CN-121995892-A - Time sequence energy-saving control large model construction method for industrial environmental control system

Abstract

The application provides a time sequence energy-saving control large model construction method for an industrial environmental control system, which is applied to the technical field of data processing. The application builds a full-link technical system around an industrial environmental control time sequence energy-saving control large model, firstly collects multi-source time sequence data through an MQTT/Modbus protocol, preprocesses the multi-source time sequence data to generate a target data set, expands the scarce data through TimeGAN, extracts multi-dimensional characteristics to build a system digital image, realizes working condition accurate prediction based on an improved PatchTST architecture and a process knowledge graph, relies on a fuzzy optimization and reinforcement learning dual-drive model output device regulation strategy, builds an edge execution layer to complete instruction landing and closed-loop feedback, ensures high concurrence and low delay operation of an industrial field through model compression and containerization deployment, and finally integrates dynamic ROI adjustment, coordinate drift correction and optical flow method to optimize model decision precision and output an energy-saving control scheme and a device state assessment result of an adaptation field.

Inventors

• Hua Haizhou
• HUANG CHUANLIN
• LIN XIU
• YAO YE
• HUA LAIZHEN
• TIAN YE
• YANG ZIXUAN
• WANG YANJIE
• ZHANG JINBAO
• XU YANG
• LIU JIAJIE

Assignees

• 中电智维(上海)科技有限公司

Dates

Publication Date

20260508

Application Date

20260410

Claims (5)

