Feeder-level non-intrusive load component identification method and system for load modeling and medium

Abstract

The invention provides a load modeling-oriented feeder-level non-intrusive load component identification method and system and a medium, and the method comprises the steps: firstly constructing a static load power model of each type of load based on parameters measured by a dynamic simulation experiment, and obtaining a labeled simulation data set of load power through a Monte Carlo method; secondly, constructing a non-intrusive load decomposition neural network model, and performing training learning by taking voltage, total power and the like as input features and taking decomposition power or proportion as label data; on the basis, in the process of identifying and decomposing the proportion of each type of load component under a substation bus or an actual feeder line of a power grid level above, the actual total power and voltage data are subjected to feature extraction and then are used as the input of the model, and a corresponding load component proportion estimation result can be output and obtained. The problems that supervised learning is difficult to carry out and the load decomposition effect is poor due to lack of label data under high-voltage-level and low-density data are solved, and a scientific decision basis is provided for regional automatic load modeling and control.

Description

Technical Field

[0001] The present invention relates to the field of smart grids, in particular to load component identification technology, and specifically to a feeder-level non-intrusive load component identification method oriented to load modeling. Background Art

[0002] The construction and development of new power systems are placing increasingly stringent demands on them. Improving power systems and structures, as well as increasing the efficiency of energy utilization, are key to resolving energy challenges. Accurate load monitoring and decomposition significantly contribute to reducing power loss, optimizing power usage, improving power quality, and enhancing power security. It is a crucial component of the new power system landscape. Establishing a load model that more accurately reflects the actual characteristics of the power grid is a prerequisite for accurate load monitoring and decomposition.

[0003] However, due to the geographical distribution and random variability of load models, it is quite difficult to obtain a comprehensive load model that can reasonably describe each load node, which has become a major challenge in improving the accuracy of load composition analysis in power systems. In order to better utilize the application effect of the automatic load modeling system and provide more accurate time-sharing and classified load model parameters for large-scale power grid simulation calculations, it is necessary to determine the real-time power load composition of the load node based on the real-time power data of the load node or feeder, combined with the low-voltage feeder load composition parameter library. Then, combined with the load model library of power users and power equipment, a statistical synthesis method is used to obtain the load model and parameters of the load node. Summary of the Invention

[0004] The present invention aims to address the problems of difficulty in conducting supervised learning and poor load decomposition due to the lack of labeled data under high voltage levels and low data density. A feeder-level non-intrusive load composition identification method for load modeling is proposed. Its purpose is to use a neural network model to perform supervised learning based on a labeled simulation data set, train and mine the deep nonlinear characteristics between the total load and the power of each type of load, and realize non-intrusive load composition identification and decomposition at the substation feeder level in the power grid.

[0005] To achieve the object of the present invention, a first aspect of the present invention provides a feeder-level non-intrusive load component identification method for load modeling, which includes the following steps:

[0006] Step 1: Obtain load model parameters of various load types through static characteristic experiments, and construct steady-state power models of various load types and the total load;

[0007] Step 2: Generate different load types and total load powers under different voltages and component ratios using the Monte Carlo method to establish a feeder-level simulated load data set; and

[0008] Step 3: Build a non-intrusive load decomposition neural network model, use voltage and total power as input features, and decomposed power or proportion as label data, perform training and learning, and obtain the trained load decomposition neural network model as the identification model;

[0009] Step 4: During the identification and decomposition of the proportion of various types of load components under the actual feeder at the substation bus or above the grid level, feature extraction of the actual total power and voltage data is performed and used as the input of the trained identification model, and the corresponding load component proportion estimation result is obtained through the model output.

[0010] As an optional implementation, in step 1, obtaining load model parameters of each type of load through a static characteristic experiment and constructing a steady-state power model of each type of load and the total load includes the following steps:

[0011] Step 11. In the power system dynamic simulation experiment system, static characteristic experiments are conducted on five typical load types: lighting (L), resistance (R), regulated power supply (SMPS), single-phase motor (IM1), and three-phase motor (IM3). A programmable variable frequency power supply is used to control the output voltage and frequency, and the steady-state voltage and steady-state power of the five types of load equipment at different voltages and frequencies are measured.

[0012] Step 12: Determine the static load model parameters of each load type based on the measured steady-state data of each load type; and

[0013] Step 13: Connect various load devices to the simulated power system in different proportions to simulate power consumption scenarios at different time periods within a day, and measure and record the steady-state data of the load group bus and various load branches, including steady-state power and steady-state voltage.

