Power distribution network intelligent evaluation and load prediction method based on whale optimization algorithm improved by differential evolution

Abstract

The invention discloses an intelligent evaluation and load prediction method for a power distribution network, and the method comprises the following steps: optimizing initial population distribution based on a differential evolution improved whale optimization algorithm (DE-WOA); setting kernel parameters of a support vector machine by using an improved whale optimization algorithm, taking a penalty factor P and a kernel function parameter g as optimization targets, solving optimal model parameters by iteratively updating parameter values each time, and finally constructing an SVM model with optimal parameters; extracting fault and load prediction characteristic data from the power distribution network data, and defining a fault grade threshold value and a load prediction related index; and inputting the extracted feature data into the SVM model with the optimal parameters, and outputting a fault level identification result or a load prediction result of the power distribution network. According to the whale optimization algorithm based on differential evolution improvement, the optimal parameters of the support vector machine can be found, and therefore the accuracy of power distribution network fault recognition and the accuracy of power load prediction are improved.

Description

Technical Field

[0001] The invention relates to an intelligent evaluation and prediction method for a power distribution network. Background Art

[0002] As the complexity of power systems continues to increase, the demand for intelligent and automated distribution networks is increasing. In the daily management of distribution networks, fault identification and power load forecasting are two key tasks. Fault identification can help quickly locate and determine the fault level, improve the response speed and power supply reliability of the power grid. Power load forecasting can effectively predict future power demand, help power companies optimize resource allocation, and avoid overload and resource waste.

[0003] Traditional distribution network fault identification and load forecasting methods are mostly based on empirical formulas or simple statistical analysis, which are difficult to handle complex nonlinear features and have limited accuracy. Support vector machine (SVM), as an effective machine learning method, has the advantage of processing high-dimensional and nonlinear data and is suitable for fault identification and load forecasting. However, the performance of SVM depends largely on the setting of its kernel parameters, so choosing a suitable parameter combination is crucial to the accuracy of the model.

[0004] To this end, the present invention proposes a whale optimization algorithm (DE-WOA) based on differential evolution to optimize the SVM model, and improves the performance of SVM in fault identification and load forecasting by optimizing the kernel parameters. This method can improve the intelligent level of distribution network operation management while ensuring a high accuracy rate, and provide scientific support for fault assessment and load forecasting of distribution networks. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent evaluation and load forecasting method suitable for distribution networks, so as to solve the problems of insufficient accuracy and difficulty in parameter optimization in the prior art in fault identification and power load forecasting. By introducing the differential evolution improved whale optimization algorithm (DE-WOA) to optimize the support vector machine (SVM) model, the present invention can improve the accuracy of fault identification and load forecasting, thereby improving the intelligent management level of the distribution network.

[0006] In order to achieve the above object, the technical solution provided by the present invention is as follows:

[0007] An intelligent evaluation and load forecasting method applicable to a distribution network comprises the following steps:

[0008] Construct an SVM model with optimal parameters: Based on the improved whale optimization algorithm (DE-WOA) based on differential evolution, optimize the initial population distribution; use the improved whale optimization algorithm for kernel parameter setting of support vector machine, take the penalty factor P and kernel function parameter g as optimization targets, use the improved whale optimization algorithm to update the parameter value in each iteration, solve the optimal model parameters, and finally construct an SVM model with optimal parameters, which can be used for distribution network fault identification and power load forecasting.

[0009] Feature point extraction: Extract feature data for fault and load forecasting from distribution network data. Fault feature data include the cumulative value of power outage times, power shortages, and power outage duration. Load feature data include historical load, meteorological, and economic data.

[0010] Data input: The extracted feature data is input into the SVM model with optimal parameters to perform fault identification or load forecasting.

[0011] Result output: The SVM model outputs the corresponding result according to the input data type, and outputs the fault level identification result of the distribution network or the power load prediction result.

