A spatial load forecasting method based on DBSCAN and improved ConvLSTM

Abstract

The application provides a spatial load prediction method based on DBSCAN and improved ConvLSTM, relates to the power distribution network technical field in a power system, solves the problem of insufficient mining of potential interaction relationships between different spatial distribution cells, and adopts the following scheme: according to collected information in a planning area, a power geographic information system is established, a type I cell layer and a type II cell layer are obtained, an abnormal value of type I cell load data is obtained, and after the abnormal value is removed, a reasonable maximum value of each type I cell load in a specified time period is formed; according to the reasonable maximum value of the type I cell load, spatial load historical data based on the type II cell is generated; the spatial load data characteristics of the type II cell are strengthened, a spatial load prediction model is constructed, and prediction data is obtained. The application can effectively identify and remove abnormal values, and constructs a SLF model facing the type II cell and synchronously considering the space-time correlation of the cell, so that the precision and generalization ability of the prediction model are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of power distribution network technology in power systems, and in particular to a spatial load forecasting method based on DBSCAN and an improved ConvLSTM. Background Technology

[0002] Spatial load forecasting (SLF) is the prediction of the magnitude and spatial distribution of future power load. A sound SLF is the basis and guarantee for accurate power system planning and the foundation for the economical operation of the power system. Determining the future power load magnitude and spatial distribution of the area to be planned allows for more accurate planning and design of substation locations and transmission line corridors, thereby establishing a rational and reliable power grid structure.

[0003] Current spatial load forecasting methods mainly rely on Type I cells to predict spatial power load. Since Type I cells are generated based on the power supply range of power supply equipment or facilities, the shape and size of Type I cells will change and their load patterns will also change as new equipment or facilities are put into operation and loads are transferred from the connection line, thus bringing serious adverse effects to SLF.

[0004] Furthermore, most relevant literature focuses on the spatial relationships between neighboring cells, lacking a concise joint prediction model that takes into account both the temporal and spatial characteristics of cell load. The exploration of potential interaction relationships between cells with different spatial distributions is insufficient and needs further resolution. Summary of the Invention

[0005] The purpose of this invention is to provide a solution to the problems that the shape and size of Type I cells change with the development and construction of the power grid, which has an adverse effect on spatial load forecasting, and that existing related forecasting models are not sufficient in exploring the potential interaction relationships between Type II cells with different spatial distributions. The invention aims to establish a scientific, reasonable, highly applicable, and accurate spatial load forecasting method.

[0006] This invention provides a spatial load forecasting method based on DBSCAN and an improved ConvLSTM, comprising:

[0007] Based on the information collected within the planning area, a power geographic information system is established to obtain Type I cell layers and Type II cell layers;

[0008] An outlier detection method is constructed using the DBSCAN algorithm to obtain outliers in the Class I cell load data and remove the reasonable maximum values ​​of each Class I cell load formed by the outliers within a specified time period.

[0009] Based on the reasonable maximum value of the type I cell load, generate spatial load data based on the type II cells;

[0010] The ConvLSTM algorithm is improved by using linear rectifier function and hyperbolic tangent function to enhance the spatial load data characteristics of the type II cells and construct a spatial load prediction model.

[0011] Based on the aforementioned space load prediction model, prediction data is obtained.

[0012] Optionally, as one embodiment, the step of establishing a power geographic information system based on the information collected within the planning area to obtain a Class I cell layer divided by the feeder power consumption range and a Class II cell layer determined by latitude, longitude, and spatial resolution specifically includes:

[0013] Collect and integrate grid structure, land use information, 10kV feeder power supply range and corresponding historical load data within the planning area;

[0014] In the power geographic information system environment, a registration base map of the planning area is constructed;

[0015] In the registered base map, the land use types of the area to be predicted are divided into eight categories according to the regional planning scheme: residential area, commercial area, cultural and entertainment area, industrial area, administrative office area, green space, municipal facilities area and special land, and a land use information layer is constructed.

