A multi-source material demand prediction and replenishment decision method for power storage

Abstract

The application belongs to the technical field of electric power warehousing, and discloses a multi-source material demand prediction and replenishment decision method for electric power warehousing, wherein a cross-system dynamic federal query system is constructed, an OPC UA protocol and a RESTful API fusion gateway are adopted, seamless docking of multiple systems such as electric power warehousing, power grid monitoring and weather warning is realized, data is standardized into a four-dimensional structure of materials, time, scene and attributes, and then implicit space-time correlation is completed through a space-time enhanced knowledge graph to form a standardized fusion data set; a hybrid prediction model combining migration learning is constructed, small sample scenes are adapted through similar scene data migration, and differentiated sub-models are designed for regular consumption materials and emergency materials, combined with time series trend decomposition and attention mechanism to strengthen the influence of key scenes; meanwhile, an adaptive data cleaning module and an automatic feature extraction mechanism are adopted, abnormal data is removed through double verification, missing values are intelligently completed, and core features are dynamically screened.

Description

Technical Field

[0001] This invention belongs to the field of power storage technology, specifically a method for multi-source material demand forecasting and replenishment decision-making in power storage. Background Technology

[0002] With the accelerated construction of new power systems, power warehousing and material management faces multiple challenges, including a surge in the types of materials, drastic fluctuations in demand, and high requirements for emergency response. The accuracy of material demand forecasting and replenishment decisions directly affects the stability of power grid operation and the control of operating costs. Currently, the relevant technologies in the field of power warehousing and material management mainly have the following technical problems: The demand for power supplies is linked to multi-dimensional information such as power grid operation data, maintenance plans, meteorological data, and historical consumption data. However, existing technologies mostly process data from each system independently, lacking a cross-system retrieval and fusion mechanism based on a unified data model. This results in insufficient data value mining and an inability to provide comprehensive support for demand forecasting. For example, traditional methods do not effectively integrate real-time operation data from smart grid monitoring systems with inventory data from warehouse management systems, making it difficult to accurately capture emergency supply demands caused by equipment failures.

[0003] Existing forecasting methods mostly employ single time series models, failing to incorporate the strong correlation between power material demand (such as the correlation between equipment lifecycle and maintenance cycle), and are unable to dynamically adapt to sudden scenarios such as changes in power grid maintenance plans and extreme weather. Furthermore, the cleaning and feature extraction of model training data rely on manual operations, which are inefficient and prone to introducing subjective errors.

[0004] Replenishment strategies are mostly based on fixed safety stock thresholds, failing to fully utilize demand trends and confidence information in forecast results, and cannot be dynamically adjusted according to factors such as real-time inventory dynamics and supplier supply capacity. Summary of the Invention

[0005] The purpose of this invention is to provide a method for multi-source material demand forecasting and replenishment decision-making for power warehousing, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-source material demand forecasting and replenishment decision-making method for power warehousing, comprising the following specific steps: Preferably, the data acquisition and adaptation stage constructs a cross-system dynamic federated query system to realize the connection of multi-source heterogeneous systems such as power storage management system, smart grid monitoring system, equipment management system, and meteorological early warning system. It adopts a dynamic adaptation gateway that integrates OPCUA protocol and RESTful API, and automatically identifies the data types of different source systems and generates adaptation query statements through a preset system ontology association library. The protocol fusion gateway has a built-in protocol conversion and adaptation layer for OPC UA and RESTful API. The OPC UA protocol is used to collect real-time binary data from power grid monitoring systems and equipment management systems, while the RESTful API is used to call structured data from meteorological early warning systems and warehouse management systems. The conversion and adaptation layer realizes field alignment and data format conversion between the two protocols through a preset mapping table. The system ontology association library contains three core association rules: inter-system field mapping rules, data type conversion rules, and business logic association rules, such as the correspondence between equipment models and material specifications.

[0007] The design incorporates a permission-based federated query control mechanism to authenticate and classify user resource requests. Once authentication is successful, cache nodes are queried first. If no match is found, parallel queries are triggered in the distributed storage cluster. Furthermore, an access frequency assessment and system capacity awareness module is introduced to calculate the query request frequency and node load in real time. When a preset threshold is reached, data prefetching inference is triggered, and prefetched objects are inferred and cached in advance based on log files, ontology, and semantically labeled resources. Standardized transformation rules are designed for different types of data. Key information in unstructured data is extracted through natural language processing. Semantic clustering algorithm is used to complete the initial fusion of multi-source data. Finally, all data is uniformly encapsulated into a four-dimensional data structure of materials, time, scene, and attributes, and stored in the raw data layer of a distributed time-series database.

[0008] Preferably, the data fusion optimization stage is based on the raw data obtained in the data acquisition and adaptation stage, and constructs a core storage engine for the fusion of power storage material demand knowledge graph and time series data, expands the graph nodes into composite entities with spatiotemporal attributes, adds time evolution attributes to the edge relationships, and forms a spatiotemporally enhanced knowledge graph. The spatiotemporal attributes of composite entities include spatial dimensions such as power grid area codes and warehouse location coordinates, and temporal dimensions such as data collection timestamps and attribute validity periods. The temporal evolution attributes of edge relationships are stored in the form of a triplet of association start time, association end time, and association strength decay coefficient. The association strength decay coefficient ranges from 0.1 to 1.0 and decreases linearly with the increase of time interval. The graph rule reasoning engine has built-in five core reasoning rules: equipment and material matching rules, scenario and demand triggering rules, region and resource allocation rules, time and consumption association rules, and fault and emergency material matching rules.

[0009] The design of the graph-time series fusion storage model adopts a hybrid storage method of graph structure metadata and time series data blocks. The graph structure metadata uses a graph database-specific storage format to store entity relationships and attribute information. The time series data blocks are partitioned according to fixed time windows to store the dynamic attribute data of the corresponding entities. The graph rule reasoning engine is used to complete the implicit spatiotemporal relationships between multi-source data. Daily partitions store data with high real-time requirements, such as power grid load and equipment operating status, with a retention period of 30 days; weekly partitions store periodic data, such as total weekly material consumption and maintenance plan execution progress, with a retention period of 1 year; monthly partitions store trend data, such as total monthly demand and supplier fulfillment capability ratings, with long-term retention; partitions adopt a naming rule of timestamp + data type, supporting fast mounting and unmounting by time range, and the data size of a single partition is controlled within 10GB to ensure read and write efficiency.