[0014] As an optional implementation, in step 12, the static load model parameters of each type of load are determined based on the measured steady-state data of each type of load, in the following manner:

[0015]

[0016] Where, P I is the constant current parameter of the active ZIP model, P Z is the constant impedance parameter of the active ZIP model, k is the number of static characteristic test groups, U k and P k are the voltage and active power measured in the kth experiment, and U0 and P0 are the voltage and active power at the steady-state point.

[0017] As an optional implementation, in step 2, generating different types of loads and total load powers under different voltages and component ratios using the Monte Carlo method to establish a feeder-level simulated load data set includes the following steps:

[0018] Step 21: Set the voltage range and use the Monte Carlo method to uniformly distribute and randomly generate voltage samples U. The number of samples is N. V ;

[0019] Step 22: Divide the five types of loads into two major parts: static and dynamic. Among them, lighting, resistance and voltage-regulated power supply belong to the static part, and single-phase motors and three-phase motors belong to the dynamic part. The static and dynamic parts are set from 0% to 100% and in steps of 5%, and the sum of the proportions of the static and dynamic parts is equal to 100%, thereby generating 21 groups of proportions; among them, in each group of static and dynamic proportions, the Monte Carlo method is used to randomly generate the weight coefficient W of each load type in a uniform distribution, and the sum of the weight coefficients of the three load types belonging to the static part is equal to the proportion of the static part of the group, and the sum of the weight coefficients of the two load types belonging to the dynamic part is equal to the proportion of the dynamic part of the group; in each group of static and dynamic proportions, the number of randomly generated weight coefficients of each load type is N. W ;as well as

[0020] Step 23: Obtain the total power under each condition of different voltage and load ratio through weighted summation to form a simulated load data set.

[0021] As an optional implementation, in step 2, the voltage sample matrix of the simulated load data set is constructed as follows:

[0022]

[0023] The weight coefficient matrix of the simulated load data set is constructed as follows:

[0024]

[0025] The total power of the simulated load data set is calculated by weighted summation as follows:

[0026]

[0027] Where, P i,t (U t ) and Q i,t (U t ) is the ZIP model of active power and reactive power with respect to voltage for various loads.

[0028] As an optional implementation, in step 3, the non-intrusive load decomposition neural network model is constructed, using voltage and total power as input features and decomposed power or proportion as label data for training and learning to obtain an identification model, including the following steps:

[0029] Step 31: Use the voltage U of the simulated load data set t , total active power P agg and total reactive power Q agg As the input feature vector, the weight coefficient matrix W is used as the target vector. A multi-layer perceptron (MLP) is used as the main body. Three fully connected layers are set as hidden layers between the input layer and the output layer to build a deep learning load decomposition neural network model. The model is trained to explore the deep nonlinear characteristics between the total load and the power proportion of each type of load; and

[0030] Step 32: Apply the load decomposition neural network model trained in step 31 to the dynamic model load group experimental data to verify the effectiveness of the decomposition model again.

[0031] As an optional implementation, in step 31, the training process of the deep learning load decomposition neural network model specifically includes the following steps:

[0032] The load decomposition model parameter training based on the multi-layer perceptron model is adopted, and the mean square error (MSE) is used as the loss function. The expression is:

[0033]

[0034] Where, represents the estimated value, y i represents the actual value, and m represents the total number of samples;

[0035] Using R 2 And accuracy ACC is used as the evaluation index, and the expression is:

[0036]

[0037] Where, represents the estimated value, y i Indicates the actual value, represents the true power value mean, and m represents the total number of samples.

[0038] As an optional implementation, in step 4, during the process of identifying and decomposing the proportions of various types of load components on the actual feeders at the substation bus or above the grid level, feature extraction is performed on the actual total power and voltage data and used as input to the trained identification model, and the corresponding load component proportion estimation result is obtained through the model output, including the following processing steps:

[0039] Step 41: Preprocess the data of the low-voltage feeder energy management system based on the marketing business system to determine the reference steady-state power at each moment and convert the actual power into per-unit power; and

[0040] Step 42: Based on the processed low-voltage feeder data, the trained and verified load decomposition neural network model is used to decompose the load components to obtain the proportion of various load components of the low-voltage feeder.

