Multivariable time sequence prediction and dynamic optimization method and device

Abstract

The invention relates to a multivariable time sequence prediction and dynamic optimization method and device, which are applied to real-time data prediction and control optimization in an industrial process. Time sequence data of a plurality of sensor modules are collected, deep learning models such as a long and short term memory network model are adopted for real-time prediction, and parameters of a control system are dynamically adjusted, so that production efficiency is improved, resource use is optimized, and energy consumption is reduced. The method is suitable for wide industrial processes and has high application value.

Description

Technical Field

[0001] This patent application belongs to the field of industrial process control technology, and more specifically, relates to a prediction and optimization method and apparatus based on time-series data, applied to the dynamic adjustment and optimization control of industrial processes. This method and apparatus can be used to process multi-source heterogeneous time-series data, improving the efficiency and quality of the production process through real-time prediction and dynamic adjustment of system parameters. Background Technology

[0002] With the development of industrial automation, more and more industrial processes involve the acquisition and processing of various real-time monitoring data. This data is typically time-series data, including multi-dimensional information such as temperature, pressure, flow rate, and composition. Existing prediction and control methods often rely on static rules or simple regression models, lacking the ability to comprehensively analyze and process multivariate time-series data in real time. How to effectively combine time-series data for prediction and dynamically optimize control parameters in the production process has become a pressing technical problem in the current industrial field. Summary of the Invention

[0003] The technical problem this invention aims to solve is to provide a multivariate time-series prediction and dynamic optimization method and apparatus, capable of predicting multidimensional time-series data in industrial processes in real time, and dynamically adjusting control system parameters based on the prediction results, thereby optimizing production efficiency, improving product quality, and reducing energy and material consumption. Specifically, the technical solution of this invention achieves accurate prediction and dynamic adjustment of key process parameters in industrial process control, enabling the production process to adapt to different operating conditions and achieve high efficiency, energy saving, and optimized resource utilization.

[0004] To solve the above problems, the technical solution adopted by the present invention is as follows: A multivariate time series prediction and dynamic optimization method includes the following steps: Step 1: Data Acquisition and Sensor Module Configuration In industrial production processes, various sensors, such as temperature sensors, pressure sensors, and flow sensors, are used. These sensors can collect data on various physical or chemical quantities in real time and transmit the data to a data processing system. To ensure the accuracy and timeliness of the data, the first step of this invention is to simultaneously collect time-series data from multiple dimensions using multiple sensor modules. The collected data may include, but is not limited to, information on temperature, pressure, flow rate, concentration, and composition. All collected sensor data will be aggregated in a central data processing unit for further preprocessing and analysis.

[0005] Step 2, Data Preprocessing Raw data often comes from different sensors at different times, which may lead to issues such as missing data, data noise, or inconsistent data timing, necessitating data preprocessing. During preprocessing, data from different sensors is time-aligned to ensure synchronous comparison across the same time dimension. Normalization techniques are then employed to ensure that data with different dimensions and ranges can be input and compared within the same model, ensuring the data meets the input requirements of subsequent deep learning models (predictive models).

[0006] Step 3: Model Training and Optimization The preprocessed data is used to train a deep learning prediction model. A Long Short-Term Memory (LSTM) network model is used as the prediction model to effectively capture long-term dependencies in time series data. The input to the prediction model is past time series data, and the output is the key parameters for prediction.

[0007] During model training, the input data is typically historical data from the past 60 minutes. This data serves as input features, helping the predictive model to forecast key parameters such as temperature, pressure, and flow rate for future time periods. Based on the changing patterns of historical data, the Long Short-Term Memory (LSTM) network model learns to identify potential temporal features and uses these features to predict future production processes. During training, the model continuously adjusts its internal parameters to reduce prediction errors.

[0008] Step 4: Model Lightweighting and Deployment A well-trained Long Short-Term Memory (LST) network model contains a large number of parameters and computational requirements. To adapt to the demands of real-time prediction and industrial environments, the trained LST network model undergoes a lightweighting process, which involves converting it to the ONNX (Open Neural Network Exchange) format. Using ONNX, the LST network model is transformed into a lightweight version suitable for edge devices (also called edge computing units or edge computing devices). This reduces the computational burden and storage pressure on the model. The lightweight LST network model can then be converted to a lightweight format for rapid inference in edge computing devices or industrial control equipment.

[0009] Once deployed to edge computing units, the predictive model can directly perform inference calculations on the real-time collected data to generate prediction results. This process of generating prediction results no longer relies on cloud servers, which can greatly reduce data transmission latency and improve response speed.

