Intelligent scheduling and optimizing method of industrial electrical automation system

Abstract

The invention discloses an intelligent scheduling and optimization method for an industrial electrical automation system, and the method comprises the following steps: deploying a distributed sensor network to collect electrical parameters, an equipment vibration spectrum and production work order data in real time, and constructing a unified feature vector based on time-space alignment and confidence weighting; a production system-energy management-external power grid three-layer interaction model is established, and a dynamic carbon emission calculation engine and process deadlock detection module is embedded; an improved NSGA-III algorithm is adopted to solve a multi-target Pareto leading edge, and energy consumption, productivity and carbon emission target priorities are adjusted in real time in combination with a dynamic weight mechanism; distributed optimization is executed through an edge-cloud federated architecture, cross-system instruction synchronization is achieved, and closed-loop dynamic feedback is formed. According to the method, the limitation of traditional single system optimization is broken through, the energy consumption is reduced by 15%-30%, the carbon emission intensity is reduced by 12%-18%, the abnormal response speed is increased to 3 seconds, and intelligent decision making and green transformation in a complex industrial scene are supported.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise management, and particularly to an intelligent scheduling and optimization method for industrial electrical automation systems, which is applicable to scenarios such as power distribution, equipment collaborative control, and dynamic optimization of production processes in industrial enterprises. Background Art

[0002] An industrial electrical automation system refers to a system that automates the control of production equipment, production lines, and the entire production process by applying electrical technology, automatic control technology, computer technology, and information technology during the industrial production process. Its core objectives are to improve production efficiency, ensure production safety, reduce human operation errors, and optimize the utilization of various resources in the production process.

[0003] Currently, the scheduling optimization of industrial electrical automation systems faces the following technical bottlenecks: 1. Static scheduling lag: Traditional methods rely on historical data and fixed rules and cannot respond in real time to fluctuations in equipment status and sudden load changes, resulting in a 20%-35% increase in the risk of downtime under abnormal working conditions; 2. Insufficient multi-objective coordination: Existing optimization models lack a dynamic weight adjustment mechanism and it is difficult to balance conflicting objectives such as energy consumption, efficiency, equipment life, and carbon emissions, with the resource utilization rate being less than 60%; 3. Cross-system data islands: The protocols between production systems, energy networks, and external power grids are heterogeneous and lack collaborative modeling, resulting in a demand response delay of more than 5 minutes and a renewable energy consumption rate of less than 40%; 4. Limitations of the computing architecture: Centralized optimization faces the contradiction between data privacy and real-time performance. Traditional edge computing solutions lack a federated learning mechanism and the model update cycle is as long as several hours, unable to meet the millisecond-level dynamic scheduling requirements. Therefore, a new intelligent scheduling and optimization method for industrial electrical automation systems is needed to solve the above problems. Summary of the Invention

[0004] The purpose of the present invention is to address the shortcomings in the prior art and propose an intelligent scheduling and optimization method for industrial electrical automation systems.

[0005] To achieve the above objective, the present invention adopts the following technical solutions:

[0006] An intelligent scheduling and optimization method for industrial electrical automation systems includes the following specific steps:

[0007] S1: Real-time collection and fusion of multi-modal data: Synchronously collect multi-modal data at the equipment layer through distributed sensing units, including electrical parameters (current, voltage, power factor), environmental parameters (temperature and humidity, vibration spectrum), and production task data (work order priority, process timing constraints), perform spatio-temporal alignment on heterogeneous data streams, and construct a unified feature vector using a confidence-weighted algorithm;

[0008] S2: Cross-system collaborative modeling: Establish a three-layer interaction model for the production system, energy management system, and external power grid;

[0009] S3: Green energy-saving oriented dynamic optimization: Construct a multi-objective optimization function: min(α·E total +β·C carbon -γ·P throughput ), where min is to take the minimum value, E total is the total system energy consumption, C carbon is the carbon emission equivalent, P throughput is the production throughput. The weight coefficients α, β, and γ are dynamically adjusted according to the power grid carbon emission factor. The improved NSGA-III algorithm is used to solve the Pareto front, and the constraint conditions include equipment safety thresholds and process sequence deadlock detection;

