Energy efficiency-oriented steel plant data flow real-time prediction system and method

Abstract

The energy efficiency-oriented steel plant data flow real-time prediction system and method of the present application, the system comprises: a data acquisition module, a data processing module, a data storage module, an energy consumption visualization module, an energy efficiency calculation module and an energy efficiency prediction module.The present application dynamically displays the energy consumption data flow of each stage and each device based on the predefined steelmaking process, real-time monitors and simulates the energy consumption data flow of the steel plant, and integrates advanced machine learning and data analysis technology for comprehensive statistical analysis, realizes real-time calculation and future prediction of the energy efficiency of the steel plant, and solves the problems of incomplete energy efficiency management and low prediction model accuracy in the prior art.The system can help enterprises reduce energy consumption and improve production efficiency, and provides a new technical path for intelligent cost management and control of industrial production.

Description

Technical Field

[0001] This invention belongs to the field of integrated energy application and big data computing technology, and relates to a real-time data flow prediction system and method for steel plants oriented towards energy efficiency. Background Technology

[0002] Energy and resources, as raw materials for industrial production, occupy an unshakeable position in the industrial production process. Rapid industrial technological development has led to an increasing demand for diversified energy and resources. Statistics show that industrial energy consumption accounts for more than 70% of total social energy consumption, resulting in increasingly prominent problems such as low energy and resource utilization rates, imbalances between energy consumption and ecological protection, and economic growth. With the advancement of Industry 4.0, the manufacturing industry is undergoing an unprecedented transformation. The concepts of intelligent manufacturing and digital factories are constantly being proposed and applied, aiming to improve energy utilization, production efficiency, and product quality through the deep integration of information technology and manufacturing technology. How to improve the energy efficiency of production systems and reduce energy consumption and production costs has become an urgent problem to be solved in industrial development.

[0003] Energy efficiency, as a crucial evaluation indicator in industrial production, not only affects factory operating costs but also directly impacts environmental protection and sustainable development. Traditional industrial systems often struggle to monitor and optimize energy consumption and improve energy efficiency in real time, resulting in low energy utilization. To address this issue, modern industrial systems are leveraging internet technology and gradually introducing real-time data flow simulation technology. Real-time data flow simulation is a method that simulates and optimizes production systems by collecting, processing, and analyzing data flows during the production process in real time. This technology can reflect the dynamic changes of industrial production systems in real time and, by combining predictive and other machine learning techniques, can promptly identify and resolve energy efficiency issues in the production process, thereby achieving efficient energy utilization and production system optimization.

[0004] Patent CN109325641A, entitled "An Industrial Energy Efficiency Management System and Method," discloses an industrial energy efficiency management system including a data acquisition module, a calculation module, and a management module. However, this invention's energy efficiency management is not comprehensive enough; it cannot display real-time simulation data streams of industrial processes and cannot predict the future energy efficiency of the enterprise. Patent CN111832161A, entitled "A Real-Time Simulation Method and System for a Comprehensive Energy System," constructs a coupled simulation model including a heating network, a natural gas network, and a power grid model. This patent application focuses on displaying real-time building energy consumption data in conjunction with meteorological data, and is not suitable for industrial production processes, nor does it address future energy efficiency analysis.

[0005] In energy efficiency prediction research, the authorized invention patent CN114004423A proposes an energy demand prediction method and system that considers regional energy efficiency synergistic optimization. It inputs the regional energy consumption data matrix and the regional GDP matrix into a deep belief network to train features and obtain the predicted energy demand value. The authorized patent CN111874182A introduces an energy efficiency prediction control system and method for hybrid power ships. By establishing an energy efficiency prediction speed model, a power prediction model, and a power dynamic management model, it obtains the suggested speed, power demand, engine output power, and propeller speed for the ship in the next time period, respectively, which facilitates the ship to improve energy utilization and sail economically.

[0006] Energy efficiency prediction and simulation technology is of great significance in modern industry, especially in energy-intensive sectors such as steel mills. By accurately predicting energy usage efficiency, this technology can help companies optimize production processes, reduce energy consumption, and decrease greenhouse gas emissions, thereby achieving sustainable development goals. Furthermore, simulation technology can model the effects of different production strategies and process improvements in a virtual environment, providing a scientific basis for corporate decision-making. This can not only significantly reduce operating costs but also improve production efficiency and equipment utilization, enhancing the company's competitiveness in the market. Therefore, energy efficiency prediction and simulation technology is crucial for promoting industrial energy conservation and emission reduction, and improving resource utilization efficiency. A comprehensive analysis of the above research and published patents reveals the following problems with existing energy efficiency prediction and simulation technologies:

[0007] 1. High latency, high cost, and complex data acquisition technologies. Industrial heterogeneous IoT achieves data acquisition through device interoperability or a unified data format. For example: using gateways to connect heterogeneous networks or protocols requires a large amount of data conversion and transmission, resulting in slow transmission speeds, poor real-time performance, and high maintenance costs for gateway devices; using protocol converters to convert between different communication protocols, but this technology cannot adapt to all protocols, cannot acquire data from devices with specific protocols, and the protocol parsing and conversion calculations of protocol converters are complex, easily leading to low transmission rates and poor performance; using a unified data model to unify various data formats on different devices increases data complexity and management difficulty, resulting in high computational overhead and high acquisition latency.

[0008] 2. The simulation system's functionality is incomplete. Some systems only collect real-time energy consumption data and do not incorporate simulations of dynamic data flow processes in industrial production. This makes it impossible to provide non-professionals with a clear understanding of industrial production processes. At the same time, other systems do not involve energy efficiency calculations, especially future energy efficiency prediction research, and therefore cannot provide factories with energy efficiency references to effectively reduce resource consumption and production costs, which is detrimental to ecological protection and sustainable development.

