1494 results about "Batch processing" patented technology

The invention provides an intelligent short message scheduling method and device based on multi-dimensional dynamic optimization, and the method comprises the steps: obtaining the performance data of a plurality of short message channels, and calculating a channel health score based on a weight dynamic adjustment model; determining a scheduling strategy according to the priority identifier of the to-be-sent message, and performing channel screening and optimal matching; executing message sending and monitoring a sending state; terminal state detection is carried out on the failure message through operator base station signaling, and a decision tree model is applied to determine a retry strategy; and performing Huffman coding compression processing on the P2-level marketing messages which fail in retry, and performing batch sending in an idle window. According to the method, a comprehensive performance evaluation index and reward function model is also constructed, and parameter optimization is performed by applying a reinforcement learning algorithm. According to the invention, multi-dimensional dynamic channel scoring, intelligent retry decision making based on terminal state perception, batch processing with balanced cost-time efficiency and a closed-loop self-optimization system are realized, the short message delivery rate is obviously improved, and the invalid retry rate and the sending cost are reduced.

An adaptive data system (ADS) for cognitive data processing is disclosed. The ADS includes an adaptive semantic preprocessor, a trigger detector, a temporal batching engine, a symbolic encoder, and a dynamic cognitive transformer engine. The adaptive semantic preprocessor is configured to receive input data from one or more databases and identify cognitive data attributes comprising one or more contextual, semantic, and temporal attributes from the received input data. The trigger detector is configured to identify semantic divergence of the identified cognitive data attributes and provide a standardized data. The temporal batching engine is configured to provide a high-dimensional cognitive data from the standardized data. The symbolic encoder compresses the high-dimensional cognitive data. The dynamic cognitive transformer engine is configured to determine decision making rules, analyze the compressed high-dimensional cognitive data based on the decision making rules and provide recommendations based on an outcome of the analysis to a user.

The invention discloses a network security information analysis and statistics method based on a regulation and control cloud platform. Protocol identification, standardized mapping and blood relationship marking of multi-source security equipment data are realized through an edge layer protocol adaptation cluster and a metadata self-registration mechanism; constructing a real-time streaming and batch processing dual-channel framework based on regulation and control of cloud platform container resources, and fusing to generate a standardized security event object with a frequency weight and a hazard coefficient; designing a main ring-emergency ring double-path decision closed loop, respectively processing conventional events and high-risk threats, and realizing knowledge base iteration; extracting features by using a containerized architecture in combination with an XGBoost + isolated forest integrated model, detecting threats and generating rules; and a space-time-system-risk four-dimensional visualization engine is constructed, attack path drawing, event evolution display and high-risk area prediction are realized, and a management closed loop is formed. According to the network security information analysis and statistics method based on the regulation and control cloud platform, the intelligence and automation level of network security information analysis and statistics is improved.

The invention provides an intelligent laboratory full-process collaborative management system and method based on multi-dimensional data fusion, and the method comprises the steps: obtaining an original data stream from laboratory instrument equipment, carrying out the format recognition and conversion of heterogeneous instrument data through a standard protocol adapter, and generating a first data set in a unified format; aiming at the first data set, executing automatic analysis by adopting data middleware, and integrating into a structured second data set based on a metadata rule of field mapping and unit conversion; according to the second data set, a multi-source data fusion model is constructed, data are classified according to instrument types through batch processing and aggregated according to timestamps, and a fused third data set is generated; aiming at the initial process configuration, adjusting a node sequence and a parameter threshold value by using a visual design tool, and generating an optimized process configuration adaptive to the diversified scene; according to the optimized process configuration, process scheduling is executed through a dynamic process engine, execution time and resource occupation are monitored in real time, and a process execution log is generated.

