A power grid material demand prediction model iteration optimization and verification method

Abstract

The application provides a power grid material demand prediction model iterative optimization and verification method, relates to the technical field of power grid material management, collects multi-dimensional influence factors of power grid material demand, performs pre-position quality control discrimination on historical demand data of power grid operation equipment based on each demand transmission path, obtains calibration driving information of the power grid material demand prediction model; the demand iteration characteristics and the change state attribute of the power grid operation scene are reconstructed and evolved, the demand response quantity in the evolution response time of the power grid material demand is obtained, the power grid material demand prediction model is executed hierarchical update based on the calibration driving information and the demand response quantity, and the updated power grid material demand prediction model is obtained; the updated power grid material demand prediction model is adapted and verified, and the power grid material demand prediction result is output. The application can realize accurate iterative verification of the power grid material demand prediction model, so as to improve the accuracy of the power grid material demand prediction result.

Description

Technical Field

[0001] This application relates to the field of power grid material management technology, and more specifically, to a method for iterative optimization and verification of a power grid material demand forecasting model. Background Technology

[0002] Power grid material management refers to a comprehensive business system in which power grid enterprises coordinate and control the entire life cycle of various materials such as transformers and cables in production activities such as power grid construction, operation and maintenance, and emergency repair. The core objective is to ensure the continuous supply of materials for the safe and stable operation of the power grid, while optimizing inventory structure, reducing capital occupation costs, and improving supply chain collaboration efficiency. With the advancement of the construction of new power systems, modern power grid material management is deeply integrated with the Internet of Things and big data technologies, and is gradually transforming into a data-driven, precise, and intelligent supply chain management model.

[0003] However, existing optimization and validation methods for power grid material demand forecasting models rely solely on full feature channels and unconstrained unidirectional output for model training and result evaluation. They fail to incorporate feature weight distribution mask injection, multi-dimensional operational constraint vector space construction, and decision boundary hyperplane mapping. This makes it difficult to accurately remove redundant feature noise and quantify the degree to which the predicted output deviates from compliance boundaries. Furthermore, it hinders the targeted constraint of underlying parameter correction paths and ensures that prediction results align with real-world business rules. Additionally, the lack of dynamic validation and constraint gradient direction decoding mechanisms based on adaptation deviation gradients easily leads to model convergence oscillations and theoretical predictions that deviate from actual operational scenarios. Therefore, achieving accurate iterative validation of power grid material demand forecasting models to improve the accuracy of power grid material demand forecasting results is a key challenge facing the industry. Summary of the Invention

[0004] This application provides an iterative optimization and verification method for a power grid material demand forecasting model, which can achieve accurate iterative verification of the power grid material demand forecasting model to improve the accuracy of power grid material demand forecasting results.

[0005] Firstly, this application provides an iterative optimization and verification method for a power grid material demand forecasting model, wherein the method includes: Collect multi-dimensional influencing factors of power grid material demand and determine the correlation and influence map between each influencing factor; The demand transmission path corresponding to the correlation influence map is determined, and the historical demand data of the power grid operating equipment is pre-controlled and screened based on each demand transmission path to obtain the calibration driving information of the power grid material demand prediction model. Determine the demand iteration characteristics of power grid material demand, reconstruct and evolve the demand iteration characteristics and the change state attributes of power grid operation scenario to obtain the demand response quantity when power grid material demand evolves, and perform hierarchical update on the power grid material demand prediction model based on the calibration driving information and the demand response quantity to obtain the updated power grid material demand prediction model. Based on the operational constraints of multi-dimensional operation scenarios, the updated power grid material demand forecasting model is adapted and verified, and the power grid material demand forecasting results are output.

[0006] In this embodiment, determining the correlation and influence map among the various influencing factors specifically includes: Based on the aforementioned multidimensional influencing factors, determine the topological characteristic sequence of the transmission influence between each influencing factor; Based on the aforementioned topological feature sequence, a community detection algorithm is used to divide the data, resulting in multiple communities affected by demand. The correlation and influence map between various influencing factors is determined based on the correlation weights of all demand effects within the community.

[0007] In this embodiment, determining the demand transmission path corresponding to the correlation influence map specifically includes: Extract directed transmission subgraphs from the aforementioned correlation influence graph; By searching layer by layer along the decreasing direction of the associated edge weights in the directed transmission subgraph, an effective demand transmission path is obtained; The demand transmission path corresponding to the correlation influence map is determined based on the node level depth of the effective demand transmission path.

[0008] In this embodiment, the calibration driving information for the power grid material demand prediction model is obtained by performing pre-qualification quality control on historical demand data of power grid operating equipment based on each demand transmission path, specifically including: Historical demand data of power grid operating equipment are extracted along each demand transmission path, and the demand variation coefficient is determined; Based on the demand variation coefficient and the quality control benchmark vector for power grid material demand forecasting, the historical demand data is mapped to a reliability level to obtain a quality control screening matrix. The historical demand data is aggregated based on the quality control screening matrix to generate calibration driving information for the power grid material demand prediction model.

[0009] In this embodiment, the historical demand data is mapped to a reliability level based on the demand variation coefficient and the quality control benchmark vector used for power grid material demand forecasting, resulting in a quality control screening matrix. This matrix specifically includes: Determine the reliable distance between the demand variation coefficient and the quality control benchmark vector; The trusted distance is reconstructed in two dimensions along the time series of the historical demand data to generate a quality control screening matrix.

[0010] In this embodiment, determining the demand iteration characteristics of power grid material demand specifically includes: Extract version change records of power grid material requirements in multiple planning batches, and calculate the demand evolution sequence of material categories between adjacent versions; The demand iteration characteristics of power grid material demand are determined based on the demand evolution state sequence.