[0009] 3. Shortcomings in energy efficiency prediction models. In industrial energy efficiency research, energy efficiency prediction models are mainly divided into two categories: traditional dynamic models and data-driven models based on intelligent algorithms. However, traditional dynamic models require professional energy dynamics derivation, are complex to model, and are difficult to meet the modeling requirements under different conditions, making it impossible to obtain high-precision energy efficiency prediction models. Data-driven models based on intelligent algorithms rely on machine learning and deep learning algorithms, which is a black box model. They have strong nonlinear fitting effects, but this model cannot guarantee high model accuracy and strong generalization ability, and the interpretability of this black box model is poor.

[0010] Research on real-time simulation and prediction technologies for energy efficiency data streams in industrial production systems remains immature. While existing energy efficiency management systems can collect energy consumption data from production systems through acquisition devices, fully utilizing this data for complex data analysis to identify problems and provide decision-making references remains a significant challenge. Their data analysis often employs basic descriptive statistics, and data visualization is limited to charts and graphs showing summaries, averages, and other trend analyses. Furthermore, data acquisition technologies suffer from high costs and high latency. Providing real-time production energy efficiency and future energy efficiency data to regulators is also a crucial consideration. While some data stream simulation systems exist that involve energy efficiency calculations, they lack the ability to predict future energy efficiency, creating barriers to managing energy resource consumption costs and monitoring costs and benefits. Therefore, helping enterprises implement data-driven supervision, reduce production costs, adapt to market demands, and achieve green manufacturing remains an urgent problem to be solved. Summary of the Invention

[0011] To address the aforementioned technical problems, the purpose of this invention is to provide a real-time data stream prediction system and method for steel plants oriented towards energy efficiency.

[0012] This invention provides a real-time data stream prediction system for steel plants oriented towards energy efficiency, comprising: a data acquisition module, a data processing module, a data storage module, an energy consumption visualization module, an energy efficiency calculation module, and an energy efficiency prediction module;

[0013] The data acquisition module is used to collect energy consumption data of energy resources and materials on various end equipment in the steelmaking process.

[0014] The data processing module is connected to the data acquisition module and the data storage module respectively, and is used to preprocess the energy consumption data collected from the edge device;

[0015] The data storage module is used to store the preprocessed energy consumption data collected over a period of time into a time-series database;

[0016] The energy consumption visualization module is used to visualize the real-time energy consumption data stream of each stage and each piece of equipment in the predefined steelmaking process, and to perform energy consumption data aggregation and statistical analysis visualization.

[0017] The energy efficiency calculation module is used to obtain energy consumption data for a period of time from the data storage module and calculate the steelmaking energy efficiency for that period of time according to the energy efficiency calculation formula.

[0018] The energy efficiency prediction module is used to construct a prediction model based on an improved particle swarm optimization multi-kernel regularized extreme learning machine. It inputs energy consumption data and steelmaking energy efficiency over a period of time into the prediction model to predict the future energy efficiency trend of the steel plant.

[0019] This invention also provides a real-time data stream prediction method for steel plants oriented towards energy efficiency, comprising the following steps:

[0020] Step 1: Receive and parse the data collection instructions sent by the server, deploy the data collection agent on the edge device, create data collection points, and complete the energy consumption data collection;

[0021] Step 2: The data acquisition agent program reports energy consumption data, performs data cleaning, data format standardization, outlier discarding, and missing value completion on the raw energy consumption data;

[0022] Step 3: Store the pre-processed energy consumption data into a time-series database;

[0023] Step 4: Read the energy consumption data from the time-series database, visualize the data flow of each stage and each piece of equipment in the steelmaking process, and summarize and statistically analyze the energy consumption data;

[0024] Step 5: Read the energy consumption data from the time series database and solve the steelmaking energy efficiency data of the steel plant within a specific time interval based on the energy efficiency calculation process;

[0025] Step 6: Combine energy consumption data and steelmaking energy efficiency data into a dataset, divide it into training and testing sets, and send it to the energy efficiency prediction module to train a prediction model based on improved particle swarm optimization multi-kernel regularized extreme learning machine, and predict the future energy efficiency of the steel plant.

[0026] The present invention provides a real-time data stream prediction system and method for steel plants oriented towards energy efficiency, which has the following beneficial effects:

[0027] (1) The system has comprehensive functions. This invention is a data flow simulation system that integrates real-time data acquisition, data statistical analysis, data flow visualization simulation, energy efficiency calculation and energy efficiency prediction. Compared with existing industrial data flow simulation systems, it has more complete functions.

[0028] (2) The data acquisition technology is more flexible and simpler. The data acquisition technology in this invention designs different data acquisition programs for different protocols, and manages these programs uniformly through a proxy program. Data acquisition programs are created according to server instructions to achieve data acquisition. Unlike existing acquisition technologies, this invention does not use protocol conversion methods, does not define a unified data type, and is compatible with and adaptable to multiple data protocols.

[0029] (3) Real-time display of industrial production process data. This invention visualizes the real-time data of each stage and equipment in the industrial production process, and also includes the summary and statistical analysis visualization of energy consumption data streams over a period of time. Compared with the prior art, this system visualizes the industrial production process data stream in real time, making the industrial production process clearer and more real-time.

[0030] (4) Real-time tracking of current and future energy efficiency. The energy efficiency calculation module of this invention uses the collected energy consumption data to solve the energy efficiency of the steel plant over a period of time. It can display the energy efficiency of the production process at any time. At the same time, the energy efficiency prediction model can predict the future energy efficiency trend of the steel plant, which makes it easier for relevant personnel of the enterprise to understand the dynamic changes in energy efficiency of the production process, and adjust the production structure based on the future energy efficiency trend to reduce energy consumption costs and achieve efficient utilization of energy resources.

