Method for predicting power load based on population flow

Abstract

The application discloses a power load prediction method based on population flow, relates to the technical field of data processing, and determines an OD flow matrix of a preset area and historical load data of the preset area; regards each target area in the preset area as a node, regards historical population flow data in the OD flow matrix as an edge, regards historical load data of each target area, historical time in the OD flow matrix, historical population data, regional characteristic data and meteorological data as attributes of the node, and constructs a space-time dynamic graph; processes the space-time dynamic graph through a population distribution prediction model to obtain predicted population data of the space-time dynamic graph; processes historical load data, historical time, meteorological data and the predicted population data of the space-time dynamic graph to obtain a predicted load distribution graph of the preset area. Through the technical scheme, the characteristic data of each target area in the time scale and the characteristic data of each target area in the space scale are matched, and the accuracy of the predicted load distribution graph can be improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a method for predicting electricity load based on population flow. Background Technology

[0002] A power system is a complete system mainly composed of power generation, transmission, transformation, distribution, and consumption. It is primarily used to convert energy into electrical energy and then transmit and distribute that energy to users. The efficient, safe, and economical operation of a power system highly depends on the accurate prediction of future power load changes. Power load forecasting, as a core component of power dispatching, market trading, and system security analysis, directly affects the stability of the power system.

[0003] In related technologies, load forecasting for power systems is typically achieved by analyzing historical load data or using models in deep learning algorithms. However, the accuracy of power load forecasting using these methods is relatively low. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting electricity load based on population flow, in order to solve the technical problems existing in related technologies.

[0005] To achieve the above objectives, the present invention provides a power load forecasting method based on population flow, comprising:

[0006] Determine the OD flow matrix of the preset area and the historical load data of each target area among multiple target areas in the preset area. The OD flow matrix includes the historical population data, historical population flow data, regional characteristic data, historical time of acquisition of historical population data and historical population flow data, and meteorological data of each target area within a preset time period.

[0007] Each target region in the preset area is regarded as a node, the historical population flow data in the OD flow matrix is ​​regarded as an edge, and the historical load data, historical time, historical population data, regional characteristic data and meteorological data of each target region, as well as the historical load data, historical time, historical population data, regional characteristic data and meteorological data in the OD flow matrix are regarded as the attribute data of each node, and a spatiotemporal dynamic map is constructed.

[0008] By processing historical population flow data, historical time, historical population data of each node, and regional characteristic data of the spatiotemporal dynamic map using a population distribution prediction model, the predicted population data of each node in the spatiotemporal dynamic map is obtained.

[0009] The historical load data, historical time, meteorological data, and predicted population data of each node in the spatiotemporal dynamic map are processed to obtain the predicted load distribution map of the preset area.

[0010] Optionally, the process of using a population distribution prediction model to process historical population flow data, historical time, historical population data of each node, and regional characteristic data of the spatiotemporal dynamic map to obtain predicted population data for each node in the spatiotemporal dynamic map includes:

[0011] By processing historical population flow data, historical time, historical population data of each node, and regional feature data of the spatiotemporal dynamic map through spatial convolutional layers, a spatial feature dynamic map is obtained.

[0012] By processing historical population flow data, historical time, historical population data of each node, and regional feature data of the spatial feature dynamic map through temporal convolutional layers, a spatiotemporal feature dynamic map is obtained.

[0013] By using an attention mechanism, historical population flow data, historical time, historical population data of each node, and regional feature data of the spatiotemporal dynamic graph are processed to obtain the predicted population data of each node in the spatiotemporal dynamic graph.

[0014] Optionally, the process of processing the historical load data, historical time, meteorological data, and predicted population data of each node in the spatiotemporal dynamic map to obtain a predicted load distribution map of the preset area includes:

[0015] The historical load data, historical time, meteorological data and predicted population data corresponding to each node in the spatiotemporal dynamic graph are processed by normalization method, and outlier data are removed to obtain normalized node feature data.

