Metering equipment warehouse division demand prediction and scheduling method based on spatio-temporal feature fusion

Abstract

A metering equipment warehouse division demand prediction and scheduling method based on spatio-temporal feature fusion comprises the steps of firstly obtaining and collecting warehouse division data, then establishing a spatio-temporal feature fusion model based on the warehouse division data, and then establishing a spatio-temporal diagram convolution prediction network, the spatio-temporal diagram convolution comprises a diagram convolution layer and a time convolution layer, through extraction and fusion of the two layered features, dynamic weight fusion is obtained, joint modeling of space-time dynamics is realized, and finally, prediction and allocation decision are implemented based on the dynamic weight fusion. According to the method, the multi-modal graph structure fusing the space-time association and the replacement rule is constructed, and a space-time joint modeling architecture and a dynamic feedback mechanism are designed, so that high-precision demand prediction and global inventory optimization are realized; the method systematically solves the core problems of insufficient spatial correlation modeling, dynamic event response lagging, low efficiency in multi-source data utilization and the like of a traditional method, and provides an efficient solution for electric power material management.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power metering, demand forecasting and industrial internet of things of power system, in particular to a metering equipment sub-warehouse demand forecasting and scheduling method based on space-time feature fusion BACKGROUND

[0002] At present, with the rapid development of China's power system, the power industry is transforming towards intelligence, and the demand forecasting of sub-warehouses of metering equipment of power system, such as electric energy meter and mutual inductor, has become an important content to ensure the efficiency of power supply service and the lean management of materials.

[0003] At present, the demand forecasting of sub-warehouses of power metering equipment is influenced by multiple factors such as regional power load growth, periodic replacement of equipment, extreme weather and metering fault anomaly, which leads to the demand of metering equipment showing significant space-time dynamics, for example, during the summer peak period, the failure rate of electric energy meter in a certain industrial and commercial intensive area will increase significantly. The fluctuation characteristics of such space-time coupling pose a more severe challenge to the demand forecasting method under the existing technology.

[0004] The demand forecasting method of metering equipment under the existing technology mainly relies on statistical models such as ARIMA, exponential smoothing or single machine learning algorithms such as decision tree. Although these methods can handle the time series data prediction of single sub-warehouse in actual application, they have the following defects when the demand of metering equipment presents significant space-time dynamics:

[0005] 1. Insufficient space-time dynamic representation, which does not effectively model the spatial correlation between sub-warehouses, i.e. the demand conduction effect of adjacent sub-warehouses due to sharing the distribution network, and external events in the time dimension, such as the impact of rules introduction / change, leading to demand lag response;

[0006] 2. Lack of rule constraints, the existing technology ignores the mathematical expression of the forced replacement period of metering equipment, leading to prediction deviation in the centralized replacement period;

[0007] 3. Weak data fusion capability, the existing technology does not consider the fusion of multi-source heterogeneous data, nor does it design deep feature interaction.

[0008] Therefore, there is an urgent need for a new demand forecasting method for sub-warehouses of power metering equipment, which can effectively fuse multi-source data, simultaneously associate space-time dynamics, business rule constraints, etc., to support accurate demand forecasting and inventory scheduling decisions. SUMMARY

[0009] In order to solve the above problems, the present application proposes a metering equipment sub-warehouse demand forecasting and scheduling method based on space-time feature fusion according to the space-time dynamic characteristics of power metering equipment sub-warehouse demand and the characteristics of multi-source heterogeneous data fusion, and the specific steps are as follows:

[0010] 1) Sub-warehouse data acquisition and collection;

[0011] 2) Based on the sub-warehouse data of step 1), a space-time feature fusion model is established;

[0012] 3) Establishing a space-time graph convolution prediction network:

[0013] The space-time graph convolution includes a graph convolution layer and a time convolution layer, and through the two hierarchical feature extraction and fusion, dynamic weight fusion is obtained, and the joint modeling of space-time dynamics is realized;

[0014] 4) Based on the dynamic weight fusion of step 3), the prediction and allocation decision is implemented.