[0010] The multi-condition combined retrieval algorithm is optimized to support multi-dimensional combined queries based on time window, material type, scene label, and spatiotemporal weight. Query redundancy is reduced by constructing a graph time-series joint index, and a fused standardized structured dataset is generated and stored in the deep processing layer of the database.

[0011] Preferably, the cleaning feature stage receives the standardized structured dataset output from the data fusion and optimization stage, constructs an adaptive data cleaning module based on native database functions, improves processing efficiency through partitioned parallel computing of distributed time-series databases, identifies outliers through user-defined functions, and adopts a dual verification method combining the isolated forest algorithm and graph database association verification. First, it quickly filters out extremely abnormal material consumption records through time-series database functions, and then verifies the business rationality of abnormal data by querying related equipment operation data and maintenance records through knowledge graph. The specific criteria for dual verification are as follows: If the abnormal data selected by the isolated forest algorithm is not supported by corresponding equipment failure records, maintenance plans or meteorological events in the knowledge graph, it is directly judged as invalid abnormal data and removed; if there is business data to support it but the data value exceeds three times the average of the same period in the past three years, it is marked as pending verification, and whether to retain it is determined in combination with the results of manual review.

[0012] We design an interpolation completion strategy based on graph temporal association. We use the entity association relationship of the knowledge graph to determine the set of similar entities of the same type and in the same scenario. We extract the historical temporal data of similar entities through the window function of the temporal database and use the weighted average method to complete the missing data. By introducing a database feature indexing mechanism, after extracting multi-dimensional core features, a feature data association table is constructed to associate features with the original data storage location and feature confidence. Feature weights are dynamically allocated through an attention mechanism to filter the core feature set that has a significant impact on the prediction results and generate indexed prediction model input data.

[0013] The weight allocation of the attention mechanism is based on the feature contribution. The feature contribution is quantified by the historical prediction error reduction rate. Features with a contribution of ≥30% are set as high weights with a weight coefficient of 0.7-1.0; features with a contribution of 10%-30% are set as medium weights with a weight coefficient of 0.3-0.7; and features with a contribution of less than 10% are set as low weights with a weight coefficient of 0.1-0.3. The dynamic adjustment trigger conditions include: new scene data access, a decrease in prediction accuracy of more than 5% for 3 consecutive rounds, and the addition or elimination of core features. During the adjustment, the proportion of high-weight features is kept at no less than 40%.

[0014] Preferably, the prediction modeling stage takes the indexed core feature data generated in the feature cleaning stage as input to construct a hybrid model that integrates time-series prediction and scene adaptation. It introduces a database-driven transfer learning mechanism to improve the prediction accuracy of small sample scenes. Based on the long short-term memory network as the basic model, it extracts time-series data fragments and association graph structures of historical similar scenes based on the dual criteria of scene label matching and time-series trend similarity calculation through the similarity retrieval module of the graph time-series fusion database. The retrieved similar data is used as the source domain data for transfer learning. The difference in data distribution between the source domain and the target domain is adjusted by the domain adaptation algorithm, and the model parameters trained in the source domain are transferred to the current prediction task. Differentiated sub-models are designed for the demand characteristics of different material types. For conventional consumable materials, an LSTM model combined with trend decomposition and time series database smoothing is used. The aggregation function of the time series database is used to denoise and smooth the prediction results to capture long-term demand trends. The time-series database smoothing adopts a moving average aggregation method. The window size is dynamically selected according to the material demand cycle. The window for routine materials with small daily consumption fluctuations, such as screws and gaskets, is set to 7 days. The window for materials with medium weekly consumption fluctuations, such as cable joints, is set to 14 days. The window for materials with large monthly consumption fluctuations, such as transformer parts, is set to 30 days. During the aggregation process, abnormal predicted values ​​that exceed twice the average value of the data in the window are automatically removed to ensure the rationality of the smoothed data.

[0015] The emergency supplies model uses LSTM combined with attention mechanism and scenario triggering. It queries the correlation strength of scenario triggering conditions in real time through graph database, dynamically adjusts attention weights, and strengthens the triggering impact of sudden scenarios such as extreme weather and equipment failure on demand. The model finally outputs material demand forecast results, confidence intervals and corresponding data source indexes at multiple time scales such as daily, weekly and monthly.

[0016] The confidence interval is derived based on the normal distribution characteristics of the model prediction error over the past 6 months. The upper limit of the 95% confidence interval is calculated as the prediction result plus 1.96 × the standard deviation of the error, and the lower limit is calculated as the prediction result minus 1.96 × the standard deviation of the error. When the prediction error exhibits a skewed distribution, the calculation method automatically switches to the 97.5 quantile (upper limit) and 2.5 quantile (lower limit) of the error distribution. The upper limit is taken from the 97.5 quantile of the error distribution, and the lower limit is taken from the 2.5 quantile, ensuring the accuracy of the confidence interval.

[0017] Preferably, the inventory threshold stage constructs a dynamic safety stock calculation model driven by database retrieval optimization based on the multi-scale prediction results and confidence intervals output from the prediction modeling stage, dividing the safety stock into two parts: basic safety stock and floating safety stock; the basic safety stock is calculated based on historical average demand data retrieved from the time-series database, the shortest supply cycle of suppliers, and warehouse capacity constraints. In the calculation of basic safety stock, historical average demand data accounts for 50% of the weight, the shortest delivery cycle of suppliers accounts for 30%, and the storage capacity constraint accounts for 20%. When the storage capacity is less than 1.2 times the historical average demand of the past 12 months, the storage capacity constraint weight is automatically increased to 40%. The popularity of data queries is quantified by combining the number of queries and the fluctuation value of query frequency for the demand data of the material in the past 7 days. When the number of queries is ≥10 times / day or the fluctuation value is ≥50%, it is judged as high popularity, and the corresponding floating inventory ratio increases by 20%-50%.