[0016] By processing the normalized node feature data through a sparse attention mechanism, a predicted load distribution map of the preset region is obtained.

[0017] Optionally, the prediction method further includes:

[0018] Determine the correction factor, which is used to characterize the deviation between the predicted population data and the corrected population data distribution within the preset area;

[0019] The predicted population data of each node in the spatiotemporal dynamic graph is multiplied by the correction factor to obtain the corrected population data of each node.

[0020] The process of processing historical load data, historical time, meteorological data, and predicted population data for each node in the spatiotemporal dynamic map to obtain a predicted load distribution map for a preset area includes:

[0021] By processing the historical load data, historical time, meteorological data, and corrected population data of each node in the spatiotemporal dynamic map using a sparse attention mechanism and a normalization method, a predicted load distribution map of the preset area is obtained.

[0022] Optionally, the correction factor includes at least one of the following: the probability that the communication company corresponding to the communication base station is located in a preset area; the probability that a person in the preset area owns multiple devices; and the probability that the proportion of elderly people and children in the preset area is high.

[0023] Optionally, the regional feature data includes the regional hierarchical classification code of the target region, the distance of the target region from the city center, and the type of historical time. The population flow data includes the resident population data of the target region, the population flowing into the target region, and the population flowing out of the target region.

[0024] Optionally, the OD flow matrix is ​​obtained by the following method:

[0025] Determine the historical population data, historical population flow data, regional characteristic data, and historical time period for acquiring the historical population data and historical population flow data for each of the multiple target regions;

[0026] Determine meteorological data for each of multiple target areas within a preset time period;

[0027] The historical time, historical population data, historical population flow data, regional characteristic data, and meteorological data corresponding to each target region are integrated to obtain the OD flow matrix.

[0028] The above technical solution constructs a spatiotemporal dynamic map within a preset area based on the OD flow matrix. A population distribution prediction model is then used to process this map and predict the population data for each target area. Subsequently, historical load data, historical time, predicted population, and meteorological data for each target area are processed to predict the load distribution map for the preset area. Processing the spatiotemporal dynamic map using the population distribution prediction model extracts temporal and spatial characteristic data, allowing for the generation of predicted population data. Finally, based on the historical load data, predicted population data, historical time, and meteorological data for each target area, the power load is predicted, resulting in the predicted load distribution map for the preset area. Matching the temporal and spatial characteristic data of each target area improves the speed and accuracy of power load prediction.

[0029] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0030] Figure 1 This is a schematic diagram illustrating a power load forecasting method based on population flow according to an exemplary embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram illustrating a population distribution prediction model according to an exemplary embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram illustrating a sparse attention mechanism according to an exemplary embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, so as to provide a better understanding of the concept of the present invention, the technical problem solved, the technical features constituting the technical solution, and the technical effects brought about.

[0034] In recent years, with the acceleration of urbanization and the rapid development of information and communication technologies, population mobility has become characterized by large scale, high frequency, and complex directions, leading to rapid local economic growth, increased energy consumption, and fluctuating electricity prices. The spatial distribution changes caused by large-scale commuting, schooling, shopping, and holiday travel have become a significant driving force behind drastic fluctuations in regional electricity load.

[0035] In related technologies, traditional load forecasting methods mainly model and predict corresponding power loads based on historical load data's time-series patterns, meteorological factors, and calendar effects. These methods have significant limitations in dealing with the nonlinear and stochastic load fluctuations caused by the large-scale grid connection of today's highly penetrated distributed energy resources, and they also insufficiently consider the dynamic social drivers causing load fluctuations. In recent years, with the rapid development of artificial intelligence, data-driven deep learning algorithms have been widely used in power system load forecasting. Deep learning algorithms can extract spatiotemporal patterns of load from massive amounts of data and are highly adaptable to handling nonlinear factors and complex power grids.