[0015] According to the metering equipment sub-warehouse demand prediction and scheduling method based on space-time feature fusion, the sub-warehouse data acquisition and collection of step 1) include equipment data, power consumption data and space data, specifically:

[0016] 1) Equipment data:

[0017] Obtain the monthly electric energy meter, historical installation quantity, inventory quantity and scrap record of the sub-warehouse of each city power supply company through the metering center;

[0018] 2) Power consumption data:

[0019] Obtain the average power consumption per household and the peak-valley load ratio of the sub-warehouse service area through the power consumption information acquisition system;

[0020] 3) Space data:

[0021] Obtain the sub-warehouse jurisdiction range and the substation / line information shared by adjacent sub-warehouses.

[0022] According to the metering equipment sub-warehouse demand prediction and scheduling method based on space-time feature fusion, the step 2) of establishing a space-time feature fusion model includes three parts: quantifying equipment replacement cycle, regional power consumption correlation and sub-warehouse space correlation graph construction, and the relationship between the quantified equipment replacement cycle, regional power consumption correlation and sub-warehouse space correlation graph construction and the sub-warehouse data obtained and collected in step 1) corresponds.

[0023] In this step, by establishing the coupling mechanism of business rule constraint and data driven prediction, the influence of core factors such as equipment replacement cycle and regional power consumption correlation on sub-warehouse demand is quantified, and the problem of insufficient modeling of power business characteristics in traditional methods is solved.

[0024] The method for forecasting and scheduling the demand of metering equipment based on spatiotemporal feature fusion according to the present invention is characterized in that the quantification of equipment replacement cycle is based on the equipment data acquired and collected in step 1) of the warehousing data acquisition, and the equipment replacement cycle constraint is established, specifically as follows:

[0025] In this step, a replacement demand function for metering equipment is constructed to calculate the total demand for equipment reaching its mandatory replacement cycle in month t, thereby quantifying the number of devices that need to be replaced due to reaching their service life, as follows:

[0026]

[0027] In the formula: K represents the total number of historical equipment installation batches. Let v be the initial installation time of the kth batch of equipment, and v be a step function. Output 1 if the number of years is greater than 8, otherwise output 0.

[0028] The method for predicting and scheduling the demand of metering equipment based on spatiotemporal feature fusion according to the present invention is characterized in that the regional electricity consumption correlation is established based on the electricity consumption data acquired and collected in step 1), as follows:

[0029] In this step, the correlation coefficient between electricity consumption growth rate and equipment demand is calculated to quantify the dynamic correlation between electricity consumption growth and equipment demand, thereby weighting the importance of electricity consumption characteristics. The closer the correlation coefficient is to 1, the greater the influence, as follows:

[0030]

[0031] In the formula, E t Let D be the average electricity consumption per household in month t. t Let Cov be the equipment installation volume in month t, and σ be the covariance. E σ D ρ represents the standard deviation of electricity consumption and installation volume, respectively. E,D Pearson correlation coefficient;

[0032] Then, regarding the above electricity consumption data E t Weighting is applied to enhance the feature representation of highly correlated regions. The weighted electricity consumption characteristics are calculated using the following formula:

[0033]

[0034] The method for predicting and scheduling the demand of metering equipment in storage compartments based on spatiotemporal feature fusion according to the present invention is characterized in that the construction of the storage compartment spatial association graph is based on the spatial data acquired and collected in step 1), and the construction of the storage compartment spatial association graph is specifically as follows:

[0035] In this step, based on the spatial data of each power supply company sub-warehouse, the spatial correlation weight between sub-warehouses is calculated by comprehensively considering the facility sharing and geographical overlap, based on the adjacency of jurisdiction area, the number of shared power grid settings; The number of shared devices represents the coupling degree of power grid operation, which can support cross-warehouse demand conduction; The proportion of overlapping area of sub-warehouse jurisdiction area reflects the convenience of device allocation caused by boundary intersection, as follows:

[0036]

[0037] In the formula, N shared (i,j) is the number of jurisdiction devices shared by sub-warehouse i and j, AreaOverlap(i,j) is the overlapping area of sub-warehouse jurisdiction area, N total (i) is the total number of jurisdiction devices of sub-warehouse i, and Area(i) is the total jurisdiction area of sub-warehouse i.

