Intelligent scheduling and energy consumption prediction system for electrical steel strip rolling process

Abstract

The application relates to the technical field of steel production, and discloses an intelligent scheduling and energy consumption prediction system for an electrical steel strip rolling process, which comprises a data acquisition module, a dynamic scheduling module, an energy consumption prediction module, an anomaly detection module and an optimization feedback module. The data acquisition module collects data by using a distributed sensor network; the dynamic scheduling module optimizes scheduling based on an improved genetic algorithm; the energy consumption prediction module predicts energy consumption by using a deep spatiotemporal convolutional neural network; the anomaly detection module detects anomalies based on an isolation forest algorithm; and the optimization feedback module adjusts a scheduling strategy by using a reinforcement learning framework. The system realizes intelligent scheduling and accurate energy consumption prediction of the rolling process, can effectively improve production efficiency, reduce energy consumption, timely find anomalies and optimize a scheduling scheme, and has remarkable economic benefits and production management advantages.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel production, specifically to an intelligent scheduling and energy consumption prediction system for the rolling process of electrical steel strips. BACKGROUND

[0002] In the modern steel production field, the rolling process of electrical steel strips is a key link that determines product quality and enterprise efficiency. With the advancement of industrial intelligence, traditional rolling processes have exposed many problems in scheduling and energy consumption management, making it difficult to meet the growing needs of the industry.

[0003] Traditional rolling process scheduling mainly relies on manual experience or simple rule algorithms, lacking effective response capabilities to complex production environments and multi-objective optimization. In the face of equipment failures, order changes and other sudden situations, it is difficult to quickly adjust the scheduling scheme, resulting in low production efficiency and low equipment utilization. For example, when scheduling rolling tasks, the load capacity of the equipment and the sequence between processes are often not fully considered, causing excessive wear of the equipment or long waiting time for the process, wasting a lot of production resources.

[0004] In terms of energy consumption management, due to the lack of accurate prediction methods, enterprises are difficult to plan energy supply and optimize production processes in advance. Traditional energy consumption prediction methods are often based on simple statistical models or empirical formulas, and cannot accurately capture the impact of various complex factors on energy consumption during the rolling process. This makes the enterprise have a lot of blindness in energy procurement and use, either because of insufficient energy reserves affecting production progress, or because of energy surplus causing waste, increasing production costs.

[0005] In data collection and processing, the previous technical means also have obvious defects. The layout of sensors is not reasonable, and key data in the rolling process cannot be fully and accurately obtained. Moreover, the collected data lack effective integration and analysis, and a large amount of valuable information is idle, which cannot provide strong support for production decisions.

[0006] In addition, with the intensification of market competition, customers have increasingly high requirements for the quality and delivery period of electrical steel strips. Enterprises need to improve production efficiency and reduce energy consumption while ensuring product quality to enhance their market competitiveness. However, the existing rolling process scheduling and energy consumption management technology cannot meet these strict requirements, and an innovative intelligent system is urgently needed to solve these problems and achieve efficient, energy-saving and intelligent production in the rolling process. SUMMARY

[0007] The purpose of the present application is to provide an intelligent scheduling and energy consumption prediction system for the rolling process of electrical steel strips to solve the problems raised in the background art.

[0008] In order to achieve the above object, the present application provides the following technical scheme: an electric steel strip rolling process intelligent scheduling and energy consumption prediction system, the system comprises:

[0009] A data acquisition module is configured to acquire temperature, pressure, speed and energy consumption data in the rolling process in real time through a distributed sensor network, the sensor network comprises a temperature sensor array, a pressure sensor array and a speed encoder, and each sensor is embedded in a key node of a rolling mill device in a spatially gridded distribution manner;

[0010] A dynamic scheduling module is configured to perform multi-objective optimization scheduling on the rolling process based on an improved genetic algorithm, the improved genetic algorithm comprises process priority encoding, dynamic resource constraint processing and a cross-variation probability self-adaptive adjustment mechanism;