[0010] S4: Cross-domain policy execution and feedback: According to the data fusion from S1-S3, cross-domain modeling, and dynamic optimization, the optimal solution is obtained. The optimal solution set is converted into executable instructions and synchronously sent to the production line PLC, energy storage system, and power grid dispatching terminal through the OPCUA protocol; Based on the equipment execution status and external power grid feedback data, the model parameters and weight coefficients in S2 are dynamically updated to form a closed-loop optimization.

[0011] As a further technical solution of the present invention, the specific content of S1 includes:

[0012] S11: Multi-modal data synchronous acquisition, using a hierarchical sensing architecture, including:

[0013] S111: Hard real-time data layer: Synchronously collect electrical parameters (current, voltage, power factor) through the FPGA embedded module at a sampling rate of ≥1kHz. The trigger signal is aligned by the IEEE 1588PTP protocol, and the clock deviation ≤0.5ms;

[0014] S112: Intermediate frequency data layer: Based on the time-driven mode, collect environmental parameters (temperature and humidity, vibration spectrum envelope) every 100ms. The data packet embeds the device location code (compliant with the ISO 24730 standard);

[0015] S113: Event-driven layer: Respond to production system events (work order switching, equipment alarm), and real-time capture associated process constraint data and equipment health status codes;

[0016] S12: Space-time alignment and confidence evaluation:

[0017] S121: Time axis alignment: Add a unified time reference label to the multi-source data stream, and use the sliding window mechanism to compensate for the transmission delay. The window size W is:

[0018] W = max(T trans-max-T trans-min , 2 ms)

[0019] where max is to take the maximum value, and T trans-max and T trans-min are the maximum and minimum values of the range of transmission delays of each data stream respectively, and 2 ms represents the transmission delay time;

[0020] S122: Spatial topology binding: Map the device location encoding to the three-dimensional factory coordinate system to construct a spatio-temporal correlation matrix M st , and the matrix element m ij represents the multi-modal data set of the i-th device within the time slice j;

[0021] S13: Confidence-weighted feature fusion: Calculate the dynamic confidence weight w k for each data source, and generate a unified input based on the weighted feature vector , where w k is the dynamic confidence weight, and v k is the normalized single-modal feature vector;

[0022] S131: Data quality factor: Calculate the initial weight by the entropy weight method based on the signal-to-noise ratio (SNR), data loss rate, and acquisition frequency stability

[0023] S132: Context-related correction: When the device is under high load or abnormal state, increase the weight of the vibration spectrum data by 30% - 50%;

[0024] S133: Time decay coefficient: Decay the weight of the historical work order data according to , where e is the natural constant, t is the time, and the decay factor λ is λ = log(1 + N switch / T), and N switch / T is the number of process switches per unit time on average.

[0025] As a further technical solution of the present invention, in the S131, calculating the initial weight by the entropy weight method specifically includes:

[0026] Index definition and data standardization:

[0027] Input parameters: m is the number of data sources (such as the number of distributed sensor nodes);

[0028] X = [x ij m×3 is the original index matrix, where each row corresponds to a data source and each column corresponds to an index;

[0029] x i1 ​is the SNR (unit: dB) of the i-th data source;

[0030] x i2 is the data loss rate (DLR) of the i-th data source,

[0031] x i3 is the acquisition frequency stability (CFS) of the i-th data source, defined as the ratio of the standard deviation of the actual sampling frequency to the target frequency:

[0032]

[0033] where f t is the actual sampling frequency, σ actual is the standard deviation of the actual sampling frequency, u target is the target frequency (e.g., 1 kHz), N is the total number of actual sampling frequency data points collected within the statistical window, and t is the index number of the data point;