[0031] (5) The energy efficiency prediction model is relatively reasonable. This invention uses a multi-core extreme learning machine algorithm based on improved particle swarm optimization to establish an energy efficiency prediction model to predict the future energy efficiency of the factory. Multi-core extreme learning machines have the characteristics of fewer learning parameters, fast training speed, and strong generalization ability, making them suitable for handling nonlinear problems with large-scale data. At the same time, the particle swarm optimization algorithm is used to solve the network parameters of the extreme learning machine, replacing the random initialization parameters of the extreme learning machine. Three improvement schemes are used to optimize the particle swarm optimization algorithm, making the constructed extreme learning machine network more reasonable, and the prediction model has stronger generalization ability, higher accuracy, and better prediction effect. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of a real-time data flow prediction system for steel plants oriented towards energy efficiency, according to the present invention.

[0033] Figure 2 This is a flowchart of a real-time data stream prediction method for steel plants oriented towards energy efficiency, according to the present invention.

[0034] Figure 3 This is an optimization schematic diagram of the multi-kernel regularized extreme learning machine model based on improved particle swarm optimization of the present invention;

[0035] Figure 4 This is a diagram of the Extreme Learning Machine network architecture;

[0036] Figure 5This is a flowchart of the multi-kernel regularized extreme learning machine prediction model based on improved particle swarm optimization of the present invention;

[0037] Figure 6 This is a graph showing the energy efficiency prediction effect of the multi-kernel regularized extreme learning machine prediction model based on improved particle swarm optimization of the present invention.

[0038] Figure 7 This is a loss diagram comparing the prediction model of the multi-kernel regularized extreme learning machine based on improved particle swarm optimization of the present invention with other models. Detailed Implementation

[0039] like Figure 1 As shown, the present invention provides a real-time data flow prediction system for steel plants oriented towards energy efficiency, comprising: a data acquisition module, a data processing module, a data storage module, an energy consumption visualization module, an energy efficiency calculation module, and an energy efficiency prediction module.

[0040] The data acquisition module is used to collect energy consumption data of energy resources and materials at various end-point devices in the steelmaking process. The data processing module is connected to both the data acquisition module and the data storage module, and is used to preprocess the energy consumption data collected from the end-point devices. The data storage module is used to store the preprocessed energy consumption data collected over a period of time into a time-series database. The energy consumption visualization module is used to visualize the real-time energy consumption data streams at each stage and on each device in the predefined steelmaking process, and to perform energy consumption data aggregation and statistical analysis visualization. The energy efficiency calculation module is used to obtain energy consumption data over a period of time from the data storage module and calculate the steelmaking energy efficiency over that period according to the energy efficiency calculation formula. The energy efficiency prediction module is used to construct a prediction model based on an improved particle swarm optimization multi-core extreme learning machine, inputting the energy consumption data and steelmaking energy efficiency over a period of time into the prediction model to predict the future energy efficiency trend of the steel plant.

[0041] The data acquisition agent is deployed on the edge devices. The server in the data acquisition module sends data acquisition instructions to the data acquisition agent through the MQTT Broker. The data acquisition agent creates data channels based on different data channel protocols (i.e., OPC UA and OPC UA), creates data points under the data channels, and completes the data acquisition tasks on each edge device. The data points represent the sensors set on each edge device in the steelmaking process, which are used to collect energy consumption data at each stage and on each device.

[0042] In the data processing module, the raw energy consumption data collected by the data acquisition module is matched with the corresponding data processing logic according to its data type, and data cleaning, data format standardization, outlier discarding and missing value completion are performed to improve the quality of energy consumption data.

[0043] In the energy consumption visualization module, based on a predefined steelmaking process, energy consumption data over a period of time is read from a time-series database or energy consumption data after real-time processing. The energy consumption data flow at each stage and on each piece of equipment in the steelmaking process is visualized, and statistical charts of energy substances are displayed.

[0044] In the energy efficiency calculation module, energy consumption data for a period of time is obtained from the time series database, mainly including energy resource consumption, electricity consumption, steel production, and recovered heat and gas. Then, the steelmaking energy efficiency data of the steel plant for a period of time is solved based on the energy efficiency calculation formula.

[0045] In the energy efficiency prediction module, an improved particle swarm optimization (PSO) and multi-kernel extreme learning machine (MLM) prediction model are combined to form a MLM prediction model based on improved PSO. Energy consumption data and steelmaking energy efficiency over a period of time are input into the prediction model to predict the future energy efficiency trend of the steel plant.

[0046] like Figure 2 As shown, a real-time data stream forecasting method for steel plants focused on energy efficiency includes the following steps:

[0047] Step 1: Receive and parse the data collection instructions sent by the server, deploy the data collection agent on the edge device, create data collection points, and complete the energy consumption data collection.

[0048] Step 2: The data acquisition agent program reports energy consumption data and performs data cleaning, data format standardization, outlier discarding, and missing value completion on the raw energy consumption data.

[0049] Step 3: Store the pre-processed energy consumption data in a time-series database.

[0050] Step 4: Read the energy consumption data from the time-series database, visualize the data flow of each stage and each piece of equipment in the steelmaking process, and summarize and statistically analyze the energy consumption data.

[0051] Step 5: Read the energy consumption data from the time-series database, and calculate the steelmaking energy efficiency data of the steel plant within a specific time interval based on the energy efficiency calculation process. The specific energy efficiency calculation process in Step 5 is as follows:

[0052] Step 5.1: Read the energy consumption data from the time-series database, including electricity consumption, consumption of various energy resources, steel production, and recovered heat and gas.

