Method, equipment and medium for quantifying power system frequency regulation speed requirements

Abstract

The present invention provides a method, device, and medium for quantifying the frequency regulation speed demand of a power system. The method comprises: constructing a sample feature vector for predicting the maximum short-term net load power fluctuation based on the load, renewable energy, and meteorological information of the power grid; simultaneously screening the training samples and the sample feature vector; and then establishing a net load short-term power maximum fluctuation point prediction model to predict the maximum short-term net load power fluctuation; further establishing a net load short-term power maximum fluctuation interval prediction model based on adaptive bandwidth kernel density estimation, which can obtain a confidence interval for the maximum short-term net load power fluctuation prediction result. Using the upper limit of the interval as the maximum short-term net load power fluctuation of the system ensures that it covers the actual maximum short-term net load fluctuation. The method proposed by the present invention can accurately quantify the maximum short-term net load power fluctuation of the system, providing a powerful reference for the rational formulation of day-ahead scheduling decision plans.

Description

Technical Field

[0001] The present invention relates to the field of power systems and automation technologies thereof, and in particular to a method for quantifying frequency regulation speed requirements of power systems. Background Art

[0002] Building a new power system dominated by renewable energy is a key approach to achieving my country's "dual carbon" goals. However, the strong random fluctuations of renewable energy pose a serious challenge to the safe and stable operation of the power grid. Rationally arranging resource operations through power system dispatch decisions is crucial for maintaining a real-time balance between supply and demand in the power system. Existing power system dispatch decision-making methods primarily include uncertain dispatch decision-making methods and deterministic dispatch decision-making methods.

[0003] Uncertainty scheduling decision-making methods can account for the impact of renewable energy uncertainty and have therefore attracted widespread attention from scholars both domestically and internationally. These methods primarily include multi-scenario methods, chance-constrained methods, and robust optimization methods. Multi-scenario optimization methods generate typical scenarios based on the random fluctuations of renewable energy sources and use the lowest expected system operating cost under these scenarios as the objective function, with operational safety under different scenarios as the constraint to formulate optimal decision plans. However, the number of scenarios reflecting the random fluctuations of renewable energy sources explodes with the increase in the dimensionality of the random variables representing the uncertainty of renewable energy sources, making the solution efficiency in large-scale power systems difficult to meet practical computational requirements. Chance-constrained methods ensure that the probability of exceeding the system's operational constraints under the influence of random fluctuations of renewable energy sources is below a set threshold, and by adjusting the threshold, a trade-off between grid operational safety and economic efficiency can be achieved. However, the chance-constrained model can only be converted to a second-order cone form when the renewable energy output satisfies a specific distribution (such as a normal distribution). Otherwise, the solution must be performed through sampling iteration and other methods, resulting in high computational complexity. Robust optimization methods ensure that power system dispatch can effectively cope with extreme scenarios caused by random fluctuations in renewable energy sources, ensuring system operational safety in the worst-case scenario. However, these methods are relatively conservative and struggle to balance system economics. Furthermore, they are generally non-convex, two-level optimization problems with high computational complexity. In summary, while uncertain dispatch decision-making methods can account for the impact of renewable energy uncertainty, these methods are generally computationally complex and struggle to meet the computational time requirements of actual power system dispatch decisions.

[0004] To meet the computational time requirements for dispatch decisions, the industry currently generally adopts deterministic dispatch decision-making methods with low computational burden. These methods address the random fluctuations of renewable energy sources by incorporating system frequency regulation capacity requirements into dispatch decisions, requiring units to reserve sufficient frequency regulation capacity. This ensures that participating units have ample output adjustment margin to account for the random fluctuations of renewable energy sources during actual operation. The frequency regulation capacity requirement is set so that the unit reserve capacity is greater than the maximum deviation of the system net load from the unit's planned output, ensuring that the units have sufficient operating margin to account for the random fluctuations of renewable energy sources. The industry typically sets frequency regulation capacity requirements based on operational experience, such as setting the frequency regulation capacity requirement at 2%-5% of the peak load for each period or at a fixed ratio of the load and renewable energy forecast values. This method of quantifying frequency regulation capacity requirements is simple and easy to use, but it fails to account for the impact of the random fluctuations of renewable energy sources. To address this issue, scholars both domestically and internationally have proposed other improved methods, including those based on the distribution of renewable energy forecast errors and data-driven approaches. However, to address the impact of the random fluctuations of renewable energy, existing methods primarily focus on the system's frequency regulation capacity requirements to ensure system regulation margin, without considering the impact of short-term system power fluctuations. In fact, even if sufficient frequency regulation capacity is reserved for the units, if the regulation speed of the units participating in frequency regulation cannot effectively keep up with the short-term fluctuations of renewable energy, the system will be unable to effectively maintain real-time power balance, threatening the safe operation of the grid. Therefore, in addition to the system frequency regulation capacity requirements, it is necessary to introduce the maximum short-term power fluctuation of the system net load to ensure that the regulation speed of the units participating in frequency regulation can effectively cope with short-term net load power fluctuations. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems in the prior art, namely, the lack of a reliable and effective method for power system frequency regulation and the lack of consideration for short-term power fluctuations in the traditional power system frequency regulation strategy.