1. 1. The time sequence energy-saving control large model construction method for the industrial environmental control system is characterized by comprising the following steps of: the method comprises the steps of collecting multi-source time sequence data of an industrial environmental control system, wherein the multi-source time sequence data comprises equipment operation parameters, environment perception data, business production data and external meteorological data, realizing real-time access and preprocessing through an MQTT/Modbus protocol, and generating a target time sequence data set; Training TimeGAN a generation model to expand industrial fault working conditions and extreme environment scene data by taking a target time sequence dataset as a training sample, formulating a fusion screening rule of original data and generated data, and extracting time characteristics, state characteristics, environment characteristics and trend characteristics from the dataset after fusion to form an industrial environmental control system digital portrait; adopting an improved PatchTST time sequence prediction framework, introducing a blocking attention mechanism to capture a long-period energy consumption rule, injecting physical constraint by combining an industrial process knowledge graph, pre-judging working condition fluctuation 1-24 hours in advance, and outputting a high-precision energy consumption and environmental state prediction result; A fuzzy optimization and reinforcement learning dual-drive decision model is constructed, energy saving rate, comfort level and equipment service life are taken as multi-objective optimization functions, and a device operation parameter regulation strategy is dynamically output through pareto optimization balance conflict targets; An edge execution layer is built based on an industrial bus and an internet of things (IoT) protocol, a decision instruction is converted into equipment physical operation, closed loop feedback is realized through a digital twin and physical entity double-track verification platform, and regulation and control effect data are collected in real time; Model compression is carried out by adopting model quantization, structured pruning and knowledge distillation technology, containerized deployment is realized based on Kubernetes+ TritonInferenceServer, and millisecond decision response and high concurrency support are ensured by combining a dynamic batch processing and elastic expansion and contraction mechanism; Based on the industrial environmental control time sequence energy-saving control large model, the environmental fluctuation interference is compensated by the dynamic ROI adjustment and coordinate drift correction model, the global working condition change trend is estimated by integrating the Lucas-Kanade optical flow method, the model decision precision is optimized, and finally the environmental control system target energy-saving control scheme and the equipment running state evaluation result which are suitable for the industrial site are output.
2. 2. The method of claim 1, wherein training TimeGAN the generation model to expand industrial fault conditions, extreme environmental scene data, formulate a fusion screening rule of raw data and generated data, extract time features, state features, environmental features and trend features from the data set after fusion, and form the industrial environmental control system digital representation using the target time sequence data set as a training sample, comprising: Combining the fault working condition of the industrial environmental control system, the scarce situation of the extreme environmental scene data and the massive data requirement of large model training, taking a target time sequence data set as a training sample to develop model training, completing the parameter optimization and the capacity adaptation of TimeGAN generated models, and generating a data generation model adapting to the industrial environmental control scene; Combining the working condition characteristics and the environment change rule of the actual operation of the industrial environmental control system, generating a model through TimeGAN after training to carry out simulation deduction, completing the expansion of the scarce data of the fault working condition and the extreme environment scene, and generating the actual expansion time sequence data fitting the industry; Combining with an authenticity verification standard and an validity screening rule of industrial environmental control time sequence data, formulating a fusion screening rule around the characteristic matching degree and scene fitting degree of the original data and the generated data, and generating an industrial environmental control exclusive time sequence data fusion screening system; based on a formulated fusion screening system, integrating and accurately screening the target time sequence original data and the generated expansion data, removing invalid and distorted data, and generating an industrial environment-control fusion time sequence data set; feature mining and extraction are carried out from the fused time sequence data set by combining the feature extraction requirement of the time sequence energy-saving control of the industrial environmental control system and the dimension requirement of the digital portrait construction, and a multi-dimensional feature set of time features, state features, environment features and trend features is generated; based on the extracted multidimensional feature set, feature modeling and systematic integration are carried out by combining the operation logic and feature association relation of the industrial environmental control system, and the digital portrait of the industrial environmental control system capable of accurately representing the operation state of the system is generated.
3. 3. The method of claim 1, wherein constructing a fuzzy optimization and reinforcement learning dual-drive decision model, taking energy saving rate, comfort level and equipment service life as multi-objective optimization functions, balancing conflict targets through pareto optimization, and dynamically outputting equipment operation parameter regulation strategies, wherein the method comprises the following steps: performing feature extraction and dimension normalization processing on the energy consumption and environmental state prediction results output by the industrial environmental control time sequence energy-saving control large model prediction module to generate a decision model input basic data set; Performing association matching processing on the input basic data set of the decision model, the operation constraint condition of the industrial environmental control system and the equipment operation threshold value to generate a multi-objective optimization decision matrix comprising an energy-saving dimension, a comfort dimension and an equipment service life dimension; Inputting a multi-objective optimization decision matrix into a fuzzy optimization and reinforcement learning dual-drive decision model to perform multi-objective cooperative operation, and generating an operation parameter optimization rule of industrial environmental control system equipment, wherein the operation parameter optimization rule comprises fan frequency adjustment logic, an air conditioner temperature setting strategy and an equipment start-stop linkage mechanism; Based on the multi-objective optimization decision matrix operation result and the equipment operation parameter optimization rule, a pareto optimization mechanism is triggered, conflict targets of all optimization dimensions are balanced, a dynamic equipment operation parameter regulation strategy is generated, and the construction of the industrial environmental control time sequence energy-saving control large model decision capability is completed.
4. 4. The method of claim 1, wherein model compression is performed by using model quantization, structured pruning and knowledge distillation techniques, containerized deployment is realized based on kubernetes+ TritonInferenceServer, dynamic batch processing and elastic expansion and contraction mechanisms are combined, millisecond decision response and high concurrency support are ensured, and the method comprises the following steps: processing the network structure, parameter scale and thrust calculation force requirements of the industrial environmental control time sequence energy-saving control large model, extracting the model compression core optimization index, and generating a model lightweight processing basic parameter set; Respectively executing model quantization, structured pruning and knowledge distillation treatment on a model layer, neurons and weight parameters corresponding to the basic parameter set to complete full-dimensional lightweight compression of the industrial environmental control time sequence energy-saving control large model; Adapting the compressed large model to a Kubernetes+ TritonInferenceServer deployment framework, carrying out container mirror image manufacture, service arrangement and interface configuration, and generating a standardized model containerized deployment system; based on the decision request concurrency quantity and reasoning response aging requirement of the industrial field, a dynamic batch processing scheduling mechanism is built, and a node elastic expansion and contraction capacity triggering rule is formulated to form a large-model high-availability operation scheduling scheme; And integrating a lightweight compression result, a containerized deployment system and an operation scheduling scheme of the model to complete engineering floor deployment of the industrial environmental control time sequence energy-saving control large model, and guaranteeing millisecond decision response and high concurrency request support of the model.
5. 5. The method of claim 4, wherein the method based on the industrial environmental control time sequence energy-saving control large model compensates the environmental fluctuation interference through dynamic ROI adjustment and coordinate drift correction model, integrates the Lucas-Kanade optical flow method to estimate the global working condition change trend, optimizes the model decision precision, and finally outputs the environmental control system target energy-saving control scheme and the equipment running state evaluation result which are suitable for the industrial site, and comprises the following steps: Based on an industrial environmental control time sequence energy-saving control large model, introducing a dynamic ROI adjustment and coordinate drift correction dual-mode cooperative mechanism to accurately compensate monitoring data deviation caused by industrial field environmental fluctuation, and accurately extracting effective characteristics of environmental control working conditions; Integrating a Lucas-Kanade optical flow method and a large model working condition analysis module, constructing a global working condition trend dynamic estimation model, and realizing real-time sensing and trend prejudgement of the working condition change of the industrial environmental control system through optical flow field characteristic calculation; Fusing the accurate working condition data after deviation compensation and the global working condition trend estimation result, constructing a large model decision accuracy dynamic optimization mechanism, performing self-adaptive optimization on model reasoning parameters, and correcting decision deviation caused by environmental interference; And (3) abutting against the industrial field environmental control process requirement and the equipment operation constraint standard, constructing a model output scheme field adaptation verification system, completing scheme iterative optimization through multi-dimensional suitability verification, and finally outputting the environmental control system target energy-saving control scheme and equipment operation state comprehensive evaluation result of the adaptation industrial field.