[0012] The present invention optimizes a support vector machine (SVM) and proposes a whale optimization algorithm (WOA) improved by differential evolution (DE) to find out the optimal parameters of the support vector machine, thereby improving the accuracy of distribution network fault classification and load prediction.

[0013] Furthermore, the fault characteristic data includes a cumulative value of power outage times, a cumulative value of power shortages, and a cumulative value of power outage duration.

[0014] The expression of the cumulative value of power outage times is as follows:

[0015]

[0016] Where N i is the cumulative value of power outages in area i, C is the number of faults, N ij is the number of power outages for the jth fault in area i.

[0017] The power failure cumulative value expression is as follows:

[0018]

[0019] In the formula, Q i is the cumulative value of power outages caused by power outages in area i, Q ij is the jth power outage in area i, Q iNThe power supply for region i.

[0020] The expression of the accumulated value of power outage duration is as follows:

[0021]

[0022] Where, T i is the cumulative value of the total power outage duration in area i, T ij is the duration of the jth power outage in region i, T iN is the total downtime of region i.

[0023] The above fault levels are divided into normal, serious and more serious. The specific classification standards are as follows:

[0024] When the cumulative value of power outages is not less than 7, the cumulative value of power shortages is greater than 80%, or the cumulative value of power outage duration is greater than 80%, the fault level is more serious; when the cumulative value of power outages is 3-6, the cumulative value of power shortages is 30%-80%, or the cumulative value of power outage duration is 30%-80%, the fault level is serious; when the cumulative value of power outages is not greater than 2, the cumulative value of power shortages is less than 30%, and the cumulative value of power outage duration is less than 30%, the fault level is normal.

[0025] Furthermore, the load characteristic data considers the statistical characteristics of historical load data, the statistical characteristics of meteorological data and the statistical characteristics of economic indicators.

[0026] Statistical characteristics of historical load data: Based on the load data over a period of time, the mean of historical load is extracted Maximum L max , minimum value L min And other statistical characteristics, expressed as:

[0027]

[0028] L max =max(L t )

[0029] L min =min(L t )

[0030] Among them, L t is the historical load value at time t, and T is the number of time periods.

[0031] Statistical characteristics of meteorological data: For meteorological data such as temperature and humidity, extract relevant statistical characteristics, such as mean Maximum value T max , minimum value T min Etc., expressed as:

[0032]

[0033] T max =max(T t )

[0034] T min =min(T t )

[0035] Among them, T t is the historical temperature value at time t.

[0036] Statistical characteristics of economic indicators: Based on economic data such as GDP and industrial production index, extract statistical characteristics, such as mean etc. to reflect the fluctuations and trends of economic factors.

[0037]

[0038] Among them, G t is the historical GDP value at time t.

[0039] Furthermore, the above-mentioned whale optimization algorithm based on differential evolution is improved to improve the initial population distribution, and the following method is used:

[0040] Generate the initial population:

[0041] X i (0)|x L；i,j ≤x i,j (0)≤x U；i,j i=1,2,...,M; j=1,2,...,N,

[0042] x i,j (0) = x L；i,j +r·(x U；i,j -x L；i,j ),

[0043] In the formula, M is the total number, N is the dimension of each individual, r is a random number in [0,1]; x U；i,j and x L；i,j denote the upper and lower bounds respectively;

[0044] Use the results to synthesize individuals that need mutation:

[0045] V i (t+1)=X r1 (t)+B·[X r2 (t)-X r3 (t)],

[0046] Where t is the number of iterations, r1, r2, r3 are random numbers greater than the total number, and B is the proportional factor;

[0047] New individuals are randomly generated according to the crossover probability, and the calculation formula for each dimension component is as follows:

[0048]

[0049] Among them, CR represents the crossover probability;

[0050] Use a greedy strategy to select suitable individuals as new individuals:

[0051]

[0052] Where f(·) is the fitness calculation function.

