Power distribution network risk assessment method based on photovoltaic output prediction error

Abstract

The invention discloses a power distribution network risk assessment method based on a photovoltaic output prediction error, and the method comprises the steps: constructing an isolated tree of an isolated Sessen algorithm, carrying out the fusion processing of the historical data of mass distributed photovoltaic output, and correcting the abnormal value of the photovoltaic output; a probability distribution model of a photovoltaic output prediction error is constructed based on a photovoltaic output typical scene, a power distribution network photovoltaic typical output scene is divided based on an adaptive K-means clustering algorithm, and distributed photovoltaic output is subjected to ultra-short-term prediction by adopting BiLSTM according to meteorological conditions to obtain a photovoltaic output prediction curve. Carrying out modeling on the photovoltaic output prediction error by adopting KDE to obtain a probability distribution curve of the photovoltaic output prediction error; in combination with a semi-invariant method and a Gram-Charger expansion series method, probability distribution of voltage of each node of the power distribution network is obtained through probabilistic load flow calculation; and calculating the node voltage out-of-limit probability of the power distribution network, and evaluating the voltage out-of-limit risk of the power distribution network. According to the method, the key uncertainty factors in the change of the running state of the power distribution network can be determined.

Description

Technical Field

[0001] The present invention relates to the field of power system risk assessment, and in particular to a distribution network risk assessment method based on photovoltaic output prediction error. Background Art

[0002] Under the guidance of the "dual carbon" goal, my country actively promotes the construction of a new power system with new energy as the main body, vigorously develops distributed new energy, and a high proportion of distributed power sources connected to the distribution network has gradually become a key feature of the future power system. However, the power output of photovoltaic power generation has obvious randomness and volatility, which is mainly affected by factors such as weather conditions and sunshine intensity. This uncertainty makes the prediction error of photovoltaic power generation an important challenge in the operation of the distribution network.

[0003] Existing distribution network risk assessment methods are usually based on historical data and traditional load forecasting models, and fail to fully consider the impact of photovoltaic power generation output forecasting errors on the stability and safety of distribution networks. Therefore, in the context of the gradual increase in photovoltaic power generation systems, traditional risk assessment methods often find it difficult to accurately assess the risk of distribution networks in actual operation, resulting in serious power supply problems or equipment overload in distribution networks when photovoltaic power generation fluctuates, which in turn affects the safe and stable operation of the power system. In addition, although the existing photovoltaic power generation prediction technology has made certain progress, it still faces the problems of low prediction accuracy and large error fluctuations. These problems have not been effectively solved in the risk assessment of distribution networks, resulting in inaccurate assessment of the actual operation risks of distribution networks. Therefore, a new method is urgently needed to comprehensively consider the photovoltaic power generation output forecasting error to more accurately assess the risk of distribution networks, thereby improving the stability and reliability of distribution networks. Summary of the invention

[0004] The purpose of the present invention is to provide a distribution network risk assessment method based on photovoltaic output prediction error, so as to ensure the voltage risk assessment accuracy of the distribution network when the photovoltaic output prediction error is disturbed and shorten the risk assessment time as much as possible.

[0005] The technical solution to achieve the purpose of the present invention is: a distribution network risk assessment method based on photovoltaic output prediction error, comprising the following steps:

[0006] Step 1: construct an isolated tree of the isolated forest algorithm, fuse and process the massive historical data of distributed photovoltaic output, and correct the abnormal values ​​of photovoltaic output;

[0007] Step 2: Based on the typical photovoltaic output scenario, a probability distribution model of photovoltaic output prediction error is constructed. The typical photovoltaic output scenario of the distribution network is divided based on the adaptive K-means clustering algorithm. The BiLSTM is used to perform ultra-short-term prediction of distributed photovoltaic output according to meteorological conditions to obtain the prediction curve of photovoltaic output. The photovoltaic output prediction error is modeled using KDE to obtain the probability distribution curve of photovoltaic output prediction error.

[0008] Step 3: Combining the semi-invariant and Gram-Charher expansion series methods, the probability distribution of the voltage at each node of the distribution network is obtained through probability flow calculation, including the node voltage probability density curve and probability distribution curve, to describe the uncertainty of the system in the form of probability distribution;

[0009] Step 4: Calculate the probability of voltage exceeding the limit at the nodes of the distribution network and evaluate the voltage exceeding the limit risk of the distribution network.