[0011] In this embodiment, the demand iteration characteristics and the changing state attributes of the power grid operation scenario are reconstructed and evolved to obtain the demand response quantity when the power grid material demand evolves, specifically including: The state transition probability in the demand iteration characteristics and the change state attributes of the power grid operation scenario are subjected to tensor product operation to obtain the joint evolution tensor. The joint evolution tensor input time-series reconstruction network is used to perform multi-step state deduction, and the demand state offset vector under each deduction step is extracted to generate the material demand evolution response trajectory. The evolution response trajectories of the material demand are aggregated to calculate the demand response quantity when the power grid material demand evolves.

[0012] In this embodiment, the power grid material demand forecasting model is updated hierarchically based on the calibration driving information and the demand response quantity to obtain the updated power grid material demand forecasting model, specifically including: A mask is constructed based on the calibration driving information and injected into the model feature embedding layer. The underlying network parameters are corrected through gradient backpropagation. The residual between the demand response and historical demand is calculated to construct a hierarchical loss function, which drives the model output layer to iteratively update the parameters of the underlying network. Conditional determinations are made on the updated underlying network parameters to form an updated power grid material demand prediction model.

[0013] In this embodiment, constructing a mask based on the calibration driving information and injecting it into the model feature embedding layer, and correcting the underlying network parameters through gradient backpropagation specifically includes: The feature weight distribution in the calibration driving information is analyzed to generate a sparse mask matrix and multiplied element by element into the feature embedding layer to mask redundant feature channels. The forward loss is calculated based on the output of the feature embedding layer after mask injection, and gradient backpropagation is performed along the unmasked feature channels to correct the parameters of the underlying network.

[0014] In this embodiment, based on the operational constraints of a multi-dimensional operational scenario, the updated power grid material demand forecasting model is adapted and verified, and the output power grid material demand forecasting results specifically include: Construct a multidimensional constraint vector space based on the various operational constraints in the multidimensional operational scenario; The multidimensional constraint vector space is mapped to a decision boundary hyperplane, and the adaptation deviation gradient between the model output and the decision boundary hyperplane is calculated to generate an adaptation verification index sequence. When the adaptation verification index sequence meets the preset convergence condition, the demand feature mapping of the model output layer is decoded along the constraint gradient direction, and the power grid material demand prediction result is output.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: This process involves collecting multi-dimensional influencing factors on power grid material demand and determining the correlation and influence maps among these factors. It then identifies the demand transmission paths corresponding to these correlation and influence maps, and performs preliminary quality control on historical demand data of power grid operating equipment based on each demand transmission path to obtain calibration driving information for the power grid material demand prediction model. Next, it determines the demand iteration characteristics of power grid material demand, reconstructs and evolves these characteristics and the changing state attributes of the power grid operation scenario to obtain the demand response quantity during the power grid material demand evolution response. Based on the calibration driving information and the demand response quantity, it performs a hierarchical update of the power grid material demand prediction model to obtain the updated power grid material demand prediction model. Finally, based on the operational constraints of the multi-dimensional operation scenario, it adapts and verifies the updated power grid material demand prediction model, outputting the power grid material demand prediction results.

[0016] Therefore, this application demonstrates that the updated power grid material demand prediction model can be adapted and verified based on the operational constraints of multi-dimensional operational scenarios, outputting power grid material demand prediction results. Specifically, determining the calibration-driven information allows for the acquisition of the directional guidance vector for model parameter correction and the gradient of the constraint boundary, effectively suppressing convergence oscillations during model iteration. This calibration-driven information deeply integrates multi-dimensional operational constraints with feature masks, providing a precise optimization direction for updating the underlying weights, avoiding gradient descent from falling into local optima or invalid search spaces. By quantifying the degree to which the predicted output deviates from the business compliance boundary, it accurately removes redundant feature noise from parameter adjustments, ensuring that each iteration strictly conforms to real-world material allocation rules, significantly improving the model's robustness and optimization efficiency in complex business scenarios. Determining the demand response quantity yields a baseline value for material demand forecasting that aligns with actual physical boundaries and operational rules. This provides a reliable quantitative reference for verifying the accuracy of the model's output. The demand response quantity maps theoretical forecasts to a multi-dimensional operational constraint vector space, effectively eliminating invalid forecast components that deviate from actual inventory capacity or budget limits. By constructing a decision boundary hyperplane, it can accurately measure the adaptation deviation between forecast results and actual business needs, thereby driving the model to perform targeted residual compensation and logical corrections, ensuring the feasibility of the final material demand forecast.

[0017] In summary, the technical solution adopted in this application can achieve accurate iterative verification of the power grid material demand forecasting model, thereby improving the accuracy of the power grid material demand forecasting results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an exemplary flowchart of an iterative optimization and verification method for a power grid material demand forecasting model provided in this application; Figure 2 This is a flowchart illustrating the process of determining the updated power grid material demand forecasting model based on the information provided in this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application provides an iterative optimization and verification method for a power grid material demand forecasting model. The core of this method involves collecting multi-dimensional influencing factors of power grid material demand and determining the correlation influence map between these factors; determining the demand transmission path corresponding to the correlation influence map, and performing pre-quality control screening on historical demand data of power grid operating equipment based on each demand transmission path to obtain calibration driving information for the power grid material demand forecasting model; determining the demand iteration characteristics of power grid material demand, reconstructing and evolving the demand iteration characteristics and the changing state attributes of the power grid operation scenario to obtain the demand response quantity when the power grid material demand evolves; performing a hierarchical update of the power grid material demand forecasting model based on the calibration driving information and the demand response quantity to obtain an updated power grid material demand forecasting model; and adapting and verifying the updated power grid material demand forecasting model based on the operational constraints of the multi-dimensional operation scenario to output the power grid material demand forecasting result.

[0022] To better understand the above technical solutions, a detailed description of the technical solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. (Refer to...) Figure 1 As shown in the figure, this is an exemplary flowchart of an iterative optimization and verification method for a power grid material demand forecasting model according to this embodiment of the application. The method includes the following steps: In step S1, multi-dimensional influencing factors of power grid material demand are collected, and the correlation and influence map between each influencing factor is determined.

