Industrial environmental control system-oriented time sequence energy-saving control large model construction method

Abstract

The invention provides an industrial environment control system-oriented time sequence energy-saving control large model construction method, which is applied to the technical field of data processing. According to the method, a full-link technical system is built around the construction of an industrial environmental control time sequence energy-saving control large model: firstly, multi-source time sequence data is collected through an MQTT / Modbus protocol and preprocessed to generate a target data set, then scarce data is expanded through a TimeGAN, and multi-dimensional features are extracted to construct a system digital portrait; based on the improved PatchTST architecture and in combination with a process knowledge graph, working condition accurate prediction is achieved, and a device regulation and control strategy is output by means of a fuzzy optimization and reinforcement learning dual-drive model; an edge execution layer is built to complete instruction landing and closed-loop feedback, and high-concurrency and low-delay operation of an industrial site is guaranteed through model compression and containerization deployment; and finally, fusing dynamic ROI adjustment, coordinate drift correction and an optical flow method to optimize model decision precision, and outputting an energy-saving control scheme adaptive to the field and an equipment state evaluation result.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for constructing a large-scale time-series energy-saving control model for industrial environmental control systems. Background Technology

[0002] Industrial environmental control systems involve multi-source time-series data, including equipment operating parameters, environmental sensing data, business production data, and external meteorological data. Differences in acquisition protocols and transmission methods among these data sources hinder efficient data fusion. Furthermore, measured data on fault conditions and extreme environmental scenarios in industrial settings are scarce, and existing datasets have limited coverage, failing to meet the demands of large-scale model training for massive, full-scenario data. This restricts the generalization ability and control accuracy of control models. The energy consumption changes and operating condition fluctuations of industrial environmental control systems exhibit significant long-cycle characteristics, making it difficult for traditional time-series prediction models to accurately capture long-cycle energy consumption patterns. Moreover, the lack of sufficient integration of physical constraints from industrial processes during model training leads to prediction results that easily deviate from actual industrial production needs, failing to provide reliable predictive basis for subsequent decision-making.

[0003] Existing environmental control systems often employ single-objective optimization strategies, making it difficult to balance multi-dimensional optimization goals such as energy efficiency, comfort, and equipment lifespan. Furthermore, decision-making models fail to adequately consider the impact of industrial environmental fluctuations and dynamic changes in operating conditions, resulting in fixed decision parameters and an inability to achieve dynamic adaptive control. Large-scale industrial environmental control models typically have a large number of parameters and high inference computing power requirements. When directly deployed to edge devices in industrial settings, they are easily limited by computing power and storage resources, leading to decision response delays and failing to meet the real-time control needs of industrial environmental control. Moreover, the protocol compatibility between decision commands and industrial environmental control equipment is insufficient, and the command execution process lacks an effective closed-loop feedback mechanism, making it difficult to verify and optimize the control effect in real time. The complex industrial environment, with sudden changes in temperature and humidity and equipment vibrations, can easily cause deviations in monitoring data. Traditional control models lack effective deviation compensation mechanisms. Simultaneously, the models struggle to perceive global operating condition trends in real time, and decision parameter adjustments lag behind operating condition fluctuations, leading to decreased control accuracy and an inability to adapt to the dynamically changing operational needs of industrial sites.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0006] According to one aspect of this application, a method for constructing a large-scale time-series energy-saving control model for industrial environmental control systems is provided, comprising: collecting multi-source time-series data of the industrial environmental control system, including equipment operating parameters, environmental perception data, business production data, and external meteorological data; achieving real-time access and preprocessing via MQTT / Modbus protocol to generate a target time-series dataset; using the target time-series dataset as training samples, training a TimeGAN generation model to expand industrial fault conditions and extreme environmental scenario data; formulating fusion and filtering rules for raw and generated data; extracting time features, state features, environmental features, and trend features from the dataset after fusion to form a digital profile of the industrial environmental control system; adopting an improved PatchTST time-series prediction architecture, introducing a block attention mechanism to capture long-cycle energy consumption patterns, and injecting physical constraints by combining an industrial process knowledge graph to predict operating condition fluctuations 1-24 hours in advance, outputting high-precision energy consumption and environmental state prediction results; and constructing a fuzzy optimization + reinforcement learning dual-driven decision model. The system employs multiple objective optimization functions, including energy saving rate, comfort, and equipment lifespan. It balances conflicting objectives through Pareto optimization and dynamically outputs equipment operating parameter control strategies. An edge execution layer is built based on industrial bus and IoT protocols, translating decision commands into physical equipment operations. A closed-loop feedback system, utilizing a dual-track verification platform of digital twin and physical entities, is implemented to collect control effect data in real time. Model compression is achieved through model quantization, structured pruning, and knowledge distillation techniques. Containerized deployment is implemented using Kubernetes and Triton Inference Server, combined with dynamic batch processing and elastic scaling mechanisms to ensure millisecond-level decision response and high concurrency support. Based on a large-scale industrial environmental control time-series energy-saving model, dynamic ROI adjustment and coordinate drift correction are used to compensate for environmental fluctuations. The Lucas-Kanade optical flow method is integrated to estimate global operating condition trends, optimizing model decision accuracy. Finally, the system outputs a target energy-saving control scheme for the environmental control system adapted to industrial environments and equipment operating status evaluation results.

[0007] Another aspect of this application is a system for constructing a time-series energy-saving control large model for industrial environmental control systems, the system being configured to execute the above-described method for constructing a time-series energy-saving control large model for industrial environmental control systems by executing the executable instructions.

[0008] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described method for constructing a time-series energy-saving control big model for industrial environmental control systems by executing the executable instructions.

[0009] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described method for constructing a time-series energy-saving control large model for industrial environmental control systems.

[0010] This application provides a method for constructing a large-scale time-series energy-saving control model for industrial environmental control systems. This method collects and preprocesses multi-source time-series data via the MQTT / Modbus protocol, then uses TimeGAN to augment scarce data to construct a digital profile of the system. An improved PatchTST architecture combined with a process knowledge graph is employed to achieve accurate prediction of operating conditions. The model outputs control strategies based on a dual-drive model of fuzzy optimization and reinforcement learning. An edge execution layer and a dual-track verification platform are built to complete closed-loop feedback. Model compression and containerized deployment ensure on-site operation. Dynamic ROI adjustment, coordinate drift correction, and optical flow methods are integrated to optimize decision-making accuracy, ultimately outputting an energy-saving control scheme and equipment evaluation results adapted to industrial sites.

[0011] This application aims to expand data through TimeGAN-generated models, enhance the generalization ability of large models, improve the accuracy of 1-24 hour operating condition prediction, and provide a reliable basis for decision-making; balance energy saving, comfort, and equipment lifespan through multi-objective optimization to avoid the drawbacks of single-objective decision-making; deploy lightweight and containerized models to meet the high-concurrency and low-latency operation requirements of industrial sites; and improve the decision-making stability of models in complex environments through multi-mechanism anti-interference optimization to achieve accurate, efficient, and energy-saving control of environmental control systems. Attached Figure Description

[0012] Figure 1 This invention provides a flowchart illustrating a method for constructing a large-scale time-series energy-saving control model for industrial environmental control systems, according to an embodiment of this application.

[0013] Figure 2 This paper presents a schematic diagram of a time-series energy-saving control large model construction system for industrial environmental control systems, according to an embodiment of this application. Detailed Implementation

[0014] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0015] In one implementation, Figure 1 A schematic diagram illustrates a process flow diagram of a time-series energy-saving control large model construction method for industrial environmental control systems according to an embodiment of this application.

[0016] S101 collects multi-source time-series data from industrial environmental control systems, including equipment operating parameters, environmental sensing data, business production data, and external meteorological data. It achieves real-time access and preprocessing through the MQTT / Modbus protocol to generate the target time-series dataset.

[0017] In one implementation, the focus is on core industrial environmental control equipment, including compressor operating frequency, outlet temperature, fan speed, and operating power of air conditioning units; fresh air volume, return air temperature, and filter pressure difference of the fresh air system; and real-time time-series data such as operating time, start / stop status, and load rate of temperature and humidity control equipment. Multi-dimensional environmental time-series data are collected within the industrial production area, including real-time temperature, relative humidity, atmospheric pressure, dust concentration, and CO2 concentration at different monitoring points within the workshop, as well as derived data such as the rate of change of environmental temperature and humidity and the uniformity of spatial environmental parameters. Production operation time-series data strongly correlated with the environmental control system are collected, including production line operating load, production cycle time, number of equipment in operation, production shift schedule, material input, and process environmental requirement thresholds for high-precision production processes. Real-time meteorological time-series data from outside the industrial plant area are collected, including outdoor temperature, humidity, wind speed, wind direction, precipitation, light intensity, atmospheric pressure, and short-term weather forecast trends.

[0018] To address the transmission characteristics of data from different acquisition dimensions, a dual-protocol adaptation mode combining MQTT and Modbus protocols is adopted to build a unified access gateway for multi-source data in industrial environmental control systems. This enables standardized, low-latency, real-time access to time-series data of different types and transmission protocols, ensuring the real-time performance and compatibility of data acquisition. Specifically, the Modbus protocol adapts to wired data transmission from industrial environmental control hardware devices, while the MQTT protocol adapts to wireless data transmission from environmental sensing terminals and external meteorological monitoring equipment. Through protocol conversion and data encapsulation, multi-source data is uniformly aggregated to the data acquisition server.

[0019] Modbus protocol data access: Industrial environmental control air conditioning units, fresh air systems and other hardware devices are equipped with Modbus-RTU / TCP communication modules. The data acquisition gateway establishes a wired connection with the device through the industrial bus. According to the Modbus protocol standard parameters of 9600bps baud rate, 8 data bits, 1 stop bit, and no parity, it sends data acquisition commands to the device to read the device's operating frequency, speed, power and other parameters in real time. The acquisition cycle is set to 5 seconds / time. After acquisition, the data is encapsulated into standardized time-series data frames.