[0053] Step 5.2: Calculate the total energy consumption of the steel plant. The total energy consumption of the steelmaking process includes both electricity consumption and fuel combustion consumption. The formula for calculating the total energy consumption is:

[0054] E 总 =P 电力 +F燃料 ×η 燃料转化

[0055] Among them, E 总 P represents the total energy consumption of the steelmaking process, expressed in kilowatt-hours. 电力 F represents the total electricity consumption during the steelmaking process, expressed in kilowatt-hours. 燃料 This refers to the consumption of various fuels (such as coal, coke, natural gas, steam, and oil) during the steelmaking process, expressed in tons or cubic meters; η 燃料转化 Fuel conversion rate, expressed in tons per kilowatt-hour or cubic meters per kilowatt-hour.

[0056] Step 5.3: Calculate the net energy consumption of the steel plant. The net energy consumption of the steelmaking process refers to the energy resource consumption excluding recovered energy or recovered gas. The formula for calculating net energy consumption is:

[0057] E 净 =E 总 -E 回收热能 -E 回收气体

[0058] Among them, E 净 E represents the net energy consumption of the steelmaking process, expressed in kilowatt-hours. 回收热能 Waste heat to be recovered and reused, measured in kilowatt-hours (kWh); E 回收气体 The energy generated from the recycled waste gas is measured in kilowatt-hours.

[0059] Step 5.4: Energy efficiency refers to the ratio of energy generated during the steelmaking process to the actual energy resources consumed, also known as the reciprocal of energy consumption per unit output. The formula for calculating steelmaking energy efficiency is:

[0060]

[0061] in, Energy efficiency is expressed in tons per kilowatt-hour (T). 产量 Total steel production recorded over a period of time, expressed in tons.

[0062] Step 6: Combine energy consumption data and steelmaking energy efficiency data into a dataset, divide it into training and testing sets, and send it to the energy efficiency prediction module to train the IPSO-MKELM prediction model based on improved particle swarm optimization and predict the future energy efficiency of the steel plant.

[0063] like Figure 3 The diagram shows an optimization schematic of a multi-kernel regularized extreme learning machine prediction model based on improved particle swarm optimization. To facilitate understanding of the algorithm used in the prediction model, the principles of extreme learning machines and multi-kernel extreme learning machines are briefly introduced before constructing the prediction model.

[0064] Extreme Learning Machine (ELM) is a single-hidden-layer feedforward neural network. Unlike traditional neural networks, it does not require iterative calculation of weights. Instead, it obtains the weights directly by solving the Moore-Penrose generalized inverse matrix.

[0065] An ELM network with L hidden layer nodes is as follows: Figure 4 Suppose there are N samples (x) i ,y i ), i = 1, 2, 3, ..., N, x i For the dataset, y i For the corresponding sample labels, the hidden layer output of the Extreme Learning Machine is as follows:

[0066] h j (x i ) = g j (w j ,b j ,x i ) = g j (w j x i +b j )

[0067] Among them, h j (x i ) represents the output of the hidden layer, g(·) is the activation function, and w j and b j Separate weights and biases.

[0068] The expression for the output layer of the Extreme Learning Machine network is as follows:

[0069]

[0070] Among them, o i For network output, L is the number of hidden layer nodes, and β is the network output. j This represents the weight matrix between the hidden layer neurons and the output layer.

[0071] The goal of Extreme Learning Machine (ELM) is to minimize training error. It obtains the weight matrix by randomly initializing weights and biases and solving the objective function. The objective function and weight matrix are as follows:

[0072] min||Hβ-Y|| 2

[0073]

[0074] Where, H=[h1(x i ),h2(x i ),...,h L (xi [)] is the hidden layer output matrix, satisfying H∈R N×L β=[β1,β2,...,β L ] T Let Y be the weight matrix, where Y = [y1, ... y2] N ] represents the sample label matrix, H -1 Let H be the Moore-Penrose generalized inverse matrix.

[0075] By introducing a regularization coefficient, we obtain the Regularized Extreme Learning Machine (RELM), whose purpose is to minimize both the training error and the output weights. The objective function and output weight matrix of RELM are as follows:

[0076]

[0077] Where C is the regularization coefficient, and when N≥L, I∈R N×N It is a unit array.

[0078] Finally, the hidden layer output is multiplied by the output weight matrix to obtain the RELM network output, as shown in the following equation:

[0079]

[0080] The linear parameter solving process of ELM is difficult to handle when the original input data is nonlinear. Introducing kernel functions into the Extreme Learning Machine can map complex nonlinear data to a high-dimensional space, making it linearly separable and significantly reducing the computational cost of processing large-scale datasets.

[0081] Based on the above, such as Figure 5 As shown, the specific process of constructing and training a multi-kernel regularized extreme learning machine prediction model based on improved particle swarm optimization in this invention is as follows:

[0082] Step 6.1: Introduce the mixture Gaussian kernel function and the second-order polynomial kernel function into the regularized extreme learning machine to obtain the multi-kernel regularized extreme learning machine model;

[0083] The mixture Gaussian kernel function and the second-order polynomial kernel function are as follows:

[0084]

[0085] Where C1 and C2 are the weights of the Gaussian kernel function and the second-order polynomial kernel function, respectively; satisfying C1 + C2 = 1, Let σ be the Gaussian kernel function, and σ be the kernel function bandwidth. i x j +c) 2Let be a second-order polynomial kernel function, 'a' be a scaling factor controlling the inner product of vectors, and 'c' be a bias term controlling the translation of the model in the feature space; N samples (x i ,y i ), i = 1, 2, 3, ..., N, x i For energy consumption data, y i For steelmaking energy efficiency data, i.e., x i Corresponding sample labels;

[0086] Using the hybrid kernel function as the kernel function for the regularized extreme learning machine, the kernel matrix is ​​obtained as follows:

[0087]

[0088] Where, H=[h1(x i ),h2(x i ),...,h L (x i h1(x) is the hidden layer output matrix. i ) represents the output of the first neuron in the hidden layer, and L represents the number of nodes in the hidden layer.

