Power distribution network load prediction method and system considering user behavior characteristics

Abstract

The invention discloses a power distribution network load prediction method and system considering user behavior characteristics, and the method specifically comprises the steps: firstly building a multi-dimensional prediction index system based on the user behavior characteristics based on the historical load operation data and natural environment meteorological data of a power distribution network; then, based on an index system, clustering analysis is carried out on historical load data through K-mean value and self-organizing mapping algorithm fusion, and different load features are extracted; then, adaptive optimization is carried out on model parameters of a time convolutional network (TCN) by adopting a myxobacteria optimization algorithm, and multi-type loads are predicted based on a TCN hyper-parameter optimization result; and finally, comparing and analyzing the prediction result by using the evaluation index. According to the method, the load prediction precision of the power distribution network is improved, the problems of active power uncertainty and voltage fluctuation caused by photovoltaic power generation are reduced, and the overall planning, fine scheduling and safe and stable operation capabilities of the power distribution network are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power dispatching and management, and in particular to a method and system for predicting load on a distribution network taking into account user behavior characteristics. Background Art

[0002] Photovoltaic power generation has random and volatile output, resulting in issues such as active power uncertainty and voltage fluctuations. These present significant challenges to the overall planning, precise scheduling, and safe and stable operation of distribution networks. By researching multi-type load forecasting technologies, it is possible to accurately grasp future load trends in high-penetration photovoltaic distribution network areas, providing data support for power companies to formulate scientific and reasonable scheduling plans, helping to achieve peak shaving and valley filling, and load balancing, thereby significantly improving the safety and stability of grid operations. For example, patent CN 116186548 A discloses a method for training a power load forecasting model and a power load forecasting method. This method obtains training factor data by screening the correlation between historical load data and influencing factor data. Finally, the training factor data and historical load data are used as training data to train the power load forecasting model, thereby improving the accuracy of power load forecasting. Patent CN 117060388 A discloses a method for training a photovoltaic load forecasting system and a photovoltaic load forecasting system. This method responds to calls from a photovoltaic monitoring cloud platform, creates a corresponding computing space, and uses historical data stored in cached data files to perform load forecasting. This method supports large-scale concurrent processing and reduces user wait time. However, since power load is affected by many complex factors, these traditional statistical analysis methods can no longer meet the growing demand for high-precision prediction.

[0003] Deep learning methods are increasingly being used in load forecasting. However, existing research has mostly focused on the impact of environmental and climate variables, often overlooking the characteristics of user electricity consumption behavior. The relatively narrow dimensions of influencing factors considered limit the model's predictive capabilities, particularly during peak demand periods, resulting in low forecast accuracy and an urgent need for further optimization. Summary of the Invention

[0004] The purpose of the present invention is to provide a distribution network load forecasting method and system that takes into account user behavior characteristics, thereby improving the distribution network load forecasting accuracy by reducing the active power uncertainty and voltage fluctuation problems caused by photovoltaic power generation, and at the same time improving the overall planning, fine scheduling and safe and stable operation capabilities of the distribution network.

[0005] The technical solution for achieving the purpose of the present invention is: a distribution network load forecasting method taking into account user behavior characteristics, comprising the following steps:

[0006] Step 1: Based on the historical load operation data of the distribution network and natural environment meteorological data, a multi-dimensional prediction indicator system based on user behavior characteristics is established;

[0007] Step 2: Based on the indicator system, cluster analysis is performed on historical load data by integrating K-means and self-organizing map algorithms to extract different load characteristics;

[0008] Step 3: Use the slime mold optimization algorithm to adaptively optimize the model parameters of the temporal convolutional network (TCN). Based on the TCN hyperparameter optimization results, predict multiple types of loads.

[0009] A distribution network load forecasting system taking into account user behavior characteristics is provided. The system is used to implement the distribution network load forecasting method taking into account user behavior characteristics. The system includes first to third modules, and the functions of each module are as follows:

[0010] The first module establishes a multi-dimensional prediction indicator system based on user behavior characteristics based on historical load operation data of the distribution network and natural environmental meteorological data;

[0011] The second module, based on the indicator system, performs cluster analysis on historical load data through the fusion of K-means and self-organizing map algorithms to extract different load characteristics;

[0012] The third module uses the slime mold optimization algorithm to adaptively optimize the model parameters of the temporal convolutional network (TCN), and predicts multiple types of loads based on the TCN hyperparameter optimization results.