[0006] To this end, a first aspect of the present invention provides a method for quantifying frequency regulation speed requirements of a power system.

[0007] A second aspect of the present invention provides a computer device.

[0008] A third aspect of the present invention provides a computer-readable storage medium.

[0009] The present invention provides a method for quantifying frequency regulation speed requirements of a power system, comprising:

[0010] Constructing a sample feature vector for predicting the maximum short-term fluctuation of net load power based on power grid operation data; wherein the power grid operation data includes load data, renewable energy output data, meteorological data, and special date data; the maximum short-term fluctuation of net load power is defined as the maximum change value of net load per unit time within a fixed scheduling period; the sample feature vector includes an input feature vector and an output feature vector;

[0011] The training samples and sample feature vectors are screened based on the similarity between the training samples and the prediction day samples and the matching degree between the input feature vectors of the training samples and the prediction day samples;

[0012] Construct a prediction model for the maximum fluctuation point of net load short-term power, and use the screened training samples and sample feature vectors to train the prediction model for the maximum fluctuation point of net load short-term power;

[0013] The predicted value of the maximum fluctuation point of the net load short-time power of the system output by the net load short-time power maximum fluctuation point prediction model is input into the net load short-time power maximum fluctuation interval prediction model to obtain the confidence interval of the net load short-time power maximum fluctuation prediction result;

[0014] The upper limit of the confidence interval of the prediction result of the maximum short-term power fluctuation of the net load is used as the system frequency regulation speed requirement.

[0015] The method for quantifying the frequency regulation speed demand of a power system according to the above technical solution of the present invention may also have the following additional technical features:

[0016] In the above technical solution, the input feature vector includes: load power characteristics, new energy power characteristics, net load power characteristics, meteorological factor characteristics and special date information;

[0017] Wherein, the special date information includes holiday type;

[0018] The output characteristic vector includes the maximum fluctuation value of the net load short-term power on the forecast day.

[0019] In the above technical solution, the screening of training samples includes:

[0020] The improved weighted Euclidean distance is used to quantify the matching degree between different samples and the predicted day sample:

[0021]

[0022]

[0023] Among them, q represents the load power characteristics, new energy power characteristics, net load power characteristics, and net load short-term power fluctuation characteristics of the day before the forecast; n represents the meteorological factor characteristics of the forecast day; r represents the holiday type of the forecast day; l Indicates the sample number; Represents the weighted Euclidean distance of each type of feature; a and b are two w-dimensional sample data respectively, z represents the weight of each dimension feature, and the subscript is the data dimension number; The final Euclidean distance calculation result of the two samples;

[0024] The L samples with the smallest Euclidean distance are used as training samples.

[0025] In the above technical solution, the screening of sample feature vectors includes:

[0026] Quantify the matching degree of each dimension feature between the training sample and the sample to be predicted:

[0027]

[0028]

[0029] in, Represents the training sample weight constructed by the normalized Euclidean distance between the training sample and the sample to be predicted. The smaller the Euclidean distance, the greater the training sample weight, and the greater the impact on the quantization result of the input feature vector matching; 、 They represent the maximum and minimum values ​​of the Euclidean distance of the training samples after screening; L represents the total number of training samples after screening; p represents the input feature dimension; Indicates the l The Euclidean distance between different input feature vectors in the sample and the corresponding input feature vector on the forecast day; Indicates the matching degree of the p-th dimension feature between the final quantized training sample and the sample to be predicted;

[0030] The T feature vectors with the highest matching degree are selected as the final input feature vectors.

[0031] In the above technical solution, the net load short-term power maximum fluctuation range prediction model includes:

[0032]

[0033]

[0034] in, represents the adaptive bandwidth of the i-th sample point; represents the global baseline bandwidth; represents the adjustment parameter; g(x) represents the density estimate at x; f(x) is the probability density function; K represents the kernel function; n represents the number of samples; represents the i-th sample point predicted by the net load short-term power maximum fluctuation point prediction model;

[0035]

[0036] Among them, the cumulative distribution function F(x) represents the probability that the random variable X is less than or equal to a specific value x;

[0037] The PPF function is obtained based on the inverse function of the cumulative distribution function F(x):

[0038]

[0039] Where b represents the quantile corresponding to the cumulative probability r, and r represents the cumulative probability.