[0010] Furthermore, in step 1, the historical data of massive distributed photovoltaic output is fused and processed to correct the abnormal value of photovoltaic output. The specific method is as follows:

[0011] Step 1.1, construct an isolated tree of the isolated Senli algorithm, and isolate the abnormal photovoltaic output data points by gradually splitting the data points;

[0012] First, n sample data points are randomly selected from the massive photovoltaic output data as a sub-sampling set, and an initial iTree is constructed, and the sub-sampling set is placed in the root node of the tree; then an attribute dimension q is randomly specified from multiple dimensions of the data, and a cutting point p is randomly generated in the current node data; then a hyperplane is generated at this cutting point, and the data space of the current node is divided into two subspaces, and the data items with the specified q attribute value less than p are placed in the left subspace of the current node; otherwise, they are placed in the right subspace; the above steps are continuously recursively called until the subspace cannot be cut further or the iTree has reached the limited height;

[0013] Step 1.2, calculate the isolated path length of each output data point in the isolation forest, and obtain the corresponding anomaly score to detect abnormal data points;

[0014] For the test data x, substitute it into each iTree in the forest, calculate the depth h(x) it falls in each tree, calculate the average value of all h(x) as E(h(x)), combine the standard average search length l(n) and the harmonic number H(x), and calculate the abnormal score of the test data;

[0015] Set the standard average search length l(n) and harmonic number H(x) to: l(n)=2H(n-1)-[2(n-1) / n] (1)

[0017] H(i)=ln(i)+0.5772 (2)

[0018] The anomaly score of the data to be tested is:

[0019]

[0020] If u(x,n) is close to 1, the data is considered to be abnormal data. For abnormal data, the nearest neighbor interpolation method is used to correct it. After correcting the abnormal photovoltaic output data points, the abnormal values ​​in the photovoltaic output data will be replaced by the corrected values, and a good photovoltaic output data set that can be used for typical photovoltaic output scene division and prediction error modeling is obtained.

[0021] Furthermore, in step 2, based on the typical scenario of photovoltaic output, a probability distribution model of photovoltaic output prediction error is constructed. The specific method is as follows:

[0022] Step 2.1: Based on the adaptive K-means clustering algorithm, the typical photovoltaic output scenarios of the distribution network are divided to obtain the typical output curve of distributed photovoltaic.

[0023] The K-means algorithm uses the Euclidean distance to evaluate the similarity between two samples. It randomly selects k points in the photovoltaic output data as the initial cluster center, and continuously updates the cluster center by calculating the Euclidean distance from the photovoltaic output value at each moment to the center until the error function reaches and stabilizes at the minimum value.

[0024] Photovoltaic output value x in the sample i With x j The Euclidean distance calculation formula between is shown in formula (4):

[0025]

[0026] The size of the square criterion error function is used to determine whether the cluster center has reached stability. When the error function reaches and stabilizes at the minimum value, it means that the cluster center of the photovoltaic output has stopped changing. The calculation formula of the square criterion error function is shown in formula (5):

[0027]

[0028] In the formula, k is the number of clusters; t i is the number of samples in the i-th class; n i is the mean of the samples of the i-th class;

[0029] The size of the silhouette coefficient is used to measure whether the clustering effect of the typical photovoltaic output scene is good. Its calculation includes cohesion and separation. The silhouette coefficient calculation formula is:

[0030]

[0031] In the formula, a(i) is the cohesion within a cluster, which is defined as the average distance between data points of the same type; b(i) is the separation between clusters, which is defined as the average distance between a sample point and data points in an adjacent cluster. From formula (6), we can see that the maximum value of the silhouette coefficient is 1 and the minimum value is 0. The closer the silhouette coefficient is to 1, the farther the two clusters are separated, the denser the distribution within the cluster, and the better the clustering effect.

[0032] Step 2.2: BiLSTM is used to make ultra-short-term predictions of distributed photovoltaic output according to meteorological conditions to obtain a prediction curve of photovoltaic output. The predicted value is compared with the actual value of the typical output scenario to form a distribution map of the prediction error.

[0033] BiLSTM consists of a double-layer LSTM. According to the BiLSTM neural network model, the short-term prediction result of photovoltaic output is obtained;

[0034] In order to evaluate the training results of the neural network, the root mean square error RMSE and the mean square error MAE are used as the evaluation indicators of the neural network for photovoltaic output prediction. The expressions of RMSE and MAE are as follows:

[0035]

[0036]

[0037] Where N represents the total number of data points; represents the predicted value of photovoltaic output; f i Indicates the actual value of photovoltaic output;

[0038]

[0039] A day is divided into a moment every 15 minutes, so there are 96 moments in a day. The predicted value at each moment is compared with the actual value of the typical output scenario of the day to form a distribution diagram of the prediction error.

[0040] Step 2.3: Based on the distribution diagram of the prediction error, KDE is used to model the photovoltaic output prediction error and obtain the probability distribution curve of the photovoltaic output prediction error;

[0041] The expression of the kernel density estimate function is as follows:

[0042]

[0043] Where N represents the total number of data points; h represents the bandwidth of the kernel density function, which is used to control the width of the kernel function and thus affect the smoothness of the density estimation; K is the kernel function, which is a probability density function with an integral of 1.