[0089] By incorporating the kernel matrix into the regularized extreme learning machine, a multi-kernel regularized extreme learning machine model is obtained. The output weight matrix of this multi-kernel regularized extreme learning machine model is as follows:

[0090]

[0091] Where C is the regularization coefficient, Y = [y1,...y N When N≥L, I∈R N×N It is a unit array.

[0092] Multiplying the hidden layer output by the output weight matrix yields the output of the multi-kernel regularized extreme learning machine model, as shown in the following equation:

[0093]

[0094] Step 6.2: Determine the particles and initialize the parameters. Set the hidden layer weights w and bias b, as well as the parameters σ, C1, a, and c of the multi-kernel regularized extreme learning machine model as particles. Initialize the particle population size M and the maximum number of iterations k. max Given the search space dimension D, the initial inertia weight ω is obtained through the inertia weight optimization scheme, and its optimization formula is as follows:

[0095]

[0096] Where k is the iteration number, N(0,1) generates random numbers following a normal distribution, and ω min For the minimum inertia weight, ω maxThis represents the maximum inertia weight.

[0097] The individual learning factor c1 and the social learning factor c2 are initialized using the learning factor optimization scheme, and the optimization formula is as follows:

[0098] c1 = 0.7 - (1.61 × cos(πω))

[0099] c2 = 2.2 - c1

[0100] Step 6.3: Initialize particle velocity and position. Randomly generate a D-dimensional vector within [0,1] as the particle position. The velocity and position of the i-th particle are given by the following formulas:

[0101] V i =[V i,1 V i,2 ,...,V i,D ] T

[0102] X i =[X i，1 ,X i，2 ,...,X i,D ] T

[0103] Then, the positions X of all initialized particles are optimized using a chaotic mapping scheme. i .

[0104] In practice, the chaotic mapping scheme in step 6.3 includes the following steps:

[0105] Step 6.3.1: Randomly generate a D-dimensional vector with elements in the range [0,1].

[0106] Step 6.3.2: Use the Logistic chaotic mapping to map the D-dimensional vector to the initial positions of the population particles, as shown in the following formula:

[0107] X s+1 =μX s (1-X s )

[0108] Where μ is the chaotic mapping parameter, s is the number of chaotic mapping iterations, and X s Let be the particle position at the s-th iteration.

[0109] Step 6.4: Combine the energy consumption data and steelmaking energy efficiency data into a dataset, and divide it into a training set and a test set.

[0110] Step 6.5: Input the training set into the multi-kernel regularized extreme learning machine model MKRELM and train the MKRELM model.

[0111] Step 6.6: Calculate fitness and determine individual extreme values ​​and the global optimum. The initial position of the particle is taken as the individual extreme value p. best,i Using the root mean square error as the fitness function, the fitness values ​​of all particles are calculated based on the multi-kernel regularized extreme learning machine model (MKRELM), and the position of the particle with the highest fitness value is selected as the initial global optimum g. best The expressions for the individual extreme value and the global optimal solution are as follows:

[0112] p best,i =[p i1 ,p i2 ,...,p iD ] T

[0113] g best =[g i1 ,g i2 ,...,g iD ] T

[0114] Step 6.7: Iteratively solve for the global optimal solution. Perform iterative optimization operations on all particles to obtain the global optimal solution.

[0115] The iterative optimization operation in step 6.7 includes the following steps:

[0116] Step 6.7.1: Update the particle velocity and position, ensuring that the particle position remains within the defined search space. The particle velocity and position update formulas are as follows:

[0117]

[0118] Where r1 and r2 are random numbers in the range [0,1]. Let the individual extreme value be the value of the k-th iteration. This is the globally optimal solution for the k-th iteration. V represents the position of the i-th particle in the k-th iteration. i k Let be the velocity of the i-th particle in the k-th iteration. Let be the vector pointing from the particle's current position to its optimal position. It is the vector from the particle's current position to the optimal position of the group.

[0119] Step 6.7.2: Calculate the particle fitness values. Based on the fitness function, i.e., the root mean square error (RMSE), calculate the fitness values ​​of all particles after the update. The calculation formula is as follows:

[0120]

[0121] in, For the predicted value, y i Label the dataset.

[0122] Step 6.7.3: Update the individual extreme value and the global optimal solution. If the fitness value is better than the individual extreme value, then update the individual extreme value p. best,i If the fitness value is better than the fitness value of the global optimum, then update the global optimum g. best .

[0123] Step 6.7.4: Determine the termination condition. If the maximum number of iterations or the predetermined fitness threshold is reached, terminate the iteration; otherwise, proceed to step 6.7.1 to continue the iteration.

[0124] Step 6.8: Obtain the optimal particle values. The optimal values ​​of all particles are obtained through iteration, that is, the optimal values ​​of the hidden layer weights and biases of the MKRELM model as well as the parameters σ, C1, a, and c are obtained.

[0125] Step 6.9: Input the test set into the MKRELM model, test the model, and obtain the prediction results.