[0036] However, the inventors discovered that when predicting power load using deep learning algorithms, single models such as convolutional neural networks, long short-term memory neural networks, empirical mode decomposition, optimized evolutionary bidirectional long short-term memory neural networks, and graph convolutional neural networks are used. Although a single model can improve the prediction accuracy of power load, it has high requirements for the quality of the input data and cannot fuse data from multiple different scales, which leads to low accuracy of the predicted power load and a reduced prediction rate.

[0037] In view of this, the present invention provides a power load forecasting method based on population flow to solve the technical problems existing in the above-mentioned related technologies.

[0038] like Figure 1 As shown, Figure 1This is a schematic diagram illustrating a power load forecasting method based on population flow according to an exemplary embodiment of the present invention, with reference to... Figure 1 The method includes;

[0039] S101: Determine the OD flow matrix of the preset area and the historical load data of each target area among multiple target areas in the preset area. The OD flow matrix includes the historical population data, historical population flow data, regional characteristic data, historical time of acquisition of historical population data and historical population flow data, and meteorological data of each target area within a preset time period.

[0040] S102: Treat each target area in the preset area as a node, the historical population flow data in the OD flow matrix as an edge, and the historical load data, historical time, historical population data, regional characteristic data and meteorological data of each target area as the attribute data of each node, and construct a spatiotemporal dynamic map.

[0041] S103: The historical population flow data, historical time, historical population data of each node and regional characteristic data of the spatiotemporal dynamic map are processed by the population distribution prediction model to obtain the predicted population data of each node in the spatiotemporal dynamic map.

[0042] S104: Process the historical load data, historical time, meteorological data, and predicted population data of each node in the spatiotemporal dynamic map to obtain the predicted load distribution map of the preset area.

[0043] The above technical solution constructs a spatiotemporal dynamic map within a preset area based on the OD flow matrix (Origin-Destination). A population distribution prediction model is then used to process this dynamic map and predict the population data for each target area. Subsequently, historical load data, historical time, predicted population, and meteorological data for each target area are processed to predict the load distribution map for the preset area. Processing the spatiotemporal dynamic map using the population distribution prediction model extracts temporal and spatial characteristic data, allowing for the generation of predicted population data. Finally, based on the historical load data, predicted population data, historical time, and meteorological data for each target area, the power load is predicted, resulting in the predicted load distribution map for the preset area. Matching the temporal and spatial characteristic data of each target area improves the speed and accuracy of power load prediction.

[0044] To enable those skilled in the art to better understand the power load forecasting method based on population flow provided by this invention, the above steps are illustrated in detail below.

[0045] For example, the target region can be any region within the predicted region. The preset region can be any region, such as a provincial-level region or a municipal-level region. When the preset region is a provincial-level region, the target region can be any urban area or county-level region under the province. When the preset region is a municipal-level region, the target region can be a residential area, commercial area, tourist area, etc., and this embodiment of the invention does not specifically limit these areas.

[0046] Historical population data refers to the number of people who have spent a cumulative total of more than three hours in the target area within a historical period. Historical population flow data refers to the population data of historical inflows, outflows, or historical permanent residents in the target area. Historical time refers to the time when historical population data and historical population flow data were collected. In this embodiment of the invention, historical population flow data may include resident population data of the target area, population data flowing into the target area, and population data flowing out of the target area. Resident population data refers to the number of people who have been on the resident user list every month for the past six months. Resident users can be divided into working resident users and living resident users. Population data flowing into the target area refers to the number of people who are not resident users but have spent a cumulative total of more than three hours in the target area within a historical period. Population data flowing out of the target area refers to the number of people remaining after subtracting the total population from the sum of resident users and inflow users.

[0047] Regional feature data can be data used to characterize the characteristics of a target region. In this embodiment of the invention, regional feature data may include the regional hierarchical classification code of the target region, the distance of the target region from the city center, and the type of historical time. The type of historical time may include categories such as Monday to Thursday, Friday, weekend, statutory holidays, adjusted workdays, local special holidays, and large-scale event days.