The invention discloses an Excel report automatic conversion and rendering method, system, device and medium, and relates to the field of data processing, visualization and document generation, the method comprises the following steps: adopting an intelligent Excel analysis engine to automatically analyze Excel file structure features, and selecting a most suitable analyzer to analyze an Excel file; based on an xlsx library and a data structured conversion mechanism, performing data processing on the analyzed Excel file, and constructing a structured data model; based on a dynamic component of React, constructing different types of visual generation components according to the structured data model, and determining a layout framework; and based on a dynamic component and a structured data model of React, dynamically loading and rendering the visual content in the layout framework by adopting a headless browser technology, configuring browser parameters suitable for generating a PDF format, and displaying a PDF document according to the browser parameters. The method has the advantages of high-flexibility Excel analysis capability, strong data processing capability, rich visual expressive force, high-quality PDF rendering and efficient batch processing capability.

The invention relates to the technical field of artificial intelligence and natural language processing, and discloses a task processing method and device, computer equipment and a storage medium, and the method comprises the steps: carrying out the dynamic batch processing of an input text task, and obtaining a conventional batch and an ultra-long batch; detecting a task belonging to an ultra-long batch in the to-be-processed text task as a priority text task, and realizing preemptive scheduling of the current text task by the priority text task through a hardware interrupt mechanism; calculating the priority text task by adopting a local-global mixed attention mechanism; calculating the text tasks belonging to the conventional batch by adopting a first sparse rate, and generating a first historical key value pair; calculating a priority text task by adopting a second sparse rate, and generating a second historical key value pair; and caching the first historical key value pair in a first cache pool, and caching the second historical key value pair in a second cache pool. The method can be applied to the field of financial science and technology businesses, and efficient differentiation processing and computing resource optimization of super-long and conventional text tasks are achieved.

The invention belongs to the field of multi-modal RGB-T data fusion, and discloses a distributed streaming multi-modal fusion-oriented adaptive gradient compression optimization method and system, and the method comprises the steps: carrying out the personalized compression of data features of different modals through employing a modal-sensitive gradient compression strategy, and enabling the modals to comprise an RGB modal and an infrared modal; introducing physical constraint time sequence alignment loss, and performing time sequence alignment on the multi-modal data subjected to personalized compression by utilizing a physical model based on a thermal diffusion equation; the dynamic batch processing strategy is sensed through the video memory, the use condition of the video memory is monitored in real time, and the batch processing size is dynamically adjusted according to the residual capacity of the video memory. According to the method, a brand new solution is provided for efficient training of distributed streaming multi-modal data, and the precision of a multi-modal data fusion task and the resource utilization rate are improved.

The invention relates to the technical field of personalized recommendation, in particular to a nonlinear enhanced decoupling type contrast hypergraph learning method for next POI recommendation. The method comprises the following steps: S1, constructing a multi-view decoupling hypergraph; s2, optimizing the multi-view decoupling hypergraph constructed in the step A to obtain a nonlinear hypergraph convolutional network; s3, based on self-adaptive fusion of user representation, learning and fusing the user preferences under the multiple views obtained in the step S2; and S4, realizing comparative learning and self-supervised learning based on an Encoder module and a GRACE module. According to the nonlinear enhanced decoupling type contrast hypergraph learning method for the next POI recommendation, by introducing ReLU and residual connection, the linear limitation of a traditional hypergraph layer is broken through, the model is endowed with higher expression ability, a KNN adjacency matrix sparse strategy is adopted, the calculation efficiency is remarkably improved on the premise that the precision is guaranteed, and the method is suitable for popularization and application. And meanwhile, batch processing InfoNCE and GRACE modules are creatively applied, so that robust cross-view collaborative learning on a large-scale data set is realized.