[0013] A mobile terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for predicting load on a distribution network taking into account user behavior characteristics is implemented.

[0014] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the distribution network load forecasting method taking into account user behavior characteristics.

[0015] Compared with the existing technology, the present invention has the following significant advantages: (1) Based on the characteristics of user electricity consumption behavior, various influencing factors are comprehensively considered, which improves the prediction ability of the model and the accuracy of load forecasting, especially the prediction accuracy during peak electricity consumption periods; (2) It reduces the problems of active power uncertainty and voltage fluctuation caused by photovoltaic power generation, and improves the overall planning, fine scheduling and safe and stable operation capabilities of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The figure is a flow chart of a method for load forecasting of a distribution network taking into account user behavior characteristics according to the present invention.

[0017] Figure 2 Schematic diagram of the process of clustering analysis on historical load data in an embodiment of the present invention.

[0018] Figure 3 This is a graph showing the results of preliminary clustering of load data using K-means in an embodiment of the present invention.

[0019] Figure 4 This is a curve chart showing the result of secondary deep feature extraction of historical load data using the SOM algorithm in an embodiment of the present invention.

[0020] Figure 5 This is a curve diagram of the first type of typical load in an embodiment of the present invention.

[0021] Figure 6 This is a curve diagram of the second type of typical load in an embodiment of the present invention.

[0022] Figure 7 This is a curve diagram of the third type of typical load in an embodiment of the present invention.

[0023] Figure 8 Schematic diagram of the process of the slime mold optimization algorithm in an embodiment of the present invention.

[0024] Figure 9 Schematic diagram of the principle of the temporal convolutional neural network in an embodiment of the present invention.

[0025] Figure 10 This is a curve diagram of the prediction results of the first type of typical load in an embodiment of the present invention.

[0026] Figure 11 This is a curve diagram of the prediction results of the second type of typical load in an embodiment of the present invention.

[0027] Figure 12 This is a curve diagram of the prediction results of the third type of typical load in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] Combine Figure 1 The present invention provides a method for load forecasting of a distribution network taking into account user behavior characteristics, comprising the following steps:

[0029] Step 1: Based on the historical load operation data of the distribution network and natural environment meteorological data, a multi-dimensional prediction indicator system based on user behavior characteristics is established;

[0030] Step 2: Based on the indicator system, cluster analysis is performed on historical load data by integrating K-means and self-organizing map algorithms to extract different load characteristics;

[0031] Step 3: Use the slime mold optimization algorithm to adaptively optimize the model parameters of the temporal convolutional network (TCN). Based on the TCN hyperparameter optimization results, predict multiple types of loads.

[0032] As a specific example, the multi-dimensional prediction indicator system based on user behavior characteristics described in step 1 includes three key dimensions: natural environment, load characteristics, and user behavior.

[0033] As a specific example, in step 1, a multi-dimensional prediction index system based on user behavior characteristics is established based on the historical load operation data of the distribution network and the natural environment meteorological data, as follows:

[0034] Step 1.1: Establish natural environment indicators, including light intensity, humidity, and ambient temperature;

[0035] Step 1.2: Establish load characteristic indicators, including daily load rate and daily peak-to-valley difference rate, as follows:

[0036] (1) Daily load rate for:

[0037] (1)

[0038] Where, is the daily load average, is the maximum daily load;

[0039] (2) Daily peak-to-valley difference for:

[0040] (2)

[0041] Where, is the minimum daily load;

[0042] Step 1.3: Establish user behavior indicators, including user time-shifting behavior characteristics, user peak avoidance behavior characteristics, user electricity costs, and user emission pollution characteristics. The specific indicators are as follows:

[0043] (1) User staggered behavior characteristics: Rationally arrange the operating time of different users or equipment to avoid peak power consumption periods. The formula is as follows:

[0044] (3)

[0045] Where, For the user's staggered behavior characteristics, 、 Peak period shift hours, postponed hours of load; Before the peak hour Hourly load average; After peak hours Hourly load average; It is the minimum load value within 7 days during the peak period; is the peak period duration;

[0046] (2) Characteristics of user peak avoidance behavior: During peak hours, users reduce their electricity consumption by reducing interruptible loads, retaining only the loads necessary for user life or equipment production operations. The reliability of peak avoidance behavior is reflected by the ratio of interruptible load to load fluctuation rate. The formula is as follows:

[0047] (4)