[0041] As an optional embodiment, in step 4, pre-processing can be performed based on the low-voltage feeder energy management system data of the marketing business system to determine the reference steady-state power at each moment and convert the actual power into per-unit power. The specific implementation includes the following steps:

[0042] Step 11: Based on the low-voltage feeder energy management system data of the marketing business system, randomly generate voltage sampling points and weighting coefficients using the Monte Carlo method according to the method of step 2, and combine them with the static load model to generate a data set;

[0043] Step 12: Classify the calculated total power according to its corresponding voltage, and determine a probability distribution curve for each voltage; and

[0044] Step 13: Using the probability analysis method, for any given voltage, the probability characteristics of the total power are calculated, and the per-unit power value with the largest probability density at each voltage is taken as the most likely per-unit power, and then the rated power is obtained.

[0045] According to a second aspect of the present invention, a computer system is provided, comprising:

[0046] one or more processors;

[0047] A memory storing operable instructions, wherein when the instructions are executed by the one or more processors, the one or more processors are caused to perform operations, wherein the operations include executing the aforementioned feeder-level non-intrusive load component identification method based on the deep learning load decomposition neural network model.

[0048] According to a third aspect of the present invention, a computer-readable medium storing software is also proposed, wherein the software includes instructions that can be executed by one or more computers, and when the instructions are executed by the one or more computers, the process of the aforementioned feeder-level non-intrusive load component identification method based on the deep learning load decomposition neural network model is performed.

[0049] Through the implementation of the above technical solutions, it can be seen that the feeder-level non-invasive load component identification method based on the deep learning load decomposition neural network model proposed in the present invention utilizes the nonlinear feature extraction capability of deep learning to construct an experimentally based simulation data set and a load decomposition model, and proposes a method for determining the low-voltage feeder load reference power, thereby realizing load decomposition at the substation bus and above power grid levels, overcoming the deficiency that traditional methods are difficult to perform load decomposition on low-density, high-voltage level low-voltage feeder data sets.

[0050] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below, as long as such concepts are not mutually inconsistent, can be considered part of the inventive subject matter of this disclosure. In addition, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0051] The foregoing and other aspects, embodiments, and features of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the present invention, such as features and / or beneficial effects of the exemplary embodiments, will become apparent from the following description or through practice of specific embodiments according to the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings are not intended to be drawn to scale. In the accompanying drawings, each identical or nearly identical component shown in various figures may be represented by the same reference numeral. For clarity, not every component is labeled in every figure. Embodiments of various aspects of the present invention will now be described by way of example and with reference to the accompanying drawings.

[0053] Figure 1 This is a flow chart for implementing a feeder-level non-intrusive load component identification method for load modeling according to an embodiment of the present invention.

[0054] Figure 2 Flowchart of constructing a simulation data set and a load decomposition model according to an embodiment of the present invention.

[0055] Figure 3 This is a flowchart of a specific implementation of the feeder load data decomposition method according to an embodiment of the present invention.

[0056] Figure 4 This is a component identification and decomposition result diagram of various load branches and bus data in multiple scenarios based on dynamic model experimental measurements in an embodiment of the present invention.

[0057] Figure 5 An example diagram of finding the most likely per-unit power and rated power corresponding to each per-unit voltage value according to an embodiment of the present invention.

[0058] Figure 6This is an example diagram of the feeder load component ratio estimation results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to better understand the technical content of the present invention, specific embodiments are given and described below with reference to the accompanying drawings.

[0060] Various aspects of the present invention are described in this disclosure with reference to the accompanying drawings, in which a number of illustrative embodiments are shown. The embodiments of the present disclosure are not necessarily intended to include all aspects of the present invention. It should be understood that the various concepts and embodiments introduced above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed herein are not limited to any embodiment. In addition, some aspects of the present disclosure may be used alone or in any appropriate combination with other aspects disclosed herein.

[0061] {Example 1}

[0062] With reference to the figures, the feeder-level non-intrusive load component identification method for load modeling according to an embodiment disclosed in the present invention includes the following steps:

[0063] Step 1: Obtain load model parameters of various load types through static characteristic experiments, and construct steady-state power models of various load types and the total load;

[0064] Step 2: Generate different load types and total load powers under different voltages and component ratios using the Monte Carlo method to establish a feeder-level simulated load data set; and

[0065] Step 3: Build a non-intrusive load decomposition neural network model, use voltage and total power as input features, and decomposed power or proportion as label data, perform training and learning, and obtain the trained load decomposition neural network model as the identification model;

[0066] Step 4: During the identification and decomposition of the proportion of various types of load components under the actual feeder at the substation bus or above the grid level, feature extraction of the actual total power and voltage data is performed and used as the input of the trained identification model, and the corresponding load component proportion estimation result is obtained through the model output.