[0053] Furthermore, the improved whale optimization algorithm is used for setting kernel parameters of the support vector machine using the following method:

[0054] (1) Use the difference improvement algorithm to initialize the parameters and individual whale positions and set the search agent:

[0055] x t；ij =[g i P i ],

[0056]

[0057] Where X(t) represents the individual matrix of the whale optimization algorithm in the iteration process, which contains the parameter positions of all individuals at the current iteration time t, and each row in the matrix represents the position vector of an individual. t； ij represents the position of the i-th individual in the j-th dimension at iteration t, which is the parameter value of the specific dimension of the specific individual in the current iteration.

[0058] (2) Define the SVM classification error rate as the fitness function, so that the search agent x in WOA t,ij Contains the model parameters to be optimized in SVM;

[0059] (3) The improved whale optimization algorithm is used to iteratively update the individual positions and find the optimal model parameters (P) corresponding to the minimum fitness generated when the optimization process is completed. d ,g d ).

[0060] Compared with the prior art, the present invention has the following advantages:

[0061] Improve parameter optimization effect: The present invention improves the whale optimization algorithm (WOA) by using the differential evolution algorithm (DE), generates a more suitable initial population, enhances the global optimization ability of WOA, and effectively optimizes the parameter settings of the support vector machine (SVM). Through this optimization, the present invention establishes a distribution network fault identification and load forecasting model that adapts to dynamic conditions and can more accurately cope with complex power grid environments.

[0062] Multifunctional application: The DE-WOA-SVM model of the present invention has both fault identification and load forecasting functions, and is more flexible than traditional methods. In terms of distribution network fault identification, the model can accurately classify fault levels and improve the accuracy of graded assessment; in terms of power load forecasting, the model effectively improves the forecasting accuracy and provides a scientific basis for power dispatching and resource optimization configuration.

[0063] The prediction accuracy is significantly improved: Through the evaluation of the effect of the method of the present invention, compared with the existing algorithms, the accuracy of this method in both fault identification and load prediction is significantly improved. This high-accuracy prediction capability can assist the distribution network monitoring system in reliable fault diagnosis and load management, ensuring the safe and stable operation of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flow chart of the intelligent evaluation method of the distribution network of the present invention;

[0065] Figure 2 Comparison of distribution network level prediction results of training samples in Example 2 of the present invention;

[0066] Figure 3 Comparison of the distribution network level prediction results of the test samples in Example 2 of the present invention;

[0067] Figure 4 The comparison of the time consumption for distribution network level prediction of the training samples in Example 2 of the present invention;

[0068] Figure 5 The comparison of the time consumption for distribution network level prediction of the test samples in Example 2 of the present invention;

[0069] Figure 6 The division of the data sample set in Embodiment 3 of the present invention;

[0070] Figure 7 The figure is a comparison of the load forecasting results in Example 3 of the present invention. DETAILED DESCRIPTION

[0071] The present invention is described in detail below in conjunction with specific embodiments and drawings.

[0072] Example 1

[0073] like Figure 1 As shown, the distribution network intelligent evaluation method based on DE-WOA-SVM of the present invention includes the following steps: 1. Classification of distribution network level

[0074] The expression of the distribution network is:

[0075] R = ∑[S(Y t ,E t )·P(E t )·P(Y t ,E t ,L)] (1)

[0076] In the formula, P(E t ) is event E t Probability of occurrence; S(Y t ,E t ) is E t and Y t The severity of the event under the working condition; P(Y t , E t ,L) is the system with event E t and load state L of Y t The probability of running under the condition.

[0077] To assess the reliability and safety of distribution networks, an index is determined by considering the likelihood and severity of potential accidents.

[0078] The mathematical expression of the cumulative value of power outage times is:

[0079]

[0080] Where N i is the cumulative value of power outages in area i; C is the number of faults; N ij is the number of power outages for the jth fault in area i.