[0018] The floating safety stock is dynamically adjusted based on the prediction confidence level and data query popularity; the full supplier fulfillment capability data of the current materials is retrieved through the graph database, the supplier distance query efficiency is optimized by using spatial index, and the safety stock threshold is dynamically adjusted in combination with the supplier's current capacity load. A database query latency feedback mechanism is introduced to monitor the data retrieval response speed in real time. When the query latency exceeds a preset threshold, the time window parameters for inventory calculation are automatically adjusted to balance calculation accuracy and real-time performance.

[0019] Preferably, the replenishment decision stage is based on the dynamic safety stock threshold determined in the inventory threshold stage, and constructs a multi-objective replenishment decision model based on graph time-series correlation weights. The objective function includes three core objectives: minimizing the total cost of inventory holding and procurement, maximizing the stockout risk coverage, and minimizing the replenishment response time. The inputs are the dynamic safety stock threshold, current real-time inventory data, multi-scale predicted demand data, and supplier full-dimensional data. The optimal replenishment plan is solved by an improved particle swarm optimization algorithm. The rules for balancing multiple objectives are as follows: In emergency scenarios such as extreme weather and equipment failure, maximizing the coverage of stockout risk has the highest priority, with a weight of 0.5; minimizing replenishment response time is the second highest priority, with a weight of 0.3; and minimizing total cost has the lowest priority, with a weight of 0.2. In typical scenarios, minimizing total cost has the highest priority, with a weight of 0.4; maximizing stockout risk coverage is the second highest priority, with a weight of 0.35; and minimizing replenishment response time has the lowest priority, with a weight of 0.25. The standard structure of the replenishment decision list is as follows: material code, material name, current inventory, dynamic safety stock, replenishment quantity, priority level, recommended supplier code, estimated arrival time, forecast data index, supplier data index, and scenario association graph identifier. The list supports export formats of Excel and JSON.

[0020] The supplier's full-dimensional data includes six core dimensions: unit price, delivery cycle, historical fulfillment rate, quality pass rate, geographical distance, and emergency response capability. Emergency response capability refers to the ability to ensure supply under extreme scenarios such as rainstorms and major equipment failures, with the delivery rate within 48 hours as the core indicator. The combined ranking of time-series association weight and entity association weight adopts a weighted calculation method of "0.4 × time-series weight + 0.6 × entity weight", with the weight value ranging from 0 to 1. They are arranged in descending order of weight, with the top 30% of replenishment tasks assigned the highest priority, 30%-70% assigned to medium priority, and the remainder assigned to low priority.

[0021] Using the association query function of the graph time-series fusion database, the time-series association weight and entity association weight of each replenishment task are calculated, and the replenishment tasks are prioritized based on the weight comprehensive sorting. The time-series correlation weight includes three calculation dimensions: the remaining time before the maintenance or emergency event (the shorter the remaining time, the higher the weight), the replenishment response time of the same period in history (the better the timeliness, the higher the weight), and the time granularity of demand forecast (daily forecasts have higher weight than weekly and monthly forecasts). The entity association weight includes four calculation dimensions: the priority of materials in supporting core power grid equipment, the risk level of outages caused by material shortages, the geographical distance between suppliers and warehouses, and the supplier's historical performance rate. The weight ratio of each dimension is dynamically calibrated through business scenario adaptation coefficients, with a calibration cycle of one month.

[0022] Finally, a structured replenishment decision list containing data traceability information is generated. Each decision in the list is associated with a corresponding predictive data index, supplier data index, and scenario association map, supporting full-link traceability of the decision-making process. At the same time, the actual data generated during the replenishment execution process is fed back to the parameter optimization stage.

[0023] Preferably, the parameter optimization stage receives execution feedback data from the replenishment decision stage, constructs a dual closed-loop feedback mechanism for collaborative optimization of database storage parameters and prediction model parameters, collects the execution results of replenishment decisions in real time and feeds them back to the application layer of the distributed time series database, and establishes a feedback data storage partition. Regularly evaluate the model dimension and database dimension based on feedback data. The model dimension evaluates the accuracy of the prediction model. If the prediction error exceeds the preset threshold, start the model retraining process, adjust the feature weights and source domain data screening parameters for transfer learning, and extract the retraining data from the graph time series fusion database through indexing. The model retraining data should prioritize core feature data from the past 12 months, including at least 3 historical scene data that are similar to the current scene (scene label matching degree ≥80%, time series trend similarity ≥75%), and the data integrity should be ≥95%, that is, the proportion of missing values ​​is <5%. The edge relationship weight calibration of the knowledge graph is based on the actual matching success rate of the feedback data. When the success rate is ≥90%, the edge weight is increased by 0.1. When the success rate is 70%-89%, the weight remains unchanged. When the success rate is <70%, the weight is decreased by 0.1. After the weight adjustment, it is always maintained in the range of 0.1-1.0.

[0024] Analyze the query frequency, storage heat distribution, read and write latency, and other indicators of replenishment data from the database dimension to optimize the database partitioning strategy, index structure and caching mechanism. Optimize the cache replacement strategy for frequently accessed replenishment decision data and dynamically adjust the partition boundaries and number of replicas for frequently written time-series data. Based on feedback data, the node attribute information and edge relationship weights of the knowledge graph are updated, and the temporal correlation and association rules of new scenarios and material demand are added to improve data retrieval efficiency and the adaptability of the prediction model. The processing effect of all the aforementioned stages, such as data collection and adaptation stage, data fusion optimization stage, feature cleaning stage, prediction modeling stage, inventory threshold stage, and replenishment decision stage, is optimized in reverse.

[0025] The beneficial effects of this invention are as follows: 1. This invention constructs a cross-system dynamic federated query system, adopts the OPC UA protocol and RESTful API fusion gateway to achieve seamless integration of multiple systems such as power warehousing, power grid monitoring, and meteorological early warning. It standardizes data into a four-dimensional structure of materials, time, scenario, and attributes, and then completes the implicit spatiotemporal correlations through a spatiotemporal enhanced knowledge graph to form a standardized fusion dataset. This breaks down data silos, fully explores the correlation value of multi-dimensional data, and avoids the problem of one-sided data support in traditional methods.