The invention discloses a dynamic batch processing method and device for reasoning requests, electronic equipment and a storage medium, relates to the technical field of computers, and aims to learn and obtain an optimal batch processing strategy by utilizing a dynamic batch processing reward algorithm of reinforcement learning, dynamically adjust the batch processing amount of the reasoning requests and improve the efficiency of batch processing of the reasoning requests. The reasoning request processing efficiency and the resource utilization rate are remarkably improved, the training and reasoning progress is accelerated, the resource utilization rate and the reasoning response speed are improved, the real-time performance index is monitored and obtained to dynamically adjust the optimal batch processing strategy in the process that the reasoning node executes the reasoning request, data changes are adapted, and the reasoning efficiency and the reasoning response speed are improved. The method is suitable for reasoning scenes requiring high throughput, low delay and high resource utilization rate, so that the technical problem of low resource utilization rate or response delay increase caused by static batch processing in the face of request quantity fluctuation or data distribution change can be solved, and resource management and scheduling are more efficient and unified.

According to the model lightweight deployment method and device based on the hardware performance of the end system, parameters such as the maximum computing power and the batch processing amount are obtained through a static hardware performance evaluation query database, and the actual computing power and the memory are calculated in real time through dynamic operation state monitoring to serve as compression targets; when the hardware resource change exceeds a threshold value, light weight is triggered, and dynamic adjustment of a compression strategy is achieved; in the lightweight process, hardware parameters and model features are fused into a low-dimensional state vector, collaborative compression parameters of a reinforcement learning strategy network output pruning rate, quantization bit width and distillation temperature are input, and finally a lightweight model adaptive to the hardware dynamic environment is generated. The automatic compression process reduces the labor cost, ensures that the lightweight model is accurately matched with the end side hardware constraint, improves the reasoning speed and reduces the energy consumption.

The invention relates to the technical field of enterprise data management, in particular to a lightweight-based enterprise data dynamic integrated management system, which is characterized in that an acquisition module is used for deploying edge computing nodes, receiving multi-source heterogeneous information streams from manufacturing execution systems and equipment logs, dynamically analyzing and standardizing the information streams, and adding metadata tags; uploading is carried out in a batch processing mode; the map construction module is used for constructing a semiconductor blood relationship map by taking the standardized key information as a blood relationship clue; a graph database is used for efficient storage, and a RESTful API interface is configured to support batch import, so that data storage and relevance expression are more flexible and efficient; the prediction module performs reasoning on the atlas by adopting a graph neural network to generate predictive risk distribution, and a correlation analysis set generated by the prediction module is stored back to the atlas in a structured manner; by introducing a multi-thread concurrent write-in and lock mechanism, the atlas supports complex combination query based on a Cypher query language, and supports multi-level and traceability query.

The invention discloses a front-end performance optimization method and system based on dynamic resource loading and rendering optimization, belongs to the technical field of webpage front-end development, and aims to solve the technical problems that an on-demand loading scheme lacks dynamic strategy adaptation, is not deeply combined with a rendering process and cannot be dynamically optimized according to a real-time environment. Comprising the steps of predicting page blocks or resource types possibly accessed by a user through a behavior prediction model on the basis of user behavior data, and dynamically calculating a loading weight of each resource on the basis of the resource types, the distance between the resources and a viewport, the user behavior data and a user behavior prediction result; based on the network state and the equipment memory capacity, according to a preset resource loading strategy, setting a loading mode and a priority of resources; splitting a long list or a component in the page block into a plurality of independent frame rendering tasks, and processing the frame rendering tasks in batches in an idle period of the browser; and pre-defined performance indexes are monitored and recorded in real time.

The invention provides a computing power service dynamic resource allocation method and system applied to AI model training. The method comprises the following steps: firstly, collecting real-time computing power resource use data (including computing node load, memory occupation and data transmission delay) and model training state data (including training task stage identification, model parameter updating frequency and training data batch processing progress) in AI model training; generating a computing power resource demand association feature set, constructing a computing power resource dynamic allocation decision model including resource allocation priority judgment, adjustment amplitude calculation and scheduling opportunity selection units based on the set, and outputting a computing power resource allocation scheme (including computing node number, memory capacity and data transmission bandwidth adjustment instructions) through the model. Resources are scheduled according to the scheme, and new data are collected to update the feature set, so that dynamic and accurate allocation of computing power resources is realized, and the resource utilization rate and the training efficiency are improved.