[0044] Furthermore, in step 3, the probability flow of the distribution network is calculated by combining the semi-invariant and Gram-Charher expansion series method. The specific method is:

[0045] Step 3.1, use the deterministic power flow calculation method to give the power flow distribution of normal operation, so as to obtain the state variable X at the reference operation point and the sensitivity matrix S;

[0046] The node power equation of the power system is:

[0047]

[0048] Where P i , Q i are the injected active power and reactive power of node i respectively; V i ,θ i are the amplitude and angle of the voltage at node i respectively; G ij , B ij are the real and imaginary parts of the admittance matrix elements respectively;

[0049] In the stochastic power flow model, the node injection quantity includes the active power and reactive power injected by the node, and the state variables include the voltage amplitude and angle of the node, which can be expressed as:

[0050]

[0051] In the formula, S0, X0 are the expected values ​​of the node injection amount S and the state variable X respectively; ΔS, ΔX are random disturbances, which are regarded as random variables obeying a certain distribution;

[0052] Step 3.2: According to the known probabilistic description characteristics of the generator set and the load, calculate the origin moment or central moment of the injected power of each node and then find out its semi-invariants of each order;

[0053] For node load power, its random component is considered to be caused by the prediction error of photovoltaic output. This random component is described by a normally distributed random variable. When the load power changes according to a certain load curve, it is simulated by a discrete random distribution.

[0054] For the normally distributed injected power, the first-order semi-invariant is equal to the expected value, the second-order semi-invariant is the variance of the normal distribution, and the third to sixth-order semi-invariants are zero. The semi-invariant is a digital characteristic of a random variable, which is obtained by the moments of random variables not higher than the corresponding order.

[0055] The central distances of random variables of various orders are expressed as:

[0056]

[0057] In the formula, μ represents the expected value; p v represents the probability value;

[0058] Semi-invariant K of each order i Distance from center M i The relationship is as follows:

[0059]

[0060] Step 3.3: Use the Gram-Charlier series to expand the probability density function F(x) and probability distribution function f(x) of the voltage distribution, and use semi-invariants to represent the coefficients of the series expansion. The expansion is as follows:

[0061]

[0062] Where: F(x) is the probability density function; f(x) is the probability distribution function; g v are normalized semi-invariants of various orders, v = 1, 2, 3, 4, 5, 6; is the standard normal distribution density function; is the Hermite polynomial;

[0063] Step 3.4, use the probability distribution of the voltage at each node of the distribution network, including the node voltage probability density curve and the cumulative distribution curve, to describe the uncertainty of the power system in the form of probability distribution.

[0064] Furthermore, in step 4, the probability of voltage exceeding the limit at the node of the distribution network is calculated, and the specific method for evaluating the voltage exceeding the limit risk of the distribution network is as follows:

[0065] The probability distribution of the voltage at each node of the distribution network is obtained through probability flow calculation. The over-limit value of the node voltage is set. When the voltage exceeds this value, it can be considered that the distribution network is at risk of over-limit. The voltage over-limit probability is calculated based on the node voltage probability distribution. The calculation formula is as follows:

[0066]

[0067] Where U i,t represents the voltage of node i at time t; u max 、u min Respectively represent the upper and lower limits of the system voltage constraint; R(U i,t ) represents the probability of system voltage exceeding the limit at time t.

[0068] A distribution network risk assessment system based on photovoltaic output prediction error implements the distribution network risk assessment method based on photovoltaic output prediction error to achieve distribution network risk assessment based on photovoltaic output prediction error.

[0069] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the distribution network risk assessment method based on photovoltaic output prediction error is implemented to achieve distribution network risk assessment based on photovoltaic output prediction error.

[0070] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the control method of the three-phase motor driver with multiple working states is implemented, and the distribution network risk assessment method based on photovoltaic output prediction error is implemented to achieve distribution network risk assessment based on photovoltaic output prediction error.

[0071] Compared with the prior art, the present invention has the following technical advantages: ① BiLSTM can provide a relatively accurate prediction error through accurate photovoltaic output prediction. After combining with KDE, these errors can be modeled for detailed probability distribution, revealing the potential patterns and characteristics of the errors, which helps to more accurately evaluate the risks brought by photovoltaic output prediction errors to the operation of the distribution network; ② Comprehensively considering power constraints, current constraints, distributed photovoltaic output constraints and other constraints, combining semi-invariants and Gram-Charher expansion series methods to calculate the probability flow of the distribution network, and obtaining the probability density function and probability distribution function of the voltage distribution according to the probability flow calculation results, which reduces the amount of random power flow convolution operations, and has faster calculation speed and higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a flow chart of a distribution network risk assessment method based on photovoltaic output prediction error.

[0073] Figure 2 This is a typical output curve of distributed photovoltaics.

[0074] Figure 3 It is the residual distribution diagram of PV output prediction.

[0075] Figure 4 It is the probability distribution diagram of PV output prediction error.

[0076] Figure 5 It is the cumulative probability curve of voltage at 33 nodes of distribution network when the error is small.

