Load scene clustering-based proxy electricity purchasing user load prediction method and system

Abstract

The invention discloses a load scene clustering-based proxy electricity purchasing user load prediction method and system. The method comprises the steps of obtaining load historical data and corresponding historical meteorological data of a proxy electricity purchasing user; carrying out feature extraction on the historical load data to obtain internal key features of the load; clustering load internal key features of the load historical data, and calculating key meteorological factors of each clustering category load; performing load prediction on the historical data of the to-be-measured load by using the general prediction model to obtain an initial load prediction result; extracting corresponding load internal key features based on the initial load prediction result; and judging a clustering category to which the corresponding load intrinsic key features belong and a key meteorological factor corresponding to the category, and inputting the historical data of the load to be measured and the corresponding key meteorological factor into the load prediction model of the corresponding category to obtain a final load prediction result. The influence of meteorological factors under different load scene categories is fully considered, and the accuracy of a load prediction result is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and in particular to a method and system for forecasting load of proxy power purchasing users based on load scenario clustering. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Proxy power purchasers refer to industrial and commercial users who have not purchased electricity directly from the power market, and are purchased from the power market by power grid companies as agents. The load of proxy power purchasers has the characteristics of seasonal changes, holiday impacts, and time differences. Therefore, in order to ensure the stability of power supply, optimize the allocation of power resources, and improve economic benefits, it is necessary to predict the load of proxy power purchasers.

[0004] The load of power purchasing agents varies greatly in size, and the characteristics of each type of user are also significantly different. The load patterns of some users vary seasonally, and climate changes in different seasons lead to fluctuations in electricity consumption, such as increased demand for cooling in summer and heating in winter. Some users have industry characteristics with time differences, and different industries have different electricity consumption patterns. For example, the manufacturing industry may have a large electricity consumption during the day, while the retail industry may have a higher electricity demand at night; in addition, the load electricity demand may also change and differ greatly under different weather scenarios; therefore, considering the changes in load electricity demand under different weather scenarios is of great significance for load forecasting and optimal scheduling of power grid systems.

[0005] However, although meteorological factors will have a certain impact on electricity load, the load data under the same meteorological scenario often do not have a high similarity, and meteorological characteristics are difficult to effectively represent the overall characteristics of the load. There may be large differences in load characteristics under similar meteorological characteristics. Most of the load forecasting methods that consider weather factors disclosed in the prior art directly perform cluster analysis on weather factors, which may lead to inaccurate load forecasting results. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a load forecasting method and system for proxy electricity purchasing users based on load scenario clustering. A load forecasting model is constructed based on all meteorological characteristics, and the intrinsic key characteristics of the load are selected and clustered. The temporal correlation between the meteorology and the load in the cluster is analyzed, and then the key meteorological factors under each category are extracted respectively, so as to realize accurate prediction of the load of proxy electricity purchasing users.

[0007] In some embodiments, the following technical solutions are adopted: A load forecasting method for proxy power purchasing users based on load scenario clustering, comprising: Obtain the load historical data and corresponding historical meteorological data of the proxy power purchasing users; Extracting features from the load history data to obtain the inherent key features of the load; Cluster the inherent key characteristics of load historical data and calculate the key meteorological factors of each cluster category load; Use the general prediction model to perform load prediction on the historical data of the load to be measured to obtain the initial load prediction result; extract the corresponding load intrinsic key characteristics based on the initial load prediction result; The cluster category to which the corresponding load intrinsic key feature belongs and the key meteorological factor corresponding to the category are determined, and the historical data of the load to be measured and the corresponding key meteorological factor are input into the load prediction model of the corresponding category to obtain the final load prediction result.

[0008] As a further solution, the key intrinsic characteristics of the load include: load peak value, load valley value, load peak-to-valley difference at noon, load daily power consumption and load ramp rate ratio.

[0009] As a further solution, the load intrinsic key features of the load history data are clustered, specifically: Determine the number of clusters K, and organize the key characteristics of the load in each historical day into cluster data; Randomly select K points in the data as centroids to achieve preliminary clustering; Calculate the distance between other points and the centroid, and aggregate them into the closest category to obtain a cluster; take each point in the cluster as the centroid, calculate the sum of the total distances between the centroid and other points, and select the point with the smallest total distance sum as the new centroid; repeat this process until the centroid no longer shifts.