[0126] Daily energy efficiency data for a steel plant from January 1, 2024 to November 3, 2024 was simulated using Python. The data was then divided into training and test sets in an 8:2 ratio according to time sequence. An extreme learning machine was trained on the training set to determine the parameters to be estimated. The test set was used for model inference. Finally, the energy efficiency of the steel plant for the next two weeks was predicted. The energy efficiency prediction results are as follows: Figure 6 As shown in the figure, the period from October 3rd to November 3rd represents the test set fitting effect. Using data from the 88 days prior to October 3rd as input, the corresponding prediction models were used to predict energy efficiency values ​​from October 3rd to November 3rd. It was observed that the IPSO-MKELM model (dashed pentagram) best matched the test set (solid black line), while the comparison models PSO-MKELM (square-dashed line), IPSO-rbf (triangle-dashed line), and IPSO-poly (circle-dashed line) showed poorer prediction fitting effects. The period from November 4th to November 18th represents the predicted energy efficiency for the next two weeks, calculated by inputting data from the 88 days prior to November 4th into the model. Figure 7 A direct comparison of the root mean square error (RMSE) of this invention with the three models reveals that the proposed IPSO-MKELM has the smallest MSE, i.e., MSE = 0.00568, indicating that the proposed prediction model has high accuracy.

[0127] Algorithm descriptions in the diagram: IPSO-MKELM (pentagon-dashed line) is the multi-kernel regularized extreme learning machine prediction algorithm based on improved particle swarm optimization proposed in this invention; PSO-MKELM (square-dashed line) is the unimproved particle swarm algorithm, which is a multi-kernel regularized extreme learning machine prediction algorithm based on the original particle swarm algorithm; IPSO-rbf (triangle-dashed line) represents a prediction algorithm that uses only Gaussian kernel functions and does not use mixed kernel functions; IPSO-poly (circle-dashed line) represents a prediction algorithm that uses only linear kernel functions and does not use mixed kernel functions.

[0128] The above description is only a preferred embodiment of the present invention and is not intended to limit the ideas of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An energy efficiency oriented steel mill data stream real-time prediction system, characterized in that, The application relates to a steelmaking energy efficiency prediction system. The system comprises a data acquisition module, a data processing module, a data storage module, an energy consumption visualization module, an energy efficiency calculation module and an energy efficiency prediction module. The data acquisition module is used for acquiring energy consumption data of energy resources and substances on each edge device in a steelmaking process. The data processing module is connected with the data acquisition module and the data storage module and is used for preprocessing the energy consumption data acquired from the edge device. The data storage module is used for storing the preprocessed energy consumption data in a time sequence database. The energy consumption visualization module is used for visualizing real-time energy consumption data flow of each stage and each device in the predefined steelmaking process, and performing energy consumption data aggregation and statistical analysis visualization. The energy efficiency calculation module is used for acquiring energy consumption data in a period of time from the data storage module, and calculating steelmaking energy efficiency in the period of time according to an energy efficiency calculation formula. The energy efficiency prediction module is used for constructing a multi-kernel regularization extreme learning machine prediction model based on an improved particle swarm optimization, inputting the energy consumption data and the steelmaking energy efficiency in a period of time into the prediction model, and predicting the future energy efficiency trend of the steel plant. Step 6.1: a mixed Gaussian kernel function and a second-order polynomial kernel function are introduced into the regularization extreme learning machine to obtain a multi-kernel regularization extreme learning machine model. The mixed Gaussian kernel function and the second-order polynomial kernel function are as follows: wherein C1 and C2 are weights of the Gaussian kernel function and the second-order polynomial kernel function, respectively; C1+C2=1 is satisfied, is a Gaussian kernel function, and σ is a kernel function bandwidth, and (ax i x j +c) 2 is a second-order polynomial kernel function, a is a scaling factor for controlling the inner product of vectors, and c is a bias term for controlling the translation of the model in the feature space; N samples (x i ,y i ), i=1, 2, 3,..., N, x i is energy consumption data, y i is steelmaking energy efficiency data, that is, x i corresponding sample labels; The mixed kernel function is used as the kernel function of the regularization extreme learning machine, and the kernel matrix is as follows: where H = [h1(x i ), h2(x i ),..., h L (x i )] is the hidden layer output matrix, h1(x i ) is the output of the first neuron of the hidden layer, and L is the number of nodes in the hidden layer; The kernel matrix is introduced into the regularization extreme learning machine to obtain the multi-kernel regularization extreme learning machine model, and the output weight matrix of the multi-kernel regularization extreme learning machine model is as follows: where C is a regularization coefficient, Y = [yl,..., yN]T, and N I∈R N×N is an identity matrix; The hidden layer output is multiplied by the output weight matrix to obtain the output of the multi-kernel regularization extreme learning machine model, as follows: Step 6.2: Determine the particles and initialize parameters, set the hidden layer weights w and bias b of the multi-core regularized extreme learning machine model and parameters σ, C1, a, c as particles, initialize the particle population size M, the maximum number of iterations k max and the search space dimension D, the initial inertia weight ω is obtained by the inertia weight optimization scheme, and the optimization formula is as follows: where k is the iteration number, N(0,1) generates a random number obeying normal distribution, ω min is the minimum inertia weight, ω max is the maximum inertia weight; The learning factor optimization scheme is used to initialize the individual learning factor c1 and the social learning factor c2, and the optimization formula is as follows: c1 = 0.7 - (1.61 * cos (pi * omega)) c2 = 2.2 - c1 Step 6.3: the particle velocity and position are initialized, a D-dimensional vector in [0, 1] is randomly generated as the position of the particle, the velocity of the i-th particle and the position of the i-th particle are as follows: V i = [V i,1 , V i,2 ,..., V i,D ] T X i = [X i，1 , X i，2 ,..., X i,D ] T Then, the position X of all particles initialized is optimized by using the chaotic mapping scheme i ; Step 6.4: the energy consumption data and the steelmaking energy efficiency data are used to form a data set, and the data set is divided into a training set and a test set; Step 6.5: the training set is input into the multi-kernel regularization extreme learning machine model, and the multi-kernel regularization extreme learning machine model is trained; Step 6.6: Calculate fitness and determine individual extreme value and global optimal solution, particle initial position as individual extreme value p best,i The root mean square error is taken as the fitness function, and the fitness values of all particles are calculated based on the multi-core regularized extreme learning machine model, and the particle position with the highest fitness value among all particles is selected as the initial global optimal solution g best The expressions of individual extreme value and global optimal solution are as follows: p best,i = [p i1 , p i2 ,..., p iD ] T g best = [g i1 , g i2 ,..., g iD ] T Step 6.7: the global optimal solution is iteratively solved, and the iteration optimization operation is performed on all particles to obtain the global optimal solution; Step 6.8: the optimal particle value is obtained, all particle optimal values are obtained through iteration, that is, the optimal values of the hidden layer weight and bias of the multi-kernel regularization extreme learning machine model and the parameters sigma, C1, a and c are obtained; Step 6.9: the test set is input into the multi-kernel regularization extreme learning machine model, the model is tested, and the prediction result is obtained.