[0038] According to the metering equipment sub-warehouse demand prediction and dispatching method based on space-time feature fusion of the application, the graph convolution layer in the space-time graph convolution prediction network is established, the graph convolution layer aggregates adjacent sub-warehouse features, and learns spatial dependence and updates node features, and the specific process is as follows:

[0039]

[0040] In the formula, The feature vector of sub-warehouse i at the lth layer, N(i) is the adjacent sub-warehouse set of sub-warehouse i; W (l) is a trainable parameter matrix, sigma is a ReLU activation function, and "⊕" is feature splicing, which combines the weighted power consumption feature and the device historical data feature;

[0041] In this step, the normalized adjacency matrix w i,j is used as the weight to aggregate the features of adjacent sub-warehouses (j∈N(i)) By learning the spatial conduction law of regional demand, the node feature of sub-warehouse i is fused with the power consumption trend of adjacent sub-warehouse j, and at the input layer of the space-time graph convolution prediction network, the weighted power consumption feature of the regional power consumption correlation in step 2) And the device historical data are spliced to form a multi-dimensional node feature vector, and the power consumption influence after weighting by the graph convolution layer makes the model adaptively pay attention to the high correlation area.

[0042] According to the metering equipment sub-warehouse demand prediction and dispatching method based on space-time feature fusion of the application, the time convolution layer in the space-time graph convolution prediction network is established, the time convolution layer extracts the seasonal and trend features of device demand through time convolution, and extracts the long-term periodic law of device demand based on dilated convolution, and the specific process is as follows:

[0043] T (l+1) = DilatedConv(T (l) , kernel_size=3)

[0044] where T (l) is the input time-series feature of the l-th layer, DilatedConv is a causal convolution with dilation factor 2 l , and the coverage time window exponentially increases with the layer number;

[0045] In this step, the time-series receptive field is expanded layer by layer to capture the annual electricity peak valley cycle, and the short-term fluctuations are skipped through the hollow convolution to enhance the robustness of the model.

[0046] The above-mentioned time-series receptive field is expanded layer by layer, for example, expanded to the second layer convolution covering 12 months of data.

[0047] The metering equipment demand prediction and scheduling method based on space-time feature fusion according to the application is characterized in that the dynamic weight fusion in the space-time graph convolution prediction network is established in step 3), the dynamic weight fusion balances the rule constraint and the data prediction result, enhances the business adaptability, and thus improves the accuracy of metering equipment demand prediction, and the specific implementation is as follows:

[0048]

[0049] wherein, α is the fusion weight, determined by fitting historical data, controlling the fusion proportion of rule constraint and data-driven prediction, is the original prediction value output by the space-time network.

[0050] The metering equipment demand prediction and scheduling method based on space-time feature fusion according to the application is characterized in that the implementation prediction and allocation decision of step 4) is as follows:

[0051] 1) Convert the predicted demand into executable inventory allocation instructions to realize closed-loop optimization of demand prediction and material scheduling, reduce the risk of stockout and transportation cost, and the formula is as follows:

[0052]

[0053] wherein, S i is the current inventory of the i-th warehouse, is only allowed to be associated with a weight ≥0.7, that is, the warehouse with high correlation participates in the allocation;

[0054] 2) Allocation suggestion generation: calculate the equipment allocation amount of the warehouse with high correlation, and give priority to guarantee the emergency demand, and the specific implementation is as follows:

[0055] When the predicted demand is greater than the current inventory Si When the metering equipment scheduling demand is triggered, the high-correlation sub-warehouse (w ij ≥0.7) is allowed to participate in the allocation to avoid a surge in long-distance transportation costs, and the Transfer i→j The instructions are generated in descending numerical order to preferentially meet the sub-warehouse with the highest risk of out-of-stock;

[0056] 3) Dynamic scheduling optimization, as follows:

[0057] The state of the sub-warehouse inventory and the demand prediction value are updated every week, the feasibility of the allocation suggestion is verified, it is confirmed whether the inventory of the sub-warehouse is sufficient, and finally the allocation suggestion is pushed to the sub-warehouse managers through the management system, and the global inventory view is updated synchronously.