[0048] The preset time can be a future time that is about to occur. The meteorological data can be meteorological data predicted within the preset time period. The meteorological data may include data such as temperature, humidity, wind speed, and precipitation, which are not specifically limited in this embodiment of the invention.

[0049] In one possible manner, the OD flow matrix is ​​obtained by the following method:

[0050] Determine the historical population data, historical population flow data, regional characteristic data, and historical time period for acquiring the historical population data and historical population flow data for each of the multiple target regions;

[0051] Determine meteorological data for each of multiple target areas within a preset time period;

[0052] The historical time, historical population data, historical population flow data, regional characteristic data, and meteorological data corresponding to each target region are integrated to obtain the OD flow matrix.

[0053] It should be understood that the OD flow matrix can be represented by the following matrix:

[0054] ;

[0055] The OD matrix contains 10 target regions, represented by letters a through j. Each row of the matrix represents various data for one target region. For example, for target region 'a', the total... a This refers to the total population of target region a, O ab ~O aj This refers to the number of people who migrate from target region a to target regions b through j. Per refers to the historical number of people who migrated from target region a to any other target region other than target regions b to j. a D refers to the historical resident population data of target region a, and A refers to the historical time period. a Refers to the regional hierarchical classification code of target region a, B a L refers to the distance of target region 'a' from the city center of the provincial capital, L refers to the time type of the historical period, and M refers to the distance of the target region 'a' from the city center of the provincial capital. a1 ~M a4 Refers to four meteorological data points for target region a within a preset time period, I ba ~I da This refers to the historical population data of population flowing from target regions b to d into target region a. The symbols for other rows follow the same pattern.

[0056] For example, after constructing the OD flow matrix, a spatiotemporal dynamic graph can be built based on the OD flow matrix. This spatiotemporal dynamic graph can be a graph structure used for modeling and analyzing spatiotemporal data. In this embodiment of the invention, target regions can be considered as nodes, historical population flow data in the OD flow matrix as edges, and historical load data, historical population data, historical time, regional characteristic data, and meteorological data of each target region as attribute data for each node, thus constructing a directed weighted spatiotemporal dynamic graph. In this spatiotemporal dynamic graph, each node includes static attribute data and dynamic attribute data. The static attribute data may include historical load data, the regional hierarchical classification code of the target region in the regional characteristic data, the distance of the target region from the city center, and the type of historical time. The dynamic attribute data may include historical population data and meteorological data within a preset time period. The edge feature data between two nodes mainly consists of historical population flow data.

[0057] For example, a population distribution prediction model can be used to predict the population corresponding to each target region. This population distribution prediction model may include spatial convolutional layers and temporal convolutional layers. When a spatiotemporal dynamic graph is input into the population distribution prediction model, the spatial and temporal characteristic data of the data corresponding to each node and edge features in the spatiotemporal dynamic graph can be extracted. These spatial and temporal characteristic data are then processed to obtain the predicted population data for each node in the spatiotemporal dynamic graph.

[0058] In one possible approach, the process of using a population distribution prediction model to process historical population flow data, historical time, historical population data for each node, and regional characteristic data of the spatiotemporal dynamic map to obtain predicted population data for each node in the spatiotemporal dynamic map includes:

[0059] By processing historical population flow data, historical time, historical population data of each node, and regional feature data of the spatiotemporal dynamic map through spatial convolutional layers, a spatial feature dynamic map is obtained.

[0060] By processing historical population flow data, historical time, historical population data of each node, and regional feature data of the spatial feature dynamic map through temporal convolutional layers, a spatiotemporal feature dynamic map is obtained.

[0061] By using an attention mechanism, historical population flow data, historical time, historical population data of each node, and regional feature data of the spatiotemporal dynamic graph are processed to obtain the predicted population data of each node in the spatiotemporal dynamic graph.