[0026] 2. This invention constructs a hybrid prediction model that integrates transfer learning. It adapts to small-sample scenarios through data transfer from similar scenarios, designs differentiated sub-models for routine consumables and emergency supplies, and strengthens the impact of key scenarios by combining time-series trend decomposition and attention mechanisms. At the same time, it adopts an adaptive data cleaning module and an automated feature extraction mechanism to remove abnormal data through double verification, intelligently fill in missing values, and dynamically select core features. This reduces subjective errors caused by manual operation, covers routine and emergency scenarios, and outputs prediction results with confidence at multiple time scales, thereby improving the accuracy and reliability of prediction.

[0027] 3. This invention constructs a dynamic safety stock model that combines basic and floating inventory levels. It dynamically adjusts thresholds based on historical demand, supplier fulfillment capabilities, and data query popularity, and introduces a query delay feedback mechanism to balance accuracy and real-time performance. Furthermore, through a multi-objective replenishment decision model, it comprehensively minimizes total cost, maximizes stockout risk coverage, and minimizes response time. It prioritizes replenishment based on correlation weights and generates a decision list that can be traced throughout the entire supply chain. This enables the replenishment strategy to accurately adapt to demand fluctuations, ensuring sufficient materials for power grid operation while effectively controlling inventory holding costs, and improving replenishment response speed and the scientific nature of decision-making. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the data acquisition adaptation process for this invention. Figure 3 This is a flowchart of the data fusion optimization process of the present invention; Figure 4 This is a flowchart of the predictive modeling process of the present invention; Figure 5 This is a flowchart illustrating the replenishment decision-making process of this invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] like Figures 1 to 5 As shown, this embodiment of the invention provides a method for multi-source material demand forecasting and replenishment decision-making for power warehousing, including the following specific steps: The data acquisition and adaptation phase involves constructing a cross-system dynamic federated query system to connect heterogeneous systems from multiple sources, such as the power storage management system, smart grid monitoring system, equipment management system, and meteorological early warning system. It employs a dynamic adaptation gateway that integrates the OPCUA protocol and RESTful API, automatically identifying data types from different source systems and generating adaptation query statements through a pre-set system ontology association library. The system ontology association library contains data dictionaries and field association rules for each system; data types include structured data, semi-structured data, and unstructured data.

[0031] Design a permission-based federated query control mechanism to authenticate and classify user resource requests. If the authentication is successful, the cache node will be queried first. If the cache node is not found, the parallel query of the distributed storage cluster will be triggered. The access control is divided into three levels: administrator, operation and maintenance, and query. The administrator level can query all system data, the operation and maintenance level can query the warehouse and equipment data of its own region, and the query level can only query public scene and non-sensitive material data. Identity verification uses a dual authentication method of account password and role token. The token is valid for 24 hours and will be automatically refreshed after expiration. The similarity judgment of the semantic clustering algorithm is based on two dimensions: field matching degree and business logic relevance. Field matching degree ≥80% and business logic relevance ≥70% are judged to be in the same cluster, realizing accurate fusion of multi-source data.

[0032] Furthermore, an access frequency assessment and system capacity perception module is introduced to calculate the query request frequency and node load in real time. When the preset threshold is reached, data prefetching inference is triggered. Based on log files, ontology and semantic annotation resources, prefetched objects are inferred and cached in advance, thereby improving data access efficiency. The core metric for access frequency assessment is the number of queries per unit time, with a default statistical window of 1 minute and a preset threshold of 50 queries per minute. The core metrics for system capacity perception include CPU utilization, memory usage, and network bandwidth utilization. The thresholds for CPU utilization are 80%, memory usage is 85%, and network bandwidth utilization is 90%. If any metric reaches its threshold, the system is considered to be approaching its capacity limit, triggering data prefetching inference.

[0033] Standardized transformation rules are designed for different types of data. Key information such as equipment failure type and maintenance time is extracted from unstructured data through natural language processing. Semantic clustering algorithm is used to complete the initial fusion of multi-source data. Finally, all data is uniformly encapsulated into a four-dimensional data structure of materials, time, scenario and attribute, and stored in the raw data layer of the distributed time series database.

[0034] The raw data layer adopts a storage structure that divides data into databases by material type and tables by time. Each material type corresponds to an independent database shard, and the time-based tables are divided by natural month. Data backup adopts a strategy of local real-time backup and off-site scheduled backup. Local backup is performed once an hour, and off-site backup is synchronized every day at midnight to ensure data security. The data retention period is set according to the data type. Real-time running data is retained for 90 days, historical consumption data is retained for 3 years, and scenario-related data is retained long-term.

[0035] The data fusion optimization stage, based on the raw data acquired in the data acquisition and adaptation stage, constructs a core storage engine that integrates the knowledge graph of power storage material demand with time-series data. The graph nodes are expanded into composite entities containing spatiotemporal attributes, including material type, equipment model, power grid area, meteorological events, maintenance items, etc. The edge relationships are given time evolution attributes, such as the temporal correlation strength and duration of the correlation between high temperature weather and transformer overload, forming a spatiotemporally enhanced knowledge graph.

[0036] The design of the graph-time-series fusion storage model adopts a hybrid storage method of graph structure metadata and time-series data blocks. The graph structure metadata uses a graph database-specific storage format to store entity relationships and attribute information. The time-series data blocks are partitioned into fixed time windows such as daily and weekly to store the dynamic attribute data of the corresponding entities, such as material consumption rate, equipment operating parameters, and power grid load data. This model realizes the associated storage and efficient retrieval of fused data. It uses the graph rule reasoning engine to complete the implicit spatiotemporal relationships between multi-source data, such as the time-series association chain between the rainy season and transmission line faults and the surge in demand for insulators.

[0037] The triggering conditions for graph rule reasoning include: new data access volume reaching 1000 records / time, existing relationship matching error exceeding 20%, and the addition of new scene tags; The reasoning execution process is rule matching → candidate association generation → business rationality verification → association relationship solidification. Candidate associations that fail verification will be marked as pending review and trigger manual review. The retrieval process of the graph time-series composite index is as follows: first-level entity identifier matching → second-level time window filtering → third-level scene tag precise positioning. During the retrieval process, high-frequency access index nodes are matched first. If no match is found, the index is traversed layer by layer downwards to ensure retrieval efficiency.