[0081] The mathematical expression of the accumulated value of power outage duration is:

[0082]

[0083] Where, T i is the cumulative value of the total power outage duration in area i; T ij is the duration of the jth power outage in region i; T iN is the total downtime of region i.

[0084] The mathematical expression of the power shortage cumulative value is:

[0085]

[0086] In the formula, Q iis the cumulative value of power outages caused by power shortage in area i; Q ij is the jth power outage in area i; Q iN The power supply for region i.

[0087] Three indicators are used to evaluate the faults in the distribution network. These indicators include the cumulative value of power outages, the cumulative value of power shortages, and the cumulative value of power outage duration. Table 1 shows the three key indicators for evaluating distribution network faults.

[0088] Table 1 Classification of distribution network fault levels

[0089]

[0090]

[0091] 2 Improved whale algorithm to optimize support vector machine

[0092] 2.1 Whale Optimization Algorithm

[0093] WOA is inspired by the smart hunting of humpback whales and mainly includes three processes: surrounding prey, bubble net attack and searching for prey.

[0094] 1) Surrounding the prey: This behavior of surrounding the prey can be expressed as:

[0095] D=|X(t)-C·X z (t)| (5)

[0096] X(t+1)=X z (t)-A·D (6)

[0097] Where t is the current iteration, X is the position vector, and x z represents the position vector of the current optimal solution,

[0098] A and C are coefficient vectors, defined as follows:

[0099]

[0100] where a decreases linearly from 2 to 0 in the iterations and r is a random vector in [0,1].

[0101] 2) Bubble net attack: Humpback whales attack their prey by circling upward and then shrinking the circle. To simulate this behavior, shrinking circling and lateral position update mechanisms are designed. The shrinking encirclement process is achieved by reducing a. A is a random vector in [0,1]. The following equation is used to simulate this behavior:

[0102] X(t+1)=X d (t)+d·e kl ·cos(2πl) (8)

[0103] Where d = |X d (t)-X(t)| represents the distance between the i-th search agent and the best search agent, k is a constant defining the spiral shape, and l is a random number between [-1,1].

[0104] The mathematical model of the spiral contraction swimming behavior of humpback whales is:

[0105]

[0106] Where p is a random number between [0,1].

[0107] 3) Searching for prey: Humpback whales analyze the positions of other individuals and randomly search for prey. During the search phase, the position of the search agent is also random, and when possible, the best position can be replaced to enhance the search ability and provide WOA with better global search capabilities:

[0108] X(t+1)=X r -A·D (10)

[0109] Among them, X r is a randomly chosen position vector.

[0110] In each iteration, when |A|≥1, a search agent is randomly selected. This process will achieve global search, thereby improving the exploration ability of WOA. When |A|<1, the location of the prey will be the location of the random whale, and the population will need to gather there to hunt the prey. However, the initial population distribution of WOA will affect the range of its feasible solutions, limiting its global optimization performance. This example uses the improved DE algorithm to initialize the search agent in a random manner.

[0111] 2.2 Improved WOA based on differential evolution

[0112] This example proposes an improved WOA (DE-WOA) algorithm based on DE to generate a suitable initial population distribution for WOA and retain excellent individuals: Using the DE algorithm, the initial solution of the population is first generated, and the position of each individual is initialized (using the following equations (11, (12)). These positions are used to define the range of the initial population. The population is obtained by randomly generating positions, and the individuals in the population are iteratively updated using mutation, crossover, and selection operations, thereby guiding the individuals in the search space to gradually approach the optimal solution. The specific steps are as follows:

[0113] The initial population is generated as follows:

[0114] X i (0)x L；i,j ≤x i,j (0)≤xU；i,j i=1,2,...,M; j=1,2,...,N (11)

[0115] x i,j (0) = x L；i,j +r·(x U；i,j -x L；i,j ) (12)

[0116] In the formula, M is the total number, N is the dimension of each individual, and r is a random number in [0,1]. U；i,j and x L；i,j represent the upper and lower bounds respectively.