[0077] Figure 6 It is the cumulative probability curve of voltage at 33 nodes of distribution network when the error is large. DETAILED DESCRIPTION

[0078] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0079] A distribution network risk assessment method based on photovoltaic output prediction error includes the following steps:

[0080] Step 1: fuse and process the massive historical data of distributed photovoltaic output to obtain output data after correcting the abnormal value of photovoltaic output.

[0081] Photovoltaic data was collected at 96 times a day for two years, among which there were many missing values ​​and zero values, which are collectively referred to as outliers. The Isolation Forest Algorithm is suitable for extracting abnormal data caused by equipment problems or human factors in photovoltaic output data. The Isolation Forest Algorithm regards sparse data points that are far away from denser groups as abnormal data. It has the advantages of high accuracy and low computational complexity, and is very suitable for data scenarios with massive data. Its main process is as follows:

[0082] Step 1.1: Build an isolation tree to isolate abnormal PV output data points as early as possible by gradually splitting the data points.

[0083] (1) Extract n sample data with or without replacement from the historical PV output data set, construct a data subset, and construct an initial binary tree iTree.

[0084] (2) A feature dimension q=96 is specified according to the daily time scale, and a cutting point p between the maximum and minimum values ​​of the feature q of the current node data is randomly selected in the current data.

[0085] (3) At point p, the current photovoltaic output data space is divided into two subspaces, left and right. Attribute value data items less than p are placed in the left subspace, and attribute value data items greater than p are placed in the right subspace.

[0086] (4) Repeat steps (2) and (3) until the subspace cannot be cut further or the iTree has reached the initially defined height limit.

[0087] Step 1.2: Calculate the isolated path length of each output data point in the isolation forest and obtain the corresponding anomaly score to detect abnormal data points.

[0088] For the test data x, substitute it into each iTree, use h(x) to represent the height it falls on each iTree, E(h(x)) is the average of all h(x), H(x) is the harmonic number, because the structure of iTree is similar to that of binary search tree, so the average path length l(n) of building a binary tree for n samples is:

[0089] l(n)=2H(n-1)-[2(n-1) / n](1)

[0090] H(i)=ln(i)+0.5772(2)

[0091] Define an abnormal score for the data to be tested as:

[0092]

[0093] The closer u(x,n) is to 1, the greater the probability that the data is abnormal data.

[0094] When u(x,n) is greater than 0.75, the data is regarded as an outlier and removed from the original photovoltaic output data set. The abnormal data is corrected using the nearest neighbor interpolation method, and the corrected value is used to replace the original outlier, resulting in a good photovoltaic output data set that can be used for typical photovoltaic output scene division and prediction error modeling.

[0095] Step 2: Based on typical photovoltaic output scenarios, a probability distribution model of photovoltaic output prediction error is constructed.

[0096] Based on the corrected historical data of photovoltaic output, the adaptive K-means clustering algorithm is used to cluster and obtain the typical scene of photovoltaic output; taking into account various meteorological conditions, BiLSTM is used to make ultra-short-term predictions of distributed photovoltaic output to obtain the ultra-short-term prediction curve of photovoltaic output, and form the initial distribution map of prediction error; finally, the probability distribution modeling of prediction error is carried out. The specific process is as follows:

[0097] Step 2.1: Based on the adaptive K-means clustering algorithm, the typical photovoltaic output scenarios of the distribution network are divided to obtain the typical output curve of distributed photovoltaics.

[0098] In the K-means algorithm, the Euclidean distance is generally used to evaluate the similarity between two samples. It randomly selects k points in the photovoltaic output data as the initial clustering center, and divides a day into a moment every 15 minutes, then k = 96. The clustering center is continuously updated by calculating the Euclidean distance from the photovoltaic output value at each moment to the center until the error function reaches and stabilizes at the minimum value.

[0099] Photovoltaic output value x in the sample i With x j The Euclidean distance calculation formula between is shown in formula (4):

[0100]

[0101] The size of the square criterion error function is used to determine whether the cluster center has reached stability. When the error function reaches and stabilizes at the minimum value, it means that the cluster center has stopped changing. The calculation formula of the square criterion error function is shown in formula (5):

[0102]

[0103] In the formula, k is the number of clusters; t iis the number of samples in the i-th class; n i is the mean of the samples of the i-th class;

[0104] The size of the silhouette coefficient can be used to measure whether the clustering effect of typical photovoltaic output scenarios is good. Its calculation includes cohesion and separation. The silhouette coefficient calculation formula is:

[0105]

[0106] In the formula, a(i) is the cohesion within a cluster, which is defined as the average distance between data points of the same type; b(i) is the separation between clusters, which is defined as the average distance between a sample point and data points in adjacent clusters. From formula (6), we can see that the maximum value of the silhouette coefficient is 1 and the minimum value is 0. The closer the silhouette coefficient is to 1, the farther the two clusters are separated, the denser the distribution within the cluster, and the better the clustering effect.