The invention discloses a TR assembly gold wire bonding process parameter prediction method based on a multilayer perceptron neural network, and belongs to the technical field of microwave device intelligent manufacturing. According to the method, an intelligent mapping model of gold wire bonding geometric parameters and radio frequency performance is constructed by fusing a multi-layer perceptron neural network and parameterized electromagnetic simulation. The method specifically comprises the following steps: generating 45 groups of samples in a process parameter space by adopting Latin hypercube sampling; obtaining an S parameter data set through batch processing electromagnetic simulation; box-Cox conversion and normalization preprocessing are carried out on the data; the method comprises the following steps: constructing an MLP neural network model of a 3-32-16-2 structure, and determining hyper-parameters by using Bayesian optimization; and after training is completed, rapid reverse mapping from target performance to process parameters is realized. According to the method, the number of traditional tests is reduced from more than 200 to 45, the predicted root-mean-square error of S21 is smaller than or equal to 0.12 dB, the determination coefficient is larger than or equal to 0.96, and the parameter backstepping time lt is obtained; according to the method, full-process automation from simulation, training, optimization to production and issuing is realized, and the development efficiency of the TR component is remarkably improved.

The invention provides a system and method for managing the lifecycle of machine learning models, from development to deployment and ongoing operation, across various environments including on-premises, cloud, and hybrid infrastructures. The system features a model build platform for data processing, feature generation, model development, training, and hyperparameter tuning. A model analytics engine extracts metadata, performs complexity analysis, and generates configuration files specifying environment settings and resource needs. A secure model repository enables version-controlled storage, while a deployment platform retrieves, validates, and deploys models in containerized environments like OpenShift or Kubernetes. The platform dynamically allocates resources, supports real-time and batch scoring, and monitors model performance with guardrails. Customizable agents provide real-time feedback and automated optimization, and the system can securely decommission models while maintaining detailed lifecycle records. The invention enhances the efficiency, security, and scalability of machine learning operations with continuous performance improvement and compliance automation.

The invention discloses a cpu + gpu-based tensor parallel computing algorithm for a large model, which comprises the following steps of: S1, combining row parallelism and column parallelism for a linear layer in the model by adopting a mixed-dimension tensor segmentation mode, and dynamically adjusting a segmentation proportion according to a model structure and hardware resources; s2, a CPU and GPU cooperative computing mechanism is constructed, part of tasks which have large video memory requirements and are relatively simple in computation are allocated to the CPU, the GPU is responsible for computing intensive tasks, communication between CPU-GPU is optimized, and overlapping of CPU-GPU communication and GPU computation is achieved; and S3, implementing a dynamic resource allocation and load balancing strategy, monitoring load conditions of the CPU and the GPU in real time, dynamically adjusting task allocation according to calculation requirements of different layers of the model and use conditions of hardware resources, and adopting a self-adaptive batch processing size adjustment strategy. The invention obviously reduces the occupation of the video memory, reduces the communication cost and improves the utilization rate of computing resources.

The invention discloses a satellite orbit forecasting method based on deep learning physical constraint loss, and the method comprises the following steps: 1, carrying out the normalization preprocessing of input data, forming a training data set and a test data set, and constructing batch processing training data; and 2, performing dimension expansion on sample data points in each window in the batch processing data formed in the step 1, constructing a multi-dimensional feature space of the sample points, and forming a batch processing input data format capable of being introduced into the model. And 3, performing forward reasoning on the batch data formed in the step 2 by using a model, and obtaining a batch processing orbit prediction value output by the model at the next moment through a CNN lightweight spatial-temporal feature extraction module and a BiLSTM bidirectional time sequence neural network module. And 4, taking the track prediction value obtained in the step 3 and the truth value label in the training set obtained in the step 1 as input, calculating to obtain a loss value of a current training iteration batch through a multi-random learning loss module fusing physical constraints, and performing reverse updating of model parameters to complete model training. And step five, through the steps two to four, performing reasoning verification on the model by using the test set formed in the step one, and comparing with a truth value in the test set to obtain a model test result.