[0040] In the above technical solution, the net load short-term power maximum fluctuation range prediction model includes:

[0041]

[0042] in, f(x) is the probability density function; h represents fixed bandwidth; K represents kernel function; n represents the number of samples; represents the i-th sample point predicted by the net load short-term power maximum fluctuation point prediction model;

[0043]

[0044] Among them, the cumulative distribution function F(x) represents the probability that the random variable X is less than or equal to a specific value x;

[0045] The PPF function is obtained based on the inverse function of the cumulative distribution function F(x):

[0046]

[0047] in, b Represents the cumulative probability corresponding to r The quantile of r represents the cumulative probability.

[0048] In the above technical solution, a deep neural network is used to construct the net load short-term power maximum fluctuation point prediction model.

[0049] In the above technical solution, a data-driven model is used to construct the net load short-term power maximum fluctuation point prediction model;

[0050] The data-driven model includes an extreme learning machine, a decision tree, a long short-term memory network or a convolutional neural network.

[0051] The present invention also provides a computer device, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded and executed by the processor, a method for quantifying the frequency regulation speed demand of the power system as described in any one of the above technical solutions is implemented.

[0052] The present invention also provides a computer-readable storage medium storing a program, which, when loaded by a processor, implements the method for quantifying the frequency regulation speed demand of a power system as described in any one of the above technical solutions.

[0053] In summary, due to the adoption of the above technical features, the beneficial effects of the present invention are:

[0054] Based on the load, new energy, and meteorological information that can actually be obtained from the power grid, the present invention constructs a sample feature vector for predicting the maximum fluctuation of net load short-term power. At the same time, the training samples and sample feature vectors are screened, and then a net load short-term power maximum fluctuation point prediction model is established, which realizes the accurate prediction of the maximum fluctuation of net load short-term power. In order to ensure the adequacy of the frequency modulation speed demand, a net load short-term power maximum fluctuation interval prediction model based on adaptive bandwidth kernel density estimation (ABKDE) is further established on the basis of the point prediction model. The confidence interval of the prediction result of the maximum fluctuation of net load short-term power can be obtained. The upper limit of the interval is used as the maximum fluctuation of the net load short-term power of the system to ensure that it can cover the actual maximum fluctuation of the net load short-term power. The method proposed by the present invention can accurately quantify the maximum fluctuation of the net load short-term power of the system, providing a powerful reference for the rational formulation of the day-ahead scheduling decision plan.

[0055] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0057] Figure 1 is a flow chart of a method for quantifying frequency regulation speed requirements of a power system according to an embodiment of the present invention;

[0058] Figure 2 2 is a schematic diagram of the deep neural network structure used in the method for quantifying the frequency regulation speed demand of the power system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0060] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0061] Refer to the following Figure 1 and Figure 2 The following describes a method for quantifying frequency regulation speed requirements of a power system according to some embodiments of the present invention.

[0062] Some embodiments of the present application provide a method for quantifying frequency regulation speed requirements of a power system.

[0063] The first embodiment of the present invention proposes a method for quantifying frequency regulation speed requirements of a power system, including steps S1-S5.

[0064] S1. Construct a sample feature vector for predicting the maximum short-term fluctuation of net load power based on the grid operation data; wherein the grid operation data includes load data, new energy output data, meteorological data and special date data; the maximum short-term fluctuation of net load power is defined as the maximum change value of net load per unit time within a fixed scheduling period; the sample feature vector includes an input feature vector and an output feature vector.

[0065] Specifically, the input feature vector should include the main factors affecting the maximum fluctuation of the system's net load short-term power. In some embodiments, the input feature vector includes: load power characteristics, new energy power characteristics, net load power characteristics, meteorological factor characteristics, and special date information; the special date information includes holiday types. It is understood that special dates refer to specific dates that are different from regular dates, such as holidays and major event days. The output feature vector includes the maximum fluctuation value of the net load short-term power on the forecast day.

[0066] In a specific embodiment, the fixed scheduling period can be set to 15 minutes, and the unit duration can be set to 1 minute. Then, the maximum short-term fluctuation of the net load power is defined as the maximum change in the net load within the 15-minute scheduling period. It should be noted that the fixed scheduling period and unit duration can be flexibly set as needed.