[0107] Step 2.2: Use BiLSTM to make ultra-short-term predictions of distributed photovoltaic output based on meteorological conditions such as solar radiation, temperature, and humidity to obtain a prediction curve for photovoltaic output. Compare the predicted value with the actual value of a typical output scenario to form a distribution map of the prediction error.

[0108] BiLSTM is composed of two layers of LSTM. According to the BiLSTM neural network model, the short-term prediction results of photovoltaic output are obtained. In order to evaluate the training results of the neural network, the root mean square error RMSE and the mean square error MAE are used as the evaluation indicators of the neural network for photovoltaic output prediction. The expressions of RMSE and MAE are as follows:

[0109]

[0110] Where N represents the total number of data points; represents the predicted value of photovoltaic output; f i Indicates the actual value of photovoltaic output;

[0111]

[0112] If a day is divided into a moment every 15 minutes, there are 96 moments in a day. The residual of the predicted value at each moment and the actual value of the typical output scenario of the day are calculated to form a distribution diagram of the prediction error.

[0113] Step 2.3: Based on the distribution diagram of the prediction error, KDE is used to model the photovoltaic output prediction error, and the probability distribution curve of the photovoltaic output prediction error is obtained. The expression of the kernel density estimation function is as follows:

[0114]

[0115] Where N is the total number of data points; h is the bandwidth of the kernel density function, which is used to control the width of the kernel function and thus affect the smoothness of the density estimation; K is the kernel function, which is a probability density function with an integral of 1;

[0116] Based on the above analysis, BiLSTM can provide relatively accurate prediction errors through accurate PV output prediction. After combining with KDE, these errors can be modeled with detailed probability distribution. The model is: a prediction error probability distribution curve close to the normal distribution.

[0117] Step 3: Calculate the probabilistic power flow of the distribution network by combining the semi-invariant and Gram-Charher expansion series method

[0118] Step 3.1: Use the deterministic power flow calculation method to give the power flow distribution of normal operation, so as to obtain the state variable X at the reference operation point and calculate the sensitivity matrix S:

[0119] The probabilistic power flow calculation method adopts a linearized model based on Newton-Raphson power flow calculation.

[0120] The node power equation of the power system is:

[0121]

[0122] Where P i , Q i are the injected active power and reactive power of node i respectively; V i ,θ i are the amplitude and angle of the voltage at node i respectively; G ij , B ij are the real and imaginary parts of the admittance matrix elements, respectively.

[0123] In the stochastic power flow model, the node injection quantity includes the active power and reactive power injected by the node, and the state variables include the voltage amplitude and angle of the node, which can be expressed as:

[0124]

[0125] Where S0 and X0 are the expected values ​​of the node injection amount S and the state variable X, respectively; ΔS and ΔX are random disturbances, which can be regarded as random variables obeying a certain distribution.

[0126] Step 3.2: According to the known probabilistic description characteristics of the generator set and load, calculate the origin moment or central moment of the injected power of each node and then find out its semi-invariants of each order:

[0127] For node load power, its random component can be regarded as caused by the prediction error of photovoltaic output. This random component can generally be described by a normally distributed random variable. When the load power changes according to a certain load curve, it can be simulated by a discrete random distribution.

[0128] For the normally distributed injected power, the first-order semi-invariant is equal to the expected value, the second-order semi-invariant is the variance of the normal distribution, and the third-order to sixth-order semi-invariants are zero. The semi-invariant is a digital characteristic of a random variable and can be obtained from the moments of random variables of the corresponding order or less.

[0129] The central distances of random variables of various orders are expressed as:

[0130]

[0131] In the formula, μ represents the expected value; p v represents the probability value;

[0132] According to formula (7), we can get the semi-invariant K of each order i Distance from center M i The relationship is as follows:

[0133]

[0134] Step 3.3: Use the Gram-Charlier series to expand the probability density function F(x) and probability distribution function f(x) of the voltage distribution, and use semi-invariants to represent the coefficients of the series expansion. The expansion is as follows:

[0135]

[0136]

[0137] Where: F(x) is the probability density function; f(x) is the probability distribution function; g v are normalized semi-invariants of various orders, v = 1, 2, 3, 4, 5, 6; is the standard normal distribution density function; is the Hermite polynomial;

[0138] Step 3.4: Consider various constraints and calculate the probabilistic power flow of the distribution network.

[0139] In the process of probabilistic power flow calculation, the following constraints are considered:

[0140] (1) Power Constraints:

[0141] In order to suppress the large power fluctuations of the distribution network, it is necessary to constrain the power of the i-th power source:

[0142]

[0143] In the formula, is the power injected into the system at time t; and They are the upper and lower power limits of the power injection system respectively.