[0020] For MQTT protocol data access, wireless sensing devices such as temperature and humidity sensors and outdoor meteorological monitoring terminals within the workshop are all connected to the industrial IoT wireless gateway. The lightweight MQTT communication protocol is used for data transmission in a publish / subscribe model. The sensing devices act as publishers, pushing collected environmental and meteorological data to the MQTT message server every 3 seconds. The data acquisition gateway acts as a subscriber, receiving and retrieving data in real time, thus completing unified access to wireless data. A data access relay module is built to parse and unify the format of raw data from Modbus and MQTT protocols. This converts raw data from different protocols and devices into a standardized data format of "device / monitoring point identifier - data item name - collection timestamp - data value - data unit," achieving unified aggregation and storage of multi-source data.

[0021] To address issues such as outliers, missing values, duplicate values, inconsistent data units, and poor temporal synchronization in the raw data after access, a full-process preprocessing operation is performed, including data cleaning, data completion, unit normalization, and temporal alignment. This process removes invalid data, standardizes data formats, and ensures the validity, consistency, and temporal sequence of the data, laying the foundation for the generation of the target time-series dataset. The preprocessing operation is implemented using a data preprocessing module, which directly interacts with the data acquisition server and employs a streaming processing approach to preprocess the real-time accessed data online.

[0022] The preprocessed multi-source time-series data are structured and stored according to the classification rules of "data dimension-collection object-time series period-data attribute" to construct the target time-series dataset of industrial environmental control system. This dataset serves as the core original sample for subsequent TimeGAN generation model training and the construction of digital profiles of industrial environmental control system. It is stored in the time-series database to support rapid data retrieval, extraction and retrieval. At the same time, a data version management mechanism is set up for the dataset to record the entire process information of data collection and preprocessing, ensuring data traceability.

[0023] S102 uses the target time series dataset as training samples to train the TimeGAN generation model to expand the data of industrial fault conditions and extreme environmental scenarios. It formulates the fusion and screening rules of the original data and the generated data. After fusion, it extracts time features, state features, environmental features and trend features from the dataset to form a digital profile of the industrial environmental control system.

[0024] In one implementation, considering the scarcity of data on fault conditions and extreme environmental scenarios in industrial environmental control systems, and the massive data requirements for large-scale model training, the target time-series dataset is used as training samples to train the model. This process optimizes the parameters and adapts the capabilities of the TimeGAN generation model, generating a data generation model suitable for industrial environmental control scenarios. Based on the target time-series dataset output by the multi-source time-series data acquisition module for industrial environmental control, the entire process of training and parameter optimization of the TimeGAN generation model is carried out. This adapts the model to the industrial environmental control scenario, creating a dedicated industrial environmental control time-series data generation model that addresses the problem of data scarcity for fault conditions and extreme environmental scenarios.

[0025] The TimeGAN generative model comprises four core modules: a generator, a discriminator, a temporal embedding module, and a supervisor. The module hierarchy is as follows: the temporal embedding module serves as the data input layer, receiving a standardized target time-series dataset and vectorizing the temporal features; the generator and discriminator form the core computational layers, with the generator generating simulated time-series data based on the embedded features, and the discriminator distinguishing between real and generated data, forming an adversarial training relationship; the supervisor acts as an auxiliary optimization layer, monitoring the temporal patterns of the generated data and correcting any deviations in the temporal features. All modules interact through fully connected layers. The output of the embedding module is connected to the inputs of the generator, discriminator, and supervisor, respectively; the output of the generator is connected to the inputs of the discriminator and supervisor; and the outputs of the discriminator and supervisor are fed back to the generator via a loss function, enabling iterative optimization of the model parameters.

[0026] The target time series dataset is divided into training and validation sets in a 9:1 ratio and input into the time series embedding module of the TimeGAN model. The embedding dimension is set to 64 dimensions to convert the time series data into high-dimensional embedding features. The input data format is "time series step size - feature dimension - data value". In the industrial environmental control scenario, the time series step size is set to 1440 (corresponding to 24-hour minute-level time series data). The feature dimension is the total dimension of the collected equipment, environment, production and meteorological data.

[0027] The generator uses a multilayer perceptron combined with a gated recurrent unit architecture, with 128 and 64 hidden layer nodes and ReLU activation function. The discriminator uses a symmetrical architecture to the generator, with 64 and 128 hidden layer nodes and LeakyReLU activation function. The supervisor has 64 hidden layer nodes and Tanh activation function. The optimizer is Adam, with an initial learning rate of 0.001, a weight decay coefficient of 0.0001, a batch size of 32, and 500 training epochs.

[0028] First, a supervisor is pre-trained using real target time-series data to learn the real patterns of industrial environmental control time-series data. The pre-training rounds are 50. Then, adversarial training is performed between the generator and the discriminator. The generator generates simulated time-series data based on embedded features, and the discriminator performs binary classification on the real data and the generated data. At the same time, the supervisor monitors the time-series features of the generated data and calculates the time-series loss between the generated data and the real data.

[0029] The total model loss consists of adversarial loss, temporal loss, and embedding loss. The adversarial loss uses cross-entropy loss, the temporal loss uses mean squared error loss, and the embedding loss uses cosine similarity loss. The weight ratio of the total loss is adversarial loss: temporal loss: embedding loss = 4:3:3. A validation set evaluation is performed every 10 training epochs. If the validation set loss does not decrease for 20 consecutive epochs, a learning rate decay strategy is adopted to reduce the learning rate to 0.5 of the original value until the number of training epochs reaches 500. After training, the model performance is evaluated using the validation set. If the similarity (cosine similarity) between the generated data and the real data is ≥0.95, the model is considered to have passed the training. The parameters of the TimeGAN generation model are optimized, and the model weight file adapted to the industrial environmental control scenario is saved, generating a data generation model specifically for the industrial environmental control scenario.

[0030] By combining the actual operating conditions and environmental change patterns of industrial environmental control systems, a trained TimeGAN generative model is used for simulation and extrapolation to expand scarce time-series data for fault conditions and extreme environmental scenarios, generating data that aligns with industrial realities. Using a TimeGAN generative model adapted to industrial environmental control scenarios, and combining the actual operating conditions and environmental change patterns of industrial environmental control systems, simulations of fault conditions and extreme environmental scenarios are performed to generate expanded time-series data that aligns with industrial realities, supplementing the scarce data types in the target time-series dataset. The trained TimeGAN generative model is input with feature vectors of typical operating conditions of the industrial environmental control system, including fault condition feature vectors (such as air conditioning compressor failure, fresh air system filter blockage, temperature and humidity sensor failure, etc.) and extreme environmental feature vectors (such as extreme outdoor high / low temperatures, heavy rain, strong winds, etc.). Initial environmental parameters and basic equipment parameters under these conditions are also input as basic constraints for the model's generated data. The model outputs full-dimensional time-series data corresponding to faults and extreme operating conditions. The output data format is consistent with the target time-series dataset, including four dimensions: equipment operating parameters, environmental perception data, business production data, and external meteorological data. The time-series step size and acquisition cycle are also consistent with the original data to ensure the compatibility of the expanded data.

[0031] The operational logic and fault evolution patterns of industrial environmental control systems are integrated into the model generation process. For example, when an air conditioning compressor fails, the time-series data generated by the model must conform to the fault evolution logic of "compressor operating frequency drops sharply → workshop temperature rises slowly → air conditioning operating power drops abnormally". When the outdoor temperature is extremely high, the data generated by the model must conform to the environmental change pattern of "outdoor temperature continues to be higher than 38°C → workshop cooling load increases sharply → air conditioning unit operating frequency rises to the rated value → workshop temperature continues to rise slowly", so that the generated extended data fits the actual operating conditions of the industrial environmental control system.

[0032] For the extreme high-temperature environment scenario in the electronics manufacturing workshop, the TimeGAN model is input with extreme environmental feature vectors and initial parameters of "outdoor temperature 40℃, initial workshop temperature 26℃, and production line operating at full load". The model, combined with the environmental control logic of the electronics manufacturing workshop, generates time-series data for this scenario: the outdoor temperature remains around 40℃, the workshop temperature gradually rises from 26℃ to 28℃, the air conditioning unit operating frequency increases from 50Hz to the rated 60Hz, the fresh air system's fresh air volume is adjusted to the maximum, and the workshop humidity gradually decreases from 50% to 45%. This expanded data accurately matches the actual operation of the environmental control system in the electronics manufacturing workshop under extreme high temperatures. For the fault condition of air conditioning fan failure, the model is input with fault feature vectors of "fan bearing jamming and initial operating frequency 30Hz". The model generates fault time-series data of "fan operating frequency gradually decreasing to 0Hz, rapid rise in local workshop temperature, and sudden drop in air conditioning operating power", reconstructing the actual evolution process of the fan failure.

[0033] By combining the authenticity verification standards and validity screening rules of industrial environmental control time-series data, and focusing on the feature matching degree and scenario fit between the original data and the generated data, a fusion screening system specifically for industrial environmental control time-series data is formulated. This system provides a unified standard for subsequent data fusion and screening. The fusion screening system comprises four core dimensions: authenticity verification rules, feature matching degree screening rules, scenario fit screening rules, and data integrity screening rules. These rules are interconnected and layered, ensuring that the fused data is authentic, valid, and relevant to the industrial environmental control scenario.

[0034] Set reasonable value ranges and change rate thresholds for various dimensions of industrial environmental control data. If the generated data exceeds the reasonable range or the change rate is abnormal, it is judged as distorted data and removed. For example, the reasonable range for the air conditioning operating frequency in a machining workshop is 0-60Hz, and the reasonable threshold for the temperature change rate is ≤1℃ / minute. If the generated extended data shows an air conditioning operating frequency of 70Hz, or a workshop temperature increase of 6℃ within 5 minutes, it is judged as distorted data and removed. Calculate the feature similarity between the generated extended data and the original target time-series data for the same type of operating condition. Use the cosine similarity algorithm and set a feature matching threshold of ≥0.9. If it is lower than this threshold, it is judged as a feature mismatch and the data is removed. For example, calculate the feature similarity between the generated extreme low-temperature environment data and the low-temperature environment data in the original data. If the similarity is 0.85, which is lower than the threshold of 0.9, it is judged as a feature mismatch and the extended data is removed.

[0035] Based on the business logic of different industrial environmental control scenarios, a scenario fit judgment standard is set. If the generated data does not match the environmental control operation logic of the corresponding scenario, it is judged as insufficient scenario fit and is removed. For example, the environmental control scenario of a food processing workshop requires that "when the humidity is below 60℃, the cooling capacity should be increased simultaneously when the temperature rises, while maintaining a stable fresh air volume". If the generated extended data for this workshop shows a situation of "temperature rises, cooling capacity decreases, and fresh air volume increases sharply", which does not match the scenario logic, it is judged as insufficient scenario fit and the data is removed.