[0067] In one specific example, load and renewable energy operational data for a specific province in China were collected from January to December 2023. Net load data for regions with high renewable energy penetration was obtained by proportionally scaling the renewable energy data. To account for the impact of meteorological factors on the maximum short-term fluctuations in the system's net load power, various meteorological values ​​for the region were also collected. The impact of special dates, such as holidays, on the forecast results was also considered, and the holiday type of the forecasted day was marked.

[0068] Specifically, to predict the maximum short-term power fluctuation of net load in the next day, an input feature vector with a granularity of 15 minutes is constructed based on the data with a fine granularity of 1 minute. The constructed input feature vector and output feature vector are shown in Table 1.

[0069]

[0070] Among them, the input feature vector includes meteorological factors, date, and historical new energy, load, and net load fluctuation characteristics, totaling 2689 dimensions; the output feature vector is the maximum short-term power fluctuation of the net load on the forecast day, totaling 96 dimensions.

[0071] S2. Screening the training samples and sample feature vectors based on the similarity between the training samples and the prediction day samples and the matching degree between the input feature vectors of the training samples and the prediction day samples.

[0072] Specifically, the prediction accuracy of the net load short-term power maximum fluctuation prediction model is closely related to the similarity between the training samples and the prediction day samples, as well as the degree of matching between the input features of the training samples and the prediction day samples. To this end, in step S2, the sample feature matching degree is first quantified and calculated, and samples with high feature matching are selected as training samples. Then, the matching degree between the input features of the training samples and the prediction day samples is quantified, and features with high matching degrees are selected as the final sample input features.

[0073] In some embodiments, screening of training samples includes:

[0074] The improved weighted Euclidean distance is used to quantify the matching degree between different samples and the predicted day sample:

[0075]

[0076]

[0077] Among them, q represents the load power characteristics, new energy power characteristics, net load power characteristics, and net load short-term power fluctuation characteristics of the day before the forecast; n represents the meteorological factor characteristics of the forecast day; r represents the holiday type of the forecast day; l Indicates the sample number; represents the weighted Euclidean distance of each type of feature; a and b are two w-dimensional sample data respectively, z represents the weight of each dimension feature (obtained by the Pearson correlation coefficient), and the subscript is the data dimension number; The final Euclidean distance calculation result of the two samples;

[0078] The L samples with the smallest Euclidean distance (highest matching degree) are used as training samples.

[0079] In the aforementioned training sample screening method, the weighted Euclidean distance is calculated for each feature type separately. The final Euclidean distance between two samples is then combined to calculate the Euclidean distance between each feature type. This is because the dimensions of features such as renewable energy, load, and meteorological factors are much higher than the dimensions of the holiday type on the forecast day. Traditional Euclidean distance tends to favor the larger-dimensional fluctuations of renewable energy, load, and net load, while easily overlooking the holiday type on the forecast day. The improved Euclidean distance calculation method described above can effectively alleviate this problem.

[0080] In some embodiments, using data that is most similar to the training day sample data for training can improve the model's training effect while saving model training time. The input feature vector constructed in this embodiment has a dimension of 2689. There may be redundancy in these input features or features that do not match the prediction day sample data well, which will mislead the model training process, increase the model training time, and reduce the model prediction accuracy. To this end, this embodiment further screens the input features of the training samples to ensure that the input features ultimately used for training can maximize the effect of model training. Specifically, the screening of sample feature vectors includes:

[0081] Quantify the matching degree of each dimension feature between the training sample and the sample to be predicted:

[0082]

[0083]

[0084] in, Represents the training sample weight constructed by the normalized Euclidean distance between the training sample and the sample to be predicted. The smaller the Euclidean distance, the greater the training sample weight, and the greater the impact on the quantization result of the input feature vector matching; 、 They represent the maximum and minimum values ​​of the Euclidean distance of the training samples after screening; L represents the total number of training samples after screening; p represents the input feature dimension; Indicates the l The Euclidean distance between different input feature vectors in the sample and the corresponding input feature vector on the forecast day; Indicates the matching degree of the p-th dimension feature between the final quantized training sample and the sample to be predicted;

[0085] The T feature vectors with the highest matching degree are selected as the final input feature vectors.

[0086] Finally, L training samples with T-dimensional input features are formed, and the input feature dimension of the sample to be predicted is also reduced to T dimensions.

[0087] In some embodiments, the meteorological factors, new energy fluctuation characteristics, load fluctuation characteristics, and maximum short-term power fluctuation of the net load contained in the training samples have different dimensions and large numerical differences, which is not conducive to the training of subsequent models. For this reason, the samples need to be normalized and preprocessed. The present disclosure uses the z-score method to normalize the input and output data of the samples. The z-score normalization method uses the sample mean and standard deviation for normalization. The data processed by z-score meets the mean of 0 and the standard deviation of 1. The specific normalization method is:

[0088]

[0089] in, is the normalized sample value, is the sample value to be normalized, is the sample mean, It should be noted that sample preprocessing is usually performed before step S2.