[0067] It can be seen that traditional supervised learning relies on real load component labels (such as user-side equipment power data), but at the high-voltage feeder level, due to the high cost of equipment deployment and data privacy protection, it is difficult to obtain a large amount of real label data, resulting in limited model generalization ability. In the implementation process of the present invention, the Monte Carlo simulation + hierarchical parameter generation method is adopted. The full voltage range (0%-100% rated voltage) and load ratio combination are uniformly generated by the Monte Carlo method to cover the extreme working conditions in the actual operation of the power grid (such as voltage sag, load surge); combined with dynamic-static load decoupling, the load is divided into static (ZIP model) and dynamic (motor type) parts, and weight coefficients are generated separately to ensure the physical rationality of the simulation data. For example, the static load weight coefficient generates 21 groups of ratio scenarios with a step size of 5%, and 5000 samples are randomly generated for each group, forming a data set containing 105,000 samples, which significantly improves data diversity. In addition, through hierarchical processing, the model complexity is reduced (such as the dynamic part only needs to process two load types), and the efficiency of model training and model performance are improved.

[0068] The method of the present invention verifies the consistency of simulated data with actual load characteristics through experiments, ensuring the effectiveness of model training. At the same time, the ZIP model and neural network are combined. The static load adopts the ZIP model (constant impedance, constant current, constant power components), and the parameters are fitted by experimental data to accurately describe the voltage-sensitive characteristics. The dynamic load (motor type) generates random weights through the Monte Carlo method to simulate the random start-stop behavior in actual operation. The neural network uses a multi-layer perceptron (MLP) as the main body to explore the deep nonlinear mapping relationship between total power, voltage, and load components. For example, it captures the difference between the inrush current and steady-state power when the motor starts. This overcomes the defect of traditional methods (such as rule-based decomposition and shallow machine learning) that are difficult to capture the complex nonlinear relationship between load components. In particular, when the voltage fluctuates and the load ratio changes dynamically, the error is large. In combination with training data of the full voltage range and load ratio combination and all working conditions, the model robustness and load decomposition identification accuracy are improved.

[0069] As an optional embodiment, combining Figure 2 The specific implementation of the aforementioned step 1 includes the following steps:

[0070] Step 11. In the power system dynamic simulation laboratory, static characteristic experiments are conducted on five typical load types: lighting (L), resistance (R), regulated power supply (SMPS), single-phase motor (IM1), and three-phase motor (IM3). A programmable variable frequency power supply is used to control the output voltage and frequency, and the steady-state voltage and steady-state power of each type of load equipment at different voltages and frequencies are measured.

[0071] Step 12: Determine the static load model parameters of each load type based on the measured steady-state data of each load type; and

[0072] Step 13: Connect various load devices to the simulated power system in different proportions to simulate power consumption scenarios at different time periods within a day, and measure and record steady-state data such as power and voltage of the load group bus and various load branches.

[0073] As an optional embodiment, the method for determining static load model parameters based on measured steady-state data of various loads is:

[0074] The static multi-form model of the load consists of three parts: constant impedance, constant current, and constant power, so it is defined as a ZIP model. Its basic form is:

[0075]

[0076] Where P is active power, Q is reactive power, U is voltage, subscript 0 is the initial operating point, P Z is the constant impedance parameter of the active ZIP model, P I is the constant current parameter of the active ZIP model, P P is the constant power parameter of the active ZIP model, Q Z is the constant impedance parameter of the reactive ZIP model, Q I is the constant current parameter of the reactive ZIP model, Q P is the constant power parameter of the reactive ZIP model.

[0077] Let P Z =A,P I =B,P P =1-AB, then:

[0078] P k =P0[A(U k / U0) 2 +B(U k / U0)+(1-AB)];

[0079] Where, k is the number of static characteristic test groups, U k and P k are the voltage and active power measured in the kth experiment, and U0 and P0 are the voltage and active power at the steady-state point.

[0080] Converting the static characteristic coefficient into a parameter optimization problem, we have:

[0081]

[0082] Where J is the value of the objective function to be optimized. Take the partial derivative of J and set it equal to 0. The simultaneous solution is:

[0083]

[0084] It should be understood that the reactive characteristic coefficient can also be calculated in a similar manner, which will not be described in detail here.