[0010] Step 5, Dynamic Optimization Control Based on the prediction results of the Long Short-Term Memory (LSTM) network model, the control system dynamically adjusts the operating parameters of the equipment according to the real-time predicted key parameters. For example, when an impending temperature increase or pressure change is predicted, the control system can automatically adjust key process parameters such as the speed, input, and flow rate of the industrial equipment to ensure the stable operation of the industrial process. In this process, the control system optimizes the industrial production process by adjusting the operating status of the industrial equipment.

[0011] include: Production efficiency optimization: Adjust process parameters based on forecast results to improve production speed or efficiency.

[0012] Energy conservation and emission reduction: Adjust energy consumption based on real-time forecasts to reduce unnecessary energy waste.

[0013] Quality control: By dynamically adjusting control parameters, we ensure the stability and consistency of the final product's quality.

[0014] Step Six: System Integration and Upper-Layer Data Interaction The prediction results in the method described in this invention directly affect the adjustments of the real-time control system, and are transmitted to the upper-level system via industrial communication protocols. These results can help the production scheduling system optimize production plans, or assist the quality control system in monitoring product quality, thereby achieving production scheduling optimization or quality control.

[0015] Through the flow of data, prediction and control in industrial processes can be seamlessly integrated with upper-level management systems. Production schedulers, engineers, and management can make decisions based on real-time predictions and dynamic optimization results, improving the flexibility and responsiveness of the overall production process.

[0016] Due to the adoption of the above technical solution, the beneficial effects achieved by this invention are: Real-time performance and efficiency: This invention uses edge computing and a lightweight long short-term memory network model, which enables real-time data prediction and control optimization in industrial settings, reducing data transmission latency and improving control response speed.

[0017] High accuracy: By learning from historical data through a long short-term memory network model, this invention can accurately capture the potential patterns in time-series data and provide accurate predictions for the production process.

[0018] Energy conservation and resource optimization: The combination of real-time forecasting and dynamic control helps optimize energy use and material consumption in industrial processes, thereby achieving the goal of energy conservation and emission reduction.

[0019] Wide applicability: This method is not only applicable to the prediction and optimization of typical industrial parameters such as temperature, pressure, and flow rate, but can also be extended to the prediction and control of other multi-dimensional time series data, and has broad application potential. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention.

[0021] Figure 2 This is a connection block diagram of the device of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the embodiments.

[0023] like Figure 1 This invention provides a multivariate time series prediction and dynamic optimization method, which can achieve accurate prediction and control of process parameters in practical industrial applications. Specifically, this invention can be applied to multiple industries such as chemical, steel, energy, and power. The following is a specific embodiment of this invention.

[0024] Step 1, Data Collection In industrial production processes, multiple sensors are used to collect multidimensional time-series data in real time. These sensors include, but are not limited to: Temperature sensor: Used to measure the temperature in pipes, equipment or reactors in real time.

[0025] Pressure sensors: used to monitor pressure changes in equipment and pipelines.

[0026] Flow sensor: Used to measure material flow rate.

[0027] Composition sensors: used to monitor the chemical composition of raw materials or products, such as concentration or proportion of substances.

[0028] These sensors can transmit analog or digital signals, which are then transmitted to a central processing system via a data acquisition module. The data acquisition module can include an embedded system or an industrial control system. The acquired data is input into the central data processing unit in the form of time-series data, providing raw data for subsequent processing and modeling.

[0029] Step 2, Data Preprocessing The raw data collected often suffers from issues such as inconsistent timing, noise, and missing values, necessitating data preprocessing. Data preprocessing includes the following two main steps: 1) Time Alignment: Since different sensors may have different update frequencies and data acquisition times, time alignment is necessary. Time alignment ensures that data from all sensors are arranged in chronological order, placing all sensor data at the same moment in the same time dimension. If data from a particular sensor is missing or not acquired in a timely manner, interpolation can be used to supplement the missing values.

[0030] 2) Data normalization: Since data collected by different sensors may have different dimensions, and the numerical ranges of data such as temperature, pressure, and flow rate are different, data normalization is required to convert all data to a uniform standard range (between 0 and 1) so that it can be input into the deep learning model for training.

[0031] Step 3: Model Training and Optimization After data preprocessing, the model training phase begins. A Long Short-Term Memory (LSTM) network model is used for prediction of multivariate time-series data. The advantage of the LSM model lies in its ability to effectively handle long-term dependencies in time-series data, making it suitable for processing time-series data from industrial processes. The input to the prediction model is past time-series data, and the output is the key parameters for prediction.