[0062] It should be understood that population distribution prediction models include temporal convolutional layers, spatial convolutional layers, and attention mechanisms. Specifically, for example... Figure 2 As shown, the spatiotemporal dynamic graph can be sequentially input into a spatial convolutional layer, a temporal convolutional layer, and an attention mechanism to obtain the predicted population data for each node in the spatiotemporal dynamic graph. This embodiment of the invention does not impose specific limitations on this.

[0063] In the actual processing, the optimizers in both spatial and temporal convolutional layers are Adam optimizers. Optimizers are one of the mathematical strategies that help the network adjust weights, biases, learning rates, and other parameters to reduce loss. The loss functions in both spatial and temporal convolutional layers include any one of the following: mean absolute error loss function, root mean square error loss function, normalized root mean square error loss function, mean absolute percentage error loss function, coefficient of variation loss function, kurtosis loss function, and Nash efficiency loss function.

[0064] In this embodiment of the invention, taking the root mean square error loss function as an example, it can be specifically expressed by the following calculation formula:

[0065] ;

[0066] Where MSE is the root mean square error, N can be the number of target regions, and z t For the t-th target region, the temporal and spatial convolutional layers can be used to predict population data, y t Historical population data can be used for the t-th target region.

[0067] Since different data corresponding to each node have different importance when calculating predicted population data, the MSE weights of historical population data, inflow population data, and outflow population data of the target area can be set to 0.5, 0.3, and 0.2, respectively. This can be expressed by the following formula:

[0068] ;

[0069] in, It can be a loss function.

[0070] Spatial convolutional layers extract spatial characteristic data of historical population flow, historical time, historical population data of each node, and regional feature data from the spatiotemporal dynamic map. Temporal convolutional layers extract temporal characteristic data of the same data, resulting in a spatiotemporal feature dynamic map. An attention mechanism is then applied to process this dynamic map to obtain predicted population data for each node. This allows for the calculation of predicted population data for each target region within a preset area, thereby improving the accuracy of predicted population data for each target region.

[0071] For example, sparse attention mechanisms can be used to optimize traditional attention mechanisms, primarily by improving model efficiency through reducing unnecessary computation and memory usage while preserving key information. A predicted load distribution map can predict the distribution of electricity load that may occur in each target area within a preset time period. After obtaining the predicted population data for each node in the spatiotemporal dynamic map, the historical load data, historical time, meteorological data, and predicted population data corresponding to each node can be processed using sparse attention mechanisms and normalization methods to obtain the predicted load distribution map.

[0072] In one possible manner, the processing of historical load data, historical time, meteorological data, and predicted population data for each node in the spatiotemporal dynamic map to obtain a predicted load distribution map for a preset area includes:

[0073] The historical load data, historical time, meteorological data and predicted population data corresponding to each node in the spatiotemporal dynamic graph are processed by normalization method, and outlier data are removed to obtain normalized node feature data.

[0074] By processing the normalized node feature data through a sparse attention mechanism, a predicted load distribution map of the preset region is obtained.

[0075] It should be understood that, in the specific data processing process of this invention, the historical load data, historical time, meteorological data, and predicted population data corresponding to each node are first processed using a normalization method, and outlier data are removed to obtain normalized node feature data, which can be expressed by the following calculation formula:

[0076] ;

[0077] Where Y is the data obtained after normalization, and x is one of the following: regional feature data, meteorological data, and predicted population data.

[0078] This invention employs a Min-Max normalization method to process historical load data, historical time data, meteorological data, and predicted population data. This transforms each value of the historical load data, historical time data, meteorological data, and predicted population data into an interval between 0 and 1. Since these data vary considerably, normalization improves the performance of the neural network during training and accelerates convergence. Before completing the prediction task, the predicted values ​​should be automatically converted back to their original dimensions to improve the readability of the prediction results. This process is called denormalization.