[0038] The inference engine executes the following steps: First, it extracts the spatiotemporal attributes and existing relationships of the entities to be inferred from the graph time-series fusion storage model; then, it matches the built-in 5 types of core inference rules to generate a set of candidate relationships; then, it verifies the temporal consistency of the candidate relationships through time-series data blocks, for example, the time difference between the time of the scene occurrence and the time of the material demand is ≤72 hours; finally, it solidifies the verified relationships into the knowledge graph, while recording the reasoning basis and timestamp.

[0039] The multi-condition combined retrieval algorithm is optimized to support multi-dimensional combined queries based on time window, material type, scene label, and spatiotemporal weight. Query redundancy is reduced by constructing a graph time-series joint index, and a fused standardized structured dataset is generated and stored in the deep processing layer of the database.

[0040] The retrieval algorithm executes the following logic: first, coarse screening is performed based on time window and material type to quickly narrow down the data range; then, it is filtered by scene tags to retain scene-related data with a matching degree of ≥80%; finally, it is sorted by spatiotemporal weight and results with a weight value of ≥0.6 are returned first; the retrieval results return the first 100 core data by default, and users can expand the upper limit of the number of queries as needed.

[0041] The graph time-series composite index adopts a three-level index structure: entity identifier, time window, and scene label. The first-level index consists of core entity identifiers such as materials, equipment, and regions; the second-level index consists of daily / weekly / monthly time windows; and the third-level index consists of scene labels such as weather, maintenance, and faults. The index is updated using an incremental update mechanism, with newly added data synchronized to the corresponding index nodes in real time. Index fragmentation is performed every 24 hours to ensure that the retrieval response time is ≤500ms.

[0042] The cleaning feature stage takes over the standardized structured dataset output from the data fusion and optimization stage, constructs an adaptive data cleaning module based on native database functions, improves processing efficiency through partitioned parallel computing of a distributed time-series database, identifies outliers through user-defined functions, and adopts a dual verification method combining the isolated forest algorithm and graph database association verification. First, it quickly filters out extremely abnormal material consumption records through time-series database functions, and then verifies the business rationality of abnormal data by querying related equipment operation data and maintenance records through a knowledge graph, such as excluding erroneously entered fault data and duplicate statistical data.

[0043] The partitioned parallel computing allocates computing tasks based on both the data time range and the material type. Each partition corresponds to an independent computing node, and the task progress is synchronized between nodes through a message queue. The task allocation rules are as follows: 1 computing node is allocated when the data volume of a single partition is ≤5GB, 2 computing nodes are allocated when it is 5-10GB, and 3 or more computing nodes are allocated when it is >10GB, so as to avoid excessive load on a single node. At the same time, a task conflict detection mechanism is set to ensure that the same data block is not processed repeatedly.

[0044] The isolated forest algorithm is configured with 100 base decision trees, the outlier threshold is set to 0.8, and samples with an outlier score ≥ 0.8 are considered outlier data. The sample size is 500 samples per batch, which is suitable for the sparsity characteristics of power material consumption data.

[0045] The outlier identification logic of user-defined functions is based on two dimensions: statistical thresholds and business rules. The statistical threshold adopts the 3σ principle, and data exceeding the mean ± 3 times the standard deviation is judged as suspected anomalies. The business rules include three core rules: the quantity of material consumption cannot be negative, the single consumption cannot exceed twice the historical maximum consumption value, and the consumption value cannot exceed 1.5 times the mean when there is no corresponding scenario record. Suspected abnormal data will trigger subsequent double verification.

[0046] To address missing data, an interpolation completion strategy based on graph temporal association is designed. The entity association relationship of the knowledge graph is used to determine the set of similar entities of the same type and in the same scenario. The historical temporal data of similar entities is extracted through the window function of the temporal database. The weighted average method is used to complete the missing data. The weights of the weighted average method are dynamically allocated according to the entity similarity. Entity similarity is calculated by combining core attribute matching degree, scene consistency, and time series data correlation. Core attributes include material type, equipment model, and application scenario. Scene consistency is determined by the overlap of scene labels. Time series data correlation is based on the Pearson correlation coefficient of the consumption trend over the past 3 months. Feature confidence is calculated based on the stability of the feature's contribution in historical predictions. The average value of the feature's contribution in the last 10 prediction rounds is taken, ranging from 0 to 1. Features with an average value ≥ 0.6 are considered to have high confidence.

[0047] By introducing a database feature indexing mechanism, multi-dimensional core features are extracted and a feature data association table is constructed. Time features include seasons, holidays, and peak maintenance periods; equipment features include remaining service life and maintenance cycle; and scenario features include weather level and power grid load level. Features are associated with the original data storage location and feature confidence level. Feature weights are dynamically allocated through an attention mechanism to select the core feature set that has a significant impact on the prediction results. This generates indexed prediction model input data, enabling end-to-end traceability and rapid retrieval of features.

[0048] The input data is a 32-dimensional standardized feature vector, with 8 dimensions of time features, including seasons, holidays, and peak maintenance periods; 10 dimensions of equipment features, including remaining lifespan and maintenance cycle; 7 dimensions of scene features, including weather level and power grid load; and 7 dimensions of attribute features, including material specifications and storage conditions. The feature values ​​are normalized to the range [0, 1]. The index identifier consists of system ID, data timestamp, and feature type code.

[0049] The predictive modeling stage takes the indexed core feature data generated in the feature cleaning stage as input to construct a hybrid model that integrates time-series prediction and scene adaptation. It introduces a database-driven transfer learning mechanism to improve the prediction accuracy of small sample scenes. Based on the Long Short-Term Memory Network (LSTM) as the basic model, it uses the similarity retrieval module of the graph time-series fusion database to extract time-series data fragments and association graph structures of historical similar scenes based on the dual criteria of scene label matching and time-series trend similarity calculation. For example, the material consumption time series and equipment association graph of the same season and the same type of maintenance project. The LSTM model contains 3 hidden layers, each with 128 hidden units. The activation function is ReLU, and the output layer uses a linear activation function. The dropout rate is set to 0.2 to suppress overfitting. The input sequence length is fixed at 30 days to adapt to the weekly and monthly demand patterns of power materials.

[0050] The calculation of time series trend similarity focuses on three core dimensions: data mean deviation ≤15%, variance deviation ≤20%, and monthly trend overlap ≥75%. Meeting two or more of these dimensions is considered as time series trend similarity. The confidence level is set at 95% by default. The upper and lower limits of the interval are determined based on the prediction error distribution of the model over the past 6 months. The larger the prediction error, the wider the confidence interval. The interval boundary value is derived by the calculation logic of "prediction result ± standard deviation of error × 1.96".