[0144] (2) Distribution network branch current constraints:

[0145] I ij ≤I ij.max (18) Where I ij.max It is the upper limit of the branch current amplitude between node i and node j.

[0146] (3) Operation constraints of distributed photovoltaic power sources:

[0147]

[0148] In the formula, are the active power and reactive power of distributed photovoltaic power sources respectively; Indicates the maximum output of distributed photovoltaic power source; the power factor of distributed photovoltaic power source The value is 0.95.

[0149] Considering the randomness of the photovoltaic output prediction error under different typical output scenarios, the probability distribution of the voltage at each node in the distribution network is obtained by combining the semi-invariant and Gram-Charher expansion series methods, including the point voltage probability density curve and the cumulative distribution curve. That is, the probability distribution of the node voltage in the distribution network is used as the calculation result of the probabilistic power flow, and the uncertainty of the system is described in the form of voltage probability distribution, and the operating status of the system is analyzed accordingly.

[0150] Step 4: Calculate the probability of voltage exceeding the limit at the nodes of the distribution network and evaluate the voltage exceeding the limit risk of the distribution network.

[0151] The probability distribution of the voltage at each node of the distribution network is obtained through probability flow calculation. The over-limit value of the node voltage is set. When the voltage exceeds this value, it can be considered that the distribution network is over-limited. The voltage over-limit probability is calculated based on the node voltage probability distribution. The calculation formula is as follows:

[0152]

[0153] Where U i,t represents the voltage of node i at time t; u max 、u min Respectively represent the upper and lower limits of the system voltage constraint; R(U i,t ) represents the probability of system voltage exceeding the limit at time t.

[0154] Adaptive K-means algorithm is used to cluster the historical data of distributed photovoltaic output. Typical distributed photovoltaic output scenarios are as follows: Figure 2 As shown. According to the silhouette coefficient curve, when the number of clusters k = 4, the silhouette coefficient is closest to 1, indicating that the clustering effect of the four scenarios is the best. The distributed typical output curve obtained by division is compared with the actual photovoltaic output curve of the distribution network, and the four typical processing curves are named sunny scene, cloudy scene, rainy scene, and rainy and snowy scene.

[0155] The present invention uses the IEEE 33-node radial distribution network as a simulation example, connects distributed photovoltaic power sources at 12 nodes, and sets the power factor to 0.95. The sunny scene is selected as the typical scene of distributed photovoltaic output on the second day, and the residual distribution of photovoltaic output prediction using the BiLSTM neural network is as follows: Figure 3 shown.

[0156] Based on KDE, the probability density distribution model of photovoltaic output error is constructed, and the probability distribution of the prediction error is as follows: Figure 3 As shown by Figure 4 It can be seen that the KDE method can include most of the observation points, has high fitting accuracy, and can better present the probability distribution results of the prediction error.

[0157] According to the above-mentioned distributed photovoltaic power prediction error curve, the semi-invariant and Gram-Charher expansion series methods are used to calculate the probability flow of the distribution network, and the Monte Carlo (MTCR) method is used to calculate the probability flow as a comparison, where the number of iterations of the MTCR method is set to 5000.

[0158] During the operation of the distribution network, weather fluctuations may occur, causing the forecast information to be inconsistent with the current state. At this time, the forecast error of photovoltaic processing becomes an uncertain factor in the operation of the distribution network. If the forecast error is too large, the gap between the forecast value and the actual value of photovoltaic output will cause a large disturbance to the operation of the distribution network, causing fluctuations in the voltage of the distribution network nodes. The voltage over-limit probability and simulation calculation time of the distribution network node 33 when the error is large and the error is small are shown in Table 1. The cumulative probability distribution of the voltage at node 33 is shown in Figure 5 , Figure 6 shown.

[0159] Table 1 Comparison of voltage over-limit probability and simulation time when large error and small error are encountered

[0160]

[0161] As can be seen from Table 2, the magnitude of the photovoltaic output prediction error will have a direct impact on the probability of voltage over-limit in the distribution network. When the error is small, the voltage over-limit probability is only about 0.01, but when the error increases, the voltage over-limit probability increases by nearly 3 times, and the over-limit probability is close to 0.3; although the probability flow calculation accuracy of the semi-invariant method and Gram-Charher expansion series is slightly lower than that of the Monte Carlo method, the simulation time is greatly shortened. Therefore, the use of the semi-invariant method and Gram-Charher expansion series for probability flow calculation greatly saves the calculation time without much reduction in calculation accuracy, and is more suitable for short-term and ultra-short-term distribution network risk real-time assessment scenarios.

