frequency modulation signal statistical variable as an example, the simulation results are as follows: Figure 4 The results show that the strategy proposed in this invention is feasible.

[0127] The present invention first extracts the statistical variables of the frequency modulation signal to characterize the characteristics of the frequency modulation signal within the real-time economic dispatch cycle; secondly, the LSTM neural network is used to predict the source-load power to obtain the source-load power changes in the future; then the Copula function is used to model the historical frequency modulation signal statistical variables and the source-load power change data to obtain the correlation model between the three; then, according to the relevant probability theory, the conditional probability density distribution model of the frequency modulation signal statistical variables based on the source-load power changes is obtained; finally, the predicted source-load power change data is substituted into the conditional probability density model of the frequency modulation signal statistical variables to obtain the frequency modulation signal statistical variables in the future. This method captures the correlation between the frequency modulation signal and the source-load power changes, and indirectly converts the direct prediction of the frequency modulation signal into a more mature and accurate source-load power prediction, thereby improving the accuracy of the prediction and reducing the probability of low-frequency load reduction action.

[0128] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the protection scope required by the present invention.

Claims

1. A method for predicting source-load correlation frequency modulation signals for preventing underfrequency load shedding, characterized in that: The following steps are involved: Step S1, extracting the statistical variable data of the frequency modulation signal to characterize the real-time economic dispatch cycle, that is, the overall characteristics of the frequency modulation signal within a five-minute time scale; Step S2: Based on the historical renewable energy power generation and load power data, the LSTM neural network is used to predict the renewable energy power generation and load power at future moments, and the renewable energy power generation and load power at adjacent moments in the predicted data are respectively subtracted to obtain the source-load power changes at adjacent moments in the future; Step S3, establishing a correlation model between the frequency modulation signal statistical variables and the source load power change through the Copula function, and obtaining the conditional probability distribution of the frequency modulation signal statistical variables based on the source load power change; Step S4: Substitute the predicted source load power change into the conditional probability distribution of the frequency modulation signal statistical variable based on the source load power change, and obtain the conditional probability distribution of the frequency modulation signal statistical variable based on the source load power change at the future moment.

2. A method for predicting source-load correlation frequency modulation signals for preventing underfrequency load shedding as claimed in claim 1, characterized in that: The step S1 includes: First, the historical data of frequency modulation signals are processed to extract relevant statistical variables within the real-time economic dispatch cycle to characterize the imbalance of distribution network operation during this period. The extracted statistical variables are: cumulative frequency modulation energy Cumulative frequency modulation mileage M n , FM amplitude quantile and the adjustment rate quantile The details are as follows: Cumulative FM energy: In the formula, represents the accumulated energy in the nth economic dispatch cycle; τ represents the time span of a single frequency modulation signal; T represents the time span of a single real-time economic dispatch cycle; j represents the jth frequency modulation signal in an economic dispatch cycle; Indicates the uplink and downlink frequency modulation signal values; Cumulative frequency tuning mileage: Where M n Indicates the accumulated frequency regulation mileage during the nth real-time economic dispatch time; AGC j represents the jth FM signal, AGC j-1 represents the j-1th FM signal; FM amplitude quantile: In the formula, is a judgment function. Its specific meaning is that if the formula in the brackets is true, its value is 1, otherwise it is 0, that is, is the uplink or downlink frequency modulation amplitude quantile; is the number of uplink or downlink signals in the frequency modulation signal sequence in the nth real-time economic dispatch cycle; α MA To select the proportion value corresponding to the quantile of the frequency modulation amplitude; Adjustment rate quantile: In the formula, represents the upward and downward quantiles of the regulation rate; α RR is the proportional value corresponding to the selected adjustment rate quantile; AGC j / j-1 is the adjustment rate of the FM signal from the jth to the j-1th; AGC j-1 / j is the adjustment rate from the j-1th to the jth frequency modulation signal; It represents the vector of the statistical variables of the frequency modulation signal in the nth time period, and its specific expression is as follows:

3. A method for predicting source-load correlation frequency modulation signals for preventing underfrequency load shedding as claimed in claim 1, characterized in that: The step S2 includes: First, the LSTM neural network used for source load power prediction consists of an input layer, a hidden layer, and an output layer, where the number of hidden layers is 1 and the number of neurons in the hidden layer is 200; The gate structures of the LSTM neural network are respectively a forget gate, an input gate and an output gate. The gate structure realizes selective forgetting and inheritance of input data, thereby avoiding the gradient explosion problem that may be caused by too much data. The specific workflow is shown as follows: The forget gate will selectively discard some state information. The specific principle is expressed as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) In the formula, f t represents the output of the forget gate at time t; σ represents the sigmoid function; W f represents the weight of the forget gate; [h t-1 ,x t ] represents the output data h of the previous moment t-1 and the output data x at time t t The vector formed by connecting; b f Represents the bias value of the forget gate; The input gate is responsible for filtering the input data and determining the cell state information that needs to be updated. The specific principle is expressed as follows: i t =σ(W i ·[h t-1 ,x t ]+b i ) In the formula, i t Represents the output of the sigmoid module; represents the output of the tanh module; W c and W i represents the weight of the input gate; b i and b c represents the bias of the input gate; tanh represents the tanh function; The neuron cell state first passes through the forget gate to selectively forget some information, and then passes through the input gate to update some cell state information, and finally obtains the current cell state C. t , which is specifically expressed as follows: In the formula, C t is the neuron cell state at time t; C t-1 represents the neuron cell state at time t-1; The output gate determines the output cell state information, and its specific principle is expressed as follows: The t =σ(W o ·[h t-1 ,x t ]+b o ) h t =O t ·tanh(C t ) In the formula, O t Represents the output of the sigmoid module; h t represents the output of the neuron at time t; W o represents the weight of the output gate; b o Represents the bias of the output gate; W o Represents the weight of the output gate; Finally, the constructed LSTM neural network is used to predict the source load power change data at future moments.

4. A method for predicting source-load correlation frequency modulation signals for preventing underfrequency load shedding as claimed in claim 1, characterized in that: The step S3 includes: First, the Copula function is used to model the correlation between the FM signal statistical variables and the source load power changes. The specific expression is as follows: In the formula, F(ξ AGC ,ΔP RE ,ΔP LD ) represents the joint probability distribution function of the frequency modulation signal statistical variable and the source load power change; is the Copula function; and They are the marginal distributions of frequency modulation signal statistical variables, renewable energy power generation changes, and load power changes; Then, the Bayesian Information Criterion (BIC) was used as the model fit goodness evaluation index to select the optimal Copula function type, which is specifically expressed as follows: BIC = kln(n)-2ln(L) In the formula, k represents the number of model parameters; n represents the number of samples; L is the likelihood function; Based on the established joint probability distribution of the frequency modulation signal statistical variables and the source-load power change, the joint probability density function of the frequency modulation signal statistical variables and the source-load power change is derived. Therefore, the joint probability density of the frequency modulation signal statistical variables and the new energy generation and load power change is expressed as follows: In the formula, f(ξ AGC ,ΔP RE ,ΔP LD ) represents the joint probability density of the frequency modulation signal statistical variables and the changes in renewable energy generation and load power; Represents the three-dimensional probability density distribution of the Copula function; and They represent the marginal probability density functions of the frequency modulation signal statistical variables, the renewable energy power generation changes, and the load power changes respectively; The conditional probability density of the frequency modulation signal statistical variable based on the source-load power change is obtained by dividing the joint probability density by the two-dimensional joint probability density of the renewable energy power change and the load power change, which is expressed as follows: In the formula, f(ξ AGC |(ΔP RE ,ΔP LD )) represents the conditional probability density of the FM signal statistical variable based on the source load power change; f(ΔP RE ,ΔP LD ) represents the two-dimensional joint probability density of the change in renewable energy power generation and load power; is the Copula density function corresponding to the two-dimensional joint probability density.

5. A method for predicting source-load correlation frequency modulation signals for preventing underfrequency load shedding as claimed in claim 1, characterized in that: The step S4 includes: Substitute the source load power change data predicted by the LSTM neural network into the conditional probability density function of the FM signal statistical variable based on the source load power change to obtain the required FM signal statistical variable prediction value, which is expressed as follows: In the formula, It represents the conditional probability density of the future time-frequency signal statistical variables based on the source load power change obtained by using the predicted source load power change data; It represents the joint probability density of the frequency modulation signal statistical variables and the changes in renewable energy power generation and load power at future moments obtained by using the predicted source-load power change data; It represents the two-dimensional joint probability density of the future renewable energy power generation change and load power change obtained by using the predicted source-load power change data.

Images