[0085] As an optional embodiment, the specific implementation of step 2 includes the following steps:

[0086] Step 21: Set the voltage range and use the Monte Carlo method to uniformly distribute and randomly generate voltage samples U. The number of samples is N. V ;

[0087] For example, in order to cover all possible values of the power system including extreme cases for full training, the voltage range is set to [0.9pu, 1.1pu], and the Monte Carlo method is used to randomly generate voltage samples U with a uniform distribution of 0.002pu steps. The number of samples is N. V =101;

[0088] Step 22: Divide the five types of loads into two major parts: static and dynamic. Among them, lighting, resistance and voltage-regulated power supply belong to the static part, and single-phase and three-phase motors belong to the dynamic part. The static and dynamic parts are stepped from 0% to 100% according to a preset step size (for example, the step size is set to 5%), and the sum of the two parts is equal to 100%, generating 21 groups of proportions; in each group of static and dynamic proportions, the Monte Carlo method is used to randomly generate the weight coefficient W of each load type in a uniform distribution, and the sum of the load type weight coefficients of the three static parts is equal to the proportion of the static part of the group, and the sum of the load type weight coefficients of the two dynamic parts is equal to the proportion of the dynamic part of the group. The number of randomly generated load type weight coefficients in each group of static and dynamic proportions is N. W =10;

[0089] For example, a set of load proportions with a static part accounting for 40% and a dynamic part accounting for 60% is generated. Then, 10 times of five load proportions such as 5%+20%+15%+24%+36% are randomly generated in this set of proportions. Each generation in this set satisfies the conditions that the sum of the proportions of the first three load types is equal to the proportion of the static part (i.e., 40%), and the sum of the proportions of the last two load types is equal to the proportion of the dynamic part (i.e., 60%).

[0090] Step 23: Obtain the total power under each condition of different voltage and load ratio through weighted summation to form a simulated load data set.

[0091] As an optional embodiment, the method for constructing the simulated load data set specifically includes:

[0092] The voltage sample matrix of the simulation data set is constructed as follows:

[0093]

[0094] The weight coefficient matrix of the simulated data set is constructed as follows:

[0095]

[0096] The total power of the simulated data set is calculated by weighted summation as follows:

[0097]

[0098] Where, P i,t (U t ) and Q i,t (U t ) is the ZIP model of active power and reactive power of various loads with respect to voltage. Therefore, the total number of simulation data set samples used in this model is:

[0099] N V ×21×N W =21210.

[0100] Preferably, the training process of the non-intrusive load decomposition neural network model based on deep learning specifically includes the following process:

[0101] The load decomposition model parameter training based on the multi-layer perceptron model is adopted, and the mean square error (MSE) is used as the loss function, which is expressed as:

[0102]

[0103] Where, represents the estimated value, y i represents the actual value, and m represents the total number of samples;

[0104] Using R 2 (R Squared) and accuracy ACC (Accuracy) are used as evaluation indicators, and the expression is:

[0105]

[0106] Where, represents the estimated value, y i Indicates the actual value, represents the true power mean, and m represents the total number of samples.

[0107] Therefore, the model converges by minimizing the MSE loss function during training. 2 The value and ACC are evaluated to ensure the validity and accuracy of the model.

[0108] As an optional embodiment, combining Figure 2 , the specific implementation of step 3 includes the following steps:

[0109] Step 31: Voltage U of the simulated data set t , total active power P agg and total reactive power Q agg As the input feature vector, the weight coefficient matrix W is used as the target vector. A multi-layer perceptron (MLP) is used as the main body, and three fully connected layers are set as hidden layers between the input layer and the output layer to build a deep learning load decomposition model. The model is trained to explore the deep nonlinear characteristics between the total load and the power proportion of each type of load; and

[0110] Step 32: Apply the load decomposition model trained in step 31 to the dynamic model load group experimental data to verify the effectiveness of the decomposition model again.

[0111] In the embodiment of the present invention, the method for constructing a simulation data set and the method for constructing a load decomposition model thereof are described based on experimental data as an example.

[0112] As an example, LED lamps are used to simulate lighting loads, kettles and incandescent lamps are used to simulate resistance loads, computer power supplies and mobile phone chargers are used to simulate regulated power supply loads, small fans are used to simulate single-phase motor loads, and asynchronous motors with DC loads are used to simulate three-phase motor loads.