[0162] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A distribution network risk assessment method based on photovoltaic output prediction error, characterized in that: The following steps are involved: Step 1: construct an isolated tree of the isolated forest algorithm, fuse and process the historical data of massive distributed photovoltaic output, and correct the abnormal values ​​of photovoltaic output; Step 2: Based on the typical photovoltaic output scenario, a probability distribution model of photovoltaic output prediction error is constructed. The typical photovoltaic output scenario of the distribution network is divided based on the adaptive K-means clustering algorithm. The BiLSTM is used to perform ultra-short-term prediction of distributed photovoltaic output according to meteorological conditions to obtain the prediction curve of photovoltaic output. The photovoltaic output prediction error is modeled using KDE to obtain the probability distribution curve of photovoltaic output prediction error. Step 3: Combining the semi-invariant and Gram-Charher expansion series methods, the probability distribution of the voltage at each node of the distribution network is obtained through probability flow calculation, including the node voltage probability density curve and probability distribution curve, to describe the uncertainty of the system in the form of probability distribution; Step 4: Calculate the probability of voltage exceeding the limit at the nodes of the distribution network and evaluate the voltage exceeding the limit risk of the distribution network.

2. A distribution network risk assessment method based on photovoltaic output prediction error as claimed in claim 1, characterized in that: In step 1, the historical data of massive distributed photovoltaic output is fused and processed to correct the abnormal values ​​of photovoltaic output. The specific method is as follows: Step 1.1, construct an isolated tree of the isolated Senli algorithm, and isolate the abnormal photovoltaic output data points by gradually splitting the data points; First, n sample data points are randomly selected from the massive photovoltaic output data as a sub-sampling set, and an initial iTree is constructed, and the sub-sampling set is placed in the root node of the tree; then an attribute dimension q is randomly specified from multiple dimensions of the data, and a cutting point p is randomly generated in the current node data; then a hyperplane is generated at this cutting point, and the data space of the current node is divided into two subspaces, and the data items with the specified q attribute value less than p are placed in the left subspace of the current node; otherwise, they are placed in the right subspace; the above steps are continuously recursively called until the subspace cannot be cut further or the iTree has reached the limited height; Step 1.2, calculate the isolated path length of each output data point in the isolation forest, and obtain the corresponding anomaly score to detect abnormal data points; For the test data x, substitute it into each iTree in the forest, calculate the depth h(x) it falls in each tree, calculate the average value of all h(x) as E(h(x)), combine the standard average search length l(n) and the harmonic number H(x), and calculate the abnormal score of the test data; Set the standard average search length l(n) and the harmonic number H(x) to: l(n) = 2H(n-1)-[2(n-1) / n] (1)H(i)=ln(i)+0.5772(2) The anomaly score of the data to be tested is: If u(x,n) is close to 1, the data is considered to be abnormal data. For abnormal data, the nearest neighbor interpolation method is used to correct it. After correcting the abnormal photovoltaic output data points, the abnormal values ​​in the photovoltaic output data will be replaced by the corrected values, and a good photovoltaic output data set that can be used for typical photovoltaic output scene division and prediction error modeling is obtained.

3. A distribution network risk assessment method based on photovoltaic output prediction error as claimed in claim 1, characterized in that: In step 2, based on the typical scenario of photovoltaic output, a probability distribution model of photovoltaic output prediction error is constructed. The specific method is as follows: Step 2.1: Based on the adaptive K-means clustering algorithm, the typical photovoltaic output scenarios of the distribution network are divided to obtain the typical output curve of distributed photovoltaic. The K-means algorithm uses the Euclidean distance to evaluate the similarity between two samples. It randomly selects k points in the photovoltaic output data as the initial cluster center, and continuously updates the cluster center by calculating the Euclidean distance from the photovoltaic output value at each moment to the center until the error function reaches and stabilizes at the minimum value. Photovoltaic output value x in the sample i With x j The Euclidean distance calculation formula between is shown in formula (4): The size of the square criterion error function is used to determine whether the cluster center has reached stability. When the error function reaches and stabilizes at the minimum value, it means that the cluster center of the photovoltaic output has stopped changing. The calculation formula of the square criterion error function is shown in formula (5): In the formula, k is the number of clusters; t i is the number of samples in the i-th class; n i is the mean of the samples of the i-th category; The size of the silhouette coefficient is used to measure whether the clustering effect of the typical photovoltaic output scene is good. Its calculation includes cohesion and separation. The silhouette coefficient calculation formula is: In the formula, a(i) is the cohesion within a cluster, which is defined as the average distance between data points of the same type; b(i) is the separation between clusters, which is defined as the average distance between a sample point and data points in an adjacent cluster. From formula (6), we can see that the maximum value of the silhouette coefficient is 1 and the minimum value is 0. The closer the silhouette coefficient is to 1, the farther the two clusters are separated, the denser the distribution within the cluster, and the better the clustering effect. Step 2.2: BiLSTM is used to make ultra-short-term predictions of distributed photovoltaic output according to meteorological conditions to obtain a prediction curve of photovoltaic output. The predicted value is compared with the actual value of the typical output scenario to form a distribution map of the prediction error. BiLSTM consists of a double-layer LSTM. According to the BiLSTM neural network model, the short-term prediction result of photovoltaic output is obtained; In order to evaluate the training results of the neural network, the root mean square error RMSE and the mean square error MAE are used as the evaluation indicators of the neural network for photovoltaic output prediction. The expressions of RMSE and MAE are as follows: Where N represents the total number of data points; represents the predicted value of photovoltaic output; f i Indicates the actual value of photovoltaic output; A day is divided into a moment every 15 minutes, so there are 96 moments in a day. The predicted value at each moment is compared with the actual value of the typical output scenario of the day to form a distribution diagram of the prediction error. Step 2.3: Based on the distribution diagram of the prediction error, KDE is used to model the photovoltaic output prediction error and obtain the probability distribution curve of the photovoltaic output prediction error; The expression of the kernel density estimate function is as follows: Where N represents the total number of data points; h represents the bandwidth of the kernel density function, which is used to control the width of the kernel function and thus affect the smoothness of the density estimation; K is the kernel function, which is a probability density function with an integral of 1.