[0058] The metering equipment sub-warehouse demand prediction and scheduling method based on spatiotemporal feature fusion has the following beneficial effects:

[0059] 1. The metering equipment sub-warehouse demand prediction and scheduling method based on spatiotemporal feature fusion, by constructing a multi-modal graph structure that fuses spatiotemporal correlation and replacement rules, designing a spatiotemporal joint modeling architecture and a dynamic feedback mechanism, realizes high-precision demand prediction and global inventory optimization.

[0060] 2. The metering equipment sub-warehouse demand prediction and scheduling method based on spatiotemporal feature fusion systematically solves the core problems of traditional methods, such as insufficient spatial correlation modeling, dynamic event response lag, and low efficiency of multi-source data utilization, and provides an efficient solution for power material management. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 The metering equipment sub-warehouse demand prediction and scheduling method based on spatiotemporal feature fusion has the following beneficial effects: DETAILED DESCRIPTION

[0062] The metering equipment sub-warehouse demand prediction and scheduling method based on spatiotemporal feature fusion has the following beneficial effects:

[0063] As shown in the drawings, the metering equipment sub-warehouse demand prediction and scheduling method based on spatiotemporal feature fusion has the following specific steps: Figure 1 1) Sub-warehouse data acquisition and collection;

[0064] 2) Based on the sub-warehouse data of step 1), a spatiotemporal feature fusion model is established;

[0065] 3) Establish a spatiotemporal graph convolution prediction network:

[0066]

[0067] ​The spatio-temporal graph convolution includes a graph convolution layer and a time convolution layer, and dynamic weight fusion is obtained through two-layer feature extraction and fusion, so as to realize joint modeling of spatio-temporal dynamics.

[0068] 4) Based on the dynamic weight fusion of step 3), the prediction and allocation decision is implemented.

[0069] The sub-warehouse data acquisition and collection of step 1) include device data, power consumption data and space data, specifically:

[0070] 1) Device data:

[0071] The monthly electric energy meter, historical installation quantity of mutual inductor, inventory quantity and scrap record in the sub-warehouse of each municipal power supply company are obtained through the metering center;

[0072] 2) Power consumption data:

[0073] The average power consumption per household and the peak-valley load ratio of the sub-warehouse service area are obtained through the power consumption information acquisition system;

[0074] 3) Space data:

[0075] The sub-warehouse jurisdiction range and the transformer substation / line information shared by adjacent sub-warehouses are obtained.

[0076] The establishment of the spatio-temporal feature fusion model of step 2) includes three parts: quantification of device replacement period, regional power consumption correlation and construction of sub-warehouse space correlation graph, and the relationship between the quantification of device replacement period, regional power consumption correlation and sub-warehouse space correlation graph construction and the sub-warehouse data obtained and collected in step 1) corresponds.

[0077] The quantification of device replacement period is based on the device data obtained and collected in step 1) to establish a device replacement period constraint, which is specifically:

[0078] In this step, a metering device replacement demand function is constructed to calculate the total amount of device demand in the tth month due to reaching the mandatory replacement period, so as to quantify the number of devices replaced due to reaching the service life, as follows:

[0079]

[0080] In the formula: K is the total batch of historical device installation, is the initial installation time of the kth batch of devices, and δ is a step function, which outputs 1 when is greater than 8 years, otherwise 0.