[0113] After static characteristic experiments and calculations, the ZIP model parameters for each load type are obtained as shown in Table 1.

[0114] The specific load composition and power setting of the load group experiment are shown in Table 2.

[0115] Table 1. ZIP model parameters for each load type

[0116]

[0117] Table 2. Composition of simulated load groups and power settings

[0118]

[0119] In an embodiment of the present invention, a deep learning load decomposition model is constructed with a multi-layer perceptron model as the main body, wherein the input layer is responsible for receiving the features of the data, and the input dimension is 3, namely voltage, total active power and total reactive power; the output layer is responsible for outputting the recognition results, the activation function is a linear function, and the output dimension is 6, namely the proportion of six load types.

[0120] Three fully connected layers are set between the input layer and the output layer as hidden layers, and the number of neurons is set to 64, 128, and 128 respectively.

[0121] The dataset is evenly split into a training set and a validation set in an 8:2 ratio. Nonlinear transformations are performed on features learned through activation functions. The ReLU (Rectified Linear Unit) function is used as the activation function for the three hidden layers to introduce nonlinear transformations. The mean squared error (MSE) is used as the loss function. The Adaptive Moment Estimation (Adam) algorithm is used as the optimizer, which automatically adjusts the learning rate to address issues such as unstable objective functions and sparse gradients. As an optional example, a dropout layer is added to the first two hidden layers. In each iteration, the neuron output is randomly set to 0 according to a specified ratio (for example, 0.5) to reduce the learning rate and avoid overfitting.

[0122] Therefore, after a certain number of training steps, training is stopped when the loss function value on the validation set no longer decreases.

[0123] The validation metrics based on the simulated dataset are shown in Table 3.

[0124] It can be seen that the load decomposition model proposed in the present invention has high accuracy for this simulation data set.

[0125] Table 3. Verification indicators of various types of loads in the simulation data set

[0126]

[0127] Combine Figure 4 Figure 2 shows the component identification and decomposition results of various load branch and bus data under multiple scenarios, based on dynamic model experimental measurements. As can be seen, the feeder-level load component identification results proposed by the present invention are very close to the true values, further verifying the effectiveness and accuracy of the identification method proposed by the present invention.

[0128] With reference to the figures, the feeder-level load component identification method according to the embodiment disclosed in the present invention is particularly applicable to the decomposition method of actual feeder load data under high voltage level and low density data.

[0129] In the process of identifying and decomposing the proportion of various types of load components under the actual feeders at the substation bus or above the power grid level, the actual total power and voltage data are feature extracted and used as the input of the trained identification model, and the corresponding load component proportion estimation results are obtained through the model output.

[0130] For example, first, the low-voltage feeder energy management system (EMS) data based on the marketing business system is preprocessed to determine the benchmark steady-state power at each moment and convert the actual power into per-unit power; then, based on the processed low-voltage feeder data, the trained and verified load decomposition model is used to decompose the load components and obtain the proportion of various load components of the low-voltage feeder.

[0131] As an optional embodiment, combining Figure 3 As shown, the above-mentioned preprocessing of low-voltage feeder data includes the following steps:

[0132] Step 11: randomly generate voltage sampling points and weighting coefficients through Monte Carlo simulation and combine them with the static load model to generate a data set. This step is the same as the method of step 2 above.

[0133] Step 12: classify the calculated total powers according to the corresponding voltages, and derive a probability distribution curve for each voltage; and

[0134] Step 13: Using probability analysis technology, for any given voltage, the probability characteristics of the calculated total power can be obtained. The per-unit power value with the largest probability density at each voltage is taken as the most likely per-unit power, and then the rated power can be obtained.

[0135] As a specific example, the voltage, active power, and reactive power data of a 110kV substation feeder in a city power grid with a sampling interval of 1 minute from August 2023 to July 2024 are used as an example for explanation.

[0136] Assume that at t0, t1, and t2, the voltages are U0, U0, and 0.98U0, respectively, and the powers are P0, 1.05P0, and 1.05P0, respectively. Therefore, the exponential load model at the three moments is:

[0137]

[0138] It can be seen that even if the rated power is P0, 1.05P0, 1.05×0.98 respectively -npu2 P0, even if the voltage remains constant, the rated power varies at any time of the day.

[0139] The per-unit power probability distribution is analyzed for each possible voltage value, and the probability distribution can be fitted very accurately as a Gaussian distribution.