[0016] Based on the power supply range of the 10kV feeder and the meshing technology, the Type I cell layer and the Type II cell layer are established sequentially.

[0017] Optionally, as one embodiment, the outlier detection method constructed using the DBSCAN algorithm to obtain outliers in the Type I cell load data specifically includes:

[0018] Step 1: Refer to the KNN algorithm to select a suitable neighborhood radius, that is, the neighborhood range of the load value of any class I cell at any time point, denoted as ε; determine the critical number of data points that can define the core load of class I cells within the neighborhood radius through clustering comparison experiments, denoted as M;

[0019] Step 2: Randomly select an unlabeled Class I cell load data, denoted as x, and check whether its load value's ε neighborhood contains a sufficient number of load data. If the number of load data is greater than or equal to M, then x is determined to be core load data, denoted as P1; otherwise, x is determined to be non-core load data, denoted as P2.

[0020] Step 3: Add the load data of each load value in the neighborhood of P1 to the normal data cluster;

[0021] Step 4: If P2 is not within the ε neighborhood of P1's load value, add P2 to the abnormal data cluster; otherwise, add it to the normal data cluster.

[0022] Step 5: Repeat steps 2 through 4 until each of the Class I cell load data has been labeled;

[0023] Step 6: After the above steps, complete normal data clusters and abnormal data clusters are obtained, and the data in the abnormal data clusters are all outliers;

[0024] Optionally, as one embodiment, the reasonable maximum value of each type I cell load formed after removing the outliers within a specified time period specifically includes:

[0025] The outlier values ​​of all the type I cell load data in the outlier data cluster are processed, and the maximum load value after removing outliers within a specified time period is the reasonable maximum value of type I cell load within the specified time period.

[0026] Optionally, as one embodiment, generating spatial load data based on the type II cells according to the reasonable maximum value of the type I cell load specifically includes:

[0027] In the power GIS environment, spatial power load gridding technology is used to transform the Type I cell load data into Type II cell load data;

[0028] Using the location information of each type II cell as the horizontal and vertical coordinates and the cell load data as the cell value, a spatial load spatiotemporal information map characterized by the type II cells is constructed using an Access database.

[0029] Optionally, as one embodiment, the improvement of the ConvLSTM algorithm using a linear rectified function and a hyperbolic tangent function specifically includes:

[0030] The ConvLSTM algorithm introduces convolution operations to replace matrix multiplication in the LSTM architecture, and the steps are as follows:

[0031] Input Gate (i t ), using formula (1):

[0032]

[0033] in: * represents the Hadamard product operation; * represents the convolution operation; σ represents the activation function; i t X is the state retention rate of the input gate; t H is the input tensor at time t; t-1 Let C be the hidden state tensor at time t-1;t-1 W is the storage unit tensor at time t-1; Xi W Hi W ci Let b be the weight matrix of the input gate. i For bias;

[0034] Forget Gate (f t Formula (2) is used:

[0035]

[0036] In the formula: f t The probability of state retention at the forget gate; W Ci W Hf W Xf These are the weight matrices for the input gate and the forget gate, respectively. f For bias;

[0037] Output gate (o) t Formula (3) is used:

[0038]

[0039] In the formula: o t W represents the output probability of the output gate at time t. Xo W Ho W co Let b be the weight matrix of the output gate. o For bias;

[0040] Cell State Update (C t Formula (4) is used:

[0041]

[0042] In the formula: C t Let be the storage unit tensor at time t; tanh is the hyperbolic tangent function; W Xc W Hc Let b be the weight matrix of the state gate. c For bias;

[0043] In the output hidden state H t Previously, ReLU was used to improve the hidden state of the original algorithm, and the hidden state update (H) was improved. t Formula (5) is used.

[0044]

[0045] Where: Ht Let be the hidden state tensor at time t.