[0011] An energy consumption prediction module is configured to perform multi-scale feature extraction on the acquired time series data using a deep spatio-temporal convolutional neural network, and predict the energy consumption value of a future process stage;

[0012] An anomaly detection module is configured to detect data anomalies in the rolling process in real time based on an isolation forest algorithm, and dynamically update an anomaly judgment threshold value in combination with a sliding window technology;

[0013] An optimization feedback module is configured to dynamically adjust a scheduling strategy through a reinforcement learning framework, the framework comprises state space modeling, reward function design and a policy gradient optimization mechanism.

[0014] Preferably, in the data acquisition module, the temperature sensor array is arranged in the following manner: miniature thermocouples are embedded on the surface of the roll along the axial and circumferential directions at equal intervals, and temperature data is uploaded to an edge computing node in real time through a wireless transmission protocol.

[0015] Preferably, in the dynamic scheduling module, the improved genetic algorithm specifically comprises the following steps:

[0016] An initial population is generated based on process priority, wherein the chromosome encoding adopts a binary and real number hybrid encoding;

[0017] A dynamic penalty function is introduced to handle device load over-limit and process conflict constraints;

[0018] An analog annealing strategy is adopted to adaptively adjust the cross and variation probabilities, and the calculation formula is:

[0019]

[0020] Wherein, P c is the cross probability after dynamic adjustment, P c0 is the initial cross probability, P m is the variation probability after dynamic adjustment, and P m0is the initial mutation probability, and a and β are attenuation coefficients, and t is the iteration number.

[0021] Preferably, in the energy consumption prediction module, the deep spatio-temporal convolutional neural network comprises the following structure:

[0022] The input layer receives time series data, and extracts local time series features through a one-dimensional convolutional layer.

[0023] The spatio-temporal attention mechanism layer dynamically allocates weights to different sensor data.

[0024] The dilated convolutional layer expands the receptive field to capture long-period dependencies.

[0025] The output layer regresses and predicts the energy consumption value through a fully connected network.

[0026] Preferably, in the anomaly detection module, the improvement of the isolation forest algorithm comprises:

[0027] Based on a sliding window, data subsets are divided, and each subset is independently constructed into an isolation tree.

[0028] Mahalanobis distance is introduced to replace Euclidean distance to eliminate the dimensional difference of sensor data.

[0029] The abnormal threshold is dynamically updated, and the calculation formula is:

[0030] T th = μ score + k·σ score

[0031] where T th is the dynamically updated abnormality judgment threshold, μ score is the mean of the historical abnormality score, σ score is the standard deviation, and k is the dynamic adjustment coefficient.

[0032] Preferably, in the optimization feedback module, the reward function of the reinforcement learning framework is designed as:

[0033]

[0034] where R is the reinforcement learning reward value, E actual is the actual energy consumption, E pred is the predicted energy consumption, T idle is the device idle time, Δ violate is the constraint violation amount, T total is the total scheduling time, and w1, w2, and w3 are weight coefficients.

[0035] Preferably, the dynamic scheduling module further comprises a conflict resolution submodule, configured to generate an alternative scheduling scheme by using a fuzzy logic rule base when a process conflict occurs, wherein the rule base comprises rolling force threshold, temperature tolerance interval, and equipment priority level.

[0036] Preferably, the energy consumption prediction module further comprises a data enhancement submodule, configured to perform the following processing on the input time series data:

[0037] generate noisy extended data through time warping technology;

[0038] synthesize virtual data under extreme working conditions by using a generative adversarial network;

[0039] remove high-frequency noise and retain low-frequency trend components by using wavelet transform.

[0040] Preferably, the anomaly detection module further comprises a false alarm filtering submodule, configured to smooth the anomaly scores output by the isolation forest through a Kalman filter, and eliminate physically infeasible abnormal points in combination with a process knowledge base.