[0090] S3. Construct a net load short-time power maximum fluctuation point prediction model, and use the screened training samples and sample feature vectors to train the net load short-time power maximum fluctuation point prediction model.

[0091] In some embodiments, a DNN (deep neural network) is used to construct a prediction model for the point of maximum short-term net load power fluctuation. Deep neural networks exhibit excellent prediction performance due to their powerful nonlinear feature extraction capabilities, efficient processing of large-scale data, and robustness in noisy data environments. Furthermore, DNN computation time can also meet scheduling requirements. Therefore, the present invention selects a DNN to construct a prediction model for maximum short-term net load power fluctuation. It should be noted that the choice of data-driven model to construct the point prediction model is not the focus of this invention; other data-driven models (such as extreme learning machines, decision trees, long short-term memory networks, and convolutional neural networks) can also be used to predict maximum short-term net load power fluctuation.

[0092] The trained DNN can be used to predict the specific value of the maximum short-term power fluctuation of the system net load in each scheduling period on the operating day.

[0093] In a specific embodiment, the deep neural network structure used in the present disclosure is as follows: Figure 2 As shown in Figure 1, it includes an input layer, multiple hidden layers, and an output layer. At the same time, the layers are connected together in a fully connected manner. The data transfer formula between layers is:

[0094]

[0095] in, Representative The output of the neurons in the layer, , the input x is expressed as ; To connect Layer and The weight matrix of the layer neurons; For the The output of the last layer of neurons (i.e. the final output of the neural network) is expressed as follows:

[0096]

[0097] Here, function s(x) is the activation function. In the hidden layer, this paper uses LReLU (Leaky Rectified Linear Unit) as the activation function. The LReLU activation function is a piecewise function. When the input value x is less than or equal to zero, the output value is equal to 0.01x; when the input value x is greater than zero, the output value is equal to the input value x.

[0098]

[0099] In the output layer, a linear activation function is used, which is expressed as follows:

[0100]

[0101] DNN training uses labeled samples to adjust parameters , so that the loss function Minimum. The loss function selected in this disclosure is the mean square error loss function, as shown below:

[0102]

[0103] in, is the number of samples; m is the sample number; is the output value of DNN; y is the actual value of the sample.

[0104] This disclosure uses a learning algorithm that combines root mean square propagation and learning rate decay to train DNN models. The root mean square propagation algorithm can maintain the moving average of the squared gradient of the parameters, allowing the step size of each parameter to be adaptively updated, thereby accelerating the convergence of the parameters. In addition, the learning rate decay strategy can gradually reduce the learning rate as training progresses, avoiding algorithm oscillation. The parameter update process of the learning algorithm used in this disclosure is as follows:

[0105]

[0106]

[0107]

[0108]

[0109]

[0110] Where, It is The gradient of θ at the time of update; It is in Parameters before the update; It is in The loss function for the update θ The partial derivative of It is in The moving average of the squared gradient of the parameters at the update; ρ is the decay rate, and its value is set to 0.9; ⊙ represents the Hadamard product; It is in The amount of change in the parameter during the update; It is in The learning rate at the update time; δ It is to prevent And set a very small constant; It is in The updated parameters; represents the initial learning rate, and its value is set to 0.001; ε Represents the decay rate of the learning rate, and its value is set to 0.95.

[0111] S4. Input the system net load short-time power maximum fluctuation point prediction value output by the net load short-time power maximum fluctuation point prediction model into the net load short-time power maximum fluctuation interval prediction model to obtain the confidence interval of the net load short-time power maximum fluctuation prediction result.

[0112] Through steps S1-S3, a point prediction is performed for the maximum short-term power fluctuation within the scheduling period. However, due to data-driven errors, the resulting point prediction of the maximum net load short-term power fluctuation may be greater than the actual maximum net load short-term power fluctuation, leading to insufficient reserved capacity for system ancillary services and reduced system frequency performance. Therefore, to prevent the point prediction value from being less than the actual value, the present disclosure uses an interval prediction model for the maximum net load short-term power fluctuation to perform interval prediction. If the upper limit of the interval prediction is selected as the frequency regulation speed requirement, the situation where the predicted maximum net load short-term power fluctuation value is less than the actual value can be greatly avoided.