[0034] Normalization processing: For positive indicators (SNR) and negative indicators (DLR, CFS), the range method is used for normalization respectively:

[0035] SNR (positive indicator):

[0036] where max(X:,1) and min(X:,1) are the maximum and minimum values of SNR among all data sources respectively;

[0037] DLR and CFS (negative indicators):

[0038] x ij is the measured value of the i-th data source for the j-th column indicator (DLR or CFS), and max(X:,j) and min(X:,j) are the maximum and minimum values of the j-th column indicator among all data sources respectively;

[0039] Obtain the normalized matrix Z = [z ij m×3 , and z ij ∈[0,1];

[0040] S131a: Calculate the index weight: For each indicator j, calculate the weight p ij :

[0041] ε = 10 -6 (To prevent division by zero)

[0042] where z ij is the normalized data of DLR and CFS, m is the number of data sources, and i is the loop index variable in the summation operation;​

[0043] S131b: Calculate the information entropy e j :

[0044]

[0045] When p ij = 0, define p ij lnp ij = 0;

[0046] S131c: Calculate the coefficient of variation and weight:

[0047] The coefficient of variation d j : d j = 1 - e j ;

[0048] The initial weight where d k is the coefficient of variation of the k-th index, and k is the loop index variable in the summation operation.

[0049] As a further technical solution of the present invention, in S2, the three-layer interaction model specifically includes: the production system layer defines the mapping relationship between equipment status and production efficiency; the energy management layer embeds a carbon emission intensity calculation module, associates the real-time electricity price of the power grid with the proportion of renewable energy; the external power grid layer obtains load forecasting and demand response signals through the API interface.

[0050] As a further technical solution of the present invention, S2 specifically includes:

[0051] S21: Modeling of the production system layer: Based on the equipment status vector S p = [load rate, health index, real-time production capacity], construct an equipment efficiency mapping function:

[0052]

[0053] where η p is the equipment efficiency, ω i is the equipment weight coefficient, a i and b i are online learning parameters, dynamically updated through the LSTM network, s i is one of the load rate, health index, and real-time production capacity, and the sigmoid function is specifically

[0054] Define the process timing constraint matrix Tc, and the element t ij represents the minimum switching time from equipment i to equipment j;

[0055] S22: Modeling of the energy management layer: Among them is the comprehensive carbon emission intensity of the system at time t, is the carbon emission factor of the power grid power supply at time t, is the output power of the local microgrid at time t, C diesel is the carbon emission factor of the diesel generator, is the output power of the diesel generator at time t is the total power consumption of the system at time t is

[0056] Deploy an edge-cloud collaborative federated learning framework, and each production line edge node trains a local energy consumption model The central server aggregates the model parameters Among them, θ g is the global model parameter vector (representing the weights of the energy consumption prediction model aggregated by the central server), K is the number of edge nodes participating in federated learning (such as the control units of different production lines in the factory), θ k is the local model parameter of the k-th edge node (obtained by training with local historical data);

[0057] S23: Interaction with the external power grid layer: Subscribe to the real-time data stream of the power grid through the IEC 61850-7-420 standard interface, including: the carbon emission intensity of the regional power grid (updated every 5 minutes), the time-of-use electricity price curve λ t and the demand response incentive signal DR t and the predicted output of renewable energy

[0058] Construct the power grid interaction revenue function: Among them, H is the optimization time domain (representing the time span for calculating the revenue), is the electricity sales volume, is the electricity purchase volume, is the predicted value of the renewable energy output (representing the predicted renewable energy power generation at time τ), and β is the renewable energy consumption penalty coefficient;

[0059] S24: Cross-layer collaborative decision-making mechanism, defining three-layer interaction constraints;

[0060] S241: Hard production timing constraint: X p ·X c ≥D min , where X p is the production plan matrix (with dimensions m×n, m is the number of devices, and n is the number of time slices), X c is the process timing constraint matrix (with dimensions n×k, k is the number of process stages), and D min is the lower limit of the delivery date;