A system uses both a present cost model and a future cost model trained on logged order data to compute a prediction of costs for delivering orders either without further delay, or with delay to allow time to potentially batch orders for delivery with other orders (and thereby reduce delivery cost). A comparison of the outputs of the present and future cost models is used to determine whether to delay assigning the order in expectation of batching order with other orders. Calculations may additionally be performed for the constituent orders of an order batch to apportion the delivery cost saving resulting from batching among the different orders. The system can analyze previously-logged data associated with prior orders to obtain features that characterize the prior orders. Using these features, and the known actual delivery costs from the prior completed deliveries, the system can train the present and future cost models.

The invention discloses a large model dynamic batch processing method based on sequence splicing. The method comprises the following steps: receiving input sequences of a plurality of users in a batch; converting the input sequence into a corresponding token sequence; carrying out heterogeneous splicing on all token sequences along a sequence length dimension to form a unified joint token; the spliced joint tokens pass through a normalization layer, standardization operation is executed on the joint tokens, and data distribution is unified; carrying out linear projection on the joint token through a shared linear transformation layer to generate a joint query vector, a joint key vector and a joint value vector, and splitting the joint query vector, the joint key vector and the joint value vector into sub-vector groups corresponding to each user; executing multi-head attention calculation to obtain attention output of the user; and carrying out linear transformation on the attention output, and inputting a transformed result into a shared MLP to carry out nonlinear feature extraction and enhancement so as to obtain a final output corresponding to each user. According to the method, the problems of efficiency bottleneck and resource consumption when a large model processes mass data are effectively solved.

The embodiment of the invention provides a data vectorization acceleration method and system, and the method comprises the steps: writing an optimizer rule into an optimizer, and carrying out the vectorization acceleration according to the optimizer rule, respectively fusing respective original operators in the first operator rule and the second operator rule with respective matched replacement operators of the first operator rule and the second operator rule to obtain a vectorized fusion operator; and converting the vectorization fusion operator into a vectorization execution operator according to the conversion logic, so that the vectorization execution operator calls a vectorization execution engine through a local interface JNI bridge to accelerate data vectorization. Therefore, through the embodiment of the invention, the problem that the calculation efficiency is relatively low in a large-scale data processing scene due to the fact that an existing Flink framework cannot fully utilize an instruction set and vectorization hardware resources in a batch processing mode can be solved.

The invention relates to the technical field of big data, in particular to a warehouse management analysis method and system based on big data and a medium, comprising heterogeneous data real-time acquisition, stream batch fusion processing, AI intelligent decision making, digital twinborn verification and hierarchical storage scheduling. Compared with the prior art in which a distributed data acquisition and delay batch processing scheme is adopted, which leads to serious data islands and lagging of decision basis, the scheme adopts an edge protocol conversion gateway and flow batch fusion processing architecture, uniformly accesses real-time data flow of industrial equipment through MQTT, captures service system changes in combination with DebeziumCDC, realizes millisecond data integration through Kafka, and realizes real-time data flow access through MQTT. Dynamic feature calculation and deep model training are synchronously completed by means of a Flink-Spark hybrid engine; the method has the advantages of total-factor real-time mapping and cross-domain data collaboration, gets through information barriers of an equipment layer, a business layer and an environment layer, and provides millisecond-level fresh global data views for storage decisions.