[0117] Use the results to synthesize individuals that need mutation:

[0118] V i (t+1)=X r1 (t)+B·[X r2 (t)-X r3 (t)] (13)

[0119] Where t is the number of iterations, r1, r2, r3 are random numbers greater than the total number, and B is the proportional factor.

[0120] The crossover process will randomly generate new individuals according to the crossover probability. The calculation formula for each dimension component is as follows:

[0121]

[0122] Where CR represents the crossover probability.

[0123] Use a greedy strategy to select suitable individuals as new individuals:

[0124]

[0125] Where f(·) is the fitness calculation function.

[0126] 2.3 Improved WOA Optimized Support Vector Machine

[0127] The parameter setting of the support vector machine should be reasonable to maintain high modeling ability in various situations. In this example, we improved the WOA algorithm and applied it to the kernel parameter selection of the support vector machine.

[0128] Support vector machine is a typical classification method, and its performance is significantly affected by the penalty factor P and the kernel function parameter g. This paper uses the improved DE-WOA method to find P and g suitable for support vector machine. The radial basis function is used as the kernel function. The specific operations of using support vector machine and DEWOA to build a classifier are as follows:

[0129] 1) Initialize the parameters and individual whale positions (including initial P and g) using the DE algorithm and set up the search agent;

[0130] x t；ij =[g i P i ] (16)

[0131]

[0132] Formula (16) defines the parameter vector [g i P i ], which contains the penalty factor P and kernel parameter g that need to be optimized in the SVM model. Formula (17) further expands formula (16) and integrates all individuals and related parameters in the DE-WOA algorithm into the X(t) matrix, which contains the position parameters of all whale individuals in each iteration.

[0133] 2) Define the SVM classification error rate as the fitness function, so that the search agent x in WOA t,ij Contains the model parameters to be optimized in SVM;

[0134] 3) Use the DE-WOA algorithm to iteratively update the individual positions and find the optimal model parameters (P d ,g d );

[0135] 4) Construct a SVM classifier model with optimal parameters.

[0136] 3. Fault level identification and load prediction

[0137] Feature extraction: Fault feature data and load feature data are extracted from the distribution network data. Fault feature data include key information such as the cumulative value of power outages, the cumulative value of power shortages, and the cumulative value of power outage duration. These feature variables are used as the core factors for distribution network classification in Table 1, and define the normal, severe, and more severe level thresholds respectively. Load feature data include historical load, meteorological, and economic data.

[0138] Data input: The extracted feature data is input into the optimized SVM classifier model. The SVM model can more accurately classify the fault level after the parameters (such as penalty factor P and kernel function parameter g) are optimized using the DE-WOA algorithm.

[0139] Classifier output: The SVM model classifies the input fault feature data and outputs the distribution network level identification results and load forecast results. Based on the learning of the training data set, the SVM matches the input features with their corresponding fault levels and predicts the level status of the current input data (for example, normal, severe, or more severe).

[0140] Fault level identification: The output of the classifier directly corresponds to the fault level defined in Table 1. According to the output of SVM, the current state and potential risk level of the distribution network can be obtained. For example, if the SVM outputs the "serious" level, it means that the current state of the distribution network meets the "serious" level range of the number of power outages, the cumulative value of power shortages, and the cumulative value of power outage duration.

[0141] By inputting distribution network fault data (number of power outages, cumulative power shortages and cumulative power outage duration) into the optimized SVM classifier model, the system can automatically evaluate and output the grade identification results of the distribution network, thereby providing scientific support for the monitoring and management of the distribution network.

[0142] Example 2

[0143] Experimental analysis

[0144] 1 Experimental data sources and preprocessing

[0145] By analyzing the factors affecting distribution network failures, the factors affecting distribution network failures are divided into four categories. This classification provides a framework for effectively evaluating and mitigating the problems related to distribution network failures. In order to extract relevant data from the source, preprocessing technology is used to screen the data types related to distribution network failures. The distribution network failure feature set is shown in Table 2.