[0061] S242: Energy flexibility constraint: Where is the grid demand response capacity margin, is the base power purchase quantity;

[0062] S243: Use the distributed ADMM algorithm to solve the cross-layer optimization problem, synchronize the Lagrange multipliers every 15 seconds, and the core iteration formula is: Where X p is the production plan variable (representing the decision variables of the production system, such as equipment start-stop plan, production capacity allocation), Z is the auxiliary variable (representing the decision variables of the energy management system or grid interaction, such as power purchase quantity, energy storage charge-discharge plan), Y is the Lagrange multiplier vector, p is the penalty coefficient, k is the iteration number, argmin is to take the minimum value of the function, and L p is the augmented Lagrangian function, and the form is: Where f(X p ) is the production layer objective function (such as maximizing equipment utilization rate), g(Z) is the energy layer objective function (such as minimizing power purchase cost), the linear term Y T (X p -Z) is to coordinate the cross-layer differences, and the quadratic penalty term forces X p to approach Z.

[0063] As a further technical solution of the present invention, in S3, using the improved NSGA-III algorithm to solve the Pareto front specifically includes:

[0064] S31: Adaptive reference point generation: According to the objective function dimension m (m≥3) and the population size N, construct a dynamic reference point set on the hyperplane, adopt the polar coordinate hierarchical division strategy, generate H equally spaced reference points along each objective axis, and the number of intervals Introduce the device safety threshold constraint direction as an additional reference point, and the weight vector W s =[0,0,...,α,...,0], where α corresponds to the penalty coefficient of the safety threshold target term; based on the historical optimization results, dynamically adjust the reference point distribution density through principal component analysis (PCA) to make the reference point cluster gather towards the high-potential solution area;

[0065] S32: Integration of constraint handling and deadlock detection: Device safety threshold constraint, process timing deadlock detection, feasibility first selection mechanism;

[0066] S33: Elite retention and diversity enhancement: Adaptive crossover and mutation, environmental selection, external archive update.

[0067] As a further technical solution of the present invention, in S32, the integration of constraint processing and deadlock detection specifically includes:

[0068] S321: Device safety threshold constraint: Define the degree of constraint violation where v j is a device state parameter (such as temperature, current harmonic distortion rate), is the safety threshold; Embed a dynamic penalty function in fitness calculation: where f(x) is the original objective function, the core index to be optimized (such as energy consumption, production efficiency, etc.), γ j = η·e βt , exponentially grows with the iteration number t, η is the initial penalty coefficient, the basic penalty weight for constraint violation, and β is the penalty growth rate, the exponential factor that controls the growth of the penalty coefficient with the iteration number;

[0069] S322: Process timing deadlock detection: Construct a resource allocation matrix A, and the element a ij = 1 indicates that process i occupies device j; After each crossover and mutation, detect whether the process sequence of the offspring individual forms a circular wait chain: Traverse the process dependency graph based on depth-first search (DFS) and mark the loop; If a deadlock is detected, trigger a repair operator, randomly select a process in the loop, and insert it into a conflict-free timing position;

[0070] S323: Feasibility first selection mechanism: In non-dominated sorting, preferentially select individuals with a constraint violation degree of ∑c j = 0; If the number of feasible solutions is insufficient, retain some low-violation solutions in ascending order of ∑c j to maintain population diversity.

[0071] As a further technical solution of the present invention, in S33, elite retention and diversity enhancement specifically include:

[0072] S331: Adaptive crossover and mutation:

[0073] The crossover probability decreases with iteration to improve late convergence, where t max is the maximum number of iterations;

[0074] The mutation probability n is the variable dimension, and the perturbation is increased when the constraint violation degree is high;

[0075] S332: Environmental selection: Combine reference point association and crowding distance to select N individuals from the combined population (parent generation + offspring generation), preferentially retain non-dominated solutions with a high association degree with the nearest reference point, and within the same reference point region, select solutions with a large crowding distance to maintain the distribution breadth;

[0076] S333: External Archive Update: The global Pareto front archive is updated every 5 generations using the ε-domination screening mechanism, and historical solutions dominated by new solutions (ε = 0.05) are removed.