[0023] In practice, the collection of multi-dimensional influencing factors of power grid material demand can be achieved in the following way: First, when collecting multi-dimensional influencing factors of power grid material demand, the data integration middleware is used to connect with the power grid enterprise resource planning system, equipment asset management system and external meteorological and macroeconomic databases. Using ETL data extraction, transformation and loading tools, raw business data such as equipment aging rate, historical fault frequency, meteorological disaster warning level and regional electricity consumption growth rate are periodically captured according to preset time granularity. Then, the K-nearest neighbor algorithm is used to impute missing values ​​in the captured raw data and the maximum and minimum normalization processing is performed to eliminate the difference in units. Finally, a standardized multi-dimensional influencing factor dataset is formed, and this multi-dimensional influencing factor dataset is used as the multi-dimensional influencing factors of power grid material demand.

[0024] It should be noted that, in this application, the multi-dimensional influencing factors refer to the set of data indicators on changes in power grid material consumption and reserve demand.

[0025] In this embodiment, determining the correlation influence spectrum among various influencing factors can be achieved through the following steps: Based on the aforementioned multidimensional influencing factors, determine the topological characteristic sequence of the transmission influence between each influencing factor; Based on the aforementioned topological feature sequence, a community detection algorithm is used to divide the data, resulting in multiple communities affected by demand. The correlation and influence map among various influencing factors is determined based on the correlation weights of all demand effects within the community.

[0026] In specific implementation, firstly, multi-dimensional influencing factors are constructed as directed graph nodes. The transfer entropy algorithm is used to calculate the information transfer between any two nodes' time series. When the information transfer exceeds a preset threshold, a directed edge is established to represent the transmission influence. The depth-first search algorithm is used to traverse the directed graph, extracting all connected paths from the starting node to the target node. The node access order and edge direction of each path are recorded. The path structure data is serialized and encoded according to timestamps, ultimately generating a topological feature sequence. Then, the topological feature sequence is restored to a network adjacency matrix. The Louvain community detection algorithm is used. Initially, each influencing factor node is regarded as an independent community. By calculating the modularity gain before and after a node moves into an adjacent community, the nodes are iteratively assigned to the community that maximizes the modularity gain. When the modularity gain is less than the convergence threshold, the iteration stops, completing community aggregation. After aggregation, node subgraphs with internal edge density greater than external edge density are extracted. The node subgraphs are encapsulated and output as multiple demand-influencing communities. Finally, node pairs within the communities affected by each demand are extracted, and the linear correlation between time series of node pairs is calculated using the Pearson correlation coefficient. The absolute value of the linear correlation is used as the influence association weight. A weight filtering threshold is set to remove weak edges with influence association weights below the threshold. The retained nodes, boundary labels of the communities affected by the demand, and directed edges with influence association weights are fused into a graph structure. The graph database is used for persistent storage and visualization rendering to finally generate an association influence graph.

[0027] It should be noted that, in this application, "transmission influence" refers to the driving effect of a change in the state of a certain factor in a multidimensional factor network causing fluctuations in the demand of other factors along a specific path; "topological feature sequence" refers to an ordered set of network feature data containing the connection structure between nodes of each influencing factor and the order of transmission of influence; "community detection algorithm" refers to a calculation method used to identify the clustering characteristics of nodes in complex networks and divide highly cohesive node groups into independent subgraph structures; "demand-influenced community" refers to a set of node subgraphs in an associated network where internal factors interact closely and have a synergistic driving effect on the demand for specific materials; "influence association weight" refers to a numerical value that quantifies the degree of interdependence and the intensity of the effect between any two influencing factor nodes within the same demand-influenced community; and "association influence graph" refers to a topological network model that globally displays the influencing factor nodes, the boundary of the demand-influenced community, and the transmission relationship and weight between nodes.

[0028] In step S2, the demand transmission path corresponding to the correlation influence map is determined, and the historical demand data of the power grid operating equipment is pre-controlled and screened based on each demand transmission path to obtain the calibration driving information of the power grid material demand prediction model.

[0029] In this embodiment, the demand transmission path corresponding to the correlation influence map can be determined in the following way: Extract directed transmission subgraphs from the aforementioned correlation influence graph; By searching layer by layer along the decreasing direction of the associated edge weights in the directed transmission subgraph, an effective demand transmission path is obtained; The demand transmission path corresponding to the correlation influence map is determined based on the node level depth of the effective demand transmission path.

[0030] In practice, firstly, the material demand nodes in the correlation influence graph are set as the target sink. A reverse breadth-first search algorithm is used to backtrack layer by layer from the target sink along the opposite direction of the directed edges, marking all ancestor nodes that can reach the target sink. Then, unmarked isolated nodes and invalid edges in the correlation influence graph are removed, and the marked nodes and their original directed edges are retained. The retained nodes and edges are reconstructed into a connected directed acyclic graph structure, thereby extracting the directed transmission subgraph. Then, during the layer-by-layer search along the decreasing weight direction of the associated edges in the directed transmission subgraph, nodes with an in-degree of zero in the directed transmission subgraph are identified as the starting source. A depth-first search algorithm is used to traverse downwards from the starting source. When expanding nodes at each level, the difference between the current edge weight and the weight of the candidate edge in the next level is calculated. Only when the weight of a candidate edge is less than the weight of the current edge is the candidate node added to the search stack and the search continues downwards. When the search stack reaches the material demand node and the entire link strictly satisfies the weight decreasing condition, the connected link is extracted and determined as an effective demand transmission path. Finally, the total number of nodes in each effective demand transmission path is counted, and the total number of nodes minus one is taken as the node level depth of the effective demand transmission path. The arithmetic mean of the node level depths of all effective demand transmission paths is calculated, and the arithmetic mean is rounded down to obtain the depth benchmark threshold. Effective demand transmission paths with a node level depth less than or equal to the depth benchmark threshold are selected as the backbone links, and finally, the demand transmission paths corresponding to the association influence graph are determined.