[0146] Table 2 Distribution system fault data collection

[0147]

[0148] The research object of this paper is 205 branches of a distribution network in a prefecture-level city. Through the study of the distribution network information management system, massive data from January to December 2023 were obtained from 6 information subsystems. At the same time, external weather data for the same period in the region was also collected.

[0149] In terms of data cleaning, we conduct data missing value analysis, data outlier analysis, and data duplication analysis, and convert the original data into a form that is easy to analyze and apply. In terms of data integration, we conduct data statistics and enter the data into the database. In addition, we use clustering methods to identify and eliminate abnormal data.

[0150] 2 Evaluation Accuracy Analysis

[0151] This example extracts the level information of 1,000 samples by analyzing the data of the prefecture-level city information management system, of which 850 fault data are selected as training samples and 150 are used for prediction.

[0152] The distribution network fault prediction process includes setting the maximum evolution to 350, initializing the particle swarm size to 15, and training and predicting the fault data. The distribution network fault level prediction results of the training samples are shown in Figure 2 The distribution network fault level prediction results of the test sample are shown in Figure 3 As shown. It can be found that the DE-WOA-SVM algorithm of the present invention can effectively predict the active situation of the distribution network, and the comprehensive accuracy rate reaches more than 95%. In terms of prediction accuracy, the DE-WOA-SVM proposed in this paper is not only better than WOA-SVM and SVM, but also better than Extreme Learning Machine (ELM) and Convolutional Neural Networks (CNN). This shows that the DE-WOA-SVM algorithm of the present invention can accurately predict the state of the distribution network, and is practical and effective.

[0153] 3 Algorithm calculation time analysis

[0154] Using the same training and prediction samples, the calculation of DE-WOA-SVM algorithm is compared with that of the other four algorithms. The time consumption of distribution network fault level prediction of training samples is as follows: Figure 4 The time consumption for predicting the fault level of the distribution network of the test sample is as follows: Figure 5 shown.

[0155] The overall prediction time of the distribution network fault level using the DE-WOA-SVM training sample of the present invention is 11.7s, and the overall prediction time of the distribution network fault level using the test sample is 8.87s. Figure 4 and Figure 5 As can be seen from the figure, the time consumption of DE-WOA-SVM is higher than that of the other four algorithms, but it does not exceed 3s, which is acceptable in practical applications and meets the timeliness requirement of the evaluation.

[0156] In summary, although the DE-WOA-SVM method of the present invention takes a little longer than the other four algorithms, it still meets the actual needs, and its accuracy is higher than 95%, which shows that the DE-WOA-SVM method can accurately predict the fault level of the distribution network, thus providing a valuable reference for the prevention and management of distribution network accidents. Future work considers studying feature selection algorithms to improve the accuracy of distribution network level identification, and exploring the specific application of distribution network level prediction in actual scenarios.

[0157] Example 3

[0158] Load Forecasting

[0159] This embodiment demonstrates a real-time power load prediction method based on the DE-WOA-SVM model. Real-time power load prediction is of great significance to the operation and management of power grids, and can help the dispatching center better cope with load fluctuations and optimize the allocation of power resources. The method of the present invention performs predictions through real-time input data to provide support for the stable operation of the power system.

[0160] 1. Historical data acquisition, division and feature extraction

[0161] Collect various load and other related data, including historical load data, current meteorological data (such as temperature, humidity, wind speed, etc.) and economic data (such as working day, non-working day identification, etc.). These data will serve as input features for load forecasting.

[0162] During the data collection process, the system cleans and preprocesses the data at each time step, including missing value filling and outlier processing, to ensure the integrity and accuracy of the input data. The data is divided proportionally before prediction. The data sample set is divided as follows: Figure 6 shown.