[0051] The retrieved similar data is used as the source domain data for transfer learning. The difference in data distribution between the source domain and the target domain is adjusted by a domain adaptation algorithm (such as the maximum mean difference algorithm) to reduce the impact of domain offset. The model parameters trained in the source domain are then transferred to the current prediction task to solve the problem of insufficient model training data for small sample scenarios such as prediction of demand for new materials and prediction of materials in rare maintenance scenarios. The maximum mean difference algorithm uses a Gaussian kernel function with a kernel width σ=0.5. During the transfer process, the training weight ratio of the source domain data and the target domain data is 0.7:0.3 to ensure model adaptability in small sample scenarios. The source domain data refers to historical similar scenarios, and the target domain data refers to the current scenario.

[0052] The model training uses the Adam optimizer with an initial learning rate of 0.001, which decays to 0.9 times the previous cycle every 10 iterations. The loss function is mean squared error, with 100 iterations and a batch size of 32. An early stopping mechanism is triggered when the loss on the validation set does not decrease for 5 consecutive cycles. The training data is divided into training, validation, and test sets in a 7:2:1 ratio. The test set uses actual power consumption data from the past 6 months.

[0053] Differentiated sub-models are designed for the demand characteristics of different material types. For conventional consumable materials, an LSTM model combined with trend decomposition and time series database smoothing is adopted. The aggregation function of the time series database is used to denoise and smooth the prediction results to capture long-term demand trends. Trend decomposition breaks down the original demand data into three components: trend, periodic, and random. The trend is extracted using the moving average function of the time series database, the periodic is extracted based on Fourier transform to extract fixed periodic components such as quarters and months, and the random is obtained through residual calculation. The scenario trigger level standard is clearly defined as follows: when the meteorological warning level is ≥4 (rainstorm, typhoon, etc.), the equipment failure level is ≥2 (core equipment shutdown, major defects), and the maintenance level is A (network-wide maintenance), the attention weight adjustment of the emergency material model will be triggered.

[0054] The emergency supplies model uses LSTM combined with attention mechanism and scenario triggering. It queries the correlation strength of scenario triggering conditions in real time through graph database, dynamically adjusts attention weights, and strengthens the triggering impact of sudden scenarios such as extreme weather and equipment failure on demand. The model finally outputs material demand forecast results, confidence intervals and corresponding data source indexes at multiple time scales such as daily, weekly and monthly.

[0055] Attention weight adjustment rules: When the correlation strength between the scene and the material is retrieved through the graph database, the weight coefficient is increased to 1.5 when the correlation strength of extreme weather such as rainstorms and typhoons with a meteorological warning level of ≥4 is ≥0.6; when the correlation strength of equipment faults such as faults reported by the power grid monitoring system with a fault level of ≥2 is ≥0.5; the weight coefficient of regular scenes is 1.0 by default.

[0056] The inventory threshold stage is based on the multi-scale prediction results and confidence intervals output by the prediction modeling stage to construct a dynamic safety stock calculation model driven by database retrieval optimization, which divides the safety stock into two parts: basic safety stock and floating safety stock. The basic safety stock is calculated based on historical average demand data retrieved from the time-series database, the shortest supply cycle of suppliers, and warehouse capacity constraints, thereby ensuring inventory guarantee capability under normal scenarios. Basic safety stock adjustment logic: When the storage capacity is ≤ historical average demand × 1.0 times, the basic safety stock is directly taken as the upper limit of the storage capacity; when the supplier's shortest delivery cycle is extended by more than 50% of the original cycle, the basic safety stock is recalculated according to the extended delivery cycle; the supplier's capacity load calculation standard is: load rate = current undelivered order quantity / maximum capacity × 100%, and the average capacity utilization rate in the past 30 days = the arithmetic average of daily actual capacity / designed capacity.

[0057] The floating safety stock is dynamically adjusted based on the prediction confidence level and data query popularity. The lower the confidence level and the higher the query popularity, the higher the proportion of floating stock, in order to cope with the risk of demand fluctuations. The full supplier fulfillment capability data of the current materials is retrieved through the graph database, including historical supply delay time, capacity reserves, quality pass rate, and geographical distribution information. Spatial indexes are used to optimize the efficiency of supplier distance queries, and the safety stock threshold is dynamically adjusted in combination with the current capacity load of suppliers. The evaluation indicators for supplier capacity load include the current order volume as a percentage of maximum capacity, the average capacity utilization rate over the past 30 days, and the order volume to be delivered in the next month. A load rate of ≥80% is considered high load, at which point the safety stock threshold is increased by 10%-20%. The preset threshold for database query latency is set to 800ms. When the response time for three consecutive queries exceeds 800ms, the time window for inventory calculation is automatically adjusted from 30 days to 15 days to balance real-time performance and calculation accuracy.

[0058] A database query latency feedback mechanism is introduced to monitor the data retrieval response speed in real time. When the query latency exceeds a preset threshold, the time window parameters for inventory calculation are automatically adjusted to balance calculation accuracy and real-time performance, ensuring the rationality, real-time performance, and feasibility of the safety stock threshold.

[0059] The replenishment decision stage is based on the dynamic safety stock threshold determined in the inventory threshold stage. It constructs a multi-objective replenishment decision model based on graph time-series correlation weights. The objective function includes three core objectives: minimizing the total cost of inventory holding and procurement, maximizing the stockout risk coverage, and minimizing the replenishment response time. The inputs are the dynamic safety stock threshold, current real-time inventory data, multi-scale predicted demand data, and supplier full-dimensional data. The optimal replenishment plan is solved by an improved particle swarm optimization algorithm (introducing a constraint handling mechanism).

[0060] The improved particle swarm optimization algorithm has a population size of 50, a maximum number of iterations of 80, an initial inertia weight of 0.9 that decreases linearly to 0.4, and learning factors of c1=1.5 and c2=1.7. The constraint handling mechanism adopts the penalty function method, which applies a penalty coefficient of 3.0 to solutions that exceed the warehouse capacity and supplier capacity limits to ensure the feasibility of the replenishment plan.