2. The energy efficiency oriented steel mill data stream real time prediction system of claim 1, wherein, The data acquisition agent is deployed on the edge device, the server in the data acquisition module sends data acquisition instructions to the data acquisition agent through the MQTTBroker, the data acquisition agent creates a data channel based on different data channel protocols, creates a data point under the data channel, and completes the data acquisition task on each edge device; The data point represents the sensor set by each edge device in the steelmaking process, which is used to collect energy consumption data at each stage and each device.

3. The energy efficiency oriented steel mill data stream real time prediction system of claim 1, wherein, In the data processing module, the raw energy consumption data collected by the data acquisition module is matched with the corresponding data processing logic according to its data type, and data cleaning, data format normalization processing, outlier discarding processing and missing value completion processing are performed to improve the quality of energy consumption data.

4. The energy efficiency oriented steel mill data stream real time prediction system of claim 1, wherein, In the energy consumption visualization module, based on the pre-defined steelmaking process, the energy consumption data within a period of time or the energy consumption data after real-time processing is read from the time series database, the energy consumption data flow of each stage and each device in the steelmaking process is visualized, and the statistical chart of energy materials is visualized.

5. The energy efficiency oriented steel mill data stream real time prediction system of claim 1, wherein In the energy efficiency calculation module, the energy consumption data within a period of time is obtained from the time series database, mainly including energy resource consumption, power consumption, steel production and recovered heat and gas, and then the steelmaking energy efficiency data of the steel plant within a period of time is solved based on the energy efficiency calculation formula.

6. An energy efficiency oriented real-time prediction method of data flow in a steel plant, characterized in that, The steps include: Step 1: receiving and analyzing the data acquisition instructions issued by the server, deploying the data acquisition agent on the edge device, creating the data acquisition point, and completing the energy consumption data acquisition; Step 2: the energy consumption data is reported by the data acquisition agent program, and the raw energy consumption data is cleaned, normalized, processed, and discarded; Step 3: store the preprocessed energy consumption data into the time series database; Step 4: read the energy consumption data in the time series database, visualize the data flow of each stage and each device in the steelmaking process, and perform energy consumption data aggregation and statistics; Step 5: read the energy consumption data in the time series database, solve the steelmaking energy efficiency data of the steel plant within a specific time interval based on the energy efficiency calculation process; Step 6: form a data set with the energy consumption data and the steelmaking energy efficiency data, divide it into a training set and a test set, and transmit it to the energy efficiency prediction module, train the multi-kernel regularized extreme learning machine prediction model based on the improved particle swarm optimization, and predict the future energy efficiency of the steel plant, the specific process is as follows: Step 6.1: introduce the mixed Gaussian kernel function and the second-order polynomial kernel function into the regularized extreme learning machine to obtain a multi-kernel regularized extreme learning machine model; The mixed Gaussian kernel function and the second-order polynomial kernel function are as follows: wherein C1 and C2 are weights of Gaussian kernel function and second-order polynomial kernel function respectively; C1+C2=1 is satisfied, is a Gaussian kernel function, and σ is a kernel function bandwidth, and (ax i is a second-order polynomial kernel function, and a is a scaling factor for controlling vector inner product, and c is a bias term for controlling model translation in feature space; N samples (x j , 2 , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , i , < The mixed kernel function is used as the kernel function of the regularized extreme learning machine, and the kernel matrix is as follows: where H = [h1(x i ), h2(x i ),...,h L (x i )] is the hidden layer output matrix, h1(x i ) is the output of the first neuron of the hidden layer, and L is the number of nodes in the hidden layer. Introduce the kernel matrix into the regularized extreme learning machine to obtain a multi-kernel regularized extreme learning machine model, and the output weight matrix of the multi-kernel regularized extreme learning machine model is as follows: where C is a regularization coefficient, Y = [yl,..., yN]T, and N ] when N ≥ L, I ∈ R N×N is the identity matrix; The hidden layer output is multiplied by the output weight matrix to obtain the output of the multi-kernel regularized extreme learning machine model, as follows: Step 6.2: Determine the particles and initialize parameters, set the hidden layer weights w and bias b of the multi-core regularized extreme learning machine model and parameters σ, C1, a, c as particles, initialize the particle population size M, the maximum number of iterations k max and the search space dimension D, the initial inertia weight ω is obtained by the inertia weight optimization scheme, and the optimization formula is as follows: where k is the iteration number, N(0,1) generates a random number obeying normal distribution, ω min is the minimum inertia weight, ω max is the maximum inertia weight; The individual learning factor c1 and the social learning factor c2 are initialized by the learning factor optimization scheme, and the optimization formula is as follows: c1 = 0.7 - (1.61 * cos(pi * omega)) c2 = 2.2 - c1 Step 6.3: Initialize particle velocity and position, randomly generate a D-dimensional vector in [0, 1] as the position of the particle, the velocity of the i-th particle and the position of the i-th particle are as follows: V i = [V i,1 , V i,2 ,..., V i,D ] T X i = [X i，1 , X i，2 ,..., X i,D ] T Then, the position X of all particles initialized is optimized by using the chaotic mapping scheme i ; Step 6.4: Form a data set with energy consumption data and steelmaking energy efficiency data, and divide it into a training set and a test set; Step 6.5: Input the training set into the multi-kernel regularized extreme learning machine model, and train the multi-kernel regularized extreme learning machine model; Step 6.6: Calculate fitness and determine individual extreme value and global optimal solution, particle initial position as individual extreme value p best,i The root mean square error is taken as the fitness function, and the fitness values of all particles are calculated based on the multi-core regularized extreme learning machine model, and the particle position with the highest fitness value among all particles is selected as the initial global optimal solution g best The expressions of individual extreme value and global optimal solution are as follows: p best,i = [p i1 , p i2 ,..., p iD ] T g best = [g i1 , g i2 ,..., g iD ] T Step 6.7: Iterative solution of global optimal solution, perform iterative optimization operation on all particles to obtain the global optimal solution; Step 6.8: Obtain the optimal particle value, obtain the optimal value of all particles by iteration, that is, obtain the optimal value of the hidden layer weight and bias of the multi-kernel regularized extreme learning machine model and the parameters σ, C1, a, c; Step 6.9: Input the test set into the multi-kernel regularized extreme learning machine model, test the model, and obtain the prediction result.