[0041] Preferably, the system further comprises a visual interaction module, configured to display the rolling process state in real time through a three-dimensional virtual twin model, wherein the model integrates sensor data, scheduling schemes, and energy consumption prediction results, and supports multi-dimensional data drilling analysis.

[0042] Compared with the prior art, the present application has the following beneficial effects:

[0043] In terms of data acquisition, the rolling mill equipment key nodes are embedded in a spatially gridded distribution manner through a distributed sensor network, so that temperature, pressure, speed, and energy consumption data can be comprehensively and accurately collected. For example, a temperature sensor array is embedded with miniature thermocouples at equal intervals along the axial and circumferential directions on the surface of the roller and transmits data wirelessly, ensuring the real-time and accuracy of temperature data and providing a reliable basis for subsequent accurate analysis. This enables enterprises to real-time master various parameters of the rolling process and timely discover potential problems, thereby laying a foundation for optimizing the production process.

[0044] The dynamic scheduling module is based on an improved genetic algorithm, has process priority coding, dynamic resource constraint processing and cross variation probability self-adaptive adjustment mechanism, and effectively solves the shortcomings of traditional scheduling methods. It generates an initial population based on process priority, uses binary and real number hybrid coding, improves the search efficiency of the algorithm, introduces a dynamic penalty function to handle device load overrun and process conflict constraints, and ensures the orderly progress of production. The simulated annealing strategy adaptively adjusts the cross and mutation probability, so that the algorithm can better converge to the optimal solution. In addition, the conflict resolution submodule generates a replacement scheduling scheme using a fuzzy logic rule base when there is a process conflict, greatly improving the flexibility and adaptability of production scheduling, reducing idle time, improving equipment utilization, and thus improving overall production efficiency and reducing production costs.

[0045] The energy consumption prediction module uses a deep spatio-temporal convolutional neural network, which can extract multi-scale features from the collected time series data through the collaborative work of one-dimensional convolutional layers, spatio-temporal attention mechanism layers, and output layers, and accurately predict the energy consumption value of the future process stage. The data enhancement submodule further improves the accuracy of the prediction and the generalization ability of the model. This enables enterprises to plan energy supply in advance, arrange production plans reasonably, avoid energy waste, reduce energy costs, and enhance the scientific nature and initiative of energy management.

[0046] The anomaly detection module is based on an improved isolation forest algorithm and dynamically updates the anomaly judgment threshold combined with the sliding window technique, which can timely and accurately detect data anomalies in the rolling process. The false alarm filtering submodule effectively eliminates false alarms through the cooperation of the Kalman filter and the flat process knowledge base, improving the reliability of anomaly detection. This helps enterprises to timely detect equipment failures and process anomalies, take appropriate measures for repair, avoid production accidents, and ensure the continuity of production and product quality.

[0047] The optimization feedback module dynamically adjusts the scheduling strategy through a reinforcement learning framework and optimizes the scheduling scheme in real time according to the actual production situation. The reward function considers factors such as the difference between actual energy consumption and predicted energy consumption, device idle time, and constraint violation amount, guiding the scheduling strategy to develop in a more optimal direction, further improving the performance and efficiency of the system. The visual interaction module displays the rolling process state in real time through a three-dimensional virtual twin model, integrates multiple data and supports multi-dimensional data drilling analysis, providing an intuitive and comprehensive production monitoring interface for operators. This facilitates operators to timely understand production conditions and make accurate decisions, improving the convenience and scientific nature of production management. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A working principle diagram of the intelligent scheduling and energy consumption prediction system for the electrical steel strip rolling process described in the present application;

[0049] Figure 2 Step chart for improving the Isolation Forest algorithm in the anomaly detection module;

[0050] Figure 3 Workflow chart for optimizing the reinforcement learning framework in the feedback module;