[0061] Using the association query function of the graph time-series fusion database, the time-series association weight and entity association weight of each replenishment task are calculated. The time-series association weight includes the urgency weight of material demand before the peak maintenance period and the replenishment timeliness weight during holidays; the entity association weight includes the matching weight of emergency materials and key equipment and the impact weight of material shortages on power grid operation; based on the comprehensive ranking of the weights, the replenishment tasks are prioritized, with emergency materials taking priority over key equipment matching materials, and key equipment matching materials taking priority over regular consumable materials.

[0062] Finally, a structured replenishment decision list containing data traceability information is generated. Each decision in the list is associated with a corresponding predictive data index, supplier data index, and scenario association map, supporting full-link traceability of the decision-making process and providing operational basis for replenishment execution. At the same time, the actual data generated during the replenishment execution process is fed back to the parameter optimization stage.

[0063] The parameter optimization stage receives execution feedback data from the replenishment decision stage, constructs a dual closed-loop feedback mechanism for collaborative optimization of database storage parameters and prediction model parameters, collects the execution results of replenishment decisions in real time and feeds them back to the application layer of the distributed time-series database, and establishes a feedback data storage partition; the execution results of replenishment decisions include actual stockout situation, replenishment cost details, supplier fulfillment status, and inventory turnover efficiency data.

[0064] The evaluation cycle for the dual closed-loop feedback mechanism is set at the weekly (model dimension) and monthly (database dimension). The specific conditions for triggering retraining in model dimensions are: mean absolute error > 15%, confidence interval coverage < 80%, and prediction bias showing an increasing trend for two consecutive weeks. Database dimension optimization is triggered by the following conditions: query latency > 1 second, cache hit rate < 60%, and storage capacity utilization > 85%. The knowledge graph update is triggered by the following conditions: adding ≥5 new material types, adding ≥10 new scene associations, and the actual matching error of existing associations being >20%.

[0065] Regularly evaluate the model and database dimensions based on feedback data. The model dimension requires evaluating the accuracy metrics of the prediction model, such as mean absolute error and root mean square error. If the prediction error exceeds the preset threshold, start the model retraining process, adjust the feature weights and source domain data selection parameters for transfer learning, and extract the retraining data from the graph time series fusion database through indexing to ensure data quality. The preset threshold for model accuracy is a mean absolute error > 12% or a root mean square error > 15%. The knowledge graph update process is as follows: feedback data collection → candidate generation of associations → historical data verification → weight calibration → update and solidification. New associations must meet the requirement that the actual matching success rate in the past 3 months is ≥ 85% before they can be solidified. Candidates that do not meet the standard will enter the observation period and will be continuously tracked for 3 rounds of evaluation before it is determined whether to update them.

[0066] At the database level, it is necessary to analyze indicators such as query frequency, storage heat distribution, and read / write latency of replenishment data, optimize the database partitioning strategy, index structure, and caching mechanism, and optimize the cache replacement strategy for frequently accessed replenishment decision data, such as using the LRUK algorithm to dynamically adjust partition boundaries and the number of replicas for frequently written time-series data, thereby avoiding node write skew and local hotspot phenomena. In the LRUK algorithm, the K value is set to 3, which means that high-frequency data accessed in the last 3 times is preferred to be retained. The cache replacement trigger condition is that the cache capacity utilization rate reaches 90%. The replacement priority is sorted in descending order of access frequency and in order of least recent access. Cache data that has been accessed less than 5 times in the last 30 days and has not been accessed in the last 7 days is first eliminated. At the same time, a cache resident flag is set for the replenishment decision data of core emergency supplies, which does not participate in the regular replacement process.

[0067] Based on feedback data, the node attribute information and edge relationship weights of the knowledge graph are updated, and the temporal correlation and association rules of new scenarios and material demand are added to improve data retrieval efficiency and the adaptability of the prediction model. The processing effect of all the aforementioned stages, such as data collection and adaptation stage, data fusion optimization stage, feature cleaning stage, prediction modeling stage, inventory threshold stage, and replenishment decision stage, is optimized in reverse.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for multi-source material demand forecasting and replenishment decision-making in power warehousing, characterized in that, The specific steps include the following: Data acquisition and adaptation phase: Connect to and collect data from multi-source heterogeneous systems including power storage and power grid monitoring. After data type identification, standardization conversion and preliminary fusion processing, the data is stored in a distributed time-series database. Data fusion and optimization stage: Based on the collected raw data, a storage engine for knowledge graph and time-series data fusion is built to form a spatiotemporally enhanced knowledge graph. A hybrid storage architecture of graph-structured metadata and time-series data blocks is adopted to optimize retrieval efficiency, complete the implicit spatiotemporal associations between multi-source data and generate a standardized structured dataset. Feature cleaning stage: Taking over the fused dataset, we complete the data noise removal, missing data completion, extract multi-dimensional core features and establish the relationship between features and original data, and select key features to generate indexed prediction model input data; Predictive modeling stage: Construct multi-scenario differentiated demand prediction models that integrate transfer learning, adapt model parameters to small sample scenarios by retrieving historical similar scenario data, design dedicated prediction sub-models for different types of material demand characteristics, and output material demand prediction results and confidence intervals. Inventory threshold stage: A dynamic safety stock calculation model is constructed based on demand forecast results and confidence intervals. The safety stock is divided into a base part and a floating part. The inventory threshold is dynamically adjusted by combining historical demand data and supplier fulfillment capability data. A database query delay feedback mechanism is introduced to balance calculation accuracy and real-time performance. Replenishment decision stage: Based on the dynamic safety stock threshold, a multi-objective replenishment decision model is constructed by combining real-time inventory, forecasted demand and supplier full-dimensional data. The correlation weight of each replenishment task is calculated and the priority is divided to generate a replenishment decision list. Parameter optimization stage: Establish a two-dimensional evaluation and optimization mechanism for the prediction model and the database. Adjust core parameters through model retraining, optimize database storage strategy, update entity attributes and relationships in the knowledge graph, and reverse optimize the processing effect of the previous stage.