7. The energy efficiency oriented steel plant data stream real time prediction method of claim 6, wherein, The energy efficiency calculation process in step 5 is specifically: Step 5.1: Read the energy consumption data in the time series database, including power consumption, various energy resource consumption, output steel quantity, and recovered heat and gas; Step 5.2: Calculate the total energy consumption of the steelmaking plant, the total energy consumption of the steelmaking process includes power consumption and fuel combustion consumption, and the total energy consumption calculation formula is: E 总 = P 电力 + F 燃料 x η 燃料转化 wherein E 总 is the total energy consumption of the steelmaking process, in kilowatt-hours; P 电力 is the total electric power consumption in the steelmaking process, in kilowatt-hours; F 燃料 is the various fuels in the steelmaking process, in tons or cubic meters; η 燃料转化 is the fuel conversion rate, in tons per kilowatt-hour or cubic meters per kilowatt-hour; Step 5.3: Calculate the net energy consumption of the steelmaking plant, the net energy consumption of the steelmaking process refers to the energy resource consumption excluding recovered energy or recovered gas, and the net energy consumption calculation formula is: E 净 = E 总 - E 回收热能 - E 回收气体 wherein E 净 is the net energy consumption of the steelmaking process, in kilowatt-hours; E 回收热能 is the recovered waste heat, in kilowatt-hours; E 回收气体 is the energy produced from the recovered waste gas, in kilowatt-hours; Step 5.4: Energy efficiency refers to the ratio of the energy resource consumption in the steelmaking process to the actual energy resource consumption, also known as the inverse of the unit output energy consumption, and the steelmaking energy efficiency calculation formula is: wherein, Energy efficiency in tons per kilowatt-hour, T 产量 Total steel production recorded over a period of time, in tons.

8. The energy efficiency oriented steel plant data stream real time prediction method of claim 6, wherein: The chaotic mapping scheme in step 6.3 includes the following steps: Step 6.3.1: Randomly generate a D-dimensional vector with elements in the range of [0, 1]; Step 6.3.2: Map the D-dimensional vector to the initial position of the population particle using Logistic chaotic mapping, and the formula is as follows: X s+1 = μX s (1 - X s ) where μ is the chaotic map parameter, s is the number of iterations of the chaotic map, X s is the particle position at the s-th iteration.

9. The energy efficiency oriented steel plant data stream real time prediction method of claim 6, wherein, The iterative optimization operation in step 6.7 includes the following steps: Step 6.7.1: Update the particle velocity and position, and ensure that the particle position is still within the given search space, and the particle velocity and position update formulas are as follows: where r1 and r2 are random numbers in [0, 1], the individual extremum for the kth iteration, the global optimal solution for the kth iteration, the position of the ith particle at the kth iteration, V i k the velocity of the ith particle at the kth iteration, the vector of the particle pointing from the current position to the individual optimal position, the vector of the particle pointing from the current position to the group optimal position; Step 6.7.2: Calculate the particle fitness value, calculate the fitness value of all particles after updating based on the fitness function, that is, the root mean square error, and the calculation formula is as follows: wherein, is the predicted value, y i is the sample label; Step 6.7.3: Update individual extreme value and global optimum solution, if the fitness value is better than the individual extreme value, then update the individual extreme value p best,i ; if the fitness value is better than the fitness value of the global optimum solution, then update the global optimum solution g best ; Step 6.7.4: Judge the termination condition, if the maximum iteration number is reached or the fitness threshold is reached, terminate the iteration, otherwise go to step 6.7.1 and continue iteration.

Images