[0081] The regional power consumption correlation is based on the power consumption data obtained and collected in step 1) to establish a regional power consumption correlation feature, which is specifically as follows:

[0082] In this step, the correlation coefficient between electricity consumption growth rate and equipment demand is calculated to quantify the dynamic correlation between electricity consumption growth and equipment demand, thereby weighting the importance of electricity consumption characteristics. The closer the correlation coefficient is to 1, the greater the influence, as follows:

[0083]

[0084] In the formula, E t Let D be the average electricity consumption per household in month t. t Let Cov be the equipment installation volume in month t, and σ be the covariance. E σ D ρ represents the standard deviation of electricity consumption and installation volume, respectively. E,D The Pearson correlation coefficient;

[0085] Then, regarding the above electricity consumption data E t Weighting is applied to enhance the feature representation of highly correlated regions. The weighted electricity consumption characteristics are calculated using the following formula:

[0086]

[0087] The spatial relationship diagram of the warehouse division is constructed based on the spatial data obtained and collected in step 1). Specifically, the spatial relationship diagram of the warehouse division is constructed as follows:

[0088] In this step, based on the spatial data of each power supply company's sub-compartments, and taking into account both facility sharing and geographical overlap, the spatial correlation weight between sub-compartments is calculated based on the adjacency of their jurisdictional areas and the number of shared grid facilities. The number of shared equipment represents the coupling degree of grid operation, indicating its ability to support cross-compartment demand transmission. The percentage of overlapping area in the jurisdictional areas of sub-compartments reflects the convenience of equipment allocation due to boundary intersections, as follows:

[0089]

[0090] In the formula, N shared (i,j) represents the number of managed devices shared by warehouses i and j, AreaOverlap(i,j) represents the overlapping area of ​​the managed areas of the warehouses, and N total (i) represents the total number of devices under the jurisdiction of sub-warehouse i, and Area(i) represents the total area under the jurisdiction of sub-warehouse i.

[0091] Step 3) Establish the graph convolutional layer in the spatiotemporal graph convolutional prediction network. This graph convolutional layer aggregates adjacent compartment features and learns spatial dependencies and updates node features. Specifically:

[0092]

[0093] In the formula, Feature vector of sub-bin i in the lth layer, N(i) is the set of adjacent sub-bins of sub-bin i; W (l) is a trainable parameter matrix, σ is a ReLU activation function, and "⊕" is feature concatenation, which combines the weighted power consumption features and device historical data features;

[0094] In this step, the normalized adjacency matrix w i,j is used as the weight to aggregate the features of adjacent sub-bins (j ∈ N(i)) By learning the spatial transmission law of regional demand, the node features of sub-bin i are fused with the power consumption trend of adjacent sub-bin j. At the same time, in the input layer of the spatio-temporal graph convolution prediction network, the weighted power consumption features of regional power consumption correlation in step 2) are concatenated with the device historical data to form a multi-dimensional node feature vector. Through the weighted power consumption influence of the graph convolution layer, the model can adaptively focus on high-correlation regions.

[0095] Step 3) establishes the time convolution layer in the spatio-temporal graph convolution prediction network. The time convolution layer extracts the seasonal and trend features of device demand through time convolution, and extracts the long-term periodicity of device demand based on dilated convolution. Specifically:

[0096] T (l+1) = DilatedConv(T (l) , kernel_size = 3)

[0097] In the formula, T (l) is the input time series feature of the lth layer, and DilatedConv is a causal convolution with a dilation factor of 2 l , and the coverage time window increases exponentially with the layer number.

[0098] In this step, the time series perception is expanded layer by layer to capture the annual power consumption peak-valley cycle, and the dilated convolution is used to skip short-term fluctuations to enhance the robustness of the model.

[0099] Step 3) establishes the dynamic weight fusion in the spatio-temporal graph convolution prediction network. The dynamic weight fusion balances the rule constraints and data prediction results to enhance business adaptability, thereby improving the accuracy of metering device demand prediction. Specifically:

[0100]

[0101] In the formula, α is the fusion weight determined by fitting historical data, which controls the fusion proportion of rule constraints and data-driven prediction, is the original prediction value output by the spatio-temporal network.