[0140] Next, we calculate the average (μ) and standard deviation (σ) of the per-unit total power at all possible voltages. As can be seen, the average (also the value with the highest probability density) of the calculated per-unit total power increases with increasing voltage, while the standard deviation decreases as the voltage approaches 1.00 pu.

[0141] Taking the daily data of a 110kV substation feeder in a certain city as an example, the actual voltage curve at all times of the day is lower than the rated voltage level (taken as 115kV). After knowing the actual voltage curve and the rated voltage level of the day, the per-unit voltage curve of the day can be obtained, and then the corresponding most likely per-unit power and rated power can be found for each per-unit voltage value, such as Figure 5 As shown in the figure, it can be seen that most rated powers are slightly higher than the actual powers. Thus, the rated powers and per-unit powers at different times of the day are obtained based on Monte Carlo simulation and probability distribution fitting. These results can be used for more accurate load decomposition and load component identification.

[0142] Figure 6 The following example shows the load component ratios for load component identification based on processed low-voltage feeder data from a certain day and the trained and validated MLP network. It can be seen that the lighting and resistor load components have the largest changes, while the single-phase and three-phase motor load components have the smallest changes. Peak loads occur around 10 a.m. and 6 p.m., which is generally in line with expectations.

[0143] {Example 2}

[0144] In conjunction with the implementation of the feeder-level non-intrusive load component identification method based on the deep learning load decomposition neural network model in the above embodiment of the present invention, according to the second aspect of the present invention, a computer system is further proposed, comprising:

[0145] one or more processors;

[0146] A memory storing operable instructions, which, when executed by the one or more processors, cause the one or more processors to perform operations, including the process of executing the aforementioned feeder-level non-intrusive load component identification method based on the deep learning load decomposition neural network model.

[0147] {Example 3}

[0148] In combination with the implementation of the feeder-level non-intrusive load composition identification method based on the deep learning load decomposition neural network model in the above embodiment of the present invention, according to the third aspect of the purpose of the present invention, a computer-readable medium storing software is also proposed, wherein the software includes instructions that can be executed by one or more computers, and the instructions, when executed by the one or more computers, perform the process of the aforementioned feeder-level non-intrusive load composition identification method based on the deep learning load decomposition neural network model.

[0149] While the present invention has been disclosed above with reference to preferred embodiments, this is not intended to limit the present invention. Persons skilled in the art will readily appreciate that various modifications and variations can be made without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A feeder-level non-intrusive load component identification method for load modeling, characterized by: The following steps are involved: Step 1: Obtain load model parameters of various load types through static characteristic experiments, and construct steady-state power models of various load types and the total load; Step 2: Generate different load types and total load power under different voltages and component ratios using the Monte Carlo method, and establish a feeder-level simulated load data set; as well as Step 3: Build a non-intrusive load decomposition neural network model, use voltage and total power as input features, and decomposed power or proportion as label data, perform training and learning, and obtain the trained load decomposition neural network model as the identification model; Step 4: During the identification and decomposition of the proportion of various types of load components under the actual feeder at the substation bus or above the grid level, feature extraction of the actual total power and voltage data is performed and used as the input of the trained identification model, and the corresponding load component proportion estimation result is obtained through the model output.

2. The feeder-level non-intrusive load component identification method for load modeling according to claim 1 is characterized in that: In step 1, the load model parameters of each type of load are obtained through static characteristic experiments, and the steady-state power models of each type of load and the total load are constructed, including the following steps: Step 11. In the power system dynamic simulation experiment system, static characteristic experiments are conducted on five typical load types: lighting (L), resistance (R), regulated power supply (SMPS), single-phase motor (IM1), and three-phase motor (IM3). A programmable variable frequency power supply is used to control the output voltage and frequency, and the steady-state voltage and steady-state power of the five types of load equipment at different voltages and frequencies are measured. Step 12: Determine the static load model parameters of each load type based on the measured steady-state data of each load type; and Step 13: Connect various load devices to the simulated power system in different proportions to simulate power consumption scenarios at different time periods within a day, and measure and record the steady-state data of the load group bus and various load branches, including steady-state power and steady-state voltage.

3. The feeder-level non-intrusive load component identification method for load modeling according to claim 2 is characterized in that: In step 12, the static load model parameters of each type of load are determined based on the measured steady-state data of each type of load, and the method is as follows: Where, P I is the constant current parameter of the active ZIP model, P Z is the constant impedance parameter of the active ZIP model, k is the number of static characteristic test groups, U k and P k are the voltage and active power measured in the kth experiment, and U0 and P0 are the voltage and active power at the steady-state point.