[0113] In some embodiments, KDE (a non-parametric method for estimating probability density functions) is selected to perform smoothing on each sample point and to affect the surrounding space. This effect is quantified by a kernel function. The KDE estimated density function is:

[0114]

[0115] in, f(x) is the probability density function; h represents the fixed bandwidth; K represents the kernel function; n represents the number of samples; It represents the i-th sample predicted by the prediction model of the maximum fluctuation point of net load short-term power.

[0116] However, while KDE can estimate the probability density of sample points to a certain extent, when the data distribution is uneven, the fixed bandwidth h in KDE may lead to problems such as oversmoothing or overfitting. To address this problem, in some embodiments, ABKDE is adopted, which introduces an adaptive bandwidth mechanism so that the bandwidth varies according to the local density of data points. That is, in areas with sparse data, the bandwidth is increased to reduce the variance of the estimate; while in areas with dense data, the bandwidth is reduced to reduce the bias of the estimate. The ABKDE estimated density function is:

[0117]

[0118]

[0119] in, represents the adaptive bandwidth of the i-th sample point; represents the global baseline bandwidth; Indicates adjustment parameters; g(x) represents the density estimate at x; f(x) is the probability density function; K represents the kernel function; n represents the number of samples; It represents the i-th sample point predicted by the prediction model of the maximum short-term power fluctuation point of the net load.

[0120] On this basis, in order to ensure that the predicted value is slightly larger than the actual value of the actual operation scheduling point, this disclosure introduces the Cumulative Distribution Function (CDF) and the Percent Point Function (PPF).

[0121] The cumulative distribution function F(x) represents the probability that the random variable X is less than or equal to a specific value x. Its expression is:

[0122]

[0123] CDF is a non-decreasing function with a value between 0 and 1. The PPF function is obtained from the inverse function of the cumulative distribution function F(x):

[0124]

[0125] Where b represents the quantile corresponding to the cumulative probability r, and r represents the cumulative probability. r is usually a value between 0 and 1. For example, with a 95% confidence interval, r is 0.975. b represents the predicted value with a 95% confidence interval.

[0126] S5. The upper limit of the confidence interval of the prediction result of the maximum short-term power fluctuation of the net load is used as the system frequency regulation speed requirement.

[0127] Through the above steps S1-S5. The present disclosure screens the training sample data based on the improved Euclidean distance and obtains the weight, introduces it into the input feature screening composed of 2689-dimensional feature vectors, and proposes a DNN-based net load short-term power maximum fluctuation point prediction model. While ensuring that the scheduling time allows, the powerful nonlinear feature extraction capability of the DNN model is utilized to achieve point prediction of the maximum short-term power fluctuation. Then, combined with the ABKDE maximum short-term power fluctuation interval prediction method, the upper limit of the net load short-term power maximum fluctuation under a certain confidence level can be obtained. If the predicted net load short-term power maximum fluctuation is taken as the upper limit value, the actual value can be greatly smaller than the predicted value, which greatly improves the safety and economy of the power system. Therefore, the present disclosure can be widely used in the day-ahead scheduling decision of the power system, provide a reference for the quantification of the system frequency regulation speed demand, and ensure that the system scheduling decision plan can effectively respond to the net load power fluctuation.

[0128] In a specific embodiment, the actual operation data of a certain power grid (historical frequency regulation mileage and meteorological data, etc.) is used to verify the effectiveness of the proposed point prediction model and the interval-based prediction model. This embodiment adopts a sliding window prediction mode to perform load forecasting, that is, a corresponding load forecasting model is trained for each day to be predicted, and the training sample data is fixed. For example: if the data on February 11 is used as the sample to be predicted, the available data includes all data from February 1 to February 10 as input samples. If the data on February 12 is predicted, the available data is all data from February 2 to February 11, and so on.

[0129] The net load short-time power maximum fluctuation point prediction model is referred to as the DNN data-driven prediction model, and the net load short-time power maximum fluctuation interval prediction model is referred to as the interval probability prediction model ABKDE.

[0130] The training method comparison is mainly based on the following schemes M1-M6:

[0131] M1: DNN data-driven prediction model, which uses improved Euclidean distance to screen training samples and input features, and selects 60 samples that are most similar to the input data of the prediction day as training samples.

[0132] M2: A DNN data-driven prediction model that has not undergone training sample screening. It selects the 60 days adjacent to the day to be predicted as training data and performs feature screening.

[0133] M3: DNN data-driven prediction model without input feature screening, the 60 samples most similar to the input data on the prediction day are selected as training samples.

[0134] M4: Add interval probability prediction model ABKDE based on M1.

[0135] M5: Based on M2, the interval probability prediction model ABKDE is added.