[0046] The technical solution of this invention has at least the following beneficial technical effects: Taking power system planning as the research background, this invention addresses the problem of mismatch between Type I cells and actual power supply range, as well as the problem of insufficient consideration of spatiotemporal information of cells in related prediction models. It proposes a SLF method based on DBSCAN and an improved ConvLSTM, which has the following advantages: 1) This invention can effectively identify and eliminate outliers, thereby accurately and efficiently determining the reasonable maximum value of the load of each Type I cell and reducing prediction errors caused by outliers; 2) This invention uses Type II cells whose cell shape is determined by latitude and longitude to characterize spatial power load and perform SLF, avoiding the problem of mismatch between the shape and size of cells and the actual power grid construction level, and is more in line with the actual requirements of power system planning; 3) This invention constructs an SLF model that simultaneously considers the spatiotemporal correlation of cells, jointly models the spatiotemporal characteristics of cells, uses convolution to implement a gating mechanism, and uses activation functions to enhance the power load data characteristics of the prediction data, thereby effectively improving the accuracy and generalization ability of the prediction model, and improving the efficiency of urban power grid planning. The solution of this invention accurately and efficiently determines the reasonable maximum value of the load of each type I cell by effectively identifying and eliminating outliers. At the same time, it constructs an SLF model that considers the spatiotemporal correlation of cells for type II cells, avoiding the problem of mismatch between the shape and size of cells and the actual power grid construction level, and effectively improving the accuracy and generalization ability of the prediction model. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the conventional art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a basic principle diagram of the prediction method provided in the embodiments of the present invention;

[0049] Figure 2 This invention provides a KNN distance curve;

[0050] Figure 3 This invention provides embodiments of the impact of different minimum data sizes on clustering categories;

[0051] Figure 4 This invention provides DBSCAN clustering results;

[0052] Figure 5 This invention provides a prediction model structure diagram;

[0053] Figure 6 This invention provides a ReLU-ConvLSTM unit module structure;

[0054] Figure 7 This invention provides quasi-measured values ​​for the target year and predicted values ​​for each method.

[0055] Figure 8 This invention provides an error distribution for prediction results. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] This paper presents a spatial load prediction method using a density-based spatial clustering algorithm (DBSCAN) and an improved convolutional long short-term memory network (ConvLSTM). The method employs spatial power load characterized by type II cells for prediction, effectively overcoming the limitation of inconsistent shapes in type I cells when characterizing spatial loads, thus significantly reducing prediction errors. A spatial load prediction model based on type II cells is constructed using ConvLSTM, simultaneously considering the temporal and spatial correlations of cell loads. This effectively improves the accuracy and generalization ability of the prediction model, providing stable and reliable data support for power system planning.

[0058] This invention provides a spatial load forecasting method based on DBSCAN and an improved ConvLSTM, which includes the following:

[0059] 1) Establish a power geographic information system

[0060] ① Collect and integrate information on the grid structure, land use, 10kV feeder power supply range, and corresponding historical load data within the planning area;

[0061] ② In the power geographic information system (GIS) environment, construct a registration base map of the planning area; in the registration base map, according to the regional planning scheme, the land use type of the area to be predicted is divided into eight categories: residential area, commercial area, cultural and entertainment area, industrial area, administrative office area, green space, municipal facilities area and special land, and a land use information layer is constructed. Then, according to the power supply range of 10kV feeder and grid technology, Class I and Class II cell layers are established in sequence.

[0062] 2) Constructing an outlier detection method using the DBSCAN algorithm

[0063] Step 1: Parameter settings: Refer to the KNN algorithm to select a suitable neighborhood radius, that is, the neighborhood range of the load value of any class I cell at any time point, denoted as ε; determine the number of critical data that can define the core load of class I cells within the neighborhood radius through clustering comparison experiments, denoted as M;

[0064] Step 2: Classification of Class I Cell Load Data: Randomly select an unlabeled Class I cell load data (denoted as x) and check whether its load value's ε neighborhood contains a sufficient number of load data. If the number of load data is greater than or equal to M, then x is determined to be core load data, denoted as P1; otherwise, x is determined to be non-core load data, denoted as P2.