[0048] Where, To avoid peak behavior characteristics of users, Producing necessary loads for users' lives or equipment; represents the load fluctuation rate, The value of is the ratio of the standard deviation to the mean of the user's typical load curve; for Load value at each moment; Indicates the number of daily load moments; Indicates the daily load average; for The lowest load in the past 30 days;

[0049] (3) User electricity cost: used to reflect the user's willingness to use electricity. The formula is as follows:

[0050] (5)

[0051] Where, The electricity cost for users, express Electricity load at all times; express Time-of-day electricity prices;

[0052] (4) User emission pollution characteristics: refers to the ratio of the difference between the user's daily pollution emissions and the maximum daily pollution emissions in the past 30 days to the total pollution emissions in the current area in the past 30 days. The formula is as follows:

[0053] (6)

[0054] Where, Emission pollution characteristics for users, The maximum daily pollution emission in the past 30 days; The current daily pollution emission of the user; It is the sum of daily pollutant emissions in the past 30 days.

[0055] As a specific example, based on the indicator system described in step 2, the historical load data is clustered and analyzed by fusing the K-means and self-organizing map algorithms to extract different load characteristics, as follows:

[0056] Step 2.1: Use K-means to preliminarily cluster the historical load data and find the optimal number of clusters. Using silhouette coefficient To determine, the larger the silhouette coefficient, the more compact the samples within the cluster, the larger the distance between clusters, and the better the clustering effect. The formula is as follows:

[0057] (7)

[0058] Where, Indicates the The average distance between a sample and other samples in a cluster; Indicates the The average distance between a sample in a cluster and samples in other clusters; Indicates the number of clusters;

[0059] Step 2.2: Use the indicator data in the multi-dimensional prediction indicator system and the K-means clustering results as input to the self-organizing map algorithm for secondary clustering analysis, as follows:

[0060] Step 2.2.1: Input sample Perform normalization; initialize and normalize the weight vector in the competitive layer, Indicates the neuron weight vectors, , Represents the total number of neuron weight vectors; for learning rate Assign initial value;

[0061] Step 2.2.2, calculate the input sample The distance between each neuron weight vector is calculated, and the closest neuron is defined as the winning neuron, and the weight vector of the winning neuron is updated. is the weight vector of the winning neuron, calculated using Euclidean distance:

[0062] (8)

[0063] Where, Represents the input sample With the The Euclidean distance between the weight vectors of neurons; Indicates the sample number, represents the total number of samples; Represents the input sample The samples; Indicates the The sample corresponding to neuron weight vectors;

[0064] Step 2.2.3: The winning neuron outputs 1, and the rest outputs 0. The weight vector of the winning neuron is adjusted according to the following formula:

[0065] (9)

[0066] Where, express The learning rate at the moment is in the range of (0,1) and gradually decreases as the weight vector is updated; for Moment The output of a neuron; Indicates the number of the winning neuron; express The weight vector of the winning neuron at the moment, express The weight vector of the winning neuron at that moment; express Moment The weight vector of each neuron, express Moment The weight vector of each neuron;

[0067] Step 2.2.4: Repeat steps 2.2.1 to 2.2.3 until the learning rate decays to 0, stop iteration, and output the results.

[0068] As a specific example, the slime mold optimization algorithm described in step 3 is used to adaptively optimize the model parameters of the temporal convolutional network (TCN). Based on the TCN hyperparameter optimization results, multiple types of loads are predicted as follows:

[0069] Step 3.1: Use the slime mold optimization algorithm to adaptively optimize the TCN model parameters, as follows:

[0070] Step 3.1.1, set in a The number of slime mold populations in the dimensional search space is , slime mold in the first The update position formula at the iteration is as follows:

[0071] (10)

[0072] Where, is the index of the current iteration, is the location of the slime mold, It is the new updated location of the slime mold; It is The position of the best solution at the iteration; and are the positions of two slime molds randomly selected from the population; represents the weight coefficient; represents the first control parameter, which is in the interval Random numbers within Used to simulate the dynamic changes when slime mold approaches the target; is the second control parameter, which is used to determine the position update mode of the slime mold; is the index of slime mold, It is The fitness of a slime mold, is the fitness of the best solution; is the maximum number of iterations; is the third control parameter, whose value decreases linearly from 1 to 0, and is used to evaluate the changes in the slime mold's use of historical data; is a range The first random number within;