[0031] It should be noted that, in this application, a directed transmission subgraph refers to a local network structure selected from the correlation influence graph that has a clear causal orientation and can drive changes in material demand; the correlation edge weight refers to a numerical parameter that quantifies the causal transmission strength and demand-driving contribution between adjacent influencing factor nodes in the graph; layer-by-layer search refers to a graph traversal operation that traverses nodes from shallow to deep according to the network topology hierarchy to select high-weight transmission links; an effective demand transmission path refers to a connected link that satisfies the weight decreasing constraint and can truly reflect the transmission effect of external factors to material demand nodes; node level depth refers to the topological hop distance traversed by a specific influencing factor node from the starting driving node in the transmission link; and the demand transmission path refers to the core causal link finally determined in the global graph to guide the feature extraction of the material demand prediction model.

[0032] In this embodiment, the calibration driving information for the power grid material demand prediction model can be obtained by pre-controlling and screening the historical demand data of power grid operating equipment based on each demand transmission path using the following steps: Historical demand data of power grid operating equipment are extracted along each demand transmission path, and the demand variation coefficient is determined; Based on the demand variation coefficient and the quality control benchmark vector for power grid material demand forecasting, the historical demand data is mapped to a reliability level to obtain a quality control screening matrix. The historical demand data is aggregated based on the quality control screening matrix to generate calibration driving information for the power grid material demand prediction model.

[0033] In practical implementation, firstly, the node identifiers included in the demand transmission path are parsed. Using a structured query language, the monthly material requisition ledgers for the corresponding nodes over the past three years are extracted from the power grid material management database to construct a historical demand data time series. When determining the demand variance coefficient, the arithmetic mean and standard deviation of this time series are calculated. The standard deviation is divided by the arithmetic mean to obtain the demand variance coefficient, which ranges from zero to positive infinity; a larger value indicates a higher degree of data dispersion. Next, a quality control benchmark vector is read from the business specification library. This benchmark vector contains the ideal variance coefficient threshold for each node under standard operating conditions. The absolute value of the difference between the demand variance coefficient and the corresponding element in the quality control benchmark vector is calculated, and this value is substituted into a Gaussian radial basis function for nonlinear transformation. The output is a confidence score between zero and one. The confidence scores corresponding to all demand transmission paths are arranged in a two-dimensional order by path number and time node to generate a quality control screening matrix. Finally, a credibility screening threshold is set, and the position indices of elements greater than the threshold are extracted by traversing the quality control screening matrix. The position indices are used to filter out high-quality data subsets from historical demand data. A weighted average algorithm is used to aggregate the high-quality data subsets, with the weight values ​​being the corresponding credibility scores. The Kullback-Leibler divergence between the aggregated data subset and the original historical demand data is calculated. The divergence value and the mean of the aggregated features are encapsulated to generate calibration driving information.

[0034] It should be noted that, in this application, historical demand data refers to a set of objective business ledgers recording the actual consumption and requisition of materials by power grid operating equipment in a specific past period; pre-processing quality control screening refers to the preprocessing operation of identifying anomalies and assessing the quality of input data before model training to ensure data reliability; power grid material demand prediction model refers to a neural network algorithm structure that uses historical characteristics to infer future power grid material consumption trends and outputs quantitative prediction values; demand variation coefficient refers to a statistical indicator that quantifies the severity of fluctuations and data dispersion in the time series of material demand under a specific transmission path; quality control benchmark vector refers to a reference numerical array that represents the ideal data distribution characteristics and reasonable fluctuation boundaries of power grid material demand under standard operating conditions; credibility mapping refers to the process of converting data fluctuation characteristics into numerical calculations that reflect the degree to which they conform to the standard distribution and are used to measure data quality; quality control screening matrix refers to a structured two-dimensional numerical array that presents the historical data quality scores and anomaly distribution locations under each transmission path in multiple dimensions; and calibration driving information refers to a set of feedback data that indicates the direction of prediction model parameter correction and the degree of feature distribution offset to guide the model's adaptive optimization.

[0035] Furthermore, in this embodiment, the quality control screening matrix is ​​obtained by performing a credibility mapping on the historical demand data based on the demand variation coefficient and the quality control benchmark vector for power grid material demand forecasting, which can be achieved through the following steps: Determine the reliable distance between the demand variation coefficient and the quality control benchmark vector; The trusted distance is reconstructed in two dimensions along the time series of the historical demand data to generate a quality control screening matrix.

[0036] In practice, firstly, the demand variation coefficient and the quality control benchmark vector are extracted. The square of the difference between corresponding elements of the demand variation coefficient and the quality control benchmark vector is calculated, multiplied by a preset dimension weighting coefficient, and then weighted and summed. The square root of the result is taken to obtain the weighted Euclidean distance. The weighted Euclidean distance is then substituted into an exponential decay function for nonlinear transformation, and a value between zero and one is output and determined as the reliable distance. The closer the reliable distance is to one, the closer the data distribution is to the standard operating conditions, and the higher the reliability level of the historical demand data. Then, the timestamps and demand transmission path identifiers corresponding to the historical demand data are extracted. A blank two-dimensional array is constructed with the demand transmission path as the row index and the time series timestamp as the column index. All reliable distances are traversed, and the values ​​are filled into the corresponding coordinate positions of the two-dimensional array according to their respective path identifiers and timestamps. The missing positions in the array are filled using linear interpolation to generate the quality control screening matrix.

[0037] It should be noted that in this application, the credible distance refers to a numerical metric that quantifies the degree to which the coefficient of variation of demand deviates from the quality control benchmark vector in order to assess the credibility level of historical demand data; the time series refers to an ordered set of observation data arranged in chronological order and recording the evolution of historical demand data of power grid materials over time; and the two-dimensional reconstruction refers to the spatial mapping operation of rearranging and combining one-dimensional distance values ​​according to the time and transmission path dimensions to construct a structured matrix feature.