[0010] As a further solution, the number of clusters K is determined as follows: According to experience, different cluster numbers K are selected for clustering, and the cluster center corresponding to each K value is determined; Calculate the average distance between each sample point and other samples in the same cluster under each cluster number K , and the average distance between each sample and the samples in the nearest neighbor cluster ; Then calculate the comprehensive evaluation coefficient P: ; ; in, is the evaluation coefficient of each sample, and N is the number of samples; The number of clusters K corresponding to the clustering result with the comprehensive evaluation coefficient P closest to 1 is selected as the optimal number of clusters.

[0011] As a further solution, the key meteorological factors of each cluster category load are calculated, specifically: Calculate the Gaussian similarity coefficient matrix between each group of historical load series and each corresponding type of meteorological series; Select the maximum value of each column in the Gaussian similarity coefficient matrix as the reference sequence; calculate the absolute difference between the Gaussian similarity coefficient matrix and the reference sequence corresponding to each element; calculate the grey correlation degree based on the maximum and minimum values ​​of the absolute difference to form a grey correlation degree matrix; Calculate the sum of each column of data in the grey correlation matrix as the sum of correlation, weight each row of the grey correlation matrix according to the sum of correlation, and then normalize the sum to obtain the weighted meteorological characteristic weight coefficient; The meteorological characteristic weight coefficients are sorted from large to small to obtain the key meteorological factors corresponding to each cluster category.

[0012] As a further solution, the universal prediction model adopts the iTransformer model, and directly trains the model using all historical load data and corresponding historical meteorological data.

[0013] As a further solution, after clustering the load intrinsic key features of the load history data, a load forecasting model is trained using the load history data under each category and the corresponding key meteorological factors, and the same number of load forecasting models as the clustering categories are obtained; Based on the clustering category to which the load intrinsic key characteristics of the initial load forecast result belong, the load forecasting model under the corresponding category is selected, and the historical data of the load to be measured and the corresponding key meteorological factors are used as input to obtain the final load forecasting result.

[0014] In other embodiments, the following technical solutions are adopted: A load forecasting system for proxy power purchasing users based on load scenario clustering, comprising: A data acquisition module is configured to acquire load historical data and corresponding historical meteorological data of the proxy electricity purchasing user; A key feature extraction module is configured to extract features from the load history data to obtain the load intrinsic key features; A key meteorological factor extraction module is configured to cluster the load intrinsic key features of the load historical data and calculate the key meteorological factors of each cluster category load; The initial load prediction module is configured to use a general prediction model to perform load prediction on the historical data of the load to be measured to obtain an initial load prediction result; and extract corresponding load intrinsic key features based on the initial load prediction result; The final load forecasting module is configured to determine the clustering category to which the corresponding load intrinsic key characteristics belong and the key meteorological factors corresponding to the category, input the historical data of the load to be measured and the corresponding key meteorological factors into the load forecasting model of the corresponding category, and obtain the final load forecasting result.

[0015] In other embodiments, the following technical solutions are adopted: A terminal device comprises a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store a plurality of instructions, wherein the instructions are suitable for being loaded by the processor and executing the above-mentioned method for predicting load of proxy electricity purchasing users based on load scenario clustering.

[0016] In other embodiments, the following technical solutions are adopted: A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor of a terminal device and executing the above-mentioned method for predicting loads of proxy electricity purchasing users based on load scenario clustering.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention extracts the intrinsic key features of the power load, clusters the intrinsic key features of the historical load data, and then extracts the key meteorological factors for each category scenario and trains the load forecasting model respectively, effectively avoiding the uncertainty error introduced into the forecasting model when the meteorological factors with low correlation are used as input.

[0018] The present invention first uses a general prediction model to obtain preliminary load forecasting results, uses the inherent key characteristics of the preliminary load forecasting results to determine the clustering category to which they belong, and then uses the load forecasting model of the corresponding category to perform load forecasting, which can fully consider the influence of meteorological factors under different load scenario categories and improve the accuracy of load forecasting results.

[0019] (2) The present invention determines the optimal number of clusters by calculating the comprehensive evaluation coefficient to avoid the situation where the number of clusters is too small, there is little similar information within the clusters, and there is a lack of consistent features; and avoid the situation where the number of clusters is too large, there are few samples within the clusters, and it is difficult to effectively construct a neural network prediction model.