4. The feeder-level non-intrusive load component identification method for load modeling according to claim 1 is characterized in that: In step 2, generating different types of loads and total load powers under different voltages and component ratios by the Monte Carlo method, and establishing a feeder-level simulated load data set, includes the following steps: Step 21: Set the voltage range and use the Monte Carlo method to uniformly distribute and randomly generate voltage samples U. The number of samples is N. V ; Step 22: Divide the five types of loads into two major parts: static and dynamic. Among them, lighting, resistance and voltage-regulated power supply belong to the static part, and single-phase motors and three-phase motors belong to the dynamic part. The static and dynamic parts are set from 0% to 100% and in steps of 5%, and the sum of the proportions of the static and dynamic parts is equal to 100%, thereby generating 21 groups of proportions; among them, in each group of static and dynamic proportions, the Monte Carlo method is used to randomly generate the weight coefficient W of each load type in a uniform distribution, and the sum of the weight coefficients of the three load types belonging to the static part is equal to the proportion of the static part of the group, and the sum of the weight coefficients of the two load types belonging to the dynamic part is equal to the proportion of the dynamic part of the group; in each group of static and dynamic proportions, the number of randomly generated weight coefficients of each load type is N. W ;as well as Step 23: Obtain the total power under each condition of different voltage and load ratio through weighted summation to form a simulated load data set.

5. The feeder-level non-intrusive load component identification method for load modeling according to claim 4 is characterized in that: In step 2, the voltage sample matrix of the simulated load data set is constructed as follows: The weight coefficient matrix of the simulated load data set is constructed as follows: The total power of the simulated load data set is calculated by weighted summation as follows: Where, P i,t (U t ) and Q i,t (U t ) is the ZIP model of active power and reactive power with respect to voltage for various loads.

6. The feeder-level non-intrusive load component identification method for load modeling according to claim 1 is characterized in that: In step 3, the non-intrusive load decomposition neural network model is constructed, using voltage and total power as input features and decomposed power or proportion as label data for training and learning to obtain an identification model, including the following steps: Step 31: Use the voltage U of the simulated load data set t , total active power P agg and total reactive power Q agg As the input feature vector, the weight coefficient matrix W is used as the target vector. A multi-layer perceptron (MLP) is used as the main body. Three fully connected layers are set as hidden layers between the input layer and the output layer to build a deep learning load decomposition neural network model. The model is trained to explore the deep nonlinear characteristics between the total load and the power proportion of each type of load; and Step 32: Apply the load decomposition neural network model trained in step 31 to the dynamic model load group experimental data to verify the effectiveness of the decomposition model again.

7. The feeder-level non-intrusive load component identification method for load modeling according to claim 6 is characterized in that: In step 31, the training process of the deep learning load decomposition neural network model specifically includes the following steps: The load decomposition model parameter training based on the multi-layer perceptron model is adopted, and the mean square error (MSE) is used as the loss function. The expression is: Where, represents the estimated value, y i represents the actual value, and m represents the total number of samples; Using R 2 And accuracy ACC is used as the evaluation index, and the expression is: Where, represents the estimated value, y i Indicates the actual value, represents the true power mean, and m represents the total number of samples.

8. The feeder-level non-intrusive load component identification method for load modeling according to claim 1 is characterized in that: In step 4, in the process of identifying and decomposing the proportion of each type of load component under the actual feeder of the substation bus or above the grid level, the actual total power and voltage data are extracted as the input of the trained identification model, and the corresponding load component proportion estimation result is obtained through the model output, including the following processing steps: Step 41: Use the low-voltage feeder energy management system data based on the marketing business system for preprocessing to determine the reference steady-state power at each moment and convert the actual power into per-unit power; as well as Step 42: Based on the processed low-voltage feeder data, the trained and verified load decomposition neural network model is used to decompose the load components to obtain the proportion of various load components of the low-voltage feeder.

9. A computer system, characterized in that: include: one or more processors; A memory storing operable instructions, wherein when the instructions are executed by the one or more processors, the one or more processors are caused to perform operations, wherein the operations include the process of executing the method according to any one of claims 1 to 8.

10. A computer-readable medium storing software, characterized in that: The software includes instructions that can be executed by one or more computers, and when the instructions are executed by the one or more computers, the process of the method according to any one of claims 1 to 8 is performed.

Images