[0065] Step 3: Form a normal data cluster: Add the load data of each load value in the neighborhood of P1 to the normal data cluster;

[0066] Step 4: Forming an anomalous data cluster: If P2 is not within the ε-neighborhood of P1's load value, add P2 to the anomalous data cluster; otherwise, add it to the normal data cluster.

[0067] Step 5: Repeat steps (2) to (4) until each Class I cell load data has been labeled;

[0068] Step 6: Determination of Class I cell load outliers: After the above steps, complete normal data clusters and outlier data clusters are obtained. All data in the outlier data clusters are outliers.

[0069] 3) Generate a spatial load spatiotemporal information map based on type II cells.

[0070] ① Process outliers in all Type I cell load data in the outlier data cluster. The maximum load value after removing outliers within a specified time period is the reasonable maximum value of Type I cell load within that time period.

[0071] ② In the power GIS environment, spatial power load gridding technology is applied to transform Type I cell load data into Type II cell load data;

[0072] ③Use the position information of each type II cell as the horizontal and vertical coordinates and the values ​​of the cell load data cells to construct a spatial load spatiotemporal information map characterized by type II cells through an Access database;

[0073] 4) Improved ConvLSTM algorithm

[0074] The ConvLSTM algorithm introduces convolution operations to replace matrix multiplication in the LSTM architecture. The calculation steps are as follows:

[0075] ① Input Gate (i t The principle is shown in formula (1):

[0076]

[0077] In the formula: * represents the Hadamard product operation; * represents the convolution operation; σ represents the activation function; i t X is the state retention rate of the input gate; t H is the input tensor at time t; t-1 Let C be the hidden state tensor at time t-1; t-1 W is the storage unit tensor at time t-1; Xi W Hi W ci Let b be the weight matrix of the input gate. i For bias;

[0078] ②Forget Gate (f t The principle of ) is shown in formula (2):

[0079]

[0080] In the formula: f t The probability of state retention at the forget gate; W Ci W Hf W Xf These are the weight matrices for the input gate and the forget gate, respectively. f For bias;

[0081] ③ Output Gate (o) t The principle is shown in formula (3):

[0082]

[0083] In the formula: o t W represents the output probability of the output gate at time t. Xo W Ho W co Let b be the weight matrix of the output gate. oFor bias;

[0084] ④ Cell State Update (C t The principle is shown in formula (4):

[0085]

[0086] In the formula: C t Let be the storage unit tensor at time t; tanh is the hyperbolic tangent function; W Xc W Hc Let b be the weight matrix of the state gate. c For bias;

[0087] ⑤ In the output hidden state H t Previously, ReLU was used to improve the hidden state of the original algorithm, and the hidden state update (H) was improved. t The principle is explained in formula (5).

[0088]

[0089] Where: H t Let be the hidden state tensor at time t;

[0090] 5) Construct a spatial load forecasting model

[0091] First, the DBSCAN algorithm is used to remove abnormal data of type I cell loads. Then, type II cell load data is generated through gridding technology. Next, convolution operation is used to extract the spatial features of type II cell loads. Then, the ConvLSTM is improved by using the linear rectification function (ReLU) to form ReLU-ConvLSTM cell modules to enhance the load data features. By combining the historical state of multiple cell modules with the current input, the type II cell load value is determined. Finally, tanh is used to output the monthly prediction results of the target year and the annual maximum value of each type II cell load is taken to form the prediction result of the target year. This achieves the joint prediction of the magnitude and spatial distribution of future spatial power load.