[0038] In step S3, the demand iteration characteristics of power grid material demand are determined, and the demand iteration characteristics and the change state attributes of power grid operation scenario are reconstructed and evolved to obtain the demand response quantity when power grid material demand evolves. Based on the calibration driving information and the demand response quantity, the power grid material demand prediction model is updated hierarchically to obtain the updated power grid material demand prediction model.

[0039] In this embodiment, the demand iteration characteristics of power grid material demand can be determined by the following steps: Extract version change records of power grid material requirements in multiple planning batches, and calculate the demand evolution sequence of material categories between adjacent versions; The demand iteration characteristics of power grid material demand are determined based on the demand evolution state sequence.

[0040] In practice, the process begins by reading the material declaration ledgers for each batch and using a text difference comparison algorithm to extract category addition, deletion, and modification logs to generate version change records. When calculating the demand evolution state sequence for material categories between adjacent versions, the change rate of adjacent batches is calculated by extracting the declaration quantity of the same category. This change rate is then input into a Markov model, and combined with a set threshold, the model is categorized into increasing, stable, and decreasing states. These states are then sorted by time to generate a demand evolution state sequence. Next, the frequency of each state transition in the sequence is statistically analyzed to construct a state transition probability matrix. The self-transition probability of stable states is extracted from the matrix and weighted and fused with the information entropy of the demand evolution state sequence to calculate a convergence index. This convergence index is then encapsulated with the state transition probability matrix to determine it as the demand iteration characteristic of power grid material demand. This demand iteration characteristic ranges from zero to one; a larger value indicates that the demand planning is more stable and convergent.

[0041] It should be noted that, in this application, "multi-round planning batches" refers to a set of multiple consecutive and progressive plan application and approval time nodes divided according to time cycles in the power grid material procurement plan; "version change record" refers to the historical trajectory data of category quantity adjustments and attribute modifications generated during the circulation of material demand plans in different planning batches; "demand evolution state sequence" is an ordered set of state transition data characterizing the changing trends of the demand quantity and structure of specific material categories between adjacent planning batches; and "demand iteration characteristics" refers to the core attribute parameters of the dynamic evolution law and convergence trend of power grid material demand in multiple rounds of planning revisions.

[0042] In this embodiment, the demand response quantity when the demand iteration characteristics and the change state attributes of the power grid operation scenario are reconstructed and evolved can be obtained by the following steps: The state transition probability in the demand iteration characteristics and the change state attributes of the power grid operation scenario are subjected to tensor product operation to obtain the joint evolution tensor. The joint evolution tensor input time-series reconstruction network is used to perform multi-step state deduction, and the demand state offset vector under each deduction step is extracted to generate the material demand evolution response trajectory. The evolution response trajectories of the material demand are aggregated to calculate the demand response quantity when the power grid material demand evolves.

[0043] In specific implementation, firstly, state transition probabilities are extracted to construct one-dimensional row vectors, and transition state attributes are extracted to construct one-dimensional column vectors. The row and column vectors are multiplied using an outer product multiplication method to generate a two-dimensional cross-feature matrix. Nonlinear activation is performed on the matrix elements, redundant zero values ​​are removed, and the remaining dense matrix is ​​reshaped into a three-dimensional array to obtain the joint evolution tensor. Then, the joint evolution tensor is flattened and input into a temporal reconstruction network containing gated recurrent units. Three incrementally increasing time span windows are set as inference step sizes. The network controls historical memory retention through update and reset gates, and the implicit states at each step size are output autoregressively. The difference between the implicit state and the initial baseline state is calculated to obtain the demand state offset vector. All offset vectors are concatenated in chronological order to generate the material demand evolution response trajectory. Finally, the magnitude of each demand state offset vector in the trajectory is extracted, and it is substituted into the Softmax function to calculate the time decay weight. The weighted summation algorithm is used to aggregate all demand state offset vectors to obtain the comprehensive offset vector. The maximum element value in the comprehensive offset vector is extracted and multiplied by the average historical material consumption benchmark of the power grid to calculate the demand response quantity when the power grid material demand evolves.

[0044] It should be noted that, in this application, the transition state attribute is a set of feature parameters that quantify the dynamic changes in business scale and operating conditions of a power grid operation scenario within a specific period; reconstructive evolution refers to the fusion calculation operation that recombines multi-dimensional features in a unified high-dimensional space to simulate the dynamic development of things over time; power grid material demand evolution response refers to the adaptive adjustment and dynamic feedback process exhibited by the scale of material consumption in response to changes in the power grid operation scenario and planning iterations; state transition probability refers to the numerical index of the likelihood of a material demand category transitioning from its current state to its target state between adjacent planning batches; tensor product operation refers to the algebraic operation that expands two low-dimensional feature vectors into a high-dimensional matrix through outer product to preserve the cross-combination relationship of multi-dimensional information; joint evolution tensor refers to the fusion of demand iteration patterns and scenario changes. Features are high-order data structures that comprehensively characterize the multi-dimensional driving mechanism of material demand; temporal reconstruction network refers to a deep neural network that captures long-term dependencies in time series and reconstructs the implicit state evolution law based on gated recurrent units; multi-step state extrapolation refers to the sequence generation operation that sets different time span windows to make rolling predictions of the implicit states of multiple consecutive time nodes in the future; demand state offset vector refers to the numerical array of spatial distance and adjustment magnitude of the predicted value of material demand deviating from the baseline evolution direction under a specific extrapolation step size; material demand evolution response trajectory refers to an ordered spatiotemporal data set that records the state offset change process of material demand under multiple consecutive extrapolation time nodes in the future; demand response quantity refers to the absolute adjustment value of material consumption finally determined by the power grid material demand under the dual driving forces of responding to scenario changes and planning iterations.