[0073] Weight coefficient represents the oscillation frequency of the biological oscillator, and the formula is as follows:

[0074] (11)

[0075] Where, It is The weight coefficient of each slime mold; is in the interval Random numbers within; is the sort index of the slime mold, and represent the best and worst fitness respectively, Used to reduce the rate of change of fitness; represents slime molds whose fitness is in the first half of the population, represents the rest of the slime molds;

[0076] Step 3.1.2: The formula for updating the position of slime mold is as follows:

[0077] (12)

[0078] Where, and Represent the upper and lower bound constraints of the search space respectively; is an interval Random numbers within; represents the proportion parameter of randomly distributed slime mold individuals in the population;

[0079] Step 3.1.3: By continuously adjusting the parameters 、 and , to achieve the search for the best result in the search space;

[0080] Step 3.2: Based on the TCN hyperparameter optimization results, predict various loads:

[0081] The TCN structure includes three key mechanisms: causal convolution, dilated convolution, and residual connection. Causal convolution relies only on current and historical data for calculation, making the model consistent with temporal causality; dilated convolution expands the receptive field by sampling the input sequence at intervals; residual connection is used to stabilize the learning process and alleviate the gradient vanishing phenomenon that occurs in deep network training.

[0082] The mathematical representation of dilated convolution is shown below:

[0083] (13)

[0084] Where, Indicates The convolution result at the moment; represents the expansion factor; Represents the size of the convolution kernel, Represents the index variable in the convolution kernel; Represents the historical load input sequence at time The value at the moment is used as the input of the convolution operation; Represents the convolution calculation performed on historical load data; represents the filter coefficients of TCN; represents the receptive field; Represents the size of the one-dimensional convolution kernel.

[0085] As a specific example, step 3 is followed by step 4, which uses evaluation indicators to compare and analyze the prediction results, as follows:

[0086] The load forecast accuracy evaluation indicators include the mean absolute error percentage and root mean square error , the formula is as follows:

[0087] (14)

[0088] (15)

[0089] Where, is the total number of load forecast samples, Index variable representing load forecast samples; For the The predicted value of the sample load; For the The true value of the sample loading;

[0090] If the mean absolute percentage error and root mean square error If both are less than the set value, the prediction results meet the requirements.

[0091] The present invention also provides a distribution network load forecasting system that takes into account user behavior characteristics. The system is used to implement the distribution network load forecasting method that takes into account user behavior characteristics. The system includes first to third modules, and the functions of each module are as follows:

[0092] The first module establishes a multi-dimensional prediction indicator system based on user behavior characteristics based on historical load operation data of the distribution network and natural environmental meteorological data;

[0093] The second module, based on the indicator system, performs cluster analysis on historical load data through the fusion of K-means and self-organizing map algorithms to extract different load characteristics;

[0094] The third module uses the slime mold optimization algorithm to adaptively optimize the model parameters of the temporal convolutional network (TCN), and predicts multiple types of loads based on the TCN hyperparameter optimization results.

[0095] The present invention also provides a mobile terminal comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for predicting the load of a distribution network taking into account user behavior characteristics is implemented.

[0096] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the distribution network load forecasting method taking into account user behavior characteristics.

[0097] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0098] Example

[0099] This example uses the historical load operation data and meteorological data of a certain region from 2019 to 2020 as an example for analysis. The load data has a total of 11 users, and the timestamps are unified to a 15-minute time scale. The data set is divided into a training set and a test set with a ratio of 7:3. The specific steps are as follows:

[0100] Step 1: Based on the historical load operation data of the distribution network and natural environment meteorological data, a multi-dimensional prediction indicator system based on user behavior characteristics is established;

[0101] This embodiment is based on two years of historical load operation data and meteorological data, and calculates through a multi-dimensional indicator system to obtain 9 indicator data sets corresponding to the load data.