[0051] Figure 4 Workflow chart for the conflict resolution submodule in the dynamic scheduling module. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0053] Please refer to Figures 1-4 The present application provides an intelligent scheduling and energy consumption prediction system for the rolling process of electrical steel strips, and the overall implementation scheme is as follows:

[0054] The data acquisition module acquires temperature, pressure, speed and energy consumption data in real time through a distributed sensor network. The sensor network is composed of a temperature sensor array, a pressure sensor array and a speed encoder, and each sensor is embedded in the key nodes of the rolling mill equipment in a spatially gridded distribution. Such distribution can comprehensively and accurately obtain various key data of the rolling mill during operation, providing a reliable basis for subsequent analysis and decision-making. For example, the temperature sensor can monitor the temperature change of the rolling mill during rolling, the pressure sensor can sense the rolling force in real time, and the speed encoder can accurately measure the rolling speed of the steel strip.

[0055] The dynamic scheduling module performs multi-objective optimization scheduling on the rolling process based on an improved genetic algorithm. The improved genetic algorithm has process priority encoding, dynamic resource constraint processing and adaptive adjustment mechanism of crossover and mutation probability. Through these mechanisms, the module can comprehensively consider various factors such as the load capacity of the equipment and the sequence between processes, so as to develop a more reasonable scheduling scheme and improve the overall efficiency of the rolling process.

[0056] The energy consumption prediction module uses a deep spatio-temporal convolutional neural network to extract multi-scale features from the collected time series data and predict the energy consumption value of the future process stage. The deep spatio-temporal convolutional neural network can effectively mine the potential features in the data, accurately capture the change rules of the data in time and space, and thus realize accurate prediction of energy consumption, providing strong support for enterprises to reasonably arrange production and control costs.

[0057] The abnormality detection module detects data abnormalities in the rolling process in real time based on an isolation forest algorithm and dynamically updates the abnormality judgment threshold in combination with a sliding window technology. The module can timely discover abnormal conditions in the rolling process, such as equipment failure, abnormal process parameters, etc., so that the staff can timely take measures for processing, thereby ensuring the smooth progress of production.

[0058] The optimization feedback module dynamically adjusts the scheduling strategy through a reinforcement learning framework. The framework includes state space modeling, reward function design, and policy gradient optimization mechanism. Through continuous learning and optimization, the scheduling strategy can be dynamically adjusted according to the actual production situation, further improving the performance and efficiency of the system.

[0059] The implementation of the present application will be further described below in conjunction with Examples 1 to 5.

[0060] Example 1:

[0061] In this embodiment, the specific implementation details of the data acquisition module and the dynamic scheduling module are further described.

[0062] For the data acquisition module, the arrangement of the temperature sensor array is crucial to the accuracy and reliability of the collected data. Miniature thermocouples are embedded at equal intervals along the axial and circumferential directions of the roll surface. This arrangement can comprehensively and uniformly monitor temperature changes at different positions on the roll surface. The miniature thermocouples can quickly perceive temperature changes and convert temperature signals into electrical signals. Through a wireless transmission protocol, these temperature data can be uploaded to the edge computing node in real time, avoiding the problems of complex wiring and easy damage caused by wired transmission, and facilitating the rapid processing and transmission of data.

[0063] In the dynamic scheduling module, the improved genetic algorithm is the core technology. When generating the initial population based on process priority, the chromosome coding uses a binary and real number hybrid coding. This coding method combines the advantages of binary coding and real number coding. Binary coding is beneficial to crossover and mutation operations in genetic algorithms, while real number coding can more intuitively represent actual parameters such as process priority, improving the search efficiency and accuracy of the algorithm.

[0064] In terms of handling device load overruns and process conflict constraints, a dynamic penalty function is introduced. When the device load exceeds the limit or conflicts occur between processes, the dynamic penalty function will punish the corresponding chromosome according to the actual situation, reducing its competitiveness in the population and prompting the algorithm to search for the optimal solution in the direction of meeting the constraint conditions.