The invention relates to the technical field of big data processing, and particularly discloses a computer data management system based on big data. The system comprises a distributed data acquisition module which captures multi-source data through an API gateway and a crawler and embeds a quality label; the intelligent storage scheduling engine is used for realizing automatic migration of cold and hot data by adopting a column-type and distributed hybrid architecture; fusing a computing framework, and dynamically coordinating flow processing and batch processing resource execution feature extraction; the self-adaptive strategy center is used for generating data partitions and encryption strategies based on reinforcement learning; and a multi-level security protection system is adopted, and attribute base decryption and dynamic desensitization are deployed. The problems of resource scheduling rigidity and security protection lag in the prior art are solved, the storage cost is reduced by more than 40%, the query delay is compressed to be within 200ms, and the method is suitable for real-time decision-making scenes in the fields of e-commerce and finance.

The invention belongs to the technical field of power system automation, and provides a self-adaptive tensor decomposition attention mechanism for a virtual power plant and a prediction optimization method thereof.The method comprises the steps that firstly, multi-modal data such as historical loads and weather of the virtual power plant are collected, and preprocessing such as normalization and feature extraction is conducted; then, an adaptive tensor decomposition attention network is built based on the processed data, an improved Xavier method is adopted to initialize a core tensor, features are fused after multi-modal data are coded, and a load prediction model is built through adaptive tensor decomposition attention calculation; and finally, adaptive core tensor updating is executed, through optimization strategies such as Hessian matrix low-rank approximation, adaptive sparse mask generation and Nesterov acceleration momentum calculation, batch processing and parallel calculation, adaptive resource allocation and core tensor updating are combined, data calculation is input, and a load prediction result is obtained. The objective of the invention is to solve the problems of high attention mechanism calculation complexity, parameter redundancy, low multi-modal data fusion efficiency, insufficient load sudden change period prediction precision and lack of an adaptive resource allocation mechanism in the existing virtual power plant load prediction technology.

The invention discloses a table-level data classification and grading method and system based on big data, and relates to the technical field of table-level data classification and grading. According to the scheme, second-level sensitive table grading response is achieved through streaming incremental calculation and dynamic semantic coupling; a finite-state machine and a Hash fingerprint technology accurately capture table structure and content mutation, and traditional batch processing delay is avoided; the multi-modal fusion model synchronously analyzes topological features and semantic vectors, and the field-level sensitivity misjudgment problem is solved; the timeliness weight mechanism guarantees real-time evaluation of a dynamic association table relationship, and is particularly suitable for a high-frequency change scene in a big promotion period; and the asynchronous architecture design isolates the calculation load, so that the processing priority of the core price list is effectively guaranteed.

The invention discloses an AI big data real-time processing and analysis method, and solves the problems of insufficient real-time performance and resource waste in traditional processing. The method comprises the steps of collecting heterogeneous data from multiple sources, dividing priorities through feature vector construction and a dynamic evaluation model, and shunting to an edge rapid processing channel, an edge-cloud collaboration channel and a cloud batch processing channel. The edge node preprocesses the high / medium priority data and compresses an analysis result in a layered manner; and the cloud receives the compression result, the middle-priority residual data and the low-priority data, a unified view is established through fusion, and the edge analysis model parameters are iteratively optimized in real time. And finally feeding back the edge preliminary analysis and the cloud depth result to the terminal. According to the method, through dynamic distribution, cooperative processing, differential compression and model iteration, real-time response of high-priority data, efficient resource allocation and continuous improvement of analysis precision are achieved, and the method is suitable for multi-scene heterogeneous data processing.

The invention discloses a large-model adaptive batch reasoning system based on time delay measurement, and the system comprises a request receiving and queue management module which is used for monitoring and receiving an LLM reasoning request sent by an external user or an application; the real-time delay sensing and batch monitoring module is used for measuring and analyzing performance indexes directly related to the current batch processing strategy in real time in the LLM reasoning process; the Token-by-Token batch processing and reasoning execution module is used for selecting a corresponding number of requests from the request queue according to the size of the current batch determined by the scheduler, organizing the requests into an effective calculation batch and submitting the effective calculation batch to the LLM reasoning core at the bottom layer to execute one or more decoding steps; and the double-stage batch adjustment decision module is used for obtaining the normalized time delay ratio. The invention also discloses a large-model adaptive batch reasoning method based on time delay measurement. According to the invention, substantive improvement and intelligent management of the overall performance of the LLM reasoning service are realized.