[0036] A data integrity threshold of ≥99% is set. If the generated expanded data contains a single time series data with a missing dimension of ≥1, or if the number of missing time series data in a single scenario is ≥0.1%, the data is considered incomplete, and that part of the data is removed. The corresponding data is then regenerated using the model. For example, if the generated air conditioner fault condition data is missing the core dimension of "air conditioner operating power", or if 20 out of 1000 time series data in that condition are missing, the integrity is 98%, which is lower than the 99% threshold. Therefore, the data is considered incomplete, removed, and regenerated.

[0037] Based on the established fusion and screening system, the original target time-series data and the generated extended time-series data are integrated and precisely screened to remove invalid and distorted data, generating an industrial environmental control fusion time-series dataset. This system, specifically designed for industrial environmental control, integrates and precisely screens the original target time-series dataset and the extended time-series data generated by the TimeGAN model, removing invalid, distorted, and feature-mismatched data to generate a high-quality industrial environmental control fusion time-series dataset, providing a solid data foundation for subsequent feature extraction. The original target time-series dataset and the generated extended time-series data are uniformly integrated according to the classification standard of "data dimension - acquisition time - scenario type." The extended data is categorized according to fault conditions and extreme environmental scenarios, and integrated into the corresponding scenario classification of the original dataset, forming a preliminary fusion dataset. Each data point is labeled "original data / generated data" for easy traceability. Following the four core rules of the fusion screening system, the preliminary fusion dataset is screened layer by layer. First, authenticity verification is performed to remove distorted data; then, feature matching degree is calculated to remove data with mismatched features; next, scene fit is verified to remove data with insufficient fit; finally, data integrity is checked to remove incomplete data, and the removed incomplete data is supplemented and generated.

[0038] The filtered fused data undergoes secondary preprocessing to maintain a normalization standard consistent with the original target time-series dataset, mapping all data values ​​to the [0,1] interval. Simultaneously, the fused data is time-aligned using the unified timestamp of the original data as a benchmark to ensure temporal consistency. The filtered and preprocessed fused data is then structured according to industrial environmental control scenario types (normal operating conditions, fault operating conditions, and extreme environmental conditions) to generate an industrial environmental control fused time-series dataset, which is stored in a time-series database. This provides a data interface for the subsequent feature extraction module, supporting rapid feature extraction and retrieval.

[0039] Combining the feature extraction requirements of time-series energy-saving control in industrial environmental control systems and the dimensional requirements of digital profile construction, feature mining and extraction are performed from fused time-series datasets to generate a multi-dimensional feature set encompassing time features, state features, environmental features, and trend features. Furthermore, deep feature mining and extraction are conducted from fused industrial environmental control time-series datasets to generate feature sets across four dimensions: time features, state features, environmental features, and trend features, providing core feature support for digital profile construction. The feature extraction process, based on the time-series energy-saving control requirements of industrial environmental control systems, focuses on features that have a core impact on energy consumption patterns, operating condition fluctuations, and environmental changes. Statistical analysis, time-series feature mining, and trend fitting methods are employed to ensure that the extracted features accurately reflect the operating status of the environmental control system.

[0040] Core features are extracted from the time dimension of the fused time-series data, including daily / weekly / monthly time-series patterns, production shift time-period characteristics, and equipment operation time-series characteristics. Periodic analysis and time-period segmentation methods are used for extraction. For example, time features are extracted from the fused data of the machining workshop, including shift time-period characteristics such as peak cooling load in the morning shift (8:00-16:00), stable load in the evening shift (16:00-24:00), and low load in the night shift (0:00-8:00), the periodic pattern of full load of the production line from Monday to Friday, and hourly and minute-level time-series characteristics of air conditioning unit operation.

[0041] Real-time operational status features of equipment, production, and environment in industrial environmental control systems are extracted, including equipment operational status features (normal / faulty, high / low load), production operational status features (full / half / idle, fast / slow production cycle), and environmental status features (normal / extreme, stable / fluctuating). Threshold judgment and status classification methods are used for extraction. For example, status features are extracted from fused data in an electronics manufacturing workshop, including equipment status features of air conditioning units ("rated load operation", "low load operation", "fault shutdown"), production status features of production lines ("full load production" and "half load production"), and environmental status features of the workshop environment ("stable temperature", "fluctuating temperature", "extreme high temperature").

[0042] The core environmental features affecting the operation of industrial environmental control systems are extracted, including indoor and outdoor environmental features. Feature dimensionality reduction and core feature screening methods are used to extract key features from multi-dimensional environmental data. For example, environmental features are extracted from fused data in a food processing workshop. Indoor environmental features include average workshop temperature, relative humidity, temperature uniformity, and humidity uniformity; outdoor environmental features include outdoor temperature, humidity, wind speed, and light intensity. Simultaneously, the temperature difference and humidity difference between indoor and outdoor environments are extracted. These features are the core influencing factors for environmental control and energy saving in food processing workshops.

[0043] This study extracts trend characteristics from various dimensions of data in industrial environmental control systems, including linear, nonlinear, abrupt, and stationary trends. Trend fitting, slope calculation, and abrupt change point detection methods are used for extraction. For example, trend features are extracted from fused data from a metallurgical workshop, including a linear upward trend in outdoor temperature, an abrupt increase in workshop temperature due to equipment failure, a stationary trend in air conditioning unit operation at rated frequencies, and a nonlinear fluctuation trend in workshop humidity with changes in fresh air volume. Simultaneously, trend characteristics of energy consumption changing with production load are extracted to provide trend basis for subsequent energy-saving control. The extracted four-dimensional features are then structurally integrated, with each feature assigned unique identifiers, feature dimensions, feature descriptions, and data sources to generate a multi-dimensional feature set for the industrial environmental control system. This feature set is stored in vector form for easy subsequent feature modeling and systematic integration.

[0044] Based on the extracted multi-dimensional feature set, feature modeling and systematic integration are performed by combining the operational logic and feature correlation relationships of the industrial environmental control system to generate a digital profile of the industrial environmental control system that accurately represents the system's operating status. Combining the operational logic of the industrial environmental control system, the inherent correlation relationships between multi-dimensional features are analyzed, a feature correlation model is constructed, the correlation strength between features is quantified, and the hierarchical relationship between core and secondary features is clarified. For example, the feature correlation model quantifies the positive correlation strength between "production line load" and "air conditioning cooling load" as 0.92, the positive correlation strength between "outdoor temperature" and "workshop temperature" as 0.88, and the correlation strength between "equipment operating status" and "energy consumption" as 0.95. Simultaneously, "energy consumption feature," "production load feature," and "equipment operating status feature" are identified as core features, while others are secondary features. Following a hierarchical structure of "core feature - secondary feature - correlated feature," the multi-dimensional feature set is systematically integrated to construct an industrial environmental control system feature system. This system covers all dimensions of the environmental control system's operational features, including equipment, environment, production, time, and trends, achieving structured and hierarchical management of features. Based on the integrated feature system, four core dimensions are constructed to form a digital profile of the industrial environmental control system: digital profile of equipment operation, digital profile of environmental status, digital profile of production linkage, and digital profile of energy consumption trend. These core dimensions are interconnected and their data is shared, together forming a complete digital profile of the environmental control system.

[0045] By mapping the features of each dimension in the feature system to the corresponding core dimensions of the digital profile, the physical operating status of the industrial environmental control system is accurately mapped to the digital profile, generating a digital profile of the industrial environmental control system. At the same time, a dynamic update mechanism for the digital profile is established, which updates the feature data of the profile in real time based on the time-series data collected by the environmental control system, ensuring that the digital profile remains synchronized with the operating status of the physical system.

[0046] S103 adopts an improved PatchTST time-series prediction architecture, introduces a block attention mechanism to capture long-cycle energy consumption patterns, and combines industrial process knowledge graphs to inject physical constraints, predicting operating condition fluctuations 1-24 hours in advance and outputting high-precision energy consumption and environmental condition prediction results.

[0047] In one implementation, the improved PatchTST time-series prediction architecture comprises six core layers: a data input layer, a time-series block layer, a block attention layer, a feature fusion layer, a physical constraint injection layer, and a prediction output layer. Each layer follows a top-down feature transfer relationship, while the physical constraint injection layer and the feature fusion layer form a bidirectional feature interaction, achieving deep integration of process knowledge. The data input layer, the bottom layer of the architecture, receives multi-dimensional feature sets from the digital profile of the industrial environmental control system, completes feature standardization and vectorization transformation, and provides a unified feature input for the upper layers. The time-series block layer aims to perform non-overlapping block processing on the input time-series feature sequences, dividing long-cycle time-series features into fixed-length time-series blocks, solving the problems of low efficiency and scattered attention in the original architecture for long-cycle data processing. The block attention layer, the core computational layer of the architecture, introduces a block attention mechanism, performs local self-attention calculations on each time-series block, and simultaneously achieves long-cycle feature association through cross-block attention interaction, accurately capturing the long-cycle energy consumption patterns of industrial environmental control.

[0048] The feature fusion layer aims to concatenate and fuse local features and cross-block related features output from the block attention layer. Through a multilayer perceptron, it achieves dimensionality enhancement and depth mining of the features, generating high-dimensional fused temporal features. The physical constraint injection layer aims to interface with the industrial process knowledge graph, transforming the process physical rules of the environmental control system into feature constraints, which are then injected into the high-dimensional features of the feature fusion layer to correct model prediction biases. The prediction output layer, the top layer of the architecture, uses a temporal prediction head to perform temporal extrapolation on the high-dimensional fused features after injecting physical constraints, outputting predictions for energy consumption and environmental conditions for the next 1-24 hours.

[0049] The feature output of the data input layer is directly connected to the input of the temporal block layer. The temporal block feature output of the temporal block layer is connected to the input of the block attention layer. The local and cross-block feature outputs of the block attention layer are connected to the feature fusion layer. The feature fusion layer and the physical constraint injection layer are bidirectionally connected through the feature mapping interface. The constrained feature output of the physical constraint injection layer is connected to the temporal prediction head of the prediction output layer. The linear transfer of features between modules is achieved through a fully connected layer, and the nonlinear transformation of features is achieved through an activation function to ensure the continuity and effectiveness of feature transfer.