[0032] 1) Input data selection The normalized data is divided into time windows, typically using data from the past 60 minutes as input features to predict key parameters such as temperature, pressure, and flow rate in the future.

[0033] 2) Training process Historical data is used to train the Long Short-Term Memory (LSTM) network model. During training, the input data includes multiple time-series variables such as temperature, pressure, and flow rate, and the target output is the predicted value for future times. Through training, the LSM network model learns the temporal patterns in the data and establishes a non-linear relationship between input and output.

[0034] During training, backpropagation is used for optimization, adjusting model parameters through error backpropagation. A loss function is used to measure the difference between predicted and actual values ​​until the model converges. The loss function includes, but is not limited to, the mean squared error (MSE) function.

[0035] Model Evaluation and Validation: The trained model is evaluated using cross-validation to ensure it can effectively predict future data and exhibits good generalization ability across different scenarios. During the evaluation process, a subset of data is used as a validation set to test the model's performance.

[0036] Step 4, Model Optimization To improve prediction accuracy, hyperparameter optimization techniques are used to adjust the hyperparameters in the Long Short-Term Memory network model, thereby further enhancing the model's performance.

[0037] 1) Model lightweighting and deployment Long Short-Term Memory (LST) network models, once trained, typically contain a large number of parameters and computational requirements. Performing inference directly on industrial equipment would strain storage and computing resources. Therefore, the trained models are lightweighted to adapt to the hardware conditions of the field.

[0038] 2) Lightweight conversion Convert the trained Long Short-Term Memory (LSTM) network model to the ONNX (Open Neural Network Exchange) format. Using ONNX, the LSM model is converted into a lightweight version suitable for edge computing devices.

[0039] 3) Computational Deployment The lightweight model is deployed to edge computing devices on-site. These devices can directly process and predict real-time data collected on-site, reducing latency and bandwidth consumption in cloud computing.

[0040] Step 5, Dynamic Optimization Control After the model is deployed, the control system will be dynamically optimized based on real-time data and model predictions. Real-time predicted key parameters such as future temperature, pressure, and flow rate will serve as inputs to the control system, automatically adjusting the operating parameters of the industrial equipment to achieve process optimization.

[0041] 1) Automatically adjust equipment parameters Based on the predictions from the Long Short-Term Memory (LSTM) network model, the control system automatically adjusts key parameters in the production process. For example, when it is predicted that the temperature of a piece of equipment is about to exceed the predetermined range, the control system will automatically adjust operating parameters such as the equipment's rotation speed, input quantity, or flow rate to ensure stable production.

[0042] 2) Optimization Objective Optimization goals include improving production efficiency, reducing energy consumption, optimizing material usage, and ensuring product quality. The control system can respond to changes in real time and automatically adjust parameters in the production process to achieve process optimization.

[0043] Step Six: System Integration and Data Interaction To enable integration with higher-level management systems, this invention provides a system integration and data interaction method that ensures real-time prediction and optimization results can be shared with the upper-level system, thereby achieving production scheduling optimization or quality control.

[0044] 1) Data Upload and Interaction The forecast results and optimization suggestions are transmitted to the upper-level production scheduling system or quality control system via industrial communication protocols. The upper-level system then adjusts the production plan, controls the production pace, and optimizes the quality inspection process based on the real-time data.

[0045] 2) Upper-level system decision support The upper-level system can obtain real-time prediction and optimization results, and can also perform macro-level scheduling and optimization of process parameters. Using this predictive data, production schedulers can adjust production line parameters in a timely manner, avoiding downtime or instability during production.

[0046] Example 1: Multivariate Time Series Prediction and Dynamic Optimization in Blast Furnace Production Process Step 1, Data Collection During blast furnace production, time-series data of key process parameters are collected through multiple sensor modules. The main data includes: Temperature sensor: measures temperature changes inside the furnace, especially the temperature in the furnace core and top areas.

[0047] Pressure sensor: Used to monitor the pressure inside the blast furnace and ensure that the pressure inside the furnace is kept within a reasonable range.

[0048] Airflow sensor: measures the airflow of the air supply system in order to regulate the combustion process.

[0049] Gas flow sensor: Real-time monitoring of gas flow during furnace charge reaction.

[0050] Material layer height sensor: Monitors changes in the material layer inside the blast furnace to ensure the stability of the furnace charge.