[0079] When dealing with outliers, data softening methods can be used to remove them, thereby reducing noise interference. In this embodiment of the invention, a moving average method can be used to smooth the data; however, this embodiment does not specifically limit the application of this method.

[0080] Then, the normalized node feature data can be processed using a sparse attention mechanism to predict the power load distribution map.

[0081] Specifically, sparse attention mechanisms can simultaneously compute data corresponding to specific locations in parallel, thereby improving training efficiency. They also enable data interaction between any two target regions. By limiting the attention range, sparse attention mechanisms reduce the number of attention weights that need to be computed, thus lowering the computational complexity of the model. This method can reduce computational complexity to O(NlogN) and also reduces the corresponding memory requirements, while still capturing important dependencies in the input sequence. In this embodiment of the invention, data can be processed using a dynamic sparsification method that adapts to complex dependencies. An adaptive Top-k selection mechanism can be selected, which can dynamically retain the most important attention scores according to each query, optimizing the feature aggregation process.

[0082] The specific calculation formula for the sparse attention mechanism is shown below:

[0083] ;

[0084] Q, K, and V are three vector matrices obtained by linear transformation of the input sequence, namely the query vector, key vector, and value vector, respectively. This refers to the adaptive Top-k selection operator, where T is the vector transpose. Let K be the dimension of the vector.

[0085] ;

[0086] Among them, t m yes The k-th maximum value in the m-th row of the vector. The scaled attention score is the difference between the m-th query vector and the n-th key vector.

[0087] Then, a multi-head attention mechanism can be used to learn in parallel the complex relationships between historical load data, historical time, meteorological data, and predicted population data. For each target region, any one of the historical load data, historical time data, meteorological data, or predicted population data in that region is used as a head, and sparse attention is calculated independently for each head. The outputs of all heads are then concatenated and fused using a linear transformation. The specific calculation formula is shown below:

[0088] ;

[0089] ;

[0090] Where pos represents the position of the element in the spatiotemporal dynamic graph, m represents the dimension index, and d... model Indicates the dimension of the model. This represents the position encoding function. Inject the order or position information of each element in the model.

[0091] Then, a feed-forward network (FFN) can be used to perform a nonlinear transformation on the data of each node. Further feature extraction is then performed using two linear layers and one ReLU activation layer, as shown in the following calculation formula, to obtain the predicted power load value for each node. After calculating the predicted power load values ​​for all nodes, a predicted load distribution map can be obtained. Here, W1, W2 and b1, b2 are the weight matrices and bias matrices of the two linear layers, respectively:

[0092] .

[0093] like Figure 3 As shown in this embodiment of the invention, the sparse attention mechanism has a 9-layer stack of encoder and decoder. The sparse attention mechanism is coupled to other modules through a fully connected configuration, and the decoder's input depends on both the input sequence and the encoder's output. The loss function employs dual-scale constraints of MSE and L2 regularization to synergistically improve prediction accuracy and stability.

[0094] In some possible ways, the prediction method further includes:

[0095] Determine the correction factor, which is used to characterize the deviation between the predicted population data and the corrected population data distribution within the preset area;

[0096] The predicted population data of each node in the spatiotemporal dynamic graph is multiplied by the correction factor to obtain the corrected population data of each node.

[0097] The process of processing historical load data, historical time, meteorological data, and predicted population data for each node in the spatiotemporal dynamic map to obtain a predicted load distribution map for a preset area includes:

[0098] By processing the historical load data, historical time, meteorological data, and corrected population data of each node in the spatiotemporal dynamic map using a sparse attention mechanism and a normalization method, a predicted load distribution map of the preset area is obtained.

[0099] It should be understood that since the historical population data and historical population flow data collected in the preset area are obtained by counting the number of people with communication devices, the calculated predicted population data may contain cases where some people have multiple communication devices or do not have communication devices. Therefore, a correction factor can be added to correct the predicted population data to obtain corrected population data.