[0092] The following uses Figures 1-8 The invention will be further illustrated by examples;

[0093] This invention is a spatial load forecasting method based on DBSCAN and an improved ConvLSTM. The basic principle diagram is shown below. Figure 1 As shown, the examples of the method include: basic data integration, establishing a power GIS, generating Class II cellular load data, constructing a spatial load prediction model, and comparative evaluation;

[0094] Based on the spatial load prediction method proposed in this invention, which uses DBSCAN and an improved ConvLSTM, a portion of the administrative region of a city in Northeast my country is used as an example. Historical load data from 2009 to 2013 are used to predict the Class II cell load value for this region in 2014. To verify the accuracy of the proposed method, the prediction results are compared and analyzed with those of traditional grey theory, linear regression, exponential smoothing, LSTM, and CNN methods. The specific details are as follows:

[0095] 1) Establish a power GIS

[0096] A portion of the administrative districts of a city in Northeast my country was selected as the area to be predicted. Historical load data from 2009 to 2013, with a sampling interval of 5 minutes, were used to predict the spatial power load in 2014.

[0097] First, a GIS for the area under study was established. Based on the regional planning scheme, the land use types of the area to be predicted were divided into eight categories: residential area, commercial area, cultural and entertainment area, industrial area, administrative office area, green space, municipal facilities area, and special-use land. Then, Class I and Class II cell layers were established according to the power supply range of the 10kV feeder and gridding technology. Finally, based on this, spatial load prediction was performed on the spatial power load data of the area in 2014.

[0098] 2) Data preprocessing

[0099] This invention uses the monthly maximum value of historical cell loads for SLF (Statistical Value Filtration). First, each type I cell load is clustered monthly. After detecting the clustering results, outlier clusters are removed. Finally, the corrected monthly maximum value of each type I cell load is calculated. The following uses the 2013 Changchun Railway Line as an example to determine the reasonable values ​​of ε and M in the outlier detection method.

[0100] When determining the value of ε, duplicate data is removed from the original data before plotting as follows. Figure 2 The KNN distance curve is shown, and the curve elbow point is adjusted with reference to the actual data. Finally, a reasonable value of ε is selected as 0.015.

[0101] Given a determined neighborhood radius, the value of M is determined empirically. When M is set to 1, each data point alone can become a core data point, making effective outlier detection impossible. When M is set to 2, an additional data point and itself can form a cluster, causing the algorithm to mistakenly identify most load data as core data, leading to decreased accuracy. Therefore, M should be greater than or equal to 3. Furthermore, as M increases, if the number of clusters also increases, it indicates that load data points that were previously partially connected between clusters are considered outliers; if the number of clusters decreases, it indicates that the number of clusters is shrinking, with some smaller clusters being marked as outliers because many smaller clusters are now considered entirely outliers. Given the relatively uniform distribution of power load data, M is generally set to less than 15. Experimental results for different clustering categories with varying M values ​​are shown below. Figure 3 As shown.

[0102] While the minimum values ​​in the line graph divide the load data into small clusters, they significantly improve the accuracy of outlier detection. The maximum values ​​in the line graph typically lead to more outliers and even misclassify normal values ​​as outliers. Since the goal of this clustering study is to select the reasonable maximum load value for 12 months from a large amount of load data, a high degree of accuracy in outlier detection is required. Therefore, M is set to the minimum value of 5 in the experiment. The final outlier detection results are as follows: Figure 4 As shown.

[0103] Then, in GIS, the monthly reasonable maximum load time series of each type I cell load from 2009 to 2013 was divided into 378 type II cells according to a square grid of equal size with a side length of 0.3km, so as to obtain the monthly quasi-measured values ​​of type II cell load from 2009 to 2013.

[0104] 3) Construct a spatial load forecasting model

[0105] The spatial load prediction method of this invention uses a ReLU-ConvLSTM modular structure. The specific principles and structure of the model are as follows: Figures 5-6 As shown in Table 1, to determine the specific structure of the prediction model, parameters were manually tuned according to the principle of prioritizing important parameters. The optimization began with the learning rate, the parameter with the greatest impact on model performance and training, followed by the model capacity, and finally, model details were adjusted. All parameters were adjusted from large to small. Specific experimental data are shown in Table 1. It should be noted that when determining the hidden layer dimensions, the three most widely applicable layers were used as a starting point. The parameters that best demonstrated the model's performance at this stage were then determined. Using these parameters as an example, comparative experiments were conducted for the second and fourth layers to determine the specific hidden layer dimensions and sizes.