[0051] Step 2, Data Preprocessing The collected raw data underwent time alignment and normalization. Data from different sensors had inconsistent timeframes, necessitating synchronization for unified input into the model for analysis. Subsequently, Min-Max normalization was used to standardize the numerical ranges of each sensor, facilitating model training.

[0052] Step 3: Model Training and Optimization The LSTM model is trained on historical data. The input to the model is multidimensional time-series data from the past 60 minutes, and the output is the predicted values ​​of key parameters such as temperature, pressure, and airflow for a certain period of time in the future (e.g., 30 minutes). During training, the LSTM model can learn the complex relationships between parameters such as temperature and pressure, as well as their trends over time.

[0053] Step 4: Model Deployment and Real-Time Inference The trained LSTM model is converted to ONNX format and deployed to an edge computing unit. This edge computing device will process the collected time-series data in real time and predict key parameters such as future furnace temperature and pressure.

[0054] Step 5, Dynamic Optimization Control Based on model predictions, the control system dynamically adjusts the blast furnace's operating parameters. When the predicted temperature reaches a set threshold, the control system automatically adjusts the blast air volume and gas flow rate to optimize the combustion process and prevent excessively high or low temperatures from affecting the quality of the molten iron. Real-time adjustment of the burden height ensures stable furnace charge levels.

[0055] Step Six: System Integration and Data Interaction The blast furnace control system interacts with the upper-level production scheduling system via the OPC protocol. Production schedulers can obtain real-time forecasts of key parameters such as temperature and pressure, further optimizing production plans, avoiding excessive equipment operation or shutdowns, and ensuring the stable operation of the blast furnace.

[0056] The forecast results here directly affect the adjustment of the control system, and are transmitted to the upper-level system through industrial communication protocols. In other words, these forecast results are used to help the production scheduling system optimize production plans or assist the quality control system in monitoring product quality. Implementation effect

[0057] Improved production efficiency: By predicting key parameters, over-operation or unstable states of equipment are avoided, thus improving the efficiency of in-furnace reactions.

[0058] Energy conservation and emission reduction: Dynamically adjusting parameters such as air volume and gas flow reduces energy waste.

[0059] Improved process stability: By adjusting operating parameters in real time, the temperature and pressure during blast furnace operation are kept within the optimal range, thus avoiding production accidents and raw material waste.

[0060] Example 2: Multivariable Time-Series Prediction and Dynamic Optimization of Chemical Reactors Step 1, Data Collection In a chemical reaction vessel, the following parameters are collected in real time by installing multiple sensors: Temperature sensor: Monitors the temperature inside the reactor to ensure that the chemical reaction proceeds within the optimal temperature range.

[0061] Pressure sensor: Monitors the pressure inside the reactor to prevent excessively high or low pressure from affecting reaction efficiency or safety.

[0062] Stirring speed sensor: measures the stirring speed of the reactor to ensure that the reactants are fully mixed.

[0063] Concentration sensor: Real-time monitoring of reactant or product concentration to ensure reaction progress and material conversion rate.

[0064] Step 2, Data Preprocessing The raw data collected from the reactor is time-aligned to ensure that data from all sensors can be compared at the same point in time. Min-Max normalization is used to convert data of different dimensions into a unified standardized range.

[0065] Step 3: Model Training and Optimization The processed data is used to train the LSTM model. The input data is typically multi-dimensional time-series data from the past hour, including parameters such as temperature, pressure, stirring speed, and concentration. The output is the predicted values ​​for the next 30 minutes. By learning the temporal relationships between the reactor parameters, the LSTM model can accurately predict the future reaction state.

[0066] Step 4: Model Deployment and Real-Time Inference The trained LSTM model is converted into a lightweight format and deployed to the edge computing unit of the reactor. This unit acquires various data from the reactor in real time, performs inference calculations through the model, and predicts key parameters such as future temperature, pressure, and stirring speed.

[0067] Step 5, Dynamic Optimization Control Based on the prediction results, the reactor's control system adjusts the operating parameters in real time. When the temperature is predicted to be too high, the system automatically reduces the heating power; when the reactant concentration reaches the set value, the system can adjust the stirring speed or the amount of gas introduced to control the reaction rate, thereby achieving precise control.

[0068] Step Six: System Integration and Data Interaction The reactor control system uploads forecast results to the upper-level monitoring system via the OPC protocol, helping operators optimize production scheduling and the reaction process. The upper-level system adjusts the production plan based on real-time data to ensure optimal product quality and production efficiency.