2. The multi-source material demand forecasting and replenishment decision-making method for power warehousing according to claim 1, characterized in that, The data acquisition and adaptation phase specifically constructs a cross-system dynamic federated query system to achieve the connection between the power storage management system, smart grid monitoring system, equipment management system, and meteorological early warning system. It adopts a dynamic adaptation gateway that integrates OPC UA protocol and RESTful API, and automatically identifies the data types of different source systems and generates adaptation query statements through a preset system ontology association library. Design a permission-based federated query control mechanism to complete the authentication and classification management of user resource requests, prioritize querying cache nodes, and trigger parallel query of distributed storage cluster if no match is found. Introduce access frequency evaluation and system capacity awareness modules to trigger data prefetching inference and cache in advance when a preset threshold is reached. Standardized transformation rules for different types of data were formulated, key information of unstructured data was extracted and preliminary fusion of multi-source data was completed, and all data was uniformly encapsulated into a four-dimensional data structure of materials, time, scenario and attributes, and stored in the raw data layer of a distributed time series database.

3. The multi-source material demand forecasting and replenishment decision-making method for power warehousing according to claim 2, characterized in that, The data fusion and optimization stage is based on the construction of a core storage engine that integrates the knowledge graph of power storage material demand with time-series data, based on the original data. The graph nodes are expanded into composite entities with spatiotemporal attributes, and the edge relationships are given time evolution attributes to form a spatiotemporally enhanced knowledge graph. The design of the graph-time series fusion storage model adopts a hybrid storage method of graph structure metadata and time series data blocks, and completes the implicit spatiotemporal correlations between multi-source data through the graph rule inference engine; The multi-condition combined retrieval algorithm is optimized to support multi-dimensional combined queries based on time window, material type, scene label, and spatiotemporal weight. Query redundancy is reduced by constructing a graph time-series joint index, and standardized structured datasets are generated and stored in the database deep processing layer.

4. The multi-source material demand forecasting and replenishment decision-making method for power warehousing according to claim 3, characterized in that, The cleaning feature stage constructs an adaptive data cleaning module based on the database's native functions, and uses a dual verification method combining the isolated forest algorithm and graph database association verification to remove abnormal data; We design an interpolation completion strategy based on graph temporal association, which uses the entity association relationship of knowledge graph to determine the set of similar entities and extracts the historical temporal data of similar entities to complete the missing data; A database feature indexing mechanism is introduced to extract multi-dimensional core features and construct a feature data association table. Feature weights are dynamically allocated through an attention mechanism, and the core feature set is selected to generate indexed prediction model input data.

5. The multi-source material demand forecasting and replenishment decision-making method for power warehousing according to claim 4, characterized in that, The predictive modeling stage takes core feature data with indexes as input, constructs a hybrid model that integrates time series prediction and scene adaptation, introduces a database-driven transfer learning mechanism, extracts historical similar scene data based on the dual criteria of scene label matching and time series trend similarity calculation, and uses it as source domain data to complete the model parameter transfer through a domain adaptation algorithm. Differentiated sub-models are designed for different types of supplies based on their demand characteristics. For routine consumable supplies, a model combining long short-term memory network with temporal trend decomposition and temporal database smoothing is used to capture long-term demand trends. For emergency supplies, a temporal prediction model combined with attention mechanism and scenario triggering mechanism is used to dynamically adjust attention weights to enhance the triggering effect of sudden scenarios on demand. The model outputs material demand forecasts at multiple time scales, including daily, weekly, and monthly levels, confidence intervals, and corresponding data source indexes.

6. The multi-source material demand forecasting and replenishment decision-making method for power warehousing according to claim 5, characterized in that, The inventory threshold stage constructs a dynamic safety stock calculation model driven by database retrieval optimization. The basic safety stock is calculated based on historical average demand data retrieved from the time-series database, the shortest supply cycle of suppliers, and warehouse capacity constraints. The floating safety stock is dynamically adjusted based on the forecast confidence level and data query popularity. The inventory threshold is dynamically corrected by retrieving supplier fulfillment capability data from the graph database and combining it with supplier capacity load. Real-time monitoring of data retrieval response speed; when query latency exceeds a preset threshold, automatic adjustment of the time window parameters for inventory calculation to balance calculation accuracy and real-time performance.

7. The multi-source material demand forecasting and replenishment decision-making method for power warehousing according to claim 6, characterized in that, The replenishment decision-making stage constructs a multi-objective replenishment decision-making model based on graph temporal correlation weights. The core objectives are to minimize the total cost of inventory holding and procurement, maximize the stockout risk coverage, and minimize the replenishment response time. The model takes into account dynamic safety stock thresholds, real-time inventory data, multi-scale predicted demand data, and full-dimensional supplier data. The optimal replenishment plan is solved by an improved particle swarm optimization algorithm. The temporal association weight and entity association weight of each replenishment task are calculated using the graph time-series fusion database association query function, and replenishment priority is divided based on weight sorting; Generate a structured replenishment decision list containing data traceability information. Each decision in the list is associated with a corresponding predictive data index, supplier data index, and scenario association graph, supporting full-chain traceability. The actual data generated during replenishment execution is fed back to the parameter optimization stage.

8. The multi-source material demand forecasting and replenishment decision-making method for power warehousing according to claim 7, characterized in that, The parameter optimization phase receives replenishment execution feedback data, constructs a dual closed-loop feedback mechanism for collaborative optimization of database storage parameters and prediction model parameters, and collects and stores the feedback data in the distributed time series database application layer. Regularly conduct model and database dimension evaluations based on feedback data: For the model dimension, evaluate the prediction accuracy, and if the error exceeds the preset threshold, start retraining and adjust the feature weights and source domain data filtering parameters; For the database dimension, analyze indicators including query frequency, storage heat, and read / write latency, and optimize partitioning strategies, index structures, and caching mechanisms. The node attributes and edge relationship weights of the knowledge graph are updated based on feedback data, and the temporal correlation relationships and rules of new scenarios and material needs are added to optimize the processing effect of the aforementioned stages.

9. The multi-source material demand forecasting and replenishment decision-making method for power warehousing according to claim 8, characterized in that, The distributed time-series database adopts a strategy that combines local real-time backup with off-site scheduled backup, and sets differentiated retention periods according to data types.

10. The multi-source material demand forecasting and replenishment decision-making method for power warehousing according to claim 9, characterized in that, The replenishment decision list supports export in both Excel and JSON formats.

Images