[0050] A customized block attention mechanism is introduced into the block attention layer of the improved PatchTST architecture. The attention calculation rules are optimized for the acquisition cycle and long-term variation patterns of industrial environmental control time-series data, enabling accurate capture and feature mining of hourly, daily, and weekly long-term energy consumption patterns of industrial environmental control systems. This solves the problems of high computational consumption and inaccurate feature capture when traditional attention mechanisms process long-term data. Based on the acquisition characteristics and long-term analysis requirements of industrial environmental control time-series data, the time-series block length is set to 24 hours (corresponding to 1440 data points acquired at the minute level), the block step size is 24 hours, and a non-overlapping block method is adopted to ensure the feature independence of each time-series block. Simultaneously, long-term correlation is achieved through cross-block attention. For each time series block, perform local self-attention calculation, calculate the attention weight of each feature point within the block, and capture the local energy consumption pattern within the block (such as the correlation between the daily production line load change and energy consumption); set the cross-block attention window to 7, that is, each time series block interacts with the three time series blocks before and after it, calculates the cross-block attention weight, and captures the long-cycle energy consumption pattern (such as the weekly energy consumption fluctuation pattern caused by the weekly production shift change).

[0051] A multi-head attention mechanism is employed, with 8 heads used to divide the feature dimension into 8 independent subspaces for parallel attention computation, improving the comprehensiveness of feature capture. The attention activation function is Softmax, the normalization method is LayerNorm, and the dropout probability is set to 0.1 to prevent overfitting. A sparse attention mechanism is introduced to sparsify low-weight feature connections in the block attention computation, retaining only feature associations with attention weights ≥ 0.05, reducing the model's computational consumption by 60% while maintaining the accuracy of capturing long-term energy consumption patterns.

[0052] The long-term time-series feature sequences of industrial environmental control (such as minute-level energy consumption features over the past 30 days) output from the data input layer are divided into non-overlapping blocks according to a set block length and step size, generating several independent time-series blocks. Local self-attention calculation is performed on each time-series block, generating attention weights by calculating the similarity of feature points within the block. These weighted features are then fused to capture local energy consumption patterns within the block, such as the energy consumption patterns of peak cooling load during the morning shift and low cooling load during the night shift. Cross-block attention calculation is performed on adjacent time-series blocks within a set cross-block attention window to capture long-term feature correlations between time-series blocks, such as the weekly energy consumption patterns of high energy consumption during full-load production from Monday to Friday and low energy consumption during low-load production on weekends. The local attention features and cross-block attention features are then weighted and fused to generate block-based attention features containing both local and long-term energy consumption patterns, which are then output to the feature fusion layer for deep mining.

[0053] By combining the equipment operation rules, production process requirements, and energy-saving logic of industrial environmental control systems, a structured and standardized industrial environmental control process knowledge graph is constructed. This graph digitizes and characterizes the physical constraints of the environmental control system, and establishes clear rules for injecting physical constraints, providing knowledge support and implementation standards for subsequent injection of physical constraints into the model. The industrial environmental control process knowledge graph uses entities, relationships, and attributes as its core elements, covering four core knowledge domains: equipment operation, production process, environmental control, and energy-saving rules. These domains are interconnected, forming a complete environmental control process knowledge system. The equipment operation knowledge domain includes entities of core environmental control equipment (air conditioning, fresh air, temperature and humidity control equipment, etc.), as well as equipment operating parameter thresholds, start-stop rules, linkage operation relationships, and lifespan constraint attributes, such as relationships and attributes like "air conditioning compressor operating frequency ≤ 60Hz" and "fresh air system and air conditioning unit linkage start-stop." The production process knowledge domain includes entities such as production lines, production processes, and production loads. It also includes the environmental requirements for temperature, humidity, and cleanliness, as well as the relationship between production load and the operation of environmental control equipment, such as relationships and attributes like "the temperature in the high-precision chip processing workshop needs to be maintained at 23±1℃ and the humidity at 50±5%" and "if the production line load increases by 10%, the air conditioning cooling load needs to be increased by 8%".

[0054] The environmental control knowledge domain includes entities such as indoor environment, outdoor environment, and environmental monitoring points. It covers the control range of indoor environmental parameters, the influence of the outdoor environment on the indoor environment, and constraints on the adjustment rate of environmental parameters, such as relationships and attributes like "workshop temperature adjustment rate ≤ 1℃ / minute" and "when the outdoor temperature is higher than 35℃, the workshop must close the windows and increase the cooling capacity." The energy-saving rules knowledge domain includes entities such as energy-saving targets, energy consumption thresholds, and energy-saving control strategies. It covers energy-saving control rules under different operating conditions and rules for balancing energy consumption and environmental comfort, such as relationships and attributes like "air conditioning cooling load reduced to 30% during non-production periods" and "under the premise of meeting production process requirements, the workshop temperature can be increased by 1℃ to reduce air conditioning energy consumption."

[0055] The physical constraints in the industrial environmental control process knowledge graph are transformed into quantifiable and embeddable model feature constraints, categorized into three types: numerical constraints, logical constraints, and correlation constraints. Numerical constraints, such as "air conditioning operating frequency ≤ 60Hz," are transformed into upper bound constraints for feature values. Logical constraints, such as "the fresh air system and air conditioning are linked for start-stop," are transformed into logical correlation constraints between features. Correlation constraints, such as "production load and cooling load are positively correlated," are transformed into weighted correlation constraints between features. A dual injection rule of "feature layer injection + inference layer constraint" is formulated. Feature layer injection involves mapping the characteristic physical constraints to the high-dimensional fused features of the feature fusion layer dimension by dimension, correcting features that exceed the constraint range. Inference layer constraint involves adding physical constraint judgment logic during the model's time-series prediction inference process. If the prediction result violates the physical constraints of the environmental control process, the model inference correction mechanism is triggered, and a new prediction result that conforms to the constraints is generated.

[0056] The characteristic physical constraints of industrial environmental control processes are injected into the feature fusion layer and prediction inference layer of the improved PatchTST time series prediction architecture according to the established injection rules. Using the industrial environmental control fusion time series dataset as training samples, the model is trained and the parameters are optimized throughout the entire process. This completes the adaptation of the model to the industrial environmental control scenario and ensures that the model prediction results conform to the time series pattern and meet the physical constraints of the industrial environmental control process.

[0057] Based on the industrial environmental control fusion time-series dataset and combined with the multi-dimensional feature set of the industrial environmental control system digital profile, the dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The input data is a combination of "time-series feature sequence - process physical constraint features". The time-series feature sequence includes multi-dimensional time-series features of equipment, environment, production, and energy consumption, while the process physical constraint features are characteristic numerical, logical, and relational constraints. The model's training output is the predicted features of energy consumption and environmental status for the next 1 / 6 / 12 / 24 hours, including core prediction indicators such as air conditioning energy consumption, fresh air energy consumption, workshop average temperature, relative humidity, and environmental uniformity. The output data format is "prediction time - prediction indicator - predicted value - constraint compliance label".

[0058] The feature fusion layer uses a two-layer multilayer perceptron with 256 and 128 hidden nodes respectively, and GELU activation function. The temporal prediction head of the prediction output layer uses a single fully connected layer, with the output dimension matching the prediction metric dimension. The optimizer is AdamW, with an initial learning rate of 0.0005, a weight decay coefficient of 0.001, a batch size of 16, and 300 training epochs. The training set is input into the improved PatchTST model, where feature segmentation, segment attention calculation, and feature fusion are performed according to the architectural hierarchy. Physical constraint features are injected into the feature fusion layer to constrain and correct the fused features, and then the prediction output layer generates preliminary prediction results.

[0059] The total model loss consists of temporal prediction loss and constraint violation loss. The temporal prediction loss uses mean squared error loss to measure the deviation between the predicted and actual results. The constraint violation loss uses L1 loss to measure the degree to which the predicted results violate physical constraints. The total loss weight ratio is temporal prediction loss: constraint violation loss = 7:3, ensuring that the model maintains both prediction accuracy and compliance with physical constraints. Every 10 training epochs, a validation set is input into the model for evaluation. If the total loss of the validation set does not decrease for 15 consecutive epochs, a learning rate decay strategy is adopted, reducing the learning rate to 0.6. After training, a test set is input into the model for performance testing. If the model's mean absolute error of prediction is ≤3% and the constraint violation rate is ≤1%, the model is considered successfully trained. The weight file of the successfully trained improved PatchTST model is saved, generating a temporal prediction model adapted to industrial environmental control scenarios. A physical constraint update interface is also reserved to support subsequent updates to constraints based on process upgrades. An improved PatchTST time-series prediction model, trained and infused with physical constraints, is used to predict the industrial environmental control system's operating conditions for the next 1-24 hours, taking the real-time operating characteristics of the system as input. Through multi-dimensional prediction accuracy verification, high-precision energy consumption and environmental state prediction results are output, providing core input data for subsequent fuzzy optimization + reinforcement learning dual-driven decision-making models.

[0060] The real-time updated features of the industrial environmental control system's digital profile (including real-time equipment operating status, environmental status, production load, and external meteorological characteristics) are input into the trained and improved PatchTST prediction model to complete the standardization and vectorization of real-time features. Based on the input real-time features, the model performs time-series predictions according to preset prediction durations (1 / 6 / 12 / 24 hours). Through a block attention mechanism combined with industrial process physical constraints, it extrapolates the future operating condition fluctuation trends for different durations, generating preliminary energy consumption and environmental status predictions. The preliminary prediction results undergo dual verification: first, accuracy verification, comparing the prediction results with actual data from similar historical operating conditions; if the prediction deviation exceeds a preset threshold (mean absolute error > 3%), the model re-predicts; second, constraint verification, checking whether the prediction results conform to the industrial environmental control process physical constraints. If constraint violations exist, the prediction results are adjusted through the correction mechanism of the physical constraint injection layer. The prediction results, after double verification, are structured and classified according to prediction duration and prediction indicators to generate standardized energy consumption and environmental status prediction results. At the same time, the analysis of operating condition fluctuation trends is output to identify key operating condition fluctuation nodes (such as production load increase, sudden rise in outdoor temperature, etc.) in the next 1-24 hours.