[0102] Step 2: Based on the index data, the historical load data is clustered and analyzed by fusion of K-means and self-organizing map algorithms to extract different load characteristics, such as Figure 2 As shown, the details are as follows:

[0103] In this embodiment, after preprocessing the original load data, cluster analysis is performed on the load data of 11 users. The optimal number of clusters obtained by K-means is 4. Figure 3 As shown in the figure, the first clustering result, load characteristics, and user behavior characteristic indicators are used as the input of SOM to perform secondary deep clustering. The result is 3 categories, as shown in the figure below. Figure 4 As shown in Figure 2. It is more practical to group loads with the same user behavior characteristics into one category. The three typical load curves obtained are shown in Figure 2. Figures 5 to 7 As shown. Combining the data characteristics of different types of load users and a comprehensive analysis of the living and production methods in the area, this embodiment divides them into three typical load types: Category 1 is industrial area load, Category 2 is residential area load, and Category 3 is commercial area load. Among them, industrial users have relatively stable load changes and little difference between peaks and valleys because their production processes are mostly continuous or follow fixed work shifts, showing the typical characteristics of the Category 1 load curve; residential users' electricity consumption behavior is greatly affected by the rhythm of daily life and is highly volatile. Electricity consumption increases during the day, especially during the morning and evening peaks, and reaches a trough during rest at night, which conforms to the characteristics of the Category 2 load curve; commercial users mainly have significant electricity consumption peaks from approximately 9:00 to 17:00 on weekdays. Due to the reduction of business activities at night and during holidays, the electricity load drops significantly, reflecting the typical pattern of the Category 3 load.

[0104] Step 3: Based on the improved time series convolutional neural network algorithm based on the fusion slime mold optimization algorithm, multi-type loads are predicted to achieve adaptive optimization of model parameters; the SMA algorithm flow chart is as follows: Figure 8 As shown, the principle diagram of the TCN algorithm is as follows Figure 9 shown.

[0105] Step 4: Use evaluation indicators to compare and analyze the prediction results.

[0106] In this embodiment, the clustered historical load data and the corresponding index data are input into the TCN model for training, and a multi-input single-output strategy is adopted. The main parameter settings of the TCN initial model are shown in Table 1. The SMA algorithm is added to the TCN algorithm to adjust the hyperparameters of the TCN model to minimize the model prediction error. The optimized hyperparameters mainly include batch_size, learning rate, number of filters and number of iterations. The TCN models of various types of loads are optimized separately through SMA to ensure that each mode can better adapt to the characteristics of the load. The hyperparameter optimization adjustment results of the three types of load models are shown in Table 2. The three types of load test sets are predicted using the trained and optimized TCN model. Some of the prediction results are shown in Table 2. Figures 10 to 12 The results show that the prediction accuracy can reach a relatively ideal state, with industrial load prediction errors ranging from -3.5 to 1.5; commercial load prediction errors ranging from -1.8 to 0.9; and residential load prediction errors ranging from -0.6 to 0.2. Finally, MAPE and RMSE were used as evaluation criteria to measure the model's prediction accuracy, and the improved prediction method proposed in this paper was compared with other prediction methods. The evaluation index comparison data is shown in Table 3. As shown in Table 3, the improved prediction algorithm based on SMA-TCN proposed in this paper has significantly improved accuracy.

[0107] Table 1 Main parameter settings of the TCN initial model

[0108]

[0109] Table 2 SMA hyperparameter adjustment values for various loads

[0110]

[0111] Table 3 Comparison of errors of different prediction methods

[0112]

[0113] The above-described embodiments are merely specific implementation methods of the present application and are intended to illustrate the technical solutions of the present application, not to limit them. The scope of protection of the present application is not limited thereto. Although the present application has been described in detail with reference to the above-described embodiments, it should be understood by those skilled in the art that any person skilled in the art can modify or easily conceive of variations to the technical solutions described in the above-described embodiments within the technical scope disclosed in the present application, or substitute equivalent features for some of the technical features thereof. Such modifications, variations, or substitutions do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application and are therefore intended to be encompassed within the scope of protection of the present application.

Claims

1. A method for load forecasting of a distribution network taking into account user behavior characteristics, characterized in that: The following steps are involved: Step 1: Based on the historical load operation data of the distribution network and natural environment meteorological data, a multi-dimensional prediction indicator system based on user behavior characteristics is established; Step 2: Based on the indicator system, cluster analysis is performed on historical load data by integrating K-means and self-organizing map algorithms to extract different load characteristics; Step 3: Use the slime mold optimization algorithm to adaptively optimize the model parameters of the temporal convolutional network (TCN). Based on the TCN hyperparameter optimization results, predict multiple types of loads.

2. The method for load forecasting of a distribution network taking into account user behavior characteristics according to claim 1, characterized in that: The multi-dimensional prediction indicator system based on user behavior characteristics described in step 1 includes three key dimensions: natural environment, load characteristics, and user behavior.