[0077] As a further technical solution of the present invention, the cross-domain policy execution and feedback in S4 specifically include:

[0078] S41: Instruction Conversion and Protocol Adaptation:

[0079] The generated optimized policy instruction set is encoded and converted according to the target system type:

[0080] Production line control instructions: Convert to IEC 61131-3 structured text (ST code) executable by PLC, and embed device start / stop timing sequence, motor speed set value, and safety interlock conditions;

[0081] Energy storage system instructions: Convert to charge / discharge power curve (SOC-P curve), and attach battery temperature protection threshold;

[0082] Grid interaction instructions: Encapsulate demand response signals (such as load reduction amount, interruptible period) through IEC 61850 protocol;

[0083] Deploy a multi-protocol conversion gateway to support real-time interoperability of OPC UA, Modbus TCP, and MQTT protocols, and ensure that instructions are synchronously sent to each terminal within 200 ms;

[0084] S42: Execution Status Monitoring and Data Feedback: Collect feedback data after policy execution through device-level sensors and the SCADA system, including the production system (actual output, device fault code, process delay time), energy system (real-time energy consumption, energy storage SOC status, PV output fluctuation), and grid interaction (demand response execution rate, real-time electricity price change, carbon emission factor update value);

[0085] Add spatio-temporal consistency tags (timestamp accuracy ±10 ms, device location code) to the feedback data to construct a traceable closed-loop data chain;

[0086] S43: Dynamic Parameter Adjustment and Model Update: Dynamically correct the weight coefficient and perform online learning of the model;

[0087] S44: Exception Handling and Policy Rollback:

[0088] Real-time detection of cross-system instruction conflicts (such as the contradiction between the production line acceleration requirement and the power grid load reduction instruction), triggering dynamic priority arbitration: If the power grid is in the peak electricity price period and the carbon emission factor > the preset threshold, forcefully enable the green energy-saving mode and limit the power of non-critical equipment; If the production delay exceeds the tolerance limit (such as > 15%), temporarily switch to the production guarantee mode and allow the use of standby diesel generators for power supply;

[0089] When the device health status monitors an anomaly (such as the harmonic component in the vibration spectrum exceeding the standard), automatically generate a preventive maintenance work order and adjust the scheduling strategy

[0090] As a further technical solution of the present invention, in S43, the dynamic parameter adjustment and model update specifically include:

[0091] S431: Dynamic correction of weight coefficients: According to the latest carbon emission factor of the power grid (such as the regional power grid carbon intensity gCO2 / kWh updated every 15 minutes) and the production urgency index, adjust the weights of the optimization function according to the following formula:

[0092]

[0093] where α t is the dynamic energy consumption weight (the weight of the energy consumption minimization target at time t, dynamically adjusted according to the power grid carbon emission intensity), α0 is the initial energy consumption weight (the preset basic weight of the energy consumption target), is the real-time power grid carbon intensity (carbon emission per unit power of the regional power grid at time t), k is the adjustment coefficient, γ t is the dynamic production efficiency weight (the weight of the production efficiency target at time t, adjusted according to the difference between the planned and actual production capacities), γ 0为 the initial production efficiency weight (the preset basic weight of the production target), is the actual production capacity in the previous period (the actual output at time t - 1), is the production plan throughput in the current period;

[0094] S432: Online model learning: Adopt an incremental federated learning framework. Each edge node updates the sub-model parameters based on local feedback data. The central server aggregates the gradients and issues the global model, and the update period ≤ 5 minutes.