[0020] (3) The present invention obtains the key meteorological factors of each category of load based on the improved grey correlation coefficient, introduces nonlinear characteristics through the Gaussian distance coefficient, extracts the temporal correlation characteristics of the meteorological and load of a large number of proxy electricity purchasing users, screens the key meteorological factors that affect load changes, and further improves the accuracy of load forecasting results.

[0021] Other features and advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of a method for load forecasting of proxy power purchasing users based on load scenario clustering in an embodiment of the present invention; Figure 2 This is a schematic diagram of solving the load climbing rate ratio in an embodiment of the present invention. , , The load peak-to-valley differences corresponding to the three power curves are shown respectively; Figure 3 A schematic diagram of solving the load peak-valley difference in an embodiment of the present invention; Figure 4 A schematic diagram of a clustering process in an embodiment of the present invention; Figure 5 Schematic diagram of the overall architecture of the iTransformer model in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0024] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0025] Embodiment 1 In one or more embodiments, a method for predicting load of proxy power purchasers based on load scenario clustering is disclosed, combined with Figure 1 , specifically including the following process: S101: Obtain historical load data and corresponding historical meteorological data of the proxy electricity purchasing user.

[0026] Specifically, the load history data includes the measured load data for each day in a historical year, usually power data, with a time resolution of one hour; Historical meteorological data includes numerical weather forecast data for each day of the historical year, including temperature, perceived dimensions, humidity, irradiance, air pressure and pollutant concentrations.

[0027] S102: Extract features from historical load data to obtain key intrinsic features of the load.

[0028] In order to achieve clustering of power load scenarios, this embodiment first conducts an in-depth analysis of the power load data and extracts representative features. According to the actual change trend of the daily power curve of the power load, the following key parameters are selected to construct the feature set: load peak, load valley, load peak-valley difference at noon, load daily power consumption, and load noon ramp rate. These features can fully reflect the daily change characteristics and dynamic behavior of the power load.

[0029] Load peak and valley values ​​are among the most important parameters in the power system, as they directly affect the dispatching and operational safety of the power grid. By analyzing the load peak-valley difference, we can further understand the maximum load fluctuation range in a day, which is crucial for the load management and peak-shaving needs of the power grid. The daily power consumption of the load provides the total energy consumption in a day, while the load noontime ramp rate reflects the growth rate of the load during the noon period, which is of great significance for predicting and responding to the power demand during the peak noon period.

[0030] The following are the extracted key intrinsic features of the load: (1) Peak load Load Peak refers to the highest point in the power demand in the power system within a certain period of time. This peak power demand usually occurs at a specific time of the day, such as during the daytime of weekdays or during the peak use of air conditioners in summer. Load Peak is crucial to the stable operation of the power system because it requires the power grid to provide sufficient power generation capacity to meet the high demand during this period to avoid power outages or unstable power supply.

[0031] Load peaks have obvious temporal and seasonal characteristics: first, in terms of timing, load peaks usually occur in specific time periods, such as summer afternoons or winter evenings, which is related to people's electricity usage habits and seasonal climate changes; in terms of predictability, through the analysis of historical data, it is possible to predict the time of load peaks and possible peak loads, so as to make adjustment preparations in advance.

[0032] The peak load formula is defined as follows: (1) in, is the load peak, is the sth load power, and S is the total load power.

[0033] (2) Load valley value Load Valley refers to the lowest point in the power system's electricity demand over a certain period of time. This term is used to describe the low period of electricity consumption, which usually occurs during periods of low electricity demand, such as late at night or early in the morning, when industrial production, commercial activities and residential electricity consumption are at a low level. Load Valley is a characteristic point on the load curve of the power system, which corresponds to the peak load and reflects the fluctuation of electricity demand during the day.

[0034] Load valleys usually occur during the period when electricity demand is lowest, and they also vary due to seasonal, weather and regional differences. At the same time, as distributed renewable energy is connected to the grid, load valley periods may coincide with peak renewable energy generation periods (such as wind power generation at night), which provides grid operators with opportunities to dispatch renewable energy.

[0035] The load valley formula is defined as follows: (2) in, is the load valley value.