[0065] The adaptive adjustment of crossover and mutation probabilities is another key feature of the algorithm. The simulated annealing strategy is used to adaptively adjust the crossover and mutation probabilities, with the calculation formula being:

[0066]

[0067] P c is the dynamic adjusted crossover probability, which determines the probability of two chromosomes performing crossover operation in the genetic algorithm. Crossover operation can exchange part of the genes between different chromosomes, thereby generating new individuals and increasing the diversity of the population. P c0 is the initial crossover probability, which is the crossover probability value set at the beginning of the algorithm. a is the decay coefficient, which is used to control the decay rate of the crossover probability with the iteration number t. As the iteration number increases, the crossover probability gradually decreases, because at the beginning of the algorithm, a larger crossover probability is needed to explore a wider solution space; while in the later stage of the algorithm, the crossover probability needs to be gradually reduced in order to avoid destroying the better solution found.

[0068] P m is the dynamic adjusted mutation probability, and mutation operation can randomly change some genes in the chromosome to prevent the algorithm from falling into a local optimal solution. P m0 is the initial mutation probability, which is the mutation probability value set at the beginning of the algorithm. β is the decay coefficient, which is used to control the change rate of the mutation probability with the iteration number t. Similar to the crossover probability, the mutation probability is larger at the beginning of the algorithm and gradually decreases later.

[0069] In addition, the dynamic scheduling module also includes a conflict resolution submodule. When there is a process conflict, this submodule generates an alternative scheduling scheme using a fuzzy logic rule base. The rolling force threshold, temperature tolerance interval, and device priority level in the rule base are important basis for developing alternative schemes. For example, when the rolling force exceeds the set threshold and the temperature is outside the tolerance interval, according to the device priority level, the process with lower priority is adjusted first to avoid device damage and ensure product quality. In this way, process conflict problems can be quickly and effectively solved in complex production environments, ensuring the continuity and stability of production.

[0070] Example 2:

[0071] The deep spatio-temporal convolutional neural network of the energy consumption prediction module has a unique structure. The input layer receives the collected time series data, which contains information about the changes of temperature, pressure, speed, etc. over time during the rolling process. The one-dimensional convolutional layer is responsible for extracting local time series features, which slides on the time series data through convolution kernels to perform convolution operation on local data, thereby capturing the trend and features of data changes in a short period of time. For example, it can find the rapid rising or falling trend of temperature in a small period of time, as well as the fluctuation of pressure.

[0072] The spatio-temporal attention mechanism layer is a key part of the network. It can dynamically assign weights to different sensor data. In actual rolling processes, the importance of different sensor data to energy consumption prediction may be different. For example, in some cases, temperature data may have a greater impact on energy consumption; while in other cases, rolling speed data may be more critical. The spatio-temporal attention mechanism layer automatically assigns appropriate weights to different sensor data by learning the characteristics and rules of the data, highlighting the role of important data and improving the accuracy of prediction.

[0073] The dilated convolution layer is used to expand the receptive field to capture long-period dependencies. The receptive field of the traditional convolution layer is limited, making it difficult to capture long-period features in the data. The dilated convolution layer introduces holes in the convolution kernel, allowing the convolution operation to obtain data information in a larger range, thereby enabling the discovery of dependencies in the data on a longer time scale. For example, it can discover the periodic energy consumption change rule that occurs every certain time in the rolling process.

[0074] The output layer uses a fully connected network to regress and predict energy consumption values. The fully connected network comprehensively processes the features extracted by the previous layers and finally outputs the predicted energy consumption values.