[0163] 2. Real-time model parameter optimization

[0164] In the real-time prediction process, the system dynamically adjusts the parameters of the support vector machine (SVM) through the differential evolution improved whale optimization algorithm (DE-WOA). The algorithm can optimize the penalty factor P and kernel function parameter g according to the data characteristics of real-time input, ensuring that the SVM model parameters are always in the optimal state, thereby improving the prediction accuracy.

[0165] 3. Application of load forecasting model

[0166] The optimized SVM model uses the real-time input feature data for prediction output. Specifically, the system inputs the feature data into the SVM regression model at each time step and outputs the power load forecast value for one or more future time steps. The model generates accurate load forecast results through rapid iteration to meet the requirements of real-time prediction.

[0167] 4. Result evaluation and feedback

[0168] To verify the accuracy of real-time prediction, the system compares the load prediction results with the actual load data and calculates the prediction error. Common evaluation indicators include Mean Absolute Percent Error (MAPE). The system feeds the prediction error back to the DE-WOA optimization module, so that the model is continuously optimized in subsequent predictions and the accuracy of real-time predictions is improved.

[0169] The calculation formula of MAPE is as follows:

[0170]

[0171] Where N is the total number of days in the data set, and are the actual load and predicted load at the jth hour on the i-th day, respectively.

[0172] 5. Application Effect

[0173] The experimental results show that the DE-WOA-SVM method of the present invention shows high accuracy in real-time load forecasting. The average forecasting performance of each month is evaluated as follows Figure 7 As shown in the figure, the proposed method has the best prediction accuracy in the third month of the first quarter of 2016, with a MAPE value of 0.85%; while it has the worst prediction accuracy in July 2016, with a MAPE value of 1.87%. In the overall predictability of the whole year, the minimum MAPE of the simple daily wavelet neural network SIWNN is 1.26%, the maximum is 2.70%, the average is 1.71%, and the standard deviation is 0.47; the minimum MAPE of the Gaussian predictor GP is 0.86%, the maximum is 2.13%, the average is 1.37%, and the standard deviation is 0.33. In contrast, the minimum MAPE of the proposed method is 0.85%, the maximum is 1.77%, the average is 1.28%, and the standard deviation is 0.26. All four indicators are the best among all the five predictors involved in the comparison. In other words, the prediction performance of the proposed method is the best among all methods. From the perspective of the average MAPE, its performance is 25.15% better than SIWNN and 6.57% better than GP. Compared with traditional load forecasting methods, it provides accurate load forecasting support for the power grid dispatching center.

Claims

1. A distribution network intelligent evaluation and load forecasting method, characterized in that: The following steps are involved: Construct an SVM model with optimal parameters: Optimize the initial population distribution based on the whale optimization algorithm improved by differential evolution; The improved whale optimization algorithm is used to set the kernel parameters of the support vector machine. The penalty factor P and the kernel function parameter g are used as optimization targets. The improved whale optimization algorithm is used to update the parameter values ​​in each iteration to solve the optimal model parameters and finally construct the SVM model with the optimal parameters. Feature point extraction: Extract feature data of fault and load forecast from distribution network data. The fault feature data includes fault features such as the cumulative value of power outage times, the cumulative value of power shortage and the duration of power outage. The load feature data includes historical load data, meteorological data and economic data. Data input: The extracted feature data is input into the SVM model with optimal parameters to perform fault identification or load forecasting; Result output: The SVM model outputs the corresponding result according to the input data type, and outputs the fault level identification result of the distribution network or the power load prediction result.

2. The distribution network intelligent evaluation and load forecasting method according to claim 1 is characterized in that: The meteorological data in the load characteristic data include temperature and humidity; the economic data in the load characteristic data include GDP and industrial production index.