[0061] The model's predictive input is deeply integrated with the real-time operational business logic of the industrial environmental control system. The real-time characteristics of the input synchronously reflect the workshop's production scheduling plan (e.g., the production line is about to operate at full capacity), external weather changes (e.g., the outdoor temperature will rise to 38°C in the next 6 hours), and equipment operating status (e.g., the air conditioning unit is currently operating at 80% load), ensuring that the fluctuations in operating conditions predicted by the model are consistent with the actual production environmental control business logic. The model's predictive output is combined with the business needs of the subsequent decision-making module. According to the control frequency of the decision-making module, the 24-hour prediction results are divided into refined prediction data with a 1-hour granularity, which facilitates the decision-making module to generate dynamic equipment operating parameter control strategies.

[0062] S104 constructs a dual-driven decision-making model of fuzzy optimization and reinforcement learning, with energy saving rate, comfort and equipment life as multi-objective optimization functions. It balances conflicting objectives through Pareto optimization and dynamically outputs equipment operation parameter control strategies.

[0063] In one implementation, feature extraction and dimensional normalization are performed on the energy consumption and environmental status prediction results output by the large-scale prediction module of the industrial environmental control time-series energy-saving control model to generate a basic dataset for the input of the decision model. Deep feature mining and dimensional normalization are performed on the energy consumption and environmental status prediction results for the next 1-24 hours output by the improved PatchTST time-series prediction module. Redundant features are eliminated, feature formats are standardized, and key decision features are strengthened to generate a standardized basic dataset for the input of the decision model, providing high-quality data support for subsequent multi-objective optimization decisions. Combining the core needs of industrial environmental control decision-making, the focus is on features that directly affect equipment control strategies. Data normalization is achieved through feature selection, format standardization, and dimensional alignment to ensure the adaptability of the input data to the decision model, while retaining the time-series trend and key fluctuation features of the prediction results. Core features relevant to decision-making are selected from the prediction results, and redundant information unrelated to equipment control is eliminated. Core features include energy consumption (predicted energy consumption of air conditioning, predicted energy consumption of fresh air, total energy consumption), environmental features (predicted average temperature, humidity, temperature uniformity, humidity uniformity), and operating conditions (predicted production load, outdoor meteorological influence coefficient, operating condition fluctuation level).

[0064] The selected core features are uniformly converted into a standardized format of "feature name-prediction period-feature value-unit-weight coefficient," where the weight coefficient is set according to the feature's influence on the decision (e.g., 0.4 for energy consumption features, 0.3 for environmental features, and 0.3 for operating conditions). Features for different prediction periods (1 / 6 / 12 / 24 hours) are dimensionally aligned to ensure consistent feature dimensions and uniform data granularity (all split by hour). Minimum-maximum normalization is used to map feature values ​​to the [0,1] interval, eliminating the impact of dimensional differences on the decision model. The standardized and normalized features are organized according to the prediction period order to generate the basic dataset for the decision model input, stored in the form of a "prediction period-feature matrix." This allows the decision model to dynamically call features by period, while also recording metadata such as feature source and weight coefficients to ensure data traceability.

[0065] The decision model input dataset, industrial environmental control system operating constraints, and equipment operation thresholds are correlated and matched to generate a multi-objective optimization decision matrix encompassing energy saving, comfort, and equipment lifespan dimensions. The correlation and matching follows a logical thread of "data adaptation - objective quantification - constraint embedding," deeply binding predicted data, system constraints, equipment thresholds, and the three optimization objectives to ensure the decision matrix comprehensively reflects the decision-making needs and constraints of industrial environmental control. Based on the actual needs of industrial environmental control scenarios, quantitative calculation methods and optimization target values ​​for energy saving rate, comfort, and equipment lifespan are developed to ensure the objectives are measurable and optimizable. Core constraints of the industrial environmental control system (such as environmental requirements of production processes and safe operation standards) and equipment operation thresholds (such as fan frequency range and air conditioning temperature adjustment limits) are collected to form a constraint list. The predicted features of the decision input dataset are correlated dimension-by-dimensionally with the three objective quantification standards and the constraint list, clarifying the contribution of each predicted feature to the optimization objective and the constraint boundary. A multi-objective optimization decision matrix is ​​constructed with "prediction period" as the row and "optimization objective-feature-constraint" as the column. The matrix elements include feature normalization value, objective contribution, upper and lower bounds of constraints, and weight coefficients. The weight coefficients are set according to "energy saving rate 0.4, comfort 0.3, equipment life 0.3" (which can be adjusted according to the scenario).

[0066] The multi-objective optimization decision matrix is ​​input into a fuzzy optimization + reinforcement learning dual-driven decision model for multi-objective collaborative computation, generating optimization rules for the operating parameters of industrial environmental control system equipment, including fan frequency adjustment logic, air conditioning temperature setting strategy, and equipment start-stop linkage mechanism. The model's core logic is "fuzzy optimization to handle uncertain constraints + reinforcement learning to achieve dynamic optimization," and it consists of three main layers: a fuzzy processing layer, a reinforcement learning interaction layer, and a rule generation layer. Each layer collaborates to achieve multi-objective decision computation. The fuzzy processing layer, as the input layer, receives the multi-objective optimization decision matrix and transforms the fuzzy constraints of the industrial environmental control system (such as "high comfort" and "low energy consumption") into quantified fuzzy sets. It calculates the degree of belonging of features to the fuzzy sets through a membership function, solving the problem of imprecise constraints in industrial scenarios. The reinforcement learning interaction layer, as the core computation layer, includes three modules: an agent, an environment simulator, and a reward function. The agent receives the fuzzy-processed feature data and simulates equipment control behavior in the environment simulator. The reward function calculates reward values ​​based on the optimization effects of energy saving rate, comfort, and equipment lifespan. The agent dynamically adjusts its decision strategy through trial and error learning.

[0067] The rule generation layer, as the output layer, transforms the optimization results of the reinforcement learning interaction layer into standardized device operation parameter optimization rules, ensuring that the rules can be directly mapped into device operation instructions. The output of the fuzzy processing layer is connected to the agent input of the reinforcement learning interaction layer, and the decision output of the reinforcement learning interaction layer is connected to the input of the rule generation layer. At the same time, the output of the rule generation layer is fed back to the environment simulator of the reinforcement learning interaction layer, forming a closed-loop learning mechanism.

[0068] Using the decision matrix corresponding to the industrial environmental control fusion time series dataset as training samples, the dataset is divided into training and validation sets in an 8:2 ratio. The training samples include multi-objective decision-making scenarios such as normal operating conditions, fault conditions, and extreme environmental conditions to ensure the model's generalization ability. The membership function of the fuzzy processing layer uses a Gaussian function, and the fuzzy set is divided into three levels: "low," "medium," and "high." The reinforcement learning agent adopts a deep Q-network (DQN) architecture with 128 and 64 hidden layer nodes, and the activation function is ReLU. The environmental simulator is built based on the industrial environmental control digital profile to simulate changes in energy consumption, environment, and equipment status after equipment regulation. The reward function is designed as: Total Reward = Energy Saving Rate Reward × 0.4 + Comfort Reward × 0.3 + Equipment Lifespan Reward × 0.3, where an energy saving rate ≥ 15% earns full marks. 0 points, deduct 1 point for every 1% below; comfort level ≥ 8 points gets the full score of 10 points, deduct 1 point for every 0.5 points below; equipment loss rate ≤ 0.05% gets the full score of 10 points, deduct 2 points for every 0.01% above; the optimizer is Adam, the initial learning rate is 0.001, the batch size is 32, and the number of training rounds is 200; every 10 training rounds, the validation set is used for evaluation. If the total reward of the validation set does not improve for 15 consecutive rounds, the learning rate decay strategy is adopted (reduced to the original 0.5); after training, if the target achievement rate of the model generation rules is ≥ 90%, the training is considered qualified.

[0069] The multi-objective optimization decision matrix is ​​input into the trained dual-drive model. The fuzzy processing layer first converts the constraints in the matrix into fuzzy quantized values. The reinforcement learning agent interacts with the environment simulator based on these quantized values ​​and finds an optimization direction that takes into account the three objectives through multiple rounds of trial and error learning. The rule generation layer converts the optimization direction into standardized equipment operation parameter optimization rules.

[0070] Based on the results of multi-objective optimization decision matrix calculations and the Pareto optimization mechanism triggered by equipment operating parameter optimization rules, conflicting objectives across various optimization dimensions are balanced to generate dynamic equipment operating parameter control strategies, thus completing the construction of a large-scale industrial environmental control time-series energy-saving control model decision-making capability. Pareto optimization uses "no room for improvement" as its core criterion, balancing conflicting objectives such as energy saving rate, comfort, and equipment lifespan by screening non-dominated solutions, ensuring that the final generated control strategy achieves multi-objective optimality while satisfying all constraints. Based on the equipment operating parameter optimization rules, within the constraint boundaries of the multi-objective optimization decision matrix, several sets of candidate solutions for equipment operating parameters are generated. Each set of candidate solutions corresponds to a complete control scheme (such as specific values ​​for air conditioning temperature, fan frequency, and fresh air volume).

[0071] The candidate solution set is evaluated for Pareto dominance. If no other solution can improve the performance of at least one objective without reducing the performance of any other objective, then the solution is Pareto optimal. All optimal solutions are selected to form a Pareto optimal solution set. Combining the real-time operating conditions of the current industrial environmental control system (such as production load, outdoor weather, and current equipment status), the most suitable solution is selected from the Pareto optimal solution set and transformed into a dynamic equipment operating parameter control strategy. This strategy includes specific parameter values, control timing, and adjustment step size. The generated dynamic control strategy is input into a digital twin platform for simulation verification. If the simulation results meet the requirements of "energy saving rate ≥ 15%, comfort level ≥ 8 points, and equipment loss rate ≤ 0.05%", the decision-making capability is deemed qualified. Simultaneously, the actual execution effect data of the control strategy is fed back to the dual-drive model for iterative optimization, completing the closed-loop construction of the industrial environmental control time-series energy-saving control large model decision-making capability.

[0072] The S105 is built on an edge execution layer based on industrial bus and IoT protocol, which transforms decision commands into physical operations of equipment. It achieves closed-loop feedback through a dual-track verification platform of digital twin and physical entity, and collects control effect data in real time.