[0036] (3) Load peak-to-valley difference at noon This embodiment uses the midday peak-to-valley difference of the load to indicate the midday drop in load, reflecting the degree of influence of distributed photovoltaic access to the distribution network on the load power. The peak value and valley value of the load in a fixed period of time are selected and subtracted to obtain the peak-to-valley difference of the midday load. The greater the peak-to-valley difference, the more obvious the load drop characteristics at noon. The calculation formula is as follows: (3) In the formula, is the peak value of load power during the noon period, is the valley value of load power during the noon period, The range is 7:00-15:00, so It is the difference between the peak value and the valley value of the load power during the period; Figure 3 A schematic diagram for solving the load peak-valley difference is given.

[0037] (4) Daily power consumption of load The daily power consumption of the load is one of the important input characteristics of the power load forecast. As a key characteristic of the load, the daily power consumption of the load reflects the daily power consumption level of the load. By analyzing the daily power consumption of the load, we can accurately grasp the power consumption characteristics of the load and predict future power demand, thereby helping the power system to carry out effective power generation and scheduling planning.

[0038] (4) In the formula, Indicates the total amount of electricity generated by distributed photovoltaics in a single day, Indicates the sampling point interval. Here, 0.25 hours is used as the sampling interval, and 96 load power values ​​can be obtained every day.

[0039] (5) Load climbing rate ratio In this embodiment, in order to accurately capture and describe the drop characteristics in the daily power curve of the power load, a novel indicator, the load ramp rate ratio, is proposed. This indicator is intended to quantify the change trend of the power load in a specific time period, especially the stage when the load power drops sharply at noon. The load ramp rate ratio can effectively reflect the drop characteristics of the load, which is of great significance for the dispatching and operation management of the power grid.

[0040] The calculation formula of the load ramp rate ratio is defined based on the overall law of the load power curve. This formula comprehensively considers the load change rate at noon. By comparing and analyzing these change rates, we can identify the drop area in the load curve. The load ramp rate ratio formula is defined as follows.

[0041] (5) In the formula, is the load peak-to-valley difference, is the climbing time, Indicates the midday downhill slope of the load.

[0042] Figure 2 A schematic diagram for solving the load ramp rate ratio is given. , , They represent the load peak-to-valley differences corresponding to the three power curves respectively, and the climbing rate is obtained by dividing them by the corresponding climbing time.

[0043] Through the comprehensive extraction and analysis of these load characteristics, different power load scenarios can be more accurately identified and classified, thus providing a scientific basis for the operation and planning of the power grid.

[0044] Furthermore, cluster analysis technology can be used to identify periods of high and low load demand, identify load power characteristics and power consumption patterns in different scenarios, and thus optimize the allocation and use of power resources, promote a better understanding of the relationship between power load and meteorological conditions, and provide support for power load forecasting and demand response.

[0045] By extracting the above parameters, a load characteristic matrix consisting of five-dimensional features, namely, load peak value, valley value, peak-to-valley difference, climbing rate ratio, and daily power consumption, can be obtained. Among them, in the load clustering analysis, the peak load can help identify the peak period in the power consumption pattern, and the value load can help identify the low peak period in the power consumption pattern. By analyzing the peak-to-valley value load, the allocation of power generation resources can be optimized to ensure that there is sufficient power supply during the peak power consumption period, so as to facilitate the adoption of peak-to-valley filling measures and improve the load rate of the power grid; the daily power consumption of the load is an important indicator to measure the power consumption level of the load within a day, which helps to identify the power consumption habits of different users and can well reflect the daily power consumption behavior characteristics of the load; the peak-to-valley difference of the load is used as a measure of the load drop caused by the influence of new energy at noon, which helps to analyze the dynamic change characteristics of the load. Clustering the load time series curve based on the above characteristics can accurately grasp the load time series characteristics and facilitate further classification and prediction of the load.

[0046] S103: Clustering the intrinsic key characteristics of the load historical data, and calculating the key meteorological factors of the load of each cluster category.

[0047] In this embodiment, based on the load intrinsic key features extracted in S102, a load daily power behavior feature set that can represent the load power level is constructed. : (6) in, They are the peak value, valley value, peak-to-valley difference, ramp rate ratio and daily power consumption of the load daily power curve.

[0048] If historical load data within a year is obtained, it corresponds to 365 sets of load intrinsic key characteristic data.