[0100] Therefore, after obtaining the predicted population data, it can be multiplied by the correction factor to obtain the corrected population data, which can be expressed by the following formula:

[0101] ;

[0102] in, To predict population data, To correct the population data, All of these are correction factors. In this embodiment of the invention, the correction factors can be the probability of the communication company corresponding to the communication base station being located in a preset area, the probability of people holding multiple devices in the preset area, and the probability of the proportion of elderly people and children in the preset area. This embodiment of the invention does not impose specific limitations on these factors.

[0103] After obtaining the corrected population data, the historical load data, historical time, meteorological data, and corrected population data of each node in the spatiotemporal dynamic map can be processed using the sparse attention mechanism and normalization method to obtain the predicted load distribution map of the preset area.

[0104] By correcting the predicted population data with correction factors, the accuracy of the corrected population data for each target region can be improved.

[0105] In the implementation process, ten typical regions in a southwestern province were selected as ten target regions within the preset area, and the training and test sets were divided in a 7:3 ratio. The model validation environment consisted of an Intel(R) Core(TM) i5-9400 CPU, Windows 11, x64-bit operating system, and MATLAB 2021a as the programming software. To ensure the objectivity of the experiment, each model underwent five independent replicate experiments, and the average of the five experiments was used as the final prediction result for that model.

[0106] To ensure the comprehensiveness of the experimental evaluation, this embodiment of the invention evaluates the prediction results using seven indicators: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Normalized Root Mean Square Error (NRMSE), Mean Absolute Percentage Error (MAPE), Coefficient of Variation (CV), Kurtosis Coefficient (KC), and Nash-Sutcliffe Efficiency (NSE).

[0107] The formula for calculating the mean absolute error is as follows:

[0108] ;

[0109] The root mean square error is calculated using the following formula:

[0110] ;

[0111] The formula for calculating the normalized root mean square error is as follows:

[0112] ;

[0113] The formula for calculating the mean absolute percentage error is as follows:

[0114] ;

[0115] Where n represents the number of target regions, z m and y m These represent the predicted and actual power load values ​​for a preset area at a given time m, respectively. This represents the average of n samples. The same symbols in the following formulas have the same meaning.

[0116] ;

[0117] Where, x m This represents the electrical load at a given time m.

[0118] ;

[0119] Among them, e m This represents the error between the predicted value and the actual value of the power load in a preset area at a given time m. and These represent the average and standard deviation of the above errors, respectively.

[0120] ;

[0121] in, This method constructs an OD (Original Demand) flow matrix using historical total population data and historical population flow data to represent the predicted power load of a preset area. Regional characteristic data, such as regional hierarchical classification codes, and meteorological data within a preset time period are also incorporated. The OD flow matrix and historical load data are then converted into a spatiotemporal dynamic map. Weights in the loss function are adjusted according to importance to obtain the predicted population data for each target area. Based on an economic model, a triple correction method using basic penetration rate, age weight, and equipment ownership rate is proposed to finally obtain the corrected population data for each target area within the preset area. This method can accurately obtain multi-dimensional characteristics such as the temporal and spatial distribution of population flow and the magnitude of flow, thereby improving the accuracy of predicted population data for target areas.