[0095] The beneficial effects of the present invention are:

[0096] 1. Multi-modal fusion breaks through the limitations of single-source data and supports high-precision dynamic decision-making: Through the spatio-temporal alignment and confidence-weighted fusion of heterogeneous sensor networks, combined with an adaptive acquisition strategy, a millisecond-level dynamic response is achieved.

[0097] 2. Cross-domain collaborative optimization of production, energy and power grid: Based on the improved NSGA-III algorithm and the three-layer interaction model, the Pareto frontier is solved in the target space of 5 dimensions or above, supporting the collaborative optimization of multiple indicators such as energy consumption, carbon emissions, and equipment life.

[0098] 3. The federated edge architecture takes into account both privacy and efficiency, breaking the bottleneck of industrial big data transmission: it adopts lightweight federated learning and OPC UA multi-protocol instruction synchronization to achieve the dual goals of data privacy protection and real-time optimization.

[0099] 4. Dynamic weighted carbon-production game mechanism to balance green goals and production needs: Dynamically adjust the objective function weight coefficient through real-time grid carbon intensity and production demand, combined with ADMM distributed optimization, to resolve the conflict between energy efficiency and production capacity during high-carbon periods. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 A flow chart of the intelligent scheduling and optimization method for the industrial electrical automation system proposed by the present invention;

[0101] Figure 2 Diagram of the algorithm for calculating weights for the entropy weight method. DETAILED DESCRIPTION

[0102] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0103] Please refer to the attached Figure 1-2 , the intelligent scheduling and optimization method of industrial electrical automation system includes the following specific steps:

[0104] S1: Real-time multimodal data acquisition and fusion: Distributed sensing units are used to synchronously collect multimodal data at the device level, including electrical parameters (current, voltage, power factor), environmental parameters (temperature and humidity, vibration spectrum), and production task data (work order priority, process timing constraints). The heterogeneous data streams are aligned in time and space, and a unified feature vector is constructed using a confidence weighted algorithm.

[0105] S2: Cross-system collaborative modeling: Establish a three-layer interaction model of the production system, energy management system and external power grid;

[0106] S3: Dynamic optimization oriented to green energy saving: Constructing multi-objective optimization function: min(α·E total +β·C carbon -γ·P throughput ), where min is the minimum value, E total is the total energy consumption of the system, C carbon is carbon emission equivalent, P throughputFor production throughput, the weight coefficients α, β, and γ are dynamically adjusted according to the power grid carbon emission factor. The improved NSGA-III algorithm is used to solve the Pareto front, and the constraint conditions include equipment safety thresholds and process sequence deadlock detection;

[0107] S4: Cross-domain policy execution and feedback: Based on the data fusion from S1 - S3, cross-domain modeling, and dynamic optimization to obtain the optimal solution, convert the optimal solution set into executable instructions, and synchronously send them to the production line PLC, energy storage system, and power grid dispatching terminal through the OPC UA protocol; Based on the equipment execution status and external power grid feedback data, dynamically update the model parameters and weight coefficients in S2 to form a closed-loop optimization.

[0108] Technical logic chain: Data fusion → Cross-domain modeling → Dynamic optimization → Closed-loop execution:

[0109] Break through the limitations of traditional single-system optimization, and establish a global optimization foundation through cross-system data coupling of production - energy - power grid (S1 - S2);

[0110] Introduce the power grid carbon emission factor to dynamically adjust the optimization weight (S3) to achieve an accurate balance between the green energy-saving goal and production requirements;

[0111] OPC UA multi-protocol collaborative control (S4) solves the problem of instruction synchronization for heterogeneous systems.