[0049] To cluster these data, the core mechanism of the existing K-Means algorithm is to continuously find the center point of each cluster, that is, the mean; this algorithm iteratively optimizes these mean points and finally achieves data clustering. However, the K-Means algorithm is sensitive to noise and outliers when processing data, which limits its performance in data sets containing noisy data.

[0050] In order to solve the above problems, combined Figure 4 , the clustering algorithm in this embodiment is as follows: S1031: Determine the value of the number of clusters K, that is, decide how many categories the data should be aggregated into; In this embodiment, firstly, different k values ​​are selected as the number of clusters for clustering based on experience, clustering results are obtained for different k values, and the cluster center corresponding to each k value is determined; Calculate the average distance between each sample point and other samples in the same cluster under each cluster number K , and the average distance between each sample and the samples in the nearest neighbor cluster ; Then calculate the comprehensive evaluation coefficient P: (7) (8) in, is the evaluation coefficient of each sample, and N is the number of samples.

[0051] The closer the comprehensive evaluation coefficient P is to 1, the better the clustering effect is. In this embodiment, the number of clusters K corresponding to the clustering result with the comprehensive evaluation coefficient P closest to 1 is selected as the optimal number of clusters. As a specific example, the optimal number of clusters selected in this embodiment is 4.

[0052] S1032: Clustering the key internal features of the load for each historical day into cluster data; S1033: Randomly select K points in the data as centroids to achieve preliminary clustering; S1034: Calculate the distances between other points and the centroid, and aggregate them into the category with the closest distance to obtain a cluster; S1035: Take each point in the cluster as the centroid, calculate the sum of the total distances between the centroid and other points, and select the point with the smallest total distance sum as the new centroid; S1036: Repeat the process of S1034-S1035 until the center of mass no longer deviates.

[0053] The centroids selected in the clustering method of this embodiment are definite existing data, and the centroid data can be used to identify key meteorological features and perform load forecasting.

[0054] This embodiment uses a clustering method to cluster the daily power consumption characteristics of the load in five dimensions, and can obtain user loads divided by category, thereby achieving effective division of load power consumption scenarios. The optimal number of clustering clusters is selected through the evaluation coefficient, thereby improving the accuracy of the clustering results.

[0055] After clustering is completed, the correlation between multivariate meteorological characteristics and load is calculated, and then the correlation characteristics between meteorological characteristics and load are characterized, and the key meteorological factors corresponding to each clustering category are screened out; for example: the service industry has obvious seasonal patterns and is greatly affected by temperature; the power load of outdoor users has a strong temporal correlation with precipitation and pollutant concentration; some industrial users with self-built distributed photovoltaic parks have obvious load drops at noon, and their load characteristics show a clear correlation with irradiance, etc. The specific implementation process is as follows: (1) Assume that there are a total of n groups of load data in a certain cluster category, and each group of load data can be composed of the load power at each time of the day; each group of load data corresponds to a group of meteorological data, and each group of meteorological data is composed of a meteorological array at each time of the day, and each meteorological array contains m types of meteorological data (such as temperature, body dimension, humidity, irradiance, air pressure and pollutant concentration, etc.).

[0056] set up is a set of load series data in this cluster category, A sequence consisting of a certain type of meteorological data (such as temperature) corresponding to the group of load data; i =1,2,…,n; j =1,2,…,m.

[0057] The similarity coefficient between the load sequence and the meteorological sequence is calculated by Gaussian index. : (9) in, The flexure point is a point at which different similarities can be explored by setting the flexure point appropriately. .

[0058] Thus, the similarity coefficient sequence between the load sequence and each corresponding meteorological sequence can be obtained. ; Then we can get the Gaussian similarity coefficient matrix between n groups of load sequence data and the corresponding meteorological sequence of each type: , Representative The load series data of the group Gaussian similarity coefficient of similar meteorological characteristics.

[0059] (2) Since the ranges of Gaussian similarity coefficients of various meteorological characteristics are not consistent, it is easy to affect the accuracy of correlation calculation; therefore, this embodiment further processes the similarity coefficient matrix: Select the maximum value of each column, that is, the maximum value of the Gaussian similarity coefficient in each data set, as the reference sequence : .