3. The distribution network intelligent fault assessment method and load forecasting method according to claim 1, characterized in that: The fault levels are classified into normal, severe and more severe; The power outage cumulative value expression is as follows: Where N i is the cumulative value of power outages in area i, C is the number of faults, N ij is the number of power outages for the jth fault in area i; The power failure cumulative value expression is as follows: In the formula, Q i is the cumulative value of power outages caused by power outages in area i, Q ij is the jth power outage in area i, Q iN Provide electricity to area i; The power outage duration cumulative value expression is as follows: Where, T i is the cumulative value of the total power outage duration in area i, T ij is the duration of the jth power outage in region i, T iN is the total downtime of region i; When the cumulative value of power outages is not less than 7, the cumulative value of power shortages is greater than 80%, or the cumulative value of power outage duration is greater than 80%, the fault level is more serious; when the cumulative value of power outages is 3-6, the cumulative value of power shortages is 30%-80%, or the cumulative value of power outage duration is 30%-80%, the fault level is serious; when the cumulative value of power outages is not greater than 2, the cumulative value of power shortages is less than 30%, and the cumulative value of power outage duration is less than 30%, the fault level is normal.

4. The distribution network intelligent fault assessment method and load forecasting method according to claim 2, characterized in that: The statistical characteristics of the historical load data are: based on the load data in the past period of time, the mean value of the historical load is extracted Maximum L max , minimum value L min , expressed as: L max =max(L t ) L min =min(L t ) Among them, L t is the historical load value at time t, and T is the number of time periods; The statistical characteristics of the meteorological data are: for temperature and humidity, relevant statistical characteristics are extracted, including the mean Maximum value T max , minimum value T min , expressed as: T max =max(T t ) T min =min(T t ) Among them, T t is the historical temperature value at time t. The statistical characteristics of the economic data are: based on GDP and industrial production index, statistical characteristics are extracted, including the mean To reflect the fluctuations and trends of economic factors, it is expressed as follows: Among them, G t is the historical GDP value at time t.

5. The distribution network intelligent evaluation and load forecasting method according to claim 1, characterized in that: The improved whale optimization algorithm based on differential evolution improves the initial population distribution by using the following method: Generate the initial population: X i (0)x L；i,j ≤x i,j (0)≤x U；i,j i=1,2,...,M;j=1,2,...,N, x i,j (0)=x L；i,j +r·(x U；i,j -x L；i,j ), In the formula, M is the total number, N is the dimension of each individual, r is a random number in [0,1]; x U；i,j and x L；i,j denote the upper and lower bounds respectively; Use the results to synthesize individuals that need mutation: V i (t+1)=X r1 (t)+B·[X r2 (t)-X r3 (t)], Where t is the number of iterations, r1, r2, r3 are random numbers greater than the total number, and B is the proportional factor; New individuals are randomly generated according to the crossover probability, and the calculation formula for each dimension component is as follows: Among them, CR represents the crossover probability; Use a greedy strategy to select suitable individuals as new individuals: Where f(·) is the fitness calculation function.

6. The distribution network intelligent evaluation and load forecasting method according to claim 5, characterized in that: The improved whale optimization algorithm is used for the kernel parameter setting of the support vector machine, using the following method: (1) Use the difference improvement algorithm to initialize the parameters and individual whale positions and set the search agent: x t；ij =[g i P i ], Where X(t) represents the individual matrix of the whale optimization algorithm in the iteration process, which contains the parameter positions of all individuals at the current iteration time t, and each row in the matrix represents the position vector of an individual. t；ij It represents the position of the i-th individual in the j-th dimension at iteration t, which is the parameter value of the specific dimension of the specific individual in the current iteration. (2) Define the SVM classification error rate as the fitness function, so that the search agent x in WOA t,ij Contains the model parameters to be optimized in SVM; (3) The improved whale optimization algorithm is used to iteratively update the individual positions and find the optimal model parameters (P) corresponding to the minimum fitness generated when the optimization process is completed. d ,g d ).

Images