[0106] Table 1 Comparison of specific parameters of the prediction model

[0107]

[0108]

[0109] The final prediction network consists of three layers of ConvLSTM units, with hidden state units of 64, 64, and 128 dimensions in each layer. The prediction network outputs the final result through a 1×1 2D convolution. The specific experimental parameters for training are shown in Table 2.

[0110] Table 2 Specific parameter settings for the prediction model

[0111]

[0112]

[0113] 4) Error Evaluation and Analysis

[0114] The maximum value of the load for each type II cell predicted in the 12 steps was taken to form the SLF result for type II cells in the target year. To demonstrate the accuracy of the SLF method of this invention, the results of methods such as grey theory, linear regression, exponential smoothing, LSTM, and convolutional neural networks (CNN) were compared and analyzed. The mean absolute error and mean relative error were used to evaluate the SLF results of each method, as shown in Table 3.

[0115] Table 3 Overall Prediction Error of Type II Cell Load in Target Year

[0116]

[0117] Table 3 shows that algorithms such as LSTM and CNN have smaller prediction errors, demonstrating the superiority of deep learning algorithms in prediction. The LSTM method only considers the temporal characteristics of the load in the SLF, while the CNN method simultaneously considers the spatiotemporal characteristics of the load, resulting in a significantly improved prediction accuracy compared to LSTM. This highlights the importance of the spatial distribution characteristics of the load for SLF prediction accuracy. The method in this invention combines the superiority of LSTM in considering temporal characteristics with the ability of CNN to consider spatial distribution characteristics, forming a joint modeling of the spatiotemporal characteristics of the SLF. By using only one model to simultaneously consider the spatiotemporal characteristics of the load, higher prediction accuracy is achieved. Finally, the prediction results of all methods are visualized to generate a spatial load map for comparison, as shown below. Figure 7 As shown.

[0118] Error distribution histograms for various methods are shown below. Figure 8As shown in the figure, the prediction errors of deep learning-based SLF methods are generally smaller. The smaller prediction errors of CNN and the method of this invention demonstrate that SLF simultaneously considering the spatiotemporal correlation between type II cells is not only theoretically sound but also produces smaller errors in practical predictions, proving the importance of simultaneously considering the spatiotemporal correlation of cells for SLF methods. Compared to the other five SLF methods, the SLF method of this invention, due to its joint modeling of the spatiotemporal characteristics of cells, is more consistent and scientific in the specific prediction process, conforming to the actual cell load transformation law, and therefore performs best in both MSE and MAE indicators.

[0119] The specific embodiments used in this invention have provided a detailed description of the invention, but are not limited to these embodiments. Any obvious modifications made by those skilled in the art based on the teachings of this invention are within the scope of protection of this invention.

[0120] The specific embodiments used in this invention have provided a detailed description of the invention, but are not limited to these embodiments. Any obvious modifications made by those skilled in the art based on the teachings of this invention are within the scope of protection of this invention.