[0045] Preferably, in this embodiment, the power grid material demand forecasting model is updated hierarchically based on the calibration driving information and the demand response quantity to obtain an updated power grid material demand forecasting model, with reference to... Figure 2 As shown in the figure, this is a flowchart illustrating the process of determining the updated power grid material demand forecasting model in some embodiments of this application. In this embodiment, determining the updated power grid material demand forecasting model can be achieved through the following steps: In step S31, a mask is constructed based on the calibration driving information and injected into the model feature embedding layer, and the underlying network parameters are corrected through gradient backpropagation. In step S32, the residual between the demand response quantity and the historical demand is calculated to construct a hierarchical loss function, which drives the model output layer to iteratively update the parameters of the underlying network. In step S33, the updated underlying network parameters are conditionally determined to form an updated power grid material demand prediction model.

[0046] In practice, firstly, the feature distribution offset in the calibration-driven information is extracted and mapped to mask weights between zero and one using the Sigmoid function. These mask weights are then element-wise multiplied with the input vector of the model's feature embedding layer to mask low-quality feature dimensions. The feature reconstruction error after mask injection is calculated, and the partial derivatives are calculated along the network topology using the backpropagation algorithm. The Adam optimizer is then used to update the underlying network parameters, thereby correcting the front-end feature extraction logic. Next, the absolute difference between the demand response and the historical true demand value is calculated as the residual. The squared term of the residual is weighted and summed with the regularization penalty term of the model output layer to construct a hierarchical loss function. This function is used to calculate the prediction error gradient of the model output layer. The gradient is then propagated back to the underlying network layer by layer using a chain rule. Stochastic gradient descent is used to iteratively update the underlying network parameters multiple times, forcing the network output to approximate the true demand response. Finally, the decrease in the hierarchical loss function under the current iteration round is calculated. If the decrease is less than the preset convergence threshold and the number of iterations reaches the maximum set number of rounds, the parameter update is determined to meet the termination condition, and the current bottom-level network parameters and model output layer weights are locked. If the condition is not met, the iteration is returned to continue. After the determination is passed, the network structure and parameters are encapsulated, and the updated power grid material demand prediction model is officially generated.

[0047] It should be noted that in this application, hierarchical update refers to a training strategy that uses differentiated driving signals to independently optimize parameters for different layers of the neural network to improve convergence accuracy; the model feature embedding layer refers to the neural network initial processing module that maps discrete material identifiers into continuous dense vectors to extract the implicit semantic features of the underlying layer; gradient backpropagation refers to the core neural network optimization mechanism that uses the chain rule to propagate the output error back along the network topology to calculate the partial derivatives of the parameters of each layer; the underlying network parameters refer to the set of weights and biases located at the front end of the network that are responsible for extracting the distribution features of basic data and directly affect the quality of the initial feature expression; the hierarchical loss function refers to a composite objective evaluation mathematical model that integrates multi-source driving signals to constrain the optimization direction of different layers of the network and balances the global error; the model output layer refers to the layer located at the end of the neural network that is responsible for decoding and mapping the high-dimensional implicit features extracted by the deep network into specific material consumption prediction values; condition judgment refers to the logical verification operation that evaluates the network parameter update status based on a preset convergence threshold and the number of iterations to determine whether to terminate the training; the updated power grid material demand prediction model refers to a material consumption inference algorithm structure with higher prediction accuracy and adaptive capability after hierarchical driving optimization with multi-source signals.

[0048] In addition, in this embodiment, constructing a mask based on the calibration driving information and injecting it into the model feature embedding layer, and correcting the underlying network parameters through gradient backpropagation can be achieved through the following steps: The feature weight distribution in the calibration driving information is analyzed to generate a sparse mask matrix and multiplied element by element into the feature embedding layer to mask redundant feature channels. The forward loss is calculated based on the output of the feature embedding layer after mask injection, and gradient backpropagation is performed along the unmasked feature channels to correct the parameters of the underlying network.

[0049] In practice, firstly, the importance scores of each dimension's features are extracted, and the feature weight distribution is obtained by normalization using the Softmax function. A weight truncation threshold is set, and weights below the threshold are set to zero while weights above the threshold are set to one, generating a sparse mask matrix. The sparse mask matrix is ​​then multiplied element-wise with the initial feature vector output from the feature embedding layer. This operation forces the corresponding values ​​to be cleared to zero, thus physically blocking the signal transmission of redundant feature channels and achieving precise pruning of feature dimensions. Next, the filtered feature vector is input into subsequent network layers to obtain predicted values. The mean squared error function is used to calculate the deviation between the predicted value and the actual required label to determine the forward loss. When performing gradient backpropagation along the unmasked feature channels, the chain rule is used to solve the partial derivative of the forward loss with respect to the underlying network parameters. Combined with the sparse mask matrix, the gradient is masked and filtered, forcing the gradient update amount of the masked channels to be zero. The Adam optimizer is then used to iteratively correct only the underlying network parameters of the unmasked channels.

[0050] It should be noted that in this application, the feature weight distribution is a numerical proportion structure of the contribution of each dimension of features in the quantification calibration driving information to the prediction of material demand; the sparse mask matrix refers to a binary filtering structure that contains zero-value elements to accurately block the transmission of specific invalid feature signals in the neural network; the feature embedding layer refers to the neural network front-end mapping module that transforms discrete material attributes into continuous vectors to extract the underlying implicit semantic features; the redundant feature channel refers to the invalid feature transmission path that carries no substantial predictive gain for the current power grid operation scenario and is prone to introducing noise interference; and the forward loss refers to a scalar evaluation value that measures the degree of deviation between the network prediction output value after mask filtering and the actual material demand label.

[0051] In step S4, the updated power grid material demand prediction model is adapted and verified based on the operational constraints of the multi-dimensional operation scenario, and the power grid material demand prediction results are output.