[0075] The energy consumption prediction module also includes a data enhancement sub-module. Time warping technology is used to generate noisy extended data. Time warping technology can simulate time series changes that may occur in actual production, such as minor time deviations in data collection, to generate extended data with certain noise, increase data diversity, and improve model generalization ability. Adversarial generative networks are used to synthesize virtual data under extreme working conditions. In actual production, extreme working conditions may rarely occur, but energy consumption prediction under these conditions is equally important. Adversarial generative networks can generate virtual data under extreme temperature, pressure, and other conditions, allowing the model to learn data characteristics under these special conditions and improve the model's prediction ability under extreme conditions. Wavelet transform is used to remove high-frequency noise and retain low-frequency trend components. Wavelet transform can effectively separate high-frequency noise and low-frequency trends in the data, remove noise interference, and allow the model to more accurately learn the essential characteristics of the data, improving prediction accuracy.

[0076] Example 3:

[0077] The anomaly detection module is based on the Isolation Forest algorithm and has been improved in many aspects in practical applications. Data subsets are divided based on a sliding window, and an isolation tree is independently constructed for each subset. The sliding window technique can dynamically select data for analysis. As time passes, the window continuously slides, and new data is included in the analysis range. Each subset independently constructs an isolation tree, which can more carefully capture the local characteristics and distribution of the data. For example, if the rolling speed fluctuates abnormally in a certain period, the isolation tree constructed by dividing the data subsets based on the sliding window can timely detect this abnormal situation.

[0078] Mahalanobis distance is introduced to replace Euclidean distance to eliminate the dimensional differences of sensor data. In actual production, the data collected by different sensors have different dimensions, such as temperature in Celsius and pressure in Pascal. Euclidean distance will produce bias when processing data with different dimensions, while Mahalanobis distance considers the covariance structure of data and can eliminate the influence of dimensional differences, more accurately measure the similarity and difference between data points, and thus improve the accuracy of anomaly detection.

[0079] The dynamic updating of the anomaly threshold is calculated by the formula:

[0080] T th = μ score + k·σ score

[0081] where T th is the dynamically updated anomaly judgment threshold, which is the key indicator for judging whether the data is abnormal. μ score is the mean of the historical anomaly score, which reflects the average level of the anomaly score in the historical data. σ score is the standard deviation, which is used to measure the dispersion of the historical anomaly score. k is a dynamic adjustment coefficient, which is adjusted according to the actual production situation and experience. For example, when the production environment is relatively stable, k can be appropriately reduced to improve the sensitivity of anomaly detection; when the production environment fluctuates greatly, k can be appropriately increased to avoid excessive false positives.

[0082] The anomaly detection module also includes a false positive filtering submodule. The anomaly score output by the Isolation Forest is smoothed by a Kalman filter, which is an optimal linear filter that can use historical information and current observations to optimally estimate the anomaly score, remove noise and fluctuations, and make the anomaly score more stable and reliable. In combination with the process knowledge base, physically impossible abnormal points are removed. The process knowledge base contains various physical constraints and empirical knowledge of the rolling process. For example, according to the process knowledge, a certain temperature and pressure combination is not possible in actual production. If the Isolation Forest algorithm detects such a data point as an anomaly, in combination with the process knowledge base, it can be determined that this is a false positive, so it is removed to improve the reliability of anomaly detection.

[0083] Example 4:

[0084] In the reinforcement learning framework of the optimization feedback module, the design of the reward function plays a key role in the optimization of the scheduling strategy. The reward function formula is:

[0085]

[0086] where R is the reinforcement learning reward value, which is an important indicator to measure the pros and cons of the scheduling strategy. actual is the actual energy consumption, i.e. the real energy consumption in the rolling process, which reflects the energy use under the current scheduling strategy. pred is the predicted energy consumption, which is the energy consumption value predicted by the energy consumption prediction module. By comparing the actual energy consumption and the predicted energy consumption, the accuracy of the prediction can be evaluated, and the control effect of the scheduling strategy on the energy consumption can also be reflected. w1 is the weight coefficient, which is used to adjust the importance of the difference between the actual energy consumption and the predicted energy consumption in the reward function. If you want to pay more attention to the accuracy of the energy consumption prediction, w1 can be appropriately increased; if you pay more attention to other factors, w1 can be correspondingly reduced.