[0060] In the formula, Represents the Gaussian similarity coefficient matrix Middle Since the Gaussian similarity coefficient ranges from 0 to 1, there is no need to perform dimensionless processing and directly calculate the matrix Grey relational degree of corresponding elements , the calculation process is as follows: (2-1) Calculation matrix With reference sequence The absolute difference of each element: ; (2-2) Select the maximum and minimum values ​​in the absolute difference matrix: (10) (11) (2-3) Finally, the grey correlation degree is obtained : (12) in, is the resolution coefficient, , which is used to adjust the differences between correlation coefficients, usually .

[0061] (2-4) The grey correlation matrix of Gaussian correlation coefficient is formed for: (13) (3) Each row represents the grey correlation of different meteorological characteristics of a set of load data. The total grey correlation of the corresponding load data can be obtained by directly summing up the matrix rows. However, in the load forecasting of proxy power purchasers, the influence of meteorological characteristics on power in different data sets is different. Directly summing up the correlation of meteorological characteristics often cannot reflect the differences of different meteorological characteristics. Therefore, this embodiment uses the sum of the correlation of each data set to predict the grey correlation of the load data. Normalize it. Assume that the sum of the correlations of each data set is ,in , and compare it with the grey correlation matrix Multiplying them together, we can get the weighted grey incidence matrix: (14) right By normalizing the sum, we can get the weighted meteorological characteristic weight coefficient: (15) By sorting the meteorological feature weight coefficients from large to small, the key meteorological factors corresponding to the clustering category can be obtained. For example, the meteorological categories corresponding to the first few data with the largest weight coefficients can be selected as the key meteorological factors.

[0062] This embodiment selects the key meteorological factors corresponding to each cluster category, and can consider different meteorological factors for different categories of data in a more targeted manner, thereby improving the accuracy of load forecasting results.

[0063] S104: Use a general prediction model to perform load prediction on the historical data of the load to be measured to obtain an initial load prediction result; and extract corresponding load intrinsic key features based on the initial load prediction result.

[0064] In this embodiment, all load historical data and their corresponding meteorological historical data are used as training sets to train the universal prediction model to obtain a trained universal prediction model.

[0065] Obtain historical data of the load to be tested, such as the load data of each time period of the previous day, and use the acquired data and the meteorological data corresponding to each time period as input to the trained general prediction model to predict the initial load prediction result, that is, the load power curve of each time period of the next day can be predicted.

[0066] Based on the predicted load power curve, the corresponding five key intrinsic features are extracted: load peak value, load valley value, load peak-to-valley difference at noon, load daily power consumption and load ramp rate ratio.

[0067] S105: Determine the cluster category to which the corresponding load intrinsic key feature belongs and the key meteorological factor corresponding to the category, input the historical data of the load to be measured and the corresponding key meteorological factor into the load prediction model of the corresponding category, and obtain the final load prediction result.

[0068] In this embodiment, after clustering is completed, a load forecasting model is constructed for each clustering category, and the load history data of each clustering category and the corresponding key meteorological factors are used as training data sets to train the corresponding load forecasting model to obtain K load forecasting models; that is, using the same clustering category T i Load historical data, train a load forecasting model M i , i represents the clustering category, i=1, 2,…, K; K is the number of clusters.

[0069] Based on the clustering result in S103, determine the clustering category T to which the key features of the initial load forecasting result obtained by the general forecasting model in S104 belong i , the load forecasting model M corresponding to this clustering category i , and the key meteorological factors corresponding to this category.

[0070] The load forecasting model M corresponding to the clustering category is obtained i After that, the historical load data to be measured (such as the load data of each period of the previous day) can be input into the trained load forecasting model M i , and obtain the final load forecast result.

[0071] As a specific example, the general prediction model and each load prediction model in this embodiment all use the iTransformer model. The advantage of this model is that the self-attention mechanism can avoid gradient explosion, better capture the periodic relationship in the time series, meet the prediction needs, and improve the prediction accuracy. Compared with the Transformer model, when merging multiple variable embeddings, there are problems such as difficulty in maintaining the correlation between data and affecting the capture of key meteorological features. The iTransformer model in this embodiment can embed the time series of each variable as an inverted embedding and use the self-attention mechanism to handle the correlation of multiple variables. The overall architecture of the iTransformer model is as follows: Figure 5 As shown, it includes three parts: embedding layer, encoder and projection layer: (1) Embedding layer The intrinsic characteristics of the classified load (peak value, valley value, peak-valley difference, ramp rate, total daily power consumption), key meteorological factors corresponding to the classified load, and historical data of the classified load are input into the embedding layer. iTransformer embeds each time series as a tag independently through inversion. The time series information of the input sequence is implicitly encoded in the neurons of the feedforward network, so there is no need for position embedding. The formula is: ;in: is the input matrix of the intrinsic characteristics of the classified load, the key meteorological factors corresponding to the classified load, and the historical data of the classified load. Embed tags for both, For embedding operation.