3. The method for load forecasting of a distribution network taking into account user behavior characteristics according to claim 2, characterized in that: Based on the historical load operation data of the distribution network and the natural environment meteorological data described in step 1, a multi-dimensional prediction index system based on user behavior characteristics is established, as follows: Step 1.1: Establish natural environment indicators, including light intensity, humidity, and ambient temperature; Step 1.2: Establish load characteristic indicators, including daily load rate and daily peak-to-valley difference rate, as follows: (1) Daily load rate for: （1） Where, is the daily load average, is the maximum daily load; (2) Daily peak-to-valley difference for: （2） Where, is the minimum daily load; Step 1.3: Establish user behavior indicators, including user staggered behavior characteristics, user peak avoidance behavior characteristics, user electricity costs, and user emission pollution characteristics, as follows: (1) User staggered behavior characteristics: Arrange the operating time of different users or equipment to avoid peak power consumption periods. The formula is as follows: （3） Where, For the user's staggered behavior characteristics, 、 Peak period shift hours, postponed hours of load; Before the peak hour Hourly load average; After peak hours Hourly load average; It is the minimum load value within 7 days during the peak period; is the peak period duration; (2) Characteristics of user peak avoidance behavior: During peak hours, users reduce their electricity consumption by reducing interruptible loads, retaining only the loads necessary for user life or equipment production operations. The reliability of peak avoidance behavior is reflected by the ratio of interruptible load to load fluctuation rate. The formula is as follows: （4） Where, To avoid peak behavior characteristics of users, Producing necessary loads for users' lives or equipment; represents the load fluctuation rate, The value of is the ratio of the standard deviation to the mean of the user's typical load curve; for Load value at each moment; Indicates the number of daily load moments; Indicates the daily load average; for The lowest load in the past 30 days; (3) User electricity cost: used to reflect the user's willingness to use electricity. The formula is as follows: （5） Where, The electricity cost for users, express Electricity load at all times; express Time-of-day electricity prices; (4) User emission pollution characteristics: refers to the ratio of the difference between the user's daily pollution emissions and the maximum daily pollution emissions in the past 30 days to the total pollution emissions in the current area in the past 30 days. The formula is as follows: （6） Where, Emission pollution characteristics for users, The maximum daily pollution emission in the past 30 days; The current daily pollution emission of the user; It is the sum of daily pollutant emissions in the past 30 days.

4. The method for load forecasting of a distribution network taking into account user behavior characteristics according to claim 3, characterized in that: Based on the indicator system described in step 2, the historical load data is clustered and analyzed by integrating the K-means and self-organizing map algorithms to extract different load characteristics, as follows: Step 2.1: Use K-means to preliminarily cluster the historical load data and find the optimal number of clusters. Using silhouette coefficient To determine, the larger the silhouette coefficient, the more compact the samples within the cluster, the larger the distance between clusters, and the better the clustering effect. The formula is as follows: （7） Where, Indicates the The average distance between a sample and other samples in a cluster; Indicates the The average distance between a sample in a cluster and samples in other clusters; Indicates the number of clusters; Step 2.2: Use the indicator data in the multidimensional prediction indicator system and the K-means clustering results as input to the self-organizing map algorithm for secondary clustering analysis.

5. The method for load forecasting of a distribution network taking into account user behavior characteristics according to claim 4, characterized in that: The secondary cluster analysis described in step 2.2 is as follows: Step 2.2.1: Input sample Perform normalization; initialize and normalize the weight vector in the competitive layer, Indicates the neuron weight vectors, , Represents the total number of neuron weight vectors; for learning rate Assign initial value; Step 2.2.2, calculate the input sample The distance between each neuron weight vector is calculated, and the closest neuron is defined as the winning neuron, and the weight vector of the winning neuron is updated. is the weight vector of the winning neuron, calculated using Euclidean distance: （8） Where, Represents the input sample With the The Euclidean distance between the weight vectors of neurons; Indicates the sample number, represents the total number of samples; Represents the input sample The samples; Indicates the The sample corresponding to neuron weight vectors; Step 2.2.3: The winning neuron outputs 1, and the rest outputs 0. The weight vector of the winning neuron is adjusted according to the following formula: （9） Where, express The learning rate at the moment is in the range of (0,1) and gradually decreases as the weight vector is updated; for Moment The output of a neuron; Indicates the number of the winning neuron; express The weight vector of the winning neuron at the moment, express The weight vector of the winning neuron at that moment; express Moment The weight vector of each neuron, express Moment The weight vector of each neuron; Step 2.2.4: Repeat steps 2.2.1 to 2.2.3 until the learning rate decays to 0, stop iteration, and output the results.