[0112] In a preferred embodiment, S1 specifically includes:

[0113] S11: Multi-modal data synchronous acquisition, using a hierarchical sensing architecture, including:

[0114] S111: Hard real-time data layer: Synchronously collect electrical parameters (current, voltage, power factor) through the FPGA embedded module at a sampling rate of ≥1kHz, and the trigger signal is aligned by the IEEE 1588PTP protocol, with a clock deviation ≤0.5ms;

[0115] S112: Intermediate frequency data layer: Based on the time-driven mode, collect environmental parameters (temperature, humidity, vibration spectrum envelope) every 100ms, and the data packet embeds the device location code (compliant with the ISO 24730 standard);

[0116] S113: Event-driven layer: Respond to production system events (work order switching, equipment alarm), and real-time capture associated process constraint data and equipment health status codes;

[0117] S12: Space-time alignment and confidence evaluation:

[0118] S121: Time axis alignment: Add a unified time reference label to the multi-source data stream, and use the sliding window mechanism to compensate for the transmission delay. The window size W is:

[0119] W = max(T trans-max - T trans-min , 2ms)

[0120] where max is to take the maximum value, T trans-max and T trans-min are respectively the maximum and minimum values of the range of data stream transmission delays, and 2ms represents the transmission delay time;

[0121] S122: Spatial topology binding: Map the device location encoding to the three-dimensional factory coordinate system to construct a spatio-temporal correlation matrix M st , and the matrix element m ij represents the multi-modal data set of the i-th device within the time slice j;

[0122] S13: Confidence-weighted feature fusion: Calculate the dynamic confidence weight w k for each data source, and generate a unified input based on the weighted feature vector , where w k is the dynamic confidence weight, and v k is the normalized single-modal feature vector;

[0123] S131: Data quality factor: Calculate the initial weight through the entropy weight method based on the signal-to-noise ratio (SNR), data loss rate, and acquisition frequency stability

[0124] Index definition and data standardization:

[0125] Input parameters: m is the number of data sources (such as the number of distributed sensor nodes);

[0126] X = [x ij m×3 is the original index matrix, with each row corresponding to a data source and each column corresponding to an index;

[0127] x i1 is the SNR (unit: dB) of the i-th data source;

[0128] x i2 is the data loss rate (DLR) of the i-th data source,

[0129] x i3 is the acquisition frequency stability (CFS) of the i-th data source, defined as the ratio of the standard deviation of the actual sampling frequency to the target frequency:

[0130]

[0131] where f t is the actual sampling frequency, and σ actual ​is the standard deviation of the actual sampling frequency, μ target is the target frequency (such as 1 kHz), N is the total number of actual sampling frequency data points collected within the statistical window, and t is the index number of the data points;

[0132] Example of calculating the acquisition frequency stability CFS

[0133] In a certain numerical control machine tool monitoring system: the target frequency μ target = 500 Hz;

[0134] Statistical window T = 5 s → N = 5 s × 500 Hz = 2500;

[0135] Actual calculation:

[0136] Result interpretation: The CFS value is much lower than the threshold (such as 0.01), indicating that the acquisition frequency is stable and the data credibility is high.

[0137] Standardization processing: For the positive index (SNR) and negative indices (DLR, CFS), the range method is used for normalization respectively:

[0138] SNR (positive index):

[0139] where max(X:,1) and min(X:,1) are the maximum and minimum values of SNR among all data sources respectively;

[0140] DLR and CFS (negative indices):

[0141] x ij is the measured value of the i-th data source for the j-th column index (DLR or CFS), and max(X:,j) and min(X:,j) are the maximum and minimum values of the j-th column index among all data sources respectively;

[0142] Obtain the standardized matrix Z = [z ij m×3 , and z ij ∈[0,1];

[0143] Example of formula application: The indicators of 3 sensor nodes are as follows in the table (the unit of SNR is dB, and DLR and CFS are unitless ratios)

[0144] Data source SNR(j = 1) DLR(j = 2) CFS(j = 3) 1 45 0.2 0.05 2 38 1.5 0.12 3 50 0.8 0.18

[0145] Normalization calculation:

[0146] 1. SNR (positive index):

[0147] ​min(X:,1) = 38, min(X:,1)) = 50

[0148] Data source 1: z11 = (45 - 38) / (50 - 38) = 7 / 12 ≈ 0.583

[0149] Data source 2: z21 = (38 - 38) / 12 = 0

[0150] Data source 3: z31 = (50 - 38) / 12 = 1.0

[0151] 2. DLR (negative index):