[0052] In this embodiment, the updated power grid material demand forecasting model is adapted and verified based on the operational constraints of the multi-dimensional operation scenario, and the power grid material demand forecasting results are output through the following steps: Construct a multidimensional constraint vector space based on the various operational constraints in the multidimensional operational scenario; The multidimensional constraint vector space is mapped to a decision boundary hyperplane, and the adaptation deviation gradient between the model output and the decision boundary hyperplane is calculated to generate an adaptation verification index sequence. When the adaptation verification index sequence meets the preset convergence condition, the demand feature mapping of the model output layer is decoded along the constraint gradient direction, and the power grid material demand prediction result is output.

[0053] In practical implementation, firstly, operational constraints such as budget, storage capacity, and transportation capacity are extracted and encoded into initial feature vectors. Principal component analysis (PCA) is then used to reduce the dimensionality of these initial feature vectors, eliminating collinear and redundant dimensions, and retaining principal components whose cumulative variance contribution rate meets a preset standard. The retained principal components are then concatenated column-wise to construct a multidimensional constraint vector space. Next, the basis vectors of the multidimensional constraint vector space are extracted, and the maximum margin classification surface is solved using a support vector machine (SVM) algorithm. A decision boundary hyperplane is then fitted and generated. When calculating the adaptation bias gradient, the vertical distance from the model's output feature points to the decision boundary hyperplane is calculated. Automatic differentiation is used to calculate the partial derivative of this distance with respect to the model's output layer parameters, yielding the adaptation bias gradient. The magnitude of the adaptation bias gradient is continuously recorded according to the validation time step, generating an adaptation validation index sequence. Finally, the average gradient magnitude of the last three consecutive time steps of the sequence is calculated. If the average is less than the set minimum threshold, the preset convergence condition is met. When decoding along the constraint gradient direction, the opposite direction of the adaptation deviation gradient is extracted as the constraint gradient direction. The hidden feature vector of the model output layer is projected along this direction to the interior of the decision boundary hyperplane. The projected demand feature mapping is decoded through the activation function of the fully connected layer to output the power grid material demand prediction result.

[0054] It should be noted that, in this application, operational constraints refer to the set of rules that limit the compliant execution of power grid material procurement and allocation operations within realistic boundaries such as funding and capacity; adaptation verification refers to the dynamic verification process of evaluating the ability of the prediction model output to continuously meet real-world business constraints under changing operational scenarios; multidimensional constraint vector space refers to the mathematical and geometric structure that transforms various discrete business rules into continuous high-dimensional coordinates to uniformly represent complex constraints; decision boundary hyperplane refers to the linear geometric segmentation surface that divides the compliant demand prediction region and the non-compliant prediction region in high-dimensional space; and adaptation deviation gradient refers to the degree to which the model's prediction output deviates from the compliant decision boundary and the fastest correction method. The vector derivative index is defined as follows: the adaptation verification index sequence refers to the ordered set of values ​​that continuously records the trend of the deviation between the model output and the compliance boundary according to the time step; the preset convergence condition refers to the threshold standard for terminating the verification when the model's predicted output has stably fallen within the compliance decision boundary range; the constraint gradient direction refers to the optimal parameter adjustment vector path that guides the model's predicted output to move closer to the interior of the compliance decision boundary hyperplane; the demand feature mapping refers to the nonlinear transformation relationship that transforms the high-dimensional implicit material demand characteristics into a specific executable procurement numerical space; the power grid material demand forecast result refers to the future consumption quantity of various power grid materials obtained after rigorous verification through multi-dimensional business constraints.

[0055] It should be noted that in this application, the operational constraints can be determined in the following way: By connecting to the financial, warehousing and supplier collaboration platforms through system interfaces, structured ledger data such as annual procurement budget, physical warehouse capacity and current inventory levels of materials at all levels, and supplier historical fulfillment rates and maximum production capacity are collected in real time. At the same time, named entity recognition algorithms in natural language processing technology are used to automatically extract business rule thresholds such as delivery cycle limits and payment ratios from unstructured material procurement management systems and project construction planning documents. Subsequently, the minimum-maximum normalization method is used to perform dimensionless processing on the above-mentioned multi-source heterogeneous numerical boundary parameters and rule thresholds to eliminate differences in data dimensions. Finally, the core parameters such as the processed budget upper limit, warehouse capacity remaining, production capacity upper limit and delivery date lower limit are vectorized and encapsulated to obtain the operational constraints under the multi-dimensional operation scenario.

[0056] Therefore, this application demonstrates that the updated power grid material demand prediction model can be adapted and verified based on the operational constraints of multi-dimensional operational scenarios, outputting power grid material demand prediction results. Specifically, determining the calibration-driven information allows for the acquisition of the directional guidance vector for model parameter correction and the gradient of the constraint boundary, effectively suppressing convergence oscillations during model iteration. This calibration-driven information deeply integrates multi-dimensional operational constraints with feature masks, providing a precise optimization direction for updating the underlying weights, avoiding gradient descent from falling into local optima or invalid search spaces. By quantifying the degree to which the predicted output deviates from the business compliance boundary, it accurately removes redundant feature noise from parameter adjustments, ensuring that each iteration strictly conforms to real-world material allocation rules, significantly improving the model's robustness and optimization efficiency in complex business scenarios. Determining the demand response quantity yields a baseline value for material demand forecasting that aligns with actual physical boundaries and operational rules. This provides a reliable quantitative reference for verifying the accuracy of the model's output. The demand response quantity maps theoretical forecasts to a multi-dimensional operational constraint vector space, effectively eliminating invalid forecast components that deviate from actual inventory capacity or budget limits. By constructing a decision boundary hyperplane, it can accurately measure the adaptation deviation between forecast results and actual business needs, thereby driving the model to perform targeted residual compensation and logical corrections, ensuring the feasibility of the final material demand forecast.

[0057] In summary, the technical solution adopted in this application can achieve accurate iterative verification of the power grid material demand forecasting model, thereby improving the accuracy of the power grid material demand forecasting results.