[0122] Subsequently, data quality is improved through various regional feature data such as time scale division and holiday label classification, as well as targeted data preprocessing such as normalization and data softening. Based on the sparse attention mechanism, the data corresponding to each target region is processed to improve prediction efficiency and accuracy. Finally, the final load prediction result is output through a feedforward neural network.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for power load forecasting based on population flow, characterized by, The method comprises the following steps: determining an OD flow matrix of a preset area and historical load data of each target area in a plurality of target areas in the preset area, wherein the OD flow matrix comprises historical population data, historical population flow data, regional characteristic data, historical time for obtaining the historical population data and the historical population flow data, and meteorological data of each target area within a preset time length; regarding each target area in the preset area as a node, regarding the historical population flow data in the OD flow matrix as an edge, and regarding the historical load data of each target area, the historical time, the historical population data, the regional characteristic data and the meteorological data in the OD flow matrix as attribute data of each node, to construct a spatio-temporal dynamic graph; processing the historical population flow data, the historical time, the historical population data and the regional characteristic data of each node of the spatio-temporal dynamic graph by a population distribution prediction model to obtain predicted population data of each node in the spatio-temporal dynamic graph; processing the historical load data, the historical time, the meteorological data and the predicted population data of each node in the spatio-temporal dynamic graph to obtain a predicted load distribution map of the preset area; the processing of the historical population flow data, the historical time, the historical population data and the regional characteristic data of each node of the spatio-temporal dynamic graph by the population distribution prediction model to obtain the predicted population data of each node in the spatio-temporal dynamic graph comprises: processing the historical population flow data, the historical time, the historical population data and the regional characteristic data of each node of the spatio-temporal dynamic graph by a spatial convolution layer to obtain a spatial feature dynamic graph; processing the historical population flow data, the historical time, the historical population data and the regional characteristic data of each node of the spatial feature dynamic graph by a time convolution layer to obtain a spatio-temporal feature dynamic graph; processing the historical population flow data, the historical time, the historical population data and the regional characteristic data of each node of the spatio-temporal feature dynamic graph by an attention mechanism to obtain the predicted population data of each node in the spatio-temporal dynamic graph; the predicted load distribution map is used to represent a predicted distribution of possible power load of each target area in the preset area within a preset time length, and the regional characteristic data comprises a type of time at which the historical time is located, and the type of time at which the historical time is located comprises Monday to Thursday, Friday, weekend, statutory holidays, make-up holidays, local special holidays and large event days.

2. The population flow based power load forecasting method of claim 1, wherein, the processing of the historical load data, the historical time, the meteorological data and the predicted population data of each node in the spatio-temporal dynamic graph to obtain the predicted load distribution map of the preset area comprises: processing the historical load data, the historical time, the meteorological data and the predicted population data corresponding to each node in the spatio-temporal dynamic graph by a normalization method, and removing outlier data to obtain normalized node feature data; processing the normalized node feature data by a sparse attention mechanism to obtain the predicted load distribution map of the preset area.

3. The population flow based power load forecasting method according to claim 1 or 2, characterized in that, the prediction method further comprises: determining a correction factor, wherein the correction factor is used to represent a deviation between the predicted population data and the corrected population data in the preset area. The prediction population data of each node in the spatiotemporal dynamic graph is multiplied by the correction factor to obtain the correction population data of each node; The historical load data, historical time, meteorological data, and prediction population data of each node in the spatiotemporal dynamic graph are processed to obtain a prediction load distribution graph of the preset area, including: The historical load data, historical time, meteorological data, and correction population data of each node in the spatiotemporal dynamic graph are processed by a sparse attention mechanism and a normalization method to obtain a prediction load distribution graph of the preset area.

4. The population flow based power load forecasting method of claim 3, wherein, The correction factor includes at least one of a probability of a communication company corresponding to a communication base station in the preset area, a probability of a person holding multiple devices in the preset area, and a proportion probability of old people and children in the preset area. 5.The population-flow-based power load forecasting method according to claim 1, wherein, The area feature data includes an area level division code of a target area and a distance of the target area from a city center, and the population flow data includes resident population data of the target area, population data flowing into the target area, and population data flowing out of the target area. 6.The population-flow-based power load forecasting method according to claim 1, wherein, The OD flow matrix is obtained by the following method: determining historical population data, historical population flow data, area feature data of each target area in a plurality of target areas, and historical time of obtaining the historical population data and the historical population flow data; determining meteorological data of each target area in the plurality of target areas within a preset time length; fusing the historical time, the historical population data, the historical population flow data, the area feature data, and the meteorological data corresponding to each target area to obtain the OD flow matrix.

Images