The invention discloses an overexposure and underexposure image enhancement method, device, equipment and medium, and the method comprises the steps: enabling an image decomposition network to respectively receive a first illumination image and a second illumination image through employing two encoder-decoder networks sharing the weight, and extracting multi-scale illumination distribution information through multi-scale connection; the image reconstruction module outputs a corresponding reflectivity component and a brightness component, the image enhancement network adopts an image enhancement sub-network to adjust illumination distribution and suppress noise based on the reflectivity component and the brightness component, and the image reconstruction module multiplies the adjusted reflectivity component with an illumination image element by element and outputs an enhanced image. According to the method, the technical effects of effectively recovering image details, reducing noise interference and improving real-time performance under the overexposure or underexposure condition are reflected, the method is particularly suitable for complex environments such as power equipment monitoring, the contrast ratio of the enhanced image is better, artifacts are fewer, the calculation burden is reduced, and efficient batch processing prediction is supported.

The invention relates to the technical field of knowledge management systems, and discloses a knowledge full-life-cycle management system and a construction method, and the system comprises a technical platform and multi-modal perception layer, a calculation and arrangement engine layer, a unified API gateway layer, a multi-modal processing module layer, a knowledge graph layer, a cognitive reasoning layer and an application layer. Real-time collection and batch processing of multi-source heterogeneous data are achieved through a standardized SDK / API; spark / Flink is combined with Kubernetes to complete the ETL (Extract Transform Load) and resource scheduling of the multi-modal data; constructing an entity-relationship-attribute knowledge graph through cross-modal feature fusion; intelligent decision support is realized based on rule reasoning, graph calculation and GNN; the application layer provides intelligent questioning and answering, decision support and personalized recommendation services; the technical problems of knowledge islands, sharing barriers, knowledge statics and the like in the prior art are solved.

The invention provides a reasoning service system. A global controller is used for dividing machine resources into an interactive resource pool and a batch processing type resource pool in advance; when a reasoning request is received, determining a service level target type of the reasoning request, and performing resource expansion and contraction on the interactive resource pool and the batch processing type resource pool according to the service level target type; the cluster controller is used for respectively dividing the interactive resource pool and the batch processing type resource pool into a cue word resource pool and a token resource pool in advance; aiming at a cue word stage and a token stage of the reasoning request, respectively distributing corresponding cue word resources and token resources from a cue word resource pool and a token resource pool, and carrying out resource expansion and contraction on the cue word resource pool and the token resource pool; and the machine controller is used for adjusting the batch processing size for processing the reasoning request according to the service level target type and the use state of the affiliated machine. And on the premise of meeting the service level target, the resource utilization rate and throughput performance of the inference service system are improved.

The invention relates to an electric power marketing business abnormity real-time detection method and system based on stream-oriented computation, and belongs to the technical field of electric power system optimizing.The method comprises the steps that data snapshots are extracted from an electric power marketing business system, difference comparison is conducted on the data snapshots and historical snapshots of an intermediate library, and standardized increment events are generated and stored; capturing an incremental event in real time through a data change capturing tool and pushing the incremental event to a message queue; a streaming computation engine consumes the event stream, sequentially performs data cleaning, association with a static dimension table and sliding window statistical feature calculation, and constructs a feature vector; and performing parallel analysis and weighted fusion on the feature vectors based on a business rule base and an online machine learning model to generate a comprehensive risk score, and outputting an abnormal event when the score exceeds a threshold value. According to the method, the problems of exception identification lagging and complex work order process in a traditional batch processing mode are solved, the crossing of the business risk from hour-level detection to minute-level real-time perception is realized, and the timeliness and accuracy of power marketing risk management and control are improved.