[0072] (2) Encoder Using the multi-head attention mechanism, we calculate the attention weights of multiple variables and assign higher weights to key features, so as to deeply mine the nonlinear features that have a significant impact on load power. The formula is as follows: (16) (17) (18) (19) (20) (twenty one) Where: is the query matrix, is the key matrix, is the value matrix, and are the dimensions of keys and values, is the dimension of the input, , , , , is the activation function, is the connection function, is the attention distribution value, is the multi-head attention distribution value, is the i-th attention distribution value; The feedforward neural network can reduce the model's sensitivity to overfitting of long sequences. The normalization layer standardizes the data to reduce the impact of differences between variables on model performance. The corresponding calculation formula is as follows: (twenty two) (twenty three) (twenty four) Where: is the normalized internal and external feature input, For normalization, is the linear processing of the feedforward neural network, is the output matrix, is the model weight, is the model bias.

[0073] (3) Projection layer The projection layer contains a multi-layer perceptron that is responsible for nonlinearly transforming the multivariate tokens processed independently by the encoder to generate the output.

[0074] This embodiment first obtains the prediction results through the general prediction model, determines the cluster category, key meteorological factors and corresponding load prediction model to which the prediction results belong, and then inputs the historical data of the load to be measured and the key meteorological factors into the corresponding load prediction model to obtain the final load prediction results. Not only does it refine the load categories, but it also considers the key meteorological factors corresponding to each load category, greatly improving the accuracy of the load prediction results.

[0075] Embodiment 2 In one or more embodiments, a load forecasting system for proxy power purchasers based on load scenario clustering is disclosed, comprising: A data acquisition module is configured to acquire load historical data and corresponding historical meteorological data of the proxy electricity purchasing user; A key feature extraction module is configured to extract features from the load history data to obtain the load intrinsic key features; The key meteorological factor extraction module is configured to cluster the intrinsic key characteristics of the load historical data and obtain the key meteorological factors of each category of load according to the correlation coefficient; The initial load prediction module is configured to use a general prediction model to perform load prediction on the historical data of the load to be measured to obtain an initial load prediction result; and extract corresponding load intrinsic key features based on the initial load prediction result; The final load forecasting module is configured to determine the clustering category to which the corresponding load intrinsic key characteristics belong and the key meteorological factors corresponding to the category, input the historical data of the load to be measured and the corresponding key meteorological factors into the load forecasting model of the corresponding category, and obtain the final load forecasting result.

[0076] The specific implementation of the above modules is the same as that in Example 1 and will not be described in detail.

[0077] Embodiment 3 In one or more embodiments, a terminal device is disclosed, which includes a processor and a memory, the processor is used to implement instructions; the memory is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executing the load forecasting method for proxy electricity purchasing users based on load scenario clustering described in Example 1.

[0078] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0079] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0080] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software.

[0081] Embodiment 4 In one or more embodiments, a computer-readable storage medium is disclosed, in which a plurality of instructions are stored, wherein the instructions are suitable for being loaded by a processor of a terminal device and executing the load forecasting method for proxy electricity purchasing users based on load scenario clustering described in Example 1.

[0082] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A load forecasting method for proxy power purchasing users based on load scenario clustering, characterized in that: include: Obtain the load historical data and corresponding historical meteorological data of the proxy power purchasing users; Extracting features from the load history data to obtain the inherent key features of the load; Cluster the inherent key characteristics of load historical data and calculate the key meteorological factors of each cluster category load; Use the general prediction model to perform load prediction on the historical data of the load to be measured, and obtain the initial load prediction result; Extract the corresponding load intrinsic key features based on the initial load forecast results; The cluster category to which the corresponding load intrinsic key feature belongs and the key meteorological factor corresponding to the category are determined, and the historical data of the load to be measured and the corresponding key meteorological factor are input into the load prediction model of the corresponding category to obtain the final load prediction result.