[0121] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0122] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A spatial load forecasting method based on DBSCAN and improved ConvLSTM, characterized in that, include: Based on the information collected within the planning area, a power geographic information system is established to obtain a Class I cell layer divided by the power supply range of the feeder and a Class II cell layer determined by latitude, longitude and spatial resolution. The Class I cells are power supply zones divided in the power geographic information system environment according to the power supply range of power supply equipment and power supply units, and the Class II cells are power supply zones divided in the power geographic information system environment by grids of equal size. An outlier detection method is constructed using the DBSCAN algorithm to obtain outliers in the load data of the Class I cells. After removing the outliers, the reasonable maximum load of each Class I cell within a specified time period is obtained. Based on the reasonable maximum value of the type I cell load, generate historical spatial load data based on the type II cells; The ConvLSTM algorithm is improved by using linear rectifier function and hyperbolic tangent function to enhance the spatial load data characteristics of the type II cells and construct a spatial load prediction model. Based on the aforementioned space load prediction model, prediction data is obtained; The method for detecting outliers using the DBSCAN algorithm to obtain outliers in the Class I cell load data specifically includes: Step 1: Refer to the KNN algorithm to select a suitable neighborhood radius, that is, the neighborhood range of the load value of any class I cell at any time point, denoted as ε; determine the critical number of data points that can define the core load of class I cells within the neighborhood radius through clustering comparison experiments, denoted as M; Step 2: Randomly select an unlabeled Class I cell load data, denoted as x, and check whether its load value's ε neighborhood contains a sufficient number of load data. If the number of load data is greater than or equal to M, then x is determined to be core load data, denoted as P1; otherwise, x is determined to be non-core load data, denoted as P2. Step 3: Add the load data of each load value in the neighborhood of P1 to the normal data cluster; Step 4: If P2 is not within the ε neighborhood of P1's load value, add P2 to the abnormal data cluster; otherwise, add it to the normal data cluster. Step 5: Repeat steps 2 through 4 until each of the Class I cell load data has been labeled; Step 6: After the above steps, complete normal data clusters and abnormal data clusters are obtained, and the data in the abnormal data clusters are all outliers; The improvement of the ConvLSTM algorithm using linear rectified functions and hyperbolic tangent functions specifically includes: The ConvLSTM algorithm introduces convolution operations to replace matrix multiplication in the LSTM architecture, and the steps are as follows: The input gate uses formula (1): in: * represents the Hadamard product operation; * represents the convolution operation; σ represents the activation function; i t X is the state retention rate of the input gate; t H is the input tensor at time t; t-1 Let C be the hidden state tensor at time t-1; t-1 W is the storage unit tensor at time t-1; Xi W Hi W ci Let b be the weight matrix of the input gate. i For bias; The forgetting gate uses formula (2): In the formula: f t The probability of state retention at the forget gate; W Ci W Hf W Xf These are the weight matrices for the input gate and the forget gate, respectively. f For bias; The output gate uses formula (3): In the formula: o t W represents the output probability of the output gate at time t. Xo W Ho W co Let b be the weight matrix of the output gate. o For bias; The state update of the memory cell is performed using formula (4): In the formula: C t Let be the storage unit tensor at time t; tanh is the hyperbolic tangent function; W Xc W Hc Let b be the weight matrix of the state gate. c For bias; Before outputting the hidden state, ReLU is used to improve the hidden state of the original algorithm. The hidden state is updated using formula (5). Where: H t Let be the hidden state tensor at time t.

2. The method according to claim 1, characterized in that, The process involves establishing a power geographic information system based on information collected within the planning area, obtaining a Class I cell layer defined by the feeder power supply range and a Class II cell layer determined by latitude, longitude, and spatial resolution. Specifically, this includes: Collect and integrate grid structure, land use information, 10kV feeder power supply range and corresponding historical load data within the planning area; In the power geographic information system environment, a registration base map of the planning area is constructed; In the registered base map, the land use types of the area to be predicted are divided into eight categories according to the regional planning scheme: residential area, commercial area, cultural and entertainment area, industrial area, administrative office area, green space, municipal facilities area and special land, and a land use information layer is constructed. Based on the power supply range of the 10kV feeder and the meshing technology, the Type I cell layer and the Type II cell layer are established sequentially.

3. The method according to claim 1, characterized in that, The reasonable maximum value of each type I cell load formed after removing the outliers within a specified time period specifically includes: The outlier values ​​of all the type I cell load data in the outlier data cluster are processed, and the maximum load value after removing outliers within a specified time period is the reasonable maximum value of type I cell load within the specified time period.

4. The method according to claim 1, characterized in that, The step of generating historical spatial load data based on the type II cells according to the reasonable maximum value of the type I cell load specifically includes: In a power GIS environment, spatial power load gridding technology is used to transform the Type I cell load data into Type II cell load data; Using the location information of each type II cell as the horizontal and vertical coordinates and the cell load data as the cell value, a spatial load spatiotemporal information map characterized by the type II cells is constructed using an Access database.

Images