[0073] In one implementation, the edge execution layer adopts a layered architecture of "edge gateway + control node + communication interface," with the functions and hierarchical relationships of each module as follows: The edge gateway module acts as the core hub, receiving decision instructions output from the large model, parsing, verifying, and distributing the instructions. It also integrates industrial bus and IoT protocol adaptation functions, supporting the conversion and transmission of multi-protocol instructions. The control node module deploys dedicated control nodes according to device type (air conditioner, fresh air system, fan, etc.), receiving instructions distributed by the edge gateway, converting them into control signals recognizable by the device, and achieving precise control of single-type devices. The communication interface module provides industrial bus interfaces (such as Modbus-RTU, Profinet) and IoT wireless interfaces (such as MQTT, LoRa), respectively adapting to wired connection devices and wireless sensing devices, ensuring the diversity and stability of instruction transmission.

[0074] The industrial bus protocol is primarily adapted to large-scale environmental control equipment (air conditioning units, fresh air systems), enabling wired transmission of commands via the industrial bus to ensure low latency and high reliability of control. The IoT protocol is adapted to small sensors, distributed actuators, and other devices, reducing wiring costs through wireless transmission and adapting to complex industrial site layouts. The decision commands output by the large model are standardized parameter control commands ("air conditioning temperature set to 25℃, fan frequency 45Hz"). After receiving the commands, the edge gateway first performs format verification (verifying command integrity and parameter rationality), and then converts the commands into control codes recognizable by the device according to the device communication protocol (Modbus protocol register control values, MQTT protocol topic messages), and sends them to the control node through the corresponding communication interface.

[0075] Based on the control nodes and communication interfaces of the edge execution layer, the converted equipment control commands are transformed into specific physical operations, ensuring that decision commands are implemented and executed, while guaranteeing the safety, stability, and accuracy of operations, in accordance with the operating specifications of industrial environmental control equipment. In accordance with the safety operating specifications of industrial environmental control equipment, equipment status verification (such as current operating status, load rate, and fault alarm status) is added before command execution. If the equipment has a fault alarm or the load exceeds the safety threshold, command execution is paused and abnormal information is reported. For continuously adjustable parameters such as temperature and frequency, a stepped adjustment strategy is adopted to avoid equipment damage or drastic environmental fluctuations caused by sudden parameter changes. The adjustment step size is set according to equipment characteristics (e.g., temperature adjustment ≤0.5℃ per step, fan frequency adjustment ≤5Hz per step). A command execution timeout threshold (e.g., no equipment response within 5 seconds) and an abnormal rollback strategy are set. If command execution times out or the equipment reports an abnormality, the system automatically rolls back to the operating parameters before adjustment and triggers an alarm to notify maintenance personnel.

[0076] A digital twin model of the industrial environmental control system is constructed, forming a dual-track verification system with the physical equipment. The simulated execution effect of commands by the digital twin is compared with the actual operating state of the physical entity to ensure that the control effect meets expectations, while providing multi-dimensional data support for closed-loop feedback. Based on the physical structure, equipment parameters, and operating logic of the industrial environmental control system, a 1:1 scale digital twin model is constructed, encompassing equipment operation simulation, environmental change simulation, and energy consumption calculation modules. The input of the digital twin model is the decision command of the large model, and the output is the simulated equipment operating state, environmental parameter changes, and energy consumption data. The operating status data of the physical equipment (such as actual temperature, frequency, and energy consumption) is collected in real time by sensors in the edge execution layer. The digital twin model synchronously receives the same decision commands and performs simulation calculations, synchronizing the operating data of the physical entity and the digital twin every 10 seconds to form a dual-track data comparison set. Compare the core indicators of the physical entity and the digital twin (such as ambient temperature, equipment energy consumption, and operating load), calculate the deviation value (the deviation threshold is set to ≤3%). If the deviation is within the threshold, it is determined that the instruction execution effect meets expectations; if the deviation exceeds the threshold, analyze the reasons for the deviation (such as the digital twin model parameters not being updated or the physical equipment being aging), and feed them back to the optimization module.

[0077] Relying on the dual-track verification platform and the perception interface of the edge execution layer, real-time data on the control effect of physical entities is collected. This data is compared with the verification results of the digital twin and the decision objectives of the large model, forming a closed loop of "decision-execution-feedback-optimization." This provides data support for the iterative optimization of the large-scale industrial environmental control time-series energy-saving control model. The control effect data covers equipment operating status data (actual equipment operating parameters, load rate, energy consumption, and operating time), environmental status data (indoor temperature, humidity, uniformity, cleanliness, etc. after control), and target achievement data (energy saving rate, comfort score, and equipment loss rate), comprehensively reflecting the command execution effect. The collected data is uploaded to the industrial time-series database in real time through the edge gateway, and simultaneously synchronized to the feedback data interface of the large model. Data transmission adopts an incremental upload strategy (only uploading data with changes exceeding the threshold) to reduce transmission bandwidth consumption. Data storage is organized according to the "equipment-time-indicator" structure, retaining nearly 6 months of historical data to support trend analysis and model optimization. The collected data on the control effect are compared with the decision objectives of the large model (energy saving rate ≥15%, comfort level ≥8 points). If the objectives are achieved, the current decision logic is maintained; if they are not achieved (energy saving rate only 12%, comfort level 7.5 points), the deviation data (actual energy consumption higher than the predicted value, environmental uniformity not meeting the standard) are fed back to the optimization module of the large model to adjust the model's prediction parameters and decision weights, thereby improving the accuracy of subsequent decisions.

[0078] S106 employs model quantization, structured pruning, and knowledge distillation techniques for model compression. It is containerized based on Kubernetes + TritonInferenceServer and combines dynamic batch processing and elastic scaling mechanisms to ensure millisecond-level decision response and high concurrency support.

[0079] In one implementation, the network structure, parameter scale, and inference computing power requirements of a large-scale industrial environmental control time-series energy-saving control model are processed to extract core optimization indicators for model compression and generate a basic parameter set for lightweight model processing. A comprehensive analysis of the network architecture, parameter scale, and inference performance of the large-scale industrial environmental control time-series energy-saving control model is conducted, extracting core optimization indicators for model compression and formulating a basic parameter set for lightweight processing adapted to industrial field deployment, providing clear goals and standards for subsequent model compression. The hierarchical structure of the large-scale model (including data input layer, feature processing layer, prediction and decision layer, and output layer), the functions of each module (such as the TimeGAN generation module, the improved PatchTST prediction module, and the dual-drive decision module), and the connections between modules are analyzed, clarifying the division between non-core redundant layers and key performance layers. The total number of parameters, the proportion of weighted parameters, and the sparsity distribution of the large-scale model are statistically analyzed to identify redundant parameters with weight values ​​close to 0 and neurons with repeated feature mappings, clarifying the potential space for parameter compression.

[0080] Based on the computing power limits of edge devices in industrial settings (such as the number of CPU cores, GPU memory, and RAM size), the computing power consumption and response time of a single decision instruction in large model inference were tested. Performance targets for the lightweight model were set (such as parameter compression of more than 50% and inference response time ≤ 100 milliseconds). Integrating the above analysis results, a lightweight basic parameter set was generated, including "compression targets, hierarchical processing rules, and performance thresholds," clarifying the compression ratio of each module, the threshold for retained core parameters, and the inference performance achievement standards.

[0081] Model quantization, structured pruning, and knowledge distillation are performed on the model layers, neurons, and weight parameters corresponding to the basic parameter set to achieve full-dimensional lightweight compression of the large-scale industrial environmental control time-series energy-saving control model. The original 32-bit floating-point (FP32) weight parameters are quantized into 8-bit integer (INT8) or 16-bit floating-point (FP16) parameters, sacrificing a small amount of precision for improved storage and computational efficiency. The least mean square error calibration method is used during quantization to reduce accuracy loss. First, statistical analysis is performed on the weight parameters of each module of the large model to determine the parameter distribution range. Then, INT8 quantization is performed on non-core modules (data preprocessing layer, some feature mapping layers), and FP16 quantization is performed on core modules such as prediction and decision-making. Finally, the decision accuracy of the quantized model is tested using an industrial environmental control verification dataset. If the accuracy loss is >3%, the key parameters are reverted to FP32.

[0082] Based on hierarchical processing rules using the basic parameter set, redundant network layers, ineffective neurons, and weakly correlated weights in large models are structurally pruned, preserving core functional components and strongly correlated feature mapping paths to avoid fragmentation of the model structure after pruning. First, sensitivity analysis is used to determine the pruning sensitivity of each network layer, and redundant layers with low sensitivity (such as repeated feature normalization layers) are directly pruned. Then, neurons in each layer are sorted by contribution, and ineffective neurons with a contribution below a threshold (e.g., 0.01) are removed. Finally, sparsity pruning is performed on the weight parameters, deleting weakly correlated weights with an absolute value below 0.005, and the model structure is reconstructed using fully connected layers to ensure structural integrity.

[0083] The compressed large model is adapted to the Kubernetes + TritonInferenceServer deployment framework, performing container image creation, service orchestration, and interface configuration to generate a standardized model containerization deployment system. The compressed large model weight files and configuration files are converted to formats supported by TritonInferenceServer (such as TensorRT and ONNX) to ensure the model can be recognized and loaded by the deployment framework. The input / output interfaces for model inference are configured, with input interfaces adapted to real-time industrial environmental control feature data (such as equipment operating parameters and environmental status data), and output interfaces adapted to decision command formats (such as equipment control parameters and strategy descriptions). Interface caching mechanisms and timeout retry logic are also set up. A base image for model deployment is built based on an Ubuntu 20.04 image, installing dependencies such as Python 3.8, CUDA 11.4, and TritonInferenceServer 2.31.0 to ensure image compatibility. The converted model files, dependency libraries, and configuration scripts are packaged into the base image, using a layered build strategy to reduce image size (such as removing compilation dependencies and compressing log files), while also setting up an image health check mechanism (regularly testing model inference availability).

[0084] Set up a Kubernetes cluster (including 1 master node and 3 worker nodes). The master node is responsible for service scheduling and resource management, and the worker nodes deploy model container instances. Configure the Deployment resource definition, set the number of replicas of the model container (initially 3), and resource limits (4 CPU cores, 8GB memory, 1GB GPU). Use a rolling update strategy to ensure deployment stability. Configure the Service resource (NodePort type) to expose the model inference interface to the industrial field network, and configure Ingress rules to implement interface access control and load balancing.