2. The method for predicting load of proxy power purchasing users based on load scenario clustering according to claim 1, characterized in that: The key intrinsic characteristics of the load include: load peak value, load valley value, load peak-to-valley difference at noon, load daily power consumption and load ramp rate ratio.

3. The method for predicting load of proxy power purchasers based on load scenario clustering according to claim 1, characterized in that: The key intrinsic characteristics of load historical data are clustered, specifically: Determine the number of clusters K, and organize the key characteristics of the load of each historical day into cluster data; Randomly select K points in the data as centroids to achieve preliminary clustering; Calculate the distance between other points and the centroid, and aggregate them into the closest category to obtain a cluster; take each point in the cluster as the centroid, calculate the sum of the total distances between the centroid and other points, and select the point with the smallest total distance sum as the new centroid; Repeat this process until the center of mass no longer shifts.

4. The method for predicting load of proxy power purchasers based on load scenario clustering according to claim 3, characterized in that: Determine the number of clusters K for clustering, specifically: According to experience, different cluster numbers K are selected for clustering, and the cluster center corresponding to each K value is determined; Calculate the average distance between each sample point and other samples in the same cluster under each cluster number K , and the average distance between each sample and the samples in the nearest neighbor cluster ; Then calculate the comprehensive evaluation coefficient P: ； ； in, is the evaluation coefficient of each sample, and N is the number of samples; The number of clusters K corresponding to the clustering result with the comprehensive evaluation coefficient P closest to 1 is selected as the optimal number of clusters.

5. The method for predicting load of proxy power purchasers based on load scenario clustering according to claim 1, characterized in that: The key meteorological factors for calculating the load of each cluster category are: Calculate the Gaussian similarity coefficient matrix between each group of historical load series and each corresponding type of meteorological series; Select the maximum value of each column in the Gaussian similarity coefficient matrix as the reference sequence; Calculate the absolute difference between the Gaussian similarity coefficient matrix and the corresponding elements of the reference sequence; calculate the grey correlation degree based on the maximum and minimum values ​​of the absolute difference to form a grey correlation degree matrix; Calculate the sum of each column of data in the grey correlation matrix as the sum of correlation, weight each row of the grey correlation matrix according to the sum of correlation, and then normalize the sum to obtain the weighted meteorological characteristic weight coefficient; The meteorological characteristic weight coefficients are sorted from large to small to obtain the key meteorological factors corresponding to each cluster category.

6. The method for predicting load of proxy power purchasers based on load scenario clustering according to claim 1, characterized in that: The general prediction model adopts the iTransformer model, and uses all historical load data and corresponding historical meteorological data to directly train the model.

7. The method for load forecasting of proxy power purchasing users based on load scenario clustering according to claim 1, characterized in that: After clustering the intrinsic key features of the load history data, a load forecasting model is trained using the load history data and the corresponding key meteorological factors in each category to obtain the same number of load forecasting models as the clustering categories. Based on the clustering category to which the load intrinsic key characteristics of the initial load forecast result belong, the load forecasting model under the corresponding category is selected, and the historical data of the load to be measured and the corresponding key meteorological factors are used as input to obtain the final load forecasting result.

8. A load forecasting system for proxy power purchasers based on load scenario clustering, characterized in that: include: A data acquisition module is configured to acquire load historical data and corresponding historical meteorological data of the proxy electricity purchasing user; A key feature extraction module is configured to extract features from the load history data to obtain the load intrinsic key features; A key meteorological factor extraction module is configured to cluster the load intrinsic key features of the load historical data and calculate the key meteorological factors of each cluster category load; The initial load prediction module is configured to use a general prediction model to perform load prediction on the historical data of the load to be measured to obtain an initial load prediction result; and extract corresponding load intrinsic key features based on the initial load prediction result; The final load forecasting module is configured to determine the clustering category to which the corresponding load intrinsic key characteristics belong and the key meteorological factors corresponding to the category, input the historical data of the load to be measured and the corresponding key meteorological factors into the load forecasting model of the corresponding category, and obtain the final load forecasting result.

9. A terminal device, comprising a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the load forecasting method for proxy electricity purchasing users based on load scenario clustering as described in any one of claims 1-7.

10. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the load forecasting method for proxy electricity purchasing users based on load scenario clustering as described in any one of claims 1-7.

Images