[0102] Step 4) implements prediction and allocation decision, which is specifically:

[0103] ​1) Transform the predicted demand into executable inventory allocation instructions, realize the closed-loop optimization of demand prediction and material scheduling, reduce the risk of stockout and transportation cost, the formula is as follows:

[0104]

[0105] In the formula, S i is the current inventory of the warehouse i, only allows the associated weight to be greater than or equal to 0.7, that is, the warehouse with high correlation participates in the allocation;

[0106] 2) Generation of allocation suggestions: calculate the device allocation amount of the warehouse with high correlation, and give priority to guarantee the emergency demand, as follows:

[0107] When the predicted demand is greater than the current inventory S i , trigger the metering device scheduling demand, allow the warehouse with high correlation (w ij ≥0.7) to participate in the allocation, avoid the surge of long-distance transportation cost, and generate instructions in descending order of Transfer i→j value, and give priority to the warehouse with the highest risk of stockout;

[0108] 3) Dynamic scheduling optimization, as follows:

[0109] Update the state of the warehouse inventory and the demand prediction value every week, and check the feasibility of the allocation suggestion, confirm whether the warehouse inventory is sufficient, and finally push it to the person in charge of each warehouse through the management system, and update the global inventory view synchronously.

[0110] Embodiment

[0111] The metering device warehouse demand prediction and scheduling method based on spatio-temporal feature fusion in the embodiment realizes accurate prediction and dynamic allocation of metering device warehouse demand through multi-modal data fusion and spatio-temporal graph convolution network, collects the device inventory quantity and regional power consumption time series data of each warehouse in real time, constructs a spatio-temporal feature matrix, captures the spatial dependence and time period law of power demand through the spatio-temporal graph convolution network, outputs the device gap prediction in the next 7 days, and generates allocation decisions based on the gap prediction and geographical position, and formulates a phased allocation plan.

[0112] In the embodiment, the high priority is:

[0113] For the warehouse with a stockout risk greater than 90%, cross-regional direct allocation is preferred (high priority uses the case of Shanghai area, Minnan to Pudong as cross-regional), to ensure delivery within 24 hours.

[0114] In the embodiment, the medium and low priority is:

[0115] Through adjacent sub-warehouse cooperative scheduling (medium and low priority uses Qingpu-Jiading in Shanghai as an adjacent area case).

[0116] The above allocation suggestion list is shown in Table 1 as follows:

[0117]

[0118] Table 1 Allocation suggestion table

[0119] The space-time feature fusion-based metering equipment sub-warehouse demand prediction and scheduling method of the application is evaluated, and the data of a power supply sub-warehouse from January 1, 2023 to January 31, 2025 is selected as an example to predict the warehouse demand data on February 7, 2025, and the prediction accuracy of the existing mainstream methods is compared, including ARIMA, XGBoost, GCN, etc. Model, to reflect the influence of prediction accuracy on allocation decision by mean absolute percentage error (MAPE), and finally the comparison results of prediction are as follows Table 2

[0120]

[0121] Table 2 Comparison of multiple model prediction results of example analysis

[0122] From the above results, it is obvious that the space-time feature fusion-based metering equipment sub-warehouse demand prediction and scheduling method of the application can significantly reduce the prediction error by fusing the space-time features of power consumption.

[0123] The space-time feature fusion-based metering equipment sub-warehouse demand prediction and scheduling method of the application can realize high-precision demand prediction and global inventory optimization by constructing a multi-modal graph structure that fuses space-time correlation and replacement rules, designing a space-time joint modeling architecture and a dynamic feedback mechanism. The application systematically solves the core problems of traditional methods, such as insufficient spatial correlation modeling, dynamic event response lag, and low efficiency of multi-source data utilization, and provides an efficient solution for power material management.

[0124] In the description of the application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore it cannot be understood as a limitation of the application.