6. The method for load forecasting of a distribution network taking into account user behavior characteristics according to claim 5, characterized in that: In step 3, the slime mold optimization algorithm is used to adaptively optimize the model parameters of the temporal convolutional network (TCN). Based on the TCN hyperparameter optimization results, multiple types of loads are predicted as follows: Step 3.1: Use the slime mold optimization algorithm to adaptively optimize the TCN model parameters, as follows: Step 3.1.1, set in a The number of slime mold populations in the dimensional search space is , slime mold in the first The update position formula at the iteration is as follows: （10） Where, is the index of the current iteration, is the location of the slime mold, It is the new updated location of the slime mold; It is The position of the best solution at the iteration; and are the positions of two slime molds randomly selected from the population; represents the weight coefficient; represents the first control parameter, which is in the interval Random numbers within Used to simulate the dynamic changes when slime mold approaches the target; is the second control parameter, which is used to determine the position update mode of the slime mold; is the index of slime mold, It is The fitness of a slime mold, is the fitness of the best solution; is the maximum number of iterations; is the third control parameter, whose value decreases linearly from 1 to 0, and is used to evaluate the changes in the slime mold's use of historical data; is a range The first random number within; Weight coefficient represents the oscillation frequency of the biological oscillator, and the formula is as follows: （11） Where, It is The weight coefficient of each slime mold; is in the interval Random numbers within; is the sort index of the slime mold, and represent the best and worst fitness respectively, Used to reduce the rate of change of fitness; represents slime molds whose fitness is in the first half of the population, represents the rest of the slime molds; Step 3.1.2: The formula for updating the position of slime mold is as follows: （12） Where, and Represent the upper and lower bound constraints of the search space respectively; is an interval Random numbers within; represents the proportion parameter of randomly distributed slime mold individuals in the population; Step 3.1.3: By continuously adjusting the parameters 、 and , to achieve the search for the best result in the search space; Step 3.2: Based on the TCN hyperparameter optimization results, predict various loads: The TCN structure includes three key mechanisms: causal convolution, dilated convolution, and residual connection. Causal convolution relies only on current and historical data for calculation, making the model consistent with temporal causality; dilated convolution expands the receptive field by sampling the input sequence at intervals; residual connection is used to stabilize the learning process and alleviate the gradient vanishing phenomenon that occurs in deep network training. The mathematical representation of dilated convolution is shown below: （13） Where, Indicates The convolution result at the moment; represents the expansion factor; Represents the size of the convolution kernel, Represents the index variable in the convolution kernel; Represents the historical load input sequence at time The value at the moment is used as the input of the convolution operation; Represents the convolution calculation performed on historical load data; represents the filter coefficients of TCN; represents the receptive field; Represents the size of the one-dimensional convolution kernel.

7. The method for load forecasting of a distribution network taking into account user behavior characteristics according to claim 6, characterized in that: Step 3 is followed by step 4, which uses evaluation indicators to compare and analyze the prediction results, as follows: The load forecast accuracy evaluation indicators include the mean absolute error percentage and root mean square error , the formula is as follows: （14） （15） Where, is the total number of load forecast samples, Index variable representing load forecast samples; For the The predicted value of the sample load; For the The true value of the sample loading; If the mean absolute percentage error and root mean square error If both are less than the set value, the prediction results meet the requirements.

8. A distribution network load forecasting system taking into account user behavior characteristics, characterized in that: The system is used to implement the distribution network load forecasting method taking into account user behavior characteristics as described in any one of claims 1 to 7. The system includes a first module to a third module, and the functions of each module are as follows: The first module establishes a multi-dimensional prediction indicator system based on user behavior characteristics based on historical load operation data of the distribution network and natural environmental meteorological data; The second module, based on the indicator system, performs cluster analysis on historical load data through the fusion of K-means and self-organizing map algorithms to extract different load characteristics; The third module uses the slime mold optimization algorithm to adaptively optimize the model parameters of the temporal convolutional network (TCN), and predicts multiple types of loads based on the TCN hyperparameter optimization results.

9. A mobile terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the distribution network load forecasting method taking into account user behavior characteristics as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the distribution network load forecasting method taking into account user behavior characteristics as described in any one of claims 1 to 7 are implemented.

Images