[0058] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0059] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0060] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A method for iterative optimization and verification of a power grid material demand forecasting model, characterized in that, The method includes: Collect multi-dimensional influencing factors of power grid material demand and determine the correlation and influence map between each influencing factor; The demand transmission path corresponding to the correlation influence map is determined, and the historical demand data of the power grid operating equipment is pre-controlled and screened based on each demand transmission path to obtain the calibration driving information of the power grid material demand prediction model. Determine the demand iteration characteristics of power grid material demand, reconstruct and evolve the demand iteration characteristics and the change state attributes of power grid operation scenario to obtain the demand response quantity when power grid material demand evolves, and perform hierarchical update on the power grid material demand prediction model based on the calibration driving information and the demand response quantity to obtain the updated power grid material demand prediction model. Based on the operational constraints of multi-dimensional operation scenarios, the updated power grid material demand forecasting model is adapted and verified, and the power grid material demand forecasting results are output.

2. The method for iterative optimization and verification of a power grid material demand forecasting model as described in claim 1, characterized in that, Determining the correlation and influence map among various influencing factors specifically includes: Based on the aforementioned multidimensional influencing factors, determine the topological characteristic sequence of the transmission influence between each influencing factor; Based on the aforementioned topological feature sequence, a community detection algorithm is used to divide the data, resulting in multiple communities affected by demand. The correlation and influence map between various influencing factors is determined based on the correlation weights of all demand effects within the community.

3. The method for iterative optimization and verification of a power grid material demand forecasting model as described in claim 1, characterized in that, Determining the demand transmission path corresponding to the aforementioned correlation influence map specifically includes: Extract directed transmission subgraphs from the aforementioned correlation influence graph; By searching layer by layer along the decreasing direction of the associated edge weights in the directed transmission subgraph, an effective demand transmission path is obtained; The demand transmission path corresponding to the correlation influence map is determined based on the node level depth of the effective demand transmission path.

4. The method for iterative optimization and verification of a power grid material demand forecasting model as described in claim 1, characterized in that, Based on the historical demand data of power grid operating equipment through various demand transmission paths, preliminary quality control and screening are performed to obtain calibration driving information for the power grid material demand forecasting model, specifically including: Historical demand data of power grid operating equipment are extracted along each demand transmission path, and the demand variation coefficient is determined; Based on the demand variation coefficient and the quality control benchmark vector for power grid material demand forecasting, the historical demand data is mapped to a reliability level to obtain a quality control screening matrix. The historical demand data is aggregated based on the quality control screening matrix to generate calibration driving information for the power grid material demand prediction model.

5. The method for iterative optimization and verification of a power grid material demand forecasting model as described in claim 4, characterized in that, Based on the demand variation coefficient and the quality control benchmark vector used for power grid material demand forecasting, a credibility mapping is performed on the historical demand data to obtain the quality control screening matrix, which specifically includes: Determine the reliable distance between the demand variation coefficient and the quality control benchmark vector; The trusted distance is reconstructed in two dimensions along the time series of the historical demand data to generate a quality control screening matrix.

6. The method for iterative optimization and verification of a power grid material demand forecasting model as described in claim 1, characterized in that, The specific characteristics of demand iteration for power grid materials include: Extract version change records of power grid material requirements in multiple planning batches, and calculate the demand evolution sequence of material categories between adjacent versions; The demand iteration characteristics of power grid material demand are determined based on the demand evolution state sequence.

7. The method for iterative optimization and verification of a power grid material demand forecasting model as described in claim 1, characterized in that, The demand response quantity is specifically derived by reconstructing and evolving the demand iteration characteristics and the changing state attributes of the power grid operation scenario to obtain the demand response quantity during the evolution of power grid material demand. The state transition probability in the demand iteration characteristics and the change state attributes of the power grid operation scenario are subjected to tensor product operation to obtain the joint evolution tensor. The joint evolution tensor input time-series reconstruction network is used to perform multi-step state deduction, and the demand state offset vector under each deduction step is extracted to generate the material demand evolution response trajectory. The evolution response trajectories of the material demand are aggregated to calculate the demand response quantity when the power grid material demand evolves.

8. The method for iterative optimization and verification of a power grid material demand forecasting model as described in claim 1, characterized in that, Based on the calibration driving information and the demand response quantity, a hierarchical update is performed on the power grid material demand forecasting model to obtain the updated power grid material demand forecasting model, which specifically includes: A mask is constructed based on the calibration driving information and injected into the model feature embedding layer. The underlying network parameters are corrected through gradient backpropagation. The residual between the demand response and historical demand is calculated to construct a hierarchical loss function, which drives the model output layer to iteratively update the parameters of the underlying network. Conditional determinations are made on the updated underlying network parameters to form an updated power grid material demand prediction model.

9. The method for iterative optimization and verification of a power grid material demand forecasting model as described in claim 8, characterized in that, Constructing a mask based on the calibration-driven information and injecting it into the model feature embedding layer, and correcting the underlying network parameters through gradient backpropagation specifically includes: The feature weight distribution in the calibration driving information is analyzed to generate a sparse mask matrix and multiplied element by element into the feature embedding layer to mask redundant feature channels. The forward loss is calculated based on the output of the feature embedding layer after mask injection, and gradient backpropagation is performed along the unmasked feature channels to correct the parameters of the underlying network.

10. The method for iterative optimization and verification of a power grid material demand forecasting model as described in claim 1, characterized in that, Based on the operational constraints of multi-dimensional operational scenarios, the updated power grid material demand forecasting model is adapted and validated, and the output power grid material demand forecasting results specifically include: Construct a multidimensional constraint vector space based on the various operational constraints in the multidimensional operational scenario; The multidimensional constraint vector space is mapped to a decision boundary hyperplane, and the adaptation deviation gradient between the model output and the decision boundary hyperplane is calculated to generate an adaptation verification index sequence. When the adaptation verification index sequence meets the preset convergence condition, the demand feature mapping of the model output layer is decoded along the constraint gradient direction, and the power grid material demand prediction result is output.

Images