[0085] Based on the concurrency of decision requests and the timeliness requirements of inference response in industrial settings, a dynamic batch processing scheduling mechanism is built, and node elastic scaling trigger rules are formulated to form a high-availability operation scheduling scheme for large models. Multiple scattered decision requests from the industrial site are batched and input into the model for inference, improving GPU utilization and inference throughput. A maximum batch size is set to avoid response latency. A batch processing trigger threshold is set (e.g., accumulating 5 requests or waiting 20 milliseconds), and the batch size range is 2-16 (dynamically adjusted according to GPU memory). When the request concurrency is <5, real-time single-request inference is performed; when the concurrency is ≥5, the batch processing mechanism is triggered, and the requests are aggregated and then batched for inference.

[0086] Leveraging Kubernetes' HPA (HorizontalPodAutoscaler) feature, the number of model container replicas is dynamically adjusted based on metrics such as CPU utilization, GPU utilization, and request queue length, achieving "scaling up under high concurrency and scaling down under low load." Scaling / scaling trigger thresholds are set (CPU utilization ≥70%, GPU utilization ≥65%, request queue length ≥20). Scaling up adds one replica at a time (maximum 8), and scaling down removes one replica at a time (minimum 2). A 3-minute cooldown period is set for scaling up / scaling to avoid frequent adjustments. A model inference caching mechanism is configured to cache decision results under the same conditions within the last 5 minutes, with a cache hit rate target of ≥30%, reducing redundant inference. A failover mechanism is established so that when a container instance on a worker node fails, Kubernetes automatically transfers requests to healthy instances and triggers a restart of the containers on the failed node.

[0087] Integrating the results of lightweight model compression, containerized deployment system, and operation scheduling scheme, the engineering deployment of a large-scale industrial environmental control time-series energy-saving control model was completed, ensuring the model's millisecond-level decision response and support for high-concurrency requests. The lightweight compressed model, containerized deployment system, and dynamic scheduling mechanism were integrated in the order of "model loading → interface startup → scheduling activation" to complete end-to-end deployment and ensure collaborative work among all stages. Simulating high-concurrency scenarios in industrial settings (e.g., 50 requests / second), the model's inference response time, throughput, concurrency support capability, and decision accuracy were tested to verify whether preset indicators were met. If problems such as excessive response latency or insufficient concurrency support occurred during testing, the dynamic batch processing parameters, scaling thresholds, or model resource limitations were adjusted until performance met the standards.

[0088] S107 is based on a large-scale industrial environmental control time-series energy-saving control model. It compensates for environmental fluctuations through dynamic ROI adjustment and coordinate drift correction models, integrates the Lucas-Kanade optical flow method to estimate the global operating condition change trend, optimizes the model decision accuracy, and finally outputs a target energy-saving control scheme for the environmental control system and equipment operation status evaluation results adapted to the industrial site.

[0089] In one implementation, based on a large-scale industrial environmental control time-series energy-saving control model, a dual-model collaborative mechanism of dynamic ROI adjustment and coordinate drift correction is introduced to accurately compensate for monitoring data deviations caused by industrial environmental fluctuations, thereby achieving precise extraction of effective features of the environmental control operating conditions. According to the fluctuation characteristics of industrial environmental control monitoring data, the effective data analysis area (ROI) is dynamically adjusted, eliminating invalid data caused by environmental fluctuations (such as instantaneous false alarms from sensors and extreme interference data), focusing on the core data range related to the operating conditions. First, based on the digital profile of the industrial environmental control system, an initial ROI range is set (e.g., temperature 20-28℃, humidity 40-60%). Then, the data fluctuation amplitude is monitored in real time. If the data fluctuation exceeds a preset threshold in a certain period (e.g., temperature fluctuation ≥3℃ within 5 minutes), the ROI range is automatically reduced to a reasonable fluctuation range (e.g., 22-26℃). Finally, invalid data outside the ROI is marked and removed, retaining the core effective data.

[0090] To address coordinate drift caused by equipment vibration and sensor aging in industrial settings (such as monitoring point data offset and time axis misalignment), a coordinate calibration algorithm is used to correct data deviations, ensuring data accuracy and temporal consistency. A coordinate drift calibration library is established based on standard data from the industrial environmental control fusion time-series dataset. The deviation between the current monitoring data and the standard data in the calibration library (such as temperature deviation ΔT and time axis offset Δt) is calculated in real time. A linear correction algorithm is then used to correct the drift data (e.g., corrected temperature = measured temperature - ΔT). A dynamic ROI adjustment model eliminates invalid data, while the coordinate drift correction model corrects data deviations. These two mechanisms form a collaborative "screening before correction" mechanism, ultimately outputting accurate and effective characteristics of environmental control conditions (such as equipment operating parameters and environmental status data).

[0091] Integrating the Lucas-Kanade optical flow method with a large-scale model operating condition analysis module, a global operating condition trend dynamic estimation model is constructed. Real-time perception and trend prediction of operating condition changes in industrial environmental control systems are achieved through optical flow field feature calculation. The Lucas-Kanade optical flow method module is connected to the large-scale model operating condition analysis layer. The input receives accurate operating condition data after dual-model compensation, and the output is connected to the large-scale model decision layer, using the operating condition trend estimation results as decision parameters. The optical flow method tracks the dynamic changes in operating condition data (such as temperature gradient changes and energy consumption growth trends), calculates the optical flow field feature vector, quantifies the rate and direction of operating condition changes, and achieves accurate prediction of global operating condition trends.

[0092] First, the compensated and accurate operating condition data is processed into a time-series serialization to construct a sequence of operating condition data frames (e.g., 1 frame every 5 minutes, including features such as temperature, energy consumption, and equipment load). Then, the Lucas-Kanade optical flow method is used to calculate the optical flow field between adjacent data frames and solve the feature point movement vectors (e.g., the movement rate v of temperature feature points and the movement direction θ of energy consumption feature points). Finally, based on the optical flow field feature vectors, the operating condition change trend in the next 1-6 hours is predicted (e.g., temperature gradually increases and energy consumption continues to increase).

[0093] By integrating accurate operating condition data after bias compensation with global operating condition trend estimation results, a dynamic optimization mechanism for the decision-making accuracy of a large model is established. This mechanism adaptively tunes the model's inference parameters to correct decision biases caused by environmental interference. Based on accurate operating condition data and guided by operating condition trend prediction results, the model's inference parameters (such as decision weights and constraint thresholds) are adjusted in real time to ensure that the model's decisions can adapt to dynamic changes in operating conditions and avoid decision biases caused by environmental interference (such as failure to meet energy-saving targets or insufficient comfort). First, a model parameter tuning library is established, containing optimal parameter combinations under different operating condition trends (such as a weight of 0.6 for air conditioning cooling load corresponding to an upward temperature trend and a weight of 0.4 corresponding to a downward temperature trend). Then, the deviation between the model's current decision results and the actual operating condition requirements is compared in real time (such as a predicted energy-saving rate of 12% vs. a target of 15%). Finally, based on the operating condition trend (such as a predicted temperature increase), suitable parameters are called from the tuning library to adjust the model's inference weights and constraint thresholds, correcting decision biases.

[0094] By aligning with industrial site environmental control process requirements and equipment operation constraints, a site-specific adaptation and verification system for the model output solution is constructed. Through multi-dimensional adaptability verification, iterative optimization of the solution is completed, ultimately outputting a target energy-saving control solution for the industrial site's environmental control system and a comprehensive evaluation result of equipment operating status. Verification includes: whether the solution meets the environmental control requirements of industrial production processes (e.g., temperature fluctuation ≤ ±0.5℃ in high-precision machining workshops); whether the equipment control parameters in the solution (e.g., fan frequency, air conditioning temperature) are within the equipment operating thresholds (e.g., fan frequency 0-50Hz); whether the energy saving rate of the solution reaches the preset target (e.g., ≥15%); and whether the solution complies with industrial safety regulations (e.g., humidity ≤65%, preventing equipment from getting damp).

[0095] First, the initial energy-saving control scheme output by the model (such as air conditioning temperature of 24℃ and fan frequency of 42Hz) is input into the digital twin platform for simulation verification; then, based on the simulation results, parameters that do not meet the process requirements are corrected (such as adjusting the temperature from 24℃ to 25℃ to meet the requirements of precision process fluctuations); finally, a small-scale pilot test is conducted on physical equipment to collect control effect data (such as actual energy saving rate of 16% and equipment operating load of 70%), and the final optimization of the scheme is completed.

[0096] Integrating the optimization results of the above steps, a target energy-saving control scheme for the environmental control system adapted to the industrial site is output. Simultaneously, a comprehensive evaluation result of equipment operating status is generated, forming a complete output of "control scheme + evaluation report". The target energy-saving control scheme includes equipment control parameters (such as air conditioning temperature, fan frequency, and fresh air volume), control timing (such as differentiated control during production / non-production periods), and linkage strategies (such as equipment start-stop linkage and dynamic load distribution). The comprehensive evaluation result of equipment operating status includes indicators such as equipment operating load, energy consumption compliance rate, fault risk level, and lifespan loss assessment, clearly defining the equipment operating status and optimization direction.

[0097] In one implementation, such as Figure 2 As shown, this application also provides a system for constructing a large-scale time-series energy-saving control model for industrial environmental control systems, including:

[0098] The industrial environmental control multi-source time-series data acquisition module 201 is used to acquire multi-source time-series data from all dimensions of the industrial environmental control system, including equipment operating parameters, environmental perception data, business production data, and external meteorological data. Real-time data access is achieved through the MQTT / Modbus protocol.

[0099] The Industrial Environmental Control Target Dataset and Digital Profile Construction Module 202 is used to preprocess the collected multi-source time-series data to generate a target time-series dataset, use it as a sample to train the TimeGAN model to expand scarce scene data, and extract multi-dimensional features after fusion and screening to form a digital profile of the industrial environmental control system.