[0125] Furthermore, the terms "first", "second", "third", etc. are used only for descriptive purposes and do not connote or imply any relative importance. In the description of the present application, it is to be understood that the terms "mounting", "connected", "connecting", etc. should be interpreted broadly, e.g. can mean fixed connection, detachable connection, or integral connection; can mean mechanical connection, or electrical connection; can mean direct connection, or indirect connection via an intermediate medium; can mean internal connection of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0126] Meanwhile, those skilled in the art should recognize that the above embodiments are only used to illustrate the present application, and are not used as a limitation to the present application. Any changes or modifications to the above described embodiments, as long as they are within the scope of the spirit of the present application, will fall within the scope of the claims of the present application.

Claims

1. A metering equipment warehouse demand prediction and scheduling method based on spatio-temporal feature fusion, the specific steps of which are as follows: 1) Warehouse data acquisition and collection; 2) Based on the warehouse data of step 1), a spatio-temporal feature fusion model is established; 3) Establishing a spatio-temporal graph convolution prediction network: The spatio-temporal graph convolution includes a graph convolution layer and a time convolution layer. Through the two hierarchical feature extraction and fusion, dynamic weight fusion is obtained, and the joint modeling of spatio-temporal dynamics is realized; 4) Based on the dynamic weight fusion of step 3), the prediction and allocation decision is implemented.

2. The method of claim 1, wherein the method further comprises: The warehouse data acquisition and collection of step 1) includes equipment data, power consumption data and spatial data, specifically: 1) Equipment data: Obtain the monthly electric energy meter, historical installation quantity, inventory quantity and scrap record of the transformer in the warehouse of each municipal power supply company through the metering center; 2) Power consumption data: Obtain the average power consumption per household and the peak-valley load ratio of the warehouse service area through the power consumption information acquisition system; 3) Spatial data: Obtain the warehouse jurisdiction range and the substation / line information shared by adjacent warehouses. 3.The method of claim 1, wherein, The establishment of the spatio-temporal feature fusion model of step 2) includes three parts: quantification of equipment replacement cycle, regional power consumption correlation and warehouse space correlation graph construction, and the relationship between the quantification of equipment replacement cycle, regional power consumption correlation and warehouse space correlation graph construction and the warehouse data obtained and collected in step 1) corresponds. 4.The method of claim 3, wherein, The quantification of equipment replacement cycle is based on the equipment data obtained and collected in step 1) to establish the equipment replacement cycle constraint, which is specifically: In this step, the metering equipment replacement demand function is constructed to calculate the total equipment demand in the tth month due to reaching the mandatory replacement cycle, thereby quantifying the number of equipment replaced due to reaching the service life, as follows: where K is the total number of historical equipment installation batches, is the initial installation time of the kth batch of equipment, and δ is a step function that outputs 1 when i.e., greater than 8 years, and 0 otherwise.

5. The method of claim 3, wherein the method further comprises: The regional power consumption correlation is based on the power consumption data obtained and collected in step 1) to establish the regional power consumption correlation feature, which is specifically as follows: In this step, the correlation coefficient between the power consumption growth rate and the equipment demand is calculated to quantify the dynamic correlation between the power consumption growth and the equipment demand, thereby weighting the importance of the power consumption feature. The closer the correlation coefficient is to 1, the greater the impact is, as follows: In the formula, E t is the average electricity consumption per household in the tth month, D t is the installation amount of equipment in the tth month, Cov is the covariance, σ E σ D are the standard deviations of the electricity consumption and the installation amount, respectively, and ρ E，D is the Pearson correlation coefficient. The above power consumption data E t is weighted, and the feature expression of the high correlation area is strengthened, The calculation formula of the weighted power consumption feature is as follows:

6. The method of claim 3, wherein the method further comprises: The warehouse space correlation graph construction is based on the spatial data obtained and collected in step 1) to construct the warehouse space correlation graph, which is specifically as follows: In this step, based on the spatial data of each power supply company warehouse, the facility sharing and geographical overlap are considered comprehensively, the spatial correlation weight between warehouses is calculated based on the adjacency of the jurisdiction area, the number of shared power grid devices, the number of shared devices representing the coupling degree of power grid operation, and the ability to support cross-warehouse demand conduction, and the proportion of overlapping area of warehouse jurisdiction area reflecting the convenience of equipment allocation caused by boundary intersection, as follows: where N shared (i,j) is the number of jurisdictional devices shared by bin i,j, AreaOverlap(i,j) is the overlapping area of jurisdictional areas of bin i,j, N total (i) is the total number of jurisdictional devices of bin i, Area(i) is the total jurisdictional area of bin i.