[0136] M6: Add interval probability prediction model ABKDE based on M3.

[0137] In order to verify the effectiveness of the net load short-term power maximum fluctuation interval prediction method M4 proposed in this disclosure, this embodiment will use the evaluation indicators R2, MAE and RMSE to discuss the accuracy of the deterministic prediction model, and use MAPE, PICP and PINAW to discuss the interval probability prediction model.

[0138] Root Mean Square Error (RMSE):

[0139]

[0140] RMSE is a standardized measure of error. Larger errors will be penalized more severely due to the square term. The smaller the value, the better.

[0141] Mean Absolute Percentage Error (MAPE):

[0142]

[0143] MAPE is the average value of the ratio of the absolute value of the prediction error to the true value. It is used to measure the accuracy of the prediction. The smaller the value, the better.

[0144] Prediction Interval Coverage Probability (PICP):

[0145]

[0146] Where, L n and U n are the upper and lower bounds of the nth predicted value, respectively. I is the indicator function, ranging from 0% to 100%. The larger the value, the better.

[0147] Prediction Interval Normalized Average Width (PINAW):

[0148]

[0149] Where, are the maximum and minimum values ​​of the true value, respectively. PINAW is an indicator that measures the width of the prediction interval. It aims to balance the accuracy and reliability of the prediction. The smaller the value, the better.

[0150] This embodiment uses a point prediction model constructed by DNN to predict the maximum fluctuation demand of the system's net load short-term power on February 11, 2023. The average RMSE and MAPE of M1-M3 on February 11, 2023 are shown in Table 2.

[0151] Table 2 Prediction accuracy of the maximum fluctuation point of short-term power of M1-M3 net load

[0152]

[0153] As can be seen from Table 2, the MAPE and RMSE of the comparison methods M1-M3 and M2-M3 are respectively greater than those of M1 after the training samples and input features are screened, verifying the effectiveness of screening the training samples and input features in improving the accuracy of the prediction model. This shows that the DNN prediction model that combines input features and training sample screening can accurately predict the maximum short-term power fluctuation of the net load. Despite this, in the 96 scheduling period points throughout the day on February 11, 2023, the maximum short-term power fluctuation of the net load predicted by method M1 still did not meet the requirement of being greater than the actual maximum short-term power fluctuation of the net load in nearly 50% of the scheduling period points, which cannot effectively guarantee the safety of the system.

[0154] Therefore, in order to ensure that the predicted value of the maximum short-term fluctuation of net load power in 96 scheduling periods is greater than the actual value as much as possible, the ABKDE interval probability prediction model is introduced. In order to reserve sufficient maximum short-term fluctuation of net load power for the system, the present invention adopts the interval prediction upper limit as the frequency modulation speed requirement. The PINAW and PICP values ​​of methods M4-M6 are shown in Table 3.

[0155] Table 3 Interval prediction accuracy of M4-M6

[0156]

[0157] Table 3 shows that, compared with methods M4 and M6, method M4 achieves a PICP of 100% for the frequency regulation speed demand, and its PINAW value is smaller than that of methods M5 and M6. This indicates that while method M4 ensures that the overall system frequency regulation speed demand on February 11th exceeds the actual value, the average frequency regulation speed forecast across 96 dispatching times is also smaller than that of methods M5 and M6. This demonstrates the impact of point prediction accuracy on interval probabilistic prediction. The higher the point prediction accuracy, the more reliable the resulting frequency regulation speed forecast is in terms of system security, and the narrower the average speed width. In contrast, the PICP values ​​of methods M5 and M6, due to insufficient point prediction accuracy, do not reach 100% for the interval prediction upper limit value exceeding the actual value within the 96 dispatching time periods. Consequently, in some dispatching time periods, the predicted maximum short-term power fluctuation does not cover the actual value, resulting in insufficient system reserved frequency regulation capacity, which in turn causes the system frequency to drop below normal.

[0158] The above calculation results show that the interval prediction quantification method based on the DNN point prediction model and ABKDE proposed in this disclosure can ensure that the upper limit value of the obtained prediction interval can effectively cover the true value of the maximum short-term power fluctuation of the net load, and the system frequency regulation speed requirement set based on the upper limit of the prediction interval can ensure the adequacy of the system setting requirements.

[0159] Some embodiments of the present invention also provide a computer device comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is loaded and executed by the processor, the method for quantifying the frequency regulation speed demand of the power system as described in any of the above embodiments is implemented.

[0160] Some embodiments of the present invention further provide a computer-readable storage medium storing a program, which, when loaded by a processor, implements the method for quantifying the frequency regulation speed demand of a power system as described in any of the above embodiments.