4. A distribution network risk assessment method based on photovoltaic output prediction error as claimed in claim 1, characterized in that: In step 3, the probability flow of the distribution network is calculated by combining the semi-invariant and Gram-Charher expansion series methods. The specific method is: Step 3.1, use the deterministic power flow calculation method to give the power flow distribution of normal operation, so as to obtain the state variable X at the reference operation point and the sensitivity matrix S; The node power equation of the power system is: Where P i , Q i are the injected active power and reactive power of node i respectively; V i ,θ i are the amplitude and angle of the voltage at node i respectively; G ij , B ij are the real and imaginary parts of the admittance matrix elements respectively; In the stochastic power flow model, the node injection quantity includes the active power and reactive power injected by the node, and the state variables include the voltage amplitude and angle of the node, which can be expressed as: In the formula, S0, X0 are the expected values ​​of the node injection amount S and the state variable X respectively; ΔS, ΔX are random disturbances, which are regarded as random variables obeying a certain distribution; Step 3.2: According to the known probabilistic description characteristics of the generator set and the load, calculate the origin moment or central moment of the injected power of each node and then find out its semi-invariants of each order; For node load power, its random component is considered to be caused by the prediction error of photovoltaic output. This random component is described by a normally distributed random variable. When the load power changes according to a certain load curve, it is simulated by a discrete random distribution. For the normally distributed injected power, the first-order semi-invariant is equal to the expected value, the second-order semi-invariant is the variance of the normal distribution, and the third to sixth-order semi-invariants are zero. The semi-invariant is a digital characteristic of a random variable, which is obtained by the moments of random variables not higher than the corresponding order. The central distances of random variables of various orders are expressed as: In the formula, μ represents the expected value; p v represents the probability value; Semi-invariant K of each order i Distance from center M i The relationship is as follows: Step 3.3: Use the Gram-Charlier series to expand the probability density function F(x) and probability distribution function f(x) of the voltage distribution, and use semi-invariants to represent the coefficients of the series expansion. The expansion is as follows: Where: F(x) is the probability density function; f(x) is the probability distribution function; g v are normalized semi-invariants of various orders, v = 1, 2, 3, 4, 5, 6; is the standard normal distribution density function; is the Hermite polynomial; Step 3.4, use the probability distribution of the voltage at each node of the distribution network, including the node voltage probability density curve and the cumulative distribution curve, to describe the uncertainty of the power system in the form of probability distribution.

5. A distribution network risk assessment method based on photovoltaic output prediction error as claimed in claim 1, characterized in that: In step 4, the probability of voltage exceeding the limit at the node of the distribution network is calculated, and the specific method for evaluating the voltage exceeding the limit risk of the distribution network is as follows: The probability distribution of the voltage at each node of the distribution network is obtained through probability flow calculation. The over-limit value of the node voltage is set. When the voltage exceeds this value, it can be considered that the distribution network is at risk of over-limit. The voltage over-limit probability is calculated based on the node voltage probability distribution. The calculation formula is as follows: Where U i,t represents the voltage of node i at time t; u max 、u min Respectively represent the upper and lower limits of the system voltage constraint; R(U i,t ) represents the probability of system voltage exceeding the limit at time t.

6. A distribution network risk assessment system based on photovoltaic output prediction error, characterized in that: Implement the distribution network risk assessment method based on photovoltaic output prediction error as described in any one of claims 1-5 to achieve distribution network risk assessment based on photovoltaic output prediction error.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the distribution network risk assessment method based on photovoltaic output prediction error according to any one of claims 1 to 5 is implemented to realize the distribution network risk assessment based on photovoltaic output prediction error.

8. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the control method of the multi-working state three-phase motor driver is implemented, and the distribution network risk assessment method based on photovoltaic output prediction error according to any one of claims 1 to 5 is implemented to realize the distribution network risk assessment based on photovoltaic output prediction error.

Images