[0100] The Industrial Environmental Control Operating Condition Trend Prediction Module 203 is used to capture long-cycle energy consumption patterns by introducing a block attention mechanism based on the improved PatchTST time series prediction architecture and injecting physical constraints by combining industrial process knowledge graphs. It can predict operating condition fluctuations 1-24 hours in advance and output high-precision energy consumption and environmental status prediction results.

[0101] The industrial environmental control multi-objective decision-making and control module 204 is used to construct a dual-drive decision-making model of fuzzy optimization and reinforcement learning. It uses energy saving rate, comfort and equipment life as multi-objective optimization functions, balances conflicting objectives through Pareto optimization, and dynamically outputs equipment operation parameter control strategies.

[0102] The Industrial Environmental Control Command Edge Execution and Feedback Module 205 is used to build an edge execution layer based on industrial bus and IoT protocol, converting decision commands into physical operations of equipment, and realizing closed-loop feedback through a dual-track verification platform of digital twin + physical entity, and collecting control effect data in real time.

[0103] The Industrial Environmental Control Large Model Engineering Deployment Module 206 is used to compress models using model quantization, structured pruning, and knowledge distillation techniques. It is containerized based on Kubernetes + TritonInferenceServer and combines dynamic batch processing and elastic scaling mechanisms to ensure millisecond-level decision response and high concurrency support.

[0104] The Industrial Environmental Control Large Model Decision Optimization and Result Output Module 207 is used to optimize the model decision accuracy based on the industrial environmental control time-series energy-saving control large model. It compensates for environmental fluctuations through dynamic ROI adjustment and coordinate drift correction, integrates the Lucas-Kanade optical flow method to estimate the global operating condition change trend, optimizes the model decision accuracy, and finally outputs the target energy-saving control scheme of the environmental control system and the equipment operation status evaluation results adapted to the industrial site.

[0105] The computer-readable storage medium provided in the above embodiments of this application and the method for constructing a time-series energy-saving control large model for industrial environmental control systems provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0106] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the time-series energy-saving control large-scale model construction method, electronic device, electronic device, and readable storage medium for industrial environmental control systems are basically similar to the embodiments of the time-series energy-saving control large-scale model construction method for industrial environmental control systems described above, and therefore are described relatively simply. Relevant parts can be referred to in the descriptions of the embodiments of the time-series energy-saving control large-scale model construction method for industrial environmental control systems described above.

Claims

1. A method for constructing a large-scale time-sequence energy-saving control model for industrial environmental control systems, characterized in that, include: Collect multi-source time-series data from industrial environmental control systems, including equipment operating parameters, environmental sensing data, business production data, and external meteorological data. Real-time access and preprocessing are achieved through the MQTT / Modbus protocol to generate the target time-series dataset. Using the target time series dataset as training samples, the TimeGAN generation model is trained to expand the data of industrial fault conditions and extreme environmental scenarios. The fusion and screening rules of the original data and the generated data are formulated. After fusion, time features, state features, environmental features and trend features are extracted from the dataset to form a digital profile of the industrial environmental control system. The improved PatchTST time series prediction architecture is adopted, and a block attention mechanism is introduced to capture long-cycle energy consumption patterns. Combined with the industrial process knowledge graph to inject physical constraints, the system can predict operating condition fluctuations 1-24 hours in advance and output high-precision energy consumption and environmental status prediction results. A dual-driven decision-making model combining fuzzy optimization and reinforcement learning is constructed. With energy saving rate, comfort, and equipment life as multi-objective optimization functions, conflicting objectives are balanced through Pareto optimization, and dynamic output of equipment operation parameter control strategies is generated. An edge execution layer is built based on industrial bus and IoT protocol to transform decision commands into physical operations of equipment. Closed-loop feedback is achieved through a dual-track verification platform of digital twin and physical entity, and real-time data on control effect is collected. Model compression is achieved using model quantization, structured pruning, and knowledge distillation techniques. Containerized deployment is implemented based on Kubernetes + TritonInferenceServer, and dynamic batch processing and elastic scaling mechanisms are combined to ensure millisecond-level decision response and high concurrency support. Based on the large-scale energy-saving control model of industrial environmental control, environmental fluctuations are compensated by dynamic ROI adjustment and coordinate drift correction model. The Lucas-Kanade optical flow method is integrated to estimate the global operating condition change trend, optimize the model decision accuracy, and finally output the target energy-saving control scheme of the environmental control system and the equipment operation status evaluation results adapted to the industrial site.

2. The method as described in claim 1, characterized in that, Using the target time-series dataset as training samples, a TimeGAN generative model is trained to expand industrial fault conditions and extreme environmental scenario data. Rules for fusion and filtering of raw and generated data are established. After fusion, time features, state features, environmental features, and trend features are extracted from the dataset to form a digital profile of the industrial environmental control system, including: In light of the scarcity of data on fault conditions and extreme environmental scenarios in industrial environmental control systems, and the massive data requirements for training large models, this paper uses the target time series dataset as training samples to conduct model training, completes parameter optimization and capability adaptation of the TimeGAN generation model, and generates a data generation model adapted to industrial environmental control scenarios. By combining the actual operating conditions and environmental change patterns of industrial environmental control systems, simulations are performed using the trained TimeGAN generation model to expand scarce data for fault conditions and extreme environmental scenarios, generating expanded time-series data that fits the actual industrial situation. Combining the authenticity verification standards and validity screening rules of industrial environmental control time series data, and formulating fusion screening rules based on the feature matching degree and scenario fit of the original data and generated data, a time series data fusion screening system specifically for industrial environmental control is generated. Based on the established fusion and screening system, the target time series raw data and the generated extended data are integrated and precisely screened to remove invalid and distorted data, and generate an industrial environmental control fusion time series dataset. Combining the feature extraction requirements of time-series energy-saving control in industrial environmental control systems and the dimensional requirements of digital profile construction, feature mining and extraction are performed from the fused time-series dataset to generate a multi-dimensional feature set including time features, state features, environmental features, and trend features. Based on the extracted multi-dimensional feature set, feature modeling and systematic integration are carried out by combining the operation logic and feature correlation of the industrial environmental control system to generate a digital profile of the industrial environmental control system that can accurately represent the system's operating status.

3. The method as described in claim 1, characterized in that, A dual-driven decision-making model combining fuzzy optimization and reinforcement learning is constructed. With energy saving rate, comfort, and equipment lifespan as multi-objective optimization functions, Pareto optimization is used to balance conflicting objectives, dynamically outputting equipment operation parameter control strategies, including: Feature extraction and dimension normalization are performed on the energy consumption and environmental state prediction results output by the large-scale prediction module of industrial environmental control time-series energy saving control to generate the basic dataset for the input of the decision model. The decision model input dataset, industrial environmental control system operation constraints, and equipment operation thresholds are correlated and matched to generate a multi-objective optimization decision matrix that includes energy-saving, comfort, and equipment lifespan dimensions. The multi-objective optimization decision matrix is ​​input into the fuzzy optimization + reinforcement learning dual-drive decision model for multi-objective collaborative computation to generate optimization rules for the operating parameters of industrial environmental control system equipment, including fan frequency adjustment logic, air conditioning temperature setting strategy, and equipment start-stop linkage mechanism. Based on the results of multi-objective optimization decision matrix calculation and the Pareto optimization mechanism triggered by equipment operating parameter optimization rules, the conflicting objectives of each optimization dimension are balanced, dynamic equipment operating parameter control strategies are generated, and the decision-making capability of the large-scale industrial environmental control time-series energy-saving control model is completed.

4. The method as described in claim 1, characterized in that, Model compression is achieved using model quantization, structured pruning, and knowledge distillation techniques. Containerized deployment is implemented based on Kubernetes + TritonInferenceServer, combined with dynamic batch processing and elastic scaling mechanisms to ensure millisecond-level decision response and high concurrency support, including: The network structure, parameter scale, and inference computing power requirements of the large-scale industrial environmental control time-series energy-saving control model are processed, the core optimization indicators of model compression are extracted, and a basic parameter set for lightweight model processing is generated. Model quantization, structured pruning, and knowledge distillation are performed on the model layers, neurons, and weight parameters corresponding to the basic parameter set to achieve full-dimensional lightweight compression of the large-scale industrial environmental control time-series energy-saving control model. The compressed large model is adapted to the Kubernetes+TritonInferenceServer deployment framework to create container images, orchestrate services and configure interfaces, and generate a standardized model containerization deployment system. Based on the concurrency of decision requests and the timeliness requirements of inference response in industrial sites, a dynamic batch processing scheduling mechanism is built, and node elastic scaling triggering rules are formulated to form a high-availability operation scheduling scheme for large models. By integrating the results of lightweight model compression, containerized deployment system and operation scheduling scheme, the engineering deployment of the large-scale industrial environmental control time-series energy-saving control model is completed, ensuring the model's millisecond-level decision response and support for high-concurrency requests.

5. The method as described in claim 4, characterized in that, Based on a large-scale industrial environmental control time-series energy-saving control model, this paper compensates for environmental fluctuations through dynamic ROI adjustment and coordinate drift correction models. It integrates the Lucas-Kanade optical flow method to estimate the global operating condition change trend, optimizes the model's decision-making accuracy, and ultimately outputs a target energy-saving control scheme for the environmental control system adapted to the industrial site, as well as equipment operating status evaluation results, including: Based on the large-scale industrial environmental control time-series energy-saving control model, a dual-model collaborative mechanism of dynamic ROI adjustment and coordinate drift correction is introduced to accurately compensate for the monitoring data deviation caused by industrial environmental fluctuations, and to accurately extract the effective features of environmental control conditions. By integrating the Lucas-Kanade optical flow method with a large-scale model operating condition analysis module, a global operating condition trend dynamic estimation model is constructed. Through optical flow field feature calculation, real-time perception and trend prediction of operating condition changes in industrial environmental control systems are achieved. By integrating the accurate operating condition data after deviation compensation with the global operating condition trend estimation results, a dynamic optimization mechanism for the decision-making accuracy of the large model is established to adaptively adjust the model inference parameters and correct the decision-making deviation caused by environmental interference. By aligning with the environmental control process requirements and equipment operation constraints of industrial sites, a field adaptation and verification system for model output solutions is constructed. Through multi-dimensional adaptability verification, the solution is iteratively optimized, and finally, the target energy-saving control solution of the environmental control system adapted to the industrial site and the comprehensive evaluation results of equipment operation status are output.

Images