[0161] In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any suitable manner in any one or more embodiments or examples.

[0162] Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for quantifying power system frequency regulation speed requirements, characterized in that: include: Constructing a sample feature vector for predicting the maximum short-term fluctuation of net load power based on power grid operation data; wherein the power grid operation data includes load data, renewable energy output data, meteorological data, and special date data; the maximum short-term fluctuation of net load power is defined as the maximum change value of net load per unit time within a fixed scheduling period; the sample feature vector includes an input feature vector and an output feature vector; The training samples and sample feature vectors are screened based on the similarity between the training samples and the prediction day samples and the matching degree between the input feature vectors of the training samples and the prediction day samples; Construct a prediction model for the maximum fluctuation point of net load short-term power, and use the screened training samples and sample feature vectors to train the prediction model for the maximum fluctuation point of net load short-term power; The predicted value of the maximum fluctuation point of the net load short-time power of the system output by the net load short-time power maximum fluctuation point prediction model is input into the net load short-time power maximum fluctuation interval prediction model to obtain the confidence interval of the net load short-time power maximum fluctuation prediction result; The upper limit of the confidence interval of the prediction result of the maximum short-term power fluctuation of the net load is used as the system frequency regulation speed requirement; The net load short-term power maximum fluctuation interval prediction model includes: in, represents the adaptive bandwidth of the i-th sample point; represents the global baseline bandwidth; Indicates adjustment parameters; g(x) represents the density estimate at x; f(x) is the probability density function; K represents the kernel function; n represents the number of samples; represents the i-th sample point predicted by the net load short-term power maximum fluctuation point prediction model; Among them, the cumulative distribution function F(x) represents the probability that the random variable X is less than or equal to a specific value x; The PPF function is obtained based on the inverse function of the cumulative distribution function F(x): Where b represents the quantile corresponding to the cumulative probability r, and r represents the cumulative probability.

2. The method for quantifying power system frequency regulation speed demand according to claim 1, characterized in that: The input feature vector includes: load power characteristics, new energy power characteristics, net load power characteristics, meteorological factor characteristics and special date information; Wherein, the special date information includes holiday type; The output characteristic vector includes the maximum fluctuation value of the net load short-term power on the forecast day.

3. The method for quantifying power system frequency regulation speed demand according to claim 2, characterized in that: The screening of training samples includes: The improved weighted Euclidean distance is used to quantify the matching degree between different samples and the predicted day sample: Among them, q represents the load power characteristics, new energy power characteristics, net load power characteristics, and net load short-term power fluctuation characteristics of the day before the forecast; n represents the meteorological factor characteristics of the forecast day; r represents the holiday type of the forecast day; l Indicates the sample number; Represents the weighted Euclidean distance of each type of feature; a and b are two w-dimensional sample data respectively, z represents the weight of each dimension feature, and the subscript is the data dimension number; The final Euclidean distance calculation result of the two samples; The L samples with the smallest Euclidean distance are used as training samples.

4. The method for quantifying power system frequency regulation speed demand according to claim 3, characterized in that: The screening of sample feature vectors includes: Quantify the matching degree of each dimension feature between the training sample and the sample to be predicted: in, Represents the training sample weight constructed by the normalized Euclidean distance between the training sample and the sample to be predicted. The smaller the Euclidean distance, the greater the training sample weight, and the greater the impact on the quantization result of the input feature vector matching; 、 They represent the maximum and minimum values ​​of the Euclidean distance of the training samples after screening; L represents the total number of training samples after screening; p represents the input feature dimension; Indicates the l The Euclidean distance between different input feature vectors in the sample and the corresponding input feature vector on the forecast day; Indicates the matching degree of the p-th dimension feature between the final quantized training sample and the sample to be predicted; Select the T feature vectors with the highest matching degree as the final input feature vectors.

5. The method for quantifying power system frequency regulation speed demand according to claim 1, characterized in that: A deep neural network is used to construct the net load short-term power maximum fluctuation point prediction model.

6. The method for quantifying power system frequency regulation speed demand according to claim 1, characterized in that: A data-driven model is used to construct a prediction model for the maximum fluctuation point of the net load short-term power; The data-driven model includes an extreme learning machine, a decision tree, a long short-term memory network or a convolutional neural network.

7. A computer device, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded and executed by the processor, the method for quantifying the frequency regulation speed demand of the power system according to any one of claims 1 to 6 is implemented.

8. A computer-readable storage medium, characterized in that A program is stored, and when the program is loaded by a processor, the method for quantifying the frequency regulation speed demand of the power system according to any one of claims 1 to 6 is implemented.

Images