[0087] T idle is the equipment idle time, T total is the total scheduling time, represents the proportion of equipment idle time in the total scheduling time. Too long equipment idle time will reduce production efficiency, so this item is used to encourage the scheduling strategy to reduce equipment idle time and improve equipment utilization. w2 is the weight coefficient, which is used to control the weight of the proportion of equipment idle time in the reward function.

[0088] Δ violate is the constraint violation amount, ∑Δ violate represents the sum of all constraint violation amounts. In actual production, there are various constraint conditions, such as equipment load limit, process sequence, etc. When the scheduling strategy violates these constraints, constraint violation amount will be generated. w3 is the weight coefficient, which is used to measure the influence degree of constraint violation on the reward function. If you want to strictly follow the constraint conditions, w3 can be set to be larger; if the strictness requirement of the constraint is relatively low, w3 can be appropriately reduced.

[0089] Through this reward function, the reinforcement learning framework can dynamically adjust the scheduling strategy according to the actual production situation. When the actual energy consumption approaches the predicted energy consumption, the equipment idle time is reduced, and the constraint violation amount is reduced, the reward value will increase, indicating that the current scheduling strategy is optimized; otherwise, the reward value will decrease, prompting the algorithm to find a better scheduling strategy, so as to realize efficient scheduling and optimization of the rolling process.

[0090] Example 5:

[0091] The visualization interaction module displays the rolling process status in real time through a three-dimensional virtual twin model. The three-dimensional virtual twin model integrates sensor data, scheduling plans, and energy consumption prediction results, providing an intuitive and comprehensive production monitoring interface for operators.

[0092] Sensor data is transmitted to the three-dimensional virtual twin model in real time, and information such as the temperature of the rolls, the pressure, and the rolling speed of the steel strip is presented in a visual manner. For example, the temperature at different positions of the rolls is displayed through different colors and numerical values, so that operators can immediately see whether there are abnormal conditions such as local overheating of the rolls. Pressure data can be displayed through charts or dynamically changing numerical values, making it easy for operators to understand the size and trend of the rolling force at any time.

[0093] The scheduling plan displays the execution order of the process and the allocation of equipment in an intuitive way in the three-dimensional virtual twin model. Operators can clearly see which device each process is performed on and the sequence between processes. If the scheduling plan changes, the model will be updated in real time, allowing operators to keep abreast of the latest production arrangements.

[0094] Energy consumption prediction results are also integrated into the three-dimensional virtual twin model. Through charts or prediction curves, operators can anticipate the energy consumption of future process stages, allowing them to reasonably arrange energy supply and production plans. For example, if it is predicted that energy consumption will increase significantly at a certain time, operators can adjust equipment parameters or optimize scheduling plans in advance to reduce energy consumption.

[0095] The model also supports multi-dimensional data drilling analysis. Operators can obtain more detailed data information by clicking on various elements in the model. For example, by clicking on a roll, they can view the temperature variation curve of the roll at different time points, pressure data, and related equipment operating status, etc. Through multi-dimensional data drilling analysis, operators can gain a deep understanding of various details in the production process, promptly identify potential problems, and make more accurate decisions, thereby improving the management level and efficiency of production.

[0096] It should be noted that in this text, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include" or "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device.