7. The method of claim 1, wherein the method further comprises: The graph convolution layer in step 3) of establishing a spatio-temporal graph convolution prediction network aggregates the features of adjacent warehouses and learns the spatial dependence relationship and updates the node features, which is specifically as follows: In the formula, Feature vector of the i-th bin in the l-th layer, N(i) is the set of adjacent bins of the i-th bin; W (l) is a trainable parameter matrix, and σ is a ReLU activation function, is feature splicing, which combines the weighted power consumption feature and the device historical data feature; In this step, the normalized adjacency matrix w i,j is taken as the weight, and the features of adjacent sub-warehouses (j ∈ N(i)) are aggregated By learning the spatial conduction law of regional demand, the node features of i sub-warehouse are fused with the electricity consumption trend of adjacent sub-warehouse j. At the same time, in the input layer of the spatio-temporal graph convolution prediction network, the weighted electricity consumption features of regional electricity consumption correlation in step 2) are formed Spliced with device historical data to form a multi-dimensional node feature vector, and through the electricity consumption influence weighted by the graph convolution layer, the model is self-adaptive to pay attention to the high correlation region. 8.The method of claim 1, wherein, The time convolution layer in step 3) of establishing a spatio-temporal graph convolution prediction network extracts the seasonal and trend features of equipment demand through time convolution, and extracts the long-term periodicity of equipment demand based on dilated convolution, which is specifically as follows: T (l+1) = DilatedConv(T (l) , kernel_size=3) In the formula, T (l) is the input timing feature of the first layer, DilatedConv is a causal convolution with a dilation factor of 2 l , and the coverage time window exponentially increases with the layer number. In this step, the annual electricity peak valley cycle is captured by gradually expanding the timing perception layer by layer, and the short-term fluctuations are skipped through hollow convolution to enhance the robustness of the model. 9.The method of claim 1, wherein, The step 3) establishes the dynamic weight fusion in the space-time graph convolution prediction network, which balances the rule constraint and the data prediction result, enhances the business adaptability, and thus improves the accuracy of the demand prediction of the metering equipment, and specifically as follows: In the formula, a is a fusion weight, which is determined by fitting historical data, and controls the fusion proportion of the rule constraint and the data-driven prediction, The original prediction value output for the spatiotemporal network. 10.The method of claim 1, wherein, The step 4) is implemented for the prediction and allocation decision, which is specifically as follows: 1) The predicted demand is converted into executable inventory allocation instructions to realize the closed-loop optimization of the demand prediction and the material scheduling, reduce the risk of shortage and the transportation cost, and the formula is as follows: In the formula, S i is the current inventory of the warehouse i, Only allow the associated weight ≥ 0.7, that is, the high correlation of the warehouse to participate in the allocation. 2) The allocation suggestion is generated: the device allocation amount of the high-correlation sub-warehouse is calculated to preferentially guarantee the emergency demand, and the specific formula is as follows: When the predicted demand exceeds the current inventory S i , trigger the metering device to schedule demand, allow high-association warehouses (w ij ≥ 0.7) to participate in allocation, avoid a sharp increase in long-distance transportation costs, and generate instructions in descending order of the Transfer i→j value, prioritizing the highest-risk warehouses. 3) Dynamic scheduling optimization, as follows: The state of the sub-warehouse inventory and the demand prediction value are updated every week, the feasibility of the allocation suggestion is verified, it is confirmed whether the sub-warehouse inventory is sufficient, and finally the management system is pushed to the sub-warehouse managers, and the global inventory view is updated synchronously.

Images