[0069] As can be seen, the following numerous implementation effects can be obtained by utilizing the method of the present invention: Improved production efficiency: By accurately predicting the reaction state and optimizing control in real time, the reaction rate and material conversion rate have been improved.

[0070] Product quality stability: Optimized reaction temperature, pressure and other parameters ensured the consistency and stability of product quality.

[0071] Energy saving and consumption reduction: By dynamically adjusting reaction parameters, energy waste and raw material loss are reduced.

Claims

1. A multivariate time series prediction and dynamic optimization method, characterized in that... Includes the following steps: S1. Data acquisition and sensor module configuration: Multiple sensor modules simultaneously and in real time acquire time-series data from multiple dimensions. All acquired time-series data are aggregated into the central data processing unit for subsequent processing. S2. Data preprocessing: Time alignment and normalization of data from different sensors. Time alignment ensures that data from different sensors can be synchronously compared on the same time dimension. Normalize the data to ensure that data with different dimensions and numerical ranges can be input and compared in the same model, and to ensure that the data can meet subsequent input requirements; S3. Model Training and Optimization: The preprocessed data is input into the prediction model to train the prediction model for deep learning. During the training process, optimization algorithms are used to optimize and adjust the model parameters, and the trained prediction model is evaluated. The input of the prediction model is past time series data, and the output is the key parameters for prediction. S4. Model Lightweighting and Deployment: The trained prediction model is lightweighted to reduce the computational burden and storage pressure on the prediction model. After the lightweight prediction model is deployed to edge devices, the edge devices can directly perform inference calculations on the real-time data collected on-site, reducing the latency and bandwidth consumption of cloud computing. S5. Dynamic optimization control: Based on real-time data and the prediction results of the prediction model, the control system dynamically adjusts the operating parameters of industrial equipment according to the key parameters predicted in real time, thereby optimizing the industrial production process. S6. System integration and upper-level data interaction ensure that real-time prediction and optimization results can be shared with the upper-level system to achieve production scheduling optimization or quality control.

2. The multivariate time series prediction and dynamic optimization method according to claim 1, characterized in that: In S1, the collected time-series data includes, but is not limited to, one or more of temperature, pressure, flow rate, concentration, and composition.

3. The multivariate time series prediction and dynamic optimization method according to claim 2, characterized in that: In S2, normalization refers to transforming all data to a uniform standard range between 0 and 1. Normalization includes, but is not limited to, Min-Max normalization.

4. The multivariate time series prediction and dynamic optimization method according to claim 2, characterized in that: In S3, the prediction model uses a long short-term memory network model to effectively capture long-term dependencies in time series data.

5. The multivariate time series prediction and dynamic optimization method according to claim 4, characterized in that: In S3, backpropagation is used for optimization during training. The model parameters are adjusted by backpropagating the error, and the loss function is used to measure the difference between the predicted value and the actual value until the prediction model converges. The loss function includes, but is not limited to, the mean squared error function. The trained prediction model is evaluated by cross-validation to ensure the prediction results and generalization ability of the prediction model in different scenarios.

6. The multivariate time series prediction and dynamic optimization method according to claim 5, characterized in that: In S4, "lightweighting the trained prediction model" refers to converting the trained Long Short-Term Memory network model into ONNX format; edge devices include, but are not limited to, one or more industrial control computers and embedded systems.

7. The multivariate time series prediction and dynamic optimization method according to claim 1, characterized in that: In S5, the algorithms selected for dynamic optimization control include, but are not limited to, PID control and fuzzy control. Specifically, dynamic optimization control means that when a temperature rise or pressure change is predicted, the control system can automatically adjust the key process parameters of industrial equipment, such as speed, input quantity, and flow rate, to ensure the stable operation of the industrial production process.

8. The multivariate time series prediction and dynamic optimization method according to claim 1, characterized in that: In S6, industrial communication protocols include, but are not limited to, one or more of OPC, Modbus, TCP / IP, and MQTT.

9. A multivariate time series prediction and dynamic optimization device, utilizing the method described in any one of claims 1-8, characterized in that... include: Multiple sensor modules are used to collect time-series data from multiple dimensions in industrial processes. The data processing module is used for data preprocessing; The prediction module contains a prediction model that performs real-time predictions on the preprocessed data. The lightweighting and deployment module is used to perform lightweighting and deployment of the trained prediction model; The dynamic optimization control module enables the control system to dynamically adjust the operating parameters of industrial equipment based on prediction results. The integration and interaction module is used to share real-time prediction and optimization results with the upper-level system to achieve production scheduling optimization or quality control.

Images