[0097] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. An intelligent scheduling and energy consumption prediction system for electrical steel strip rolling process, characterized in that, The application relates to a rolling process optimization system, which comprises the following modules: a data acquisition module for collecting temperature, pressure, speed and energy consumption data in a rolling process in real time through a distributed sensor network, wherein the sensor network comprises a temperature sensor array, a pressure sensor array and a speed encoder, and each sensor is embedded in a key node of a rolling mill equipment in a spatially gridded distribution mode; a dynamic scheduling module for multi-objective optimization scheduling of the rolling process based on an improved genetic algorithm, wherein the improved genetic algorithm comprises process priority coding, dynamic resource constraint processing and a self-adaptive adjustment mechanism of crossover and mutation probabilities; an energy consumption prediction module for multi-scale feature extraction of collected time series data by adopting a deep spatio-temporal convolutional neural network, and prediction of energy consumption values in future process stages; an anomaly detection module for real-time detection of data anomalies in the rolling process based on an isolation forest algorithm, and dynamic updating of an anomaly judgment threshold value in combination with a sliding window technology; an optimization feedback module for dynamic adjustment of a scheduling strategy through a reinforcement learning framework, wherein the framework comprises state space modeling, reward function design and a policy gradient optimization mechanism; in the dynamic scheduling module, the improved genetic algorithm specifically comprises the following steps: generating an initial population based on process priorities, wherein binary and real number hybrid coding is adopted for chromosome coding; introducing a dynamic penalty function to process equipment load over-limit and process conflict constraints; adopting a simulated annealing strategy to adaptively adjust the crossover and mutation probabilities, and the calculation formula is as follows: wherein, is the dynamic adjusted crossover probability, is the initial crossover probability, is the dynamic adjusted mutation probability, is the initial mutation probability, , is the decay coefficient, is the iteration number; the dynamic scheduling module further comprises a conflict resolution submodule for generating a replacement scheduling scheme by adopting a fuzzy logic rule base when a process conflict occurs, and the rule base comprises rolling force threshold values, temperature tolerance intervals and equipment priority levels.

2. The system of claim 1, wherein, in the data acquisition module, the temperature sensor array is arranged in the following mode: miniature thermocouples are embedded on the surface of a rolling mill roller in an equidistant manner along the axial and circumferential directions, and temperature data are uploaded to an edge computing node in real time through a wireless transmission protocol.

3. The system of claim 1, wherein, in the energy consumption prediction module, the deep spatio-temporal convolutional neural network comprises the following structure: an input layer receives time series data, and local time series features are extracted through a one-dimensional convolutional layer; a spatio-temporal attention mechanism layer dynamically allocates weights of different sensor data; a cavity convolutional layer expands a receptive field to capture long-period dependencies; and an output layer regresses and predicts energy consumption values through a fully connected network.

4. The system of claim 1, wherein, in the anomaly detection module, the improvement of the isolation forest algorithm comprises the following steps: dividing data subsets based on a sliding window, and independently constructing an isolation tree for each subset; introducing Mahalanobis distance to replace Euclidean distance to eliminate dimension differences of sensor data; dynamically updating an anomaly threshold value, and the calculation formula is as follows: wherein, is a dynamically updated abnormality determination threshold, is a historical abnormality score mean, is a standard deviation, is a dynamic adjustment factor.

5. The system of claim 1, wherein, in the optimization feedback module, the reward function design of the reinforcement learning framework is as follows: wherein, is a reinforcement learning reward value, is an actual energy consumption, is a predicted energy consumption, is a device idle time, is a constraint violation amount, is a total scheduling time, , , is a weight coefficient.

6. The system of claim 3, wherein, the energy consumption prediction module further comprises a data enhancement submodule for the following processing of input time series data: generating extended data with noise through a time warping technology; adopting an adversarial generative network to synthesize virtual data under extreme working conditions; applying wavelet transform to remove high-frequency noise and retain low-frequency trend components.

7. The system of claim 4, wherein, the anomaly detection module further comprises a false alarm filtering submodule for smoothing abnormal scores output by the isolation forest through a Kalman filter, and eliminating physically infeasible abnormal points in combination with a process knowledge base.

8. The system of claim 1, wherein, The system further comprises a visual interactive module for real-time display of the rolling process state through a three-dimensional virtual twin model, the model integrating sensor data, scheduling scheme and energy consumption prediction results and supporting multi-dimensional data drilling analysis.

Images