Description

Time sequence energy-saving control large model construction method for industrial environmental control system Technical Field The invention relates to the technical field of data processing, in particular to a time sequence energy-saving control large model construction method for an industrial environmental control system. Background Meanwhile, the actual measured data of fault working conditions and extreme environment scenes in the industrial scenes are scarce, the coverage scenes of the existing data sets are limited, the requirements of large model training on massive and full scene data can not be met, and the generalization capability and control precision of a control model are restricted. The energy consumption change and the working condition fluctuation of the industrial environmental control system have obvious long-period characteristics, the traditional time sequence prediction model is difficult to accurately capture the long-period energy consumption rule, and the physical constraint of an industrial process is not fully integrated in the model training process, so that the prediction result is easy to deviate from the actual industrial production requirement, and a reliable pre-judgment basis cannot be provided for subsequent decisions. The existing environmental control system is controlled by adopting a single target optimization strategy, so that the multi-dimensional optimization targets such as energy saving rate, comfort level, equipment service life and the like are difficult to balance, meanwhile, the influence of industrial field environment fluctuation and working condition dynamic change is not fully considered by a decision model, and the dynamic self-adaptive regulation and control cannot be realized by the decision parameter immobilization. The industrial environmental control large model is large in parameter quantity and high in inference calculation force requirement, is easy to be limited by calculation force and storage resources when being directly deployed to industrial field edge equipment, causes delay of decision response, cannot meet the industrial environmental control real-time control requirement, is insufficient in protocol suitability of a decision instruction and the industrial environmental control equipment, lacks an effective closed loop feedback mechanism in the instruction execution process, and is difficult to verify and optimize in real time in regulation effect. Meanwhile, the model is difficult to sense the overall working condition change trend in real time, and decision parameter adjustment is delayed to working condition fluctuation, so that the control precision is reduced, and the running requirement of dynamic change of the industrial field cannot be adapted. It should be noted that the information disclosed in the above background section is only for enhancing understanding of the background of the present disclosure and thus may include information that does not constitute prior art known to those of ordinary skill in the art. Disclosure of Invention Other features and advantages of the application will be apparent from the following detailed description, or may be learned by the practice of the application. According to one aspect of the application, a time sequence energy-saving control large model construction method for an industrial environmental control system is provided, which comprises the steps of collecting multi-source time sequence data of the industrial environmental control system, including equipment operation parameters, environment sensing data, business production data and external meteorological data, realizing real-time access and preprocessing through an MQTT/Modbus protocol, and generating a target time sequence data set; the method comprises the steps of taking a target time sequence data set as a training sample, training TimeGAN to generate model expansion industrial fault working condition and extreme environment scene data, formulating fusion screening rules of original data and generated data, extracting time characteristics, state characteristics, environment characteristics and trend characteristics from the data set after fusion to form an industrial environmental control system digital image, adopting an improved PatchTST time sequence prediction framework, introducing a blocking attention mechanism to capture long-period energy consumption rules, injecting physical constraints by combining an industrial process knowledge map, predicting working condition fluctuation for 1-24 hours in advance, outputting high-precision energy consumption and environment state prediction results, constructing a fuzzy optimization and reinforcement learning dual-drive decision model, taking energy conservation rate, comfort and equipment service life as multi-objective optimization functions, optimizing balance conflict targets by pareto, dynamically outputting equipment operation parameter regulation and control strategies, bu