[0152] min(X:,2) = 0.2 min(X:,2) = 1.5

[0153] Data source 1: z12 = (1.5 - 0.2) / (1.5 - 0.2) = 1.0

[0154] Data source 2: z22 = (1.5 - 1.5) / 1.3 = 0

[0155] Data source 3: z32 = (1.5 - 0.8) / 1.3 ≈ 0.538

[0156] 3. CFS (negative index):

[0157] min(X:,3) = 0.05 min(X:,3) = 0.18

[0158] Data source 1: z13 = (0.18 - 0.05) / (0.18 - 0.05) = 1.0

[0159] Data source 2: z23 = (0.18 - 0.12) / 0.13 ≈ 0.462

[0160] Data source 3: z33 = (0.18 - 0.18) / 0.13 = 0

[0161] S131a: Calculate the index proportion: For each index j, calculate the proportion p of each i data source ij :

[0162] ε = 10 -6 (To prevent division by zero)

[0163] where z ij is the standardized data of DLR and CFS, m is the number of data sources, and i is the loop index variable in the summation operation;

[0164] S131b: Calculate the information entropy e j : When p ij = 0, define pij lnp ij = 0;

[0165] S131c: Calculate the coefficient of variation and weights:

[0166] Coefficient of variation d j : d j = 1 - e j ;

[0167] Initial weight

[0168] where d k is the coefficient of variation of the k-th index, and k is the loop index variable in the summation operation;

[0169] Example of the calculation process

[0170] Assume that the coefficients of variation of three indices are respectively:

[0171] d1 = 0.2 (SNR), d2 = 0.5 (DLR), d3 = 0.3 (CFS).

[0172] Denominator calculation:

[0173] Weights of each index:

[0174]

[0175] Example of calculating the initial weight by the entropy weight method:

[0176] Input data: Index values of 3 data sources

[0177] Data source SNR (dB) DLR (%) CFS 1 45 0.2 0.05 2 38 1.5 0.12 3 50 0.8 0.18

[0178] Normalization result:

[0179] SNR column (positive): Z = [0.583, 0.0, 1.0]

[0180] DLR column (negative): Z = [0.923, 0.0, 0.538]

[0181] CFS column (negative): Z = [1.0, 0.615, 0.0]

[0182] Entropy weight calculation:

[0183] Information entropy e = [0.899, 0.775, 0.940]

[0184] Coefficient of variation d = [0.101, 0.225, 0.060]

[0185] Weight w base = [0.101 / 0.386, 0.225 / 0.386, 0.060 / 0.386]

[0186] = [0.262, 0.583, 0.155]

[0187] Conclusion: The data loss rate (DLR) has the largest difference (weight 58.3%), and has the most significant impact on the confidence level.

[0188] S132: Context-related correction: When the device is in a high-load or abnormal state, the weight of the vibration spectrum data is increased by 30% - 50%;

[0189] S133: Time decay coefficient: The weight of the historical work order data is decayed, where e is the natural constant, t is the time, and the decay factor λ is λ = log(1 + N switch / T), N switch / T is the number of process switches per unit time on average.

[0190] Behavior verification of the time decay formula

[0191]

[0192] In a preferred embodiment, in S2, the three-layer interaction model specifically includes: the production system layer defines the mapping relationship between device status and production efficiency; the energy management layer embeds a carbon emission intensity calculation module, associates the real-time electricity price of the power grid with the proportion of renewable energy; the external power grid layer obtains load forecasting and demand response signals through the API interface.

[0193] In a preferred embodiment, S2 specifically includes:

[0194] S21: Modeling of the production system layer: Based on the device status vector S p = [load rate, health index, real-time production capacity], construct a device efficiency mapping function:

[0195]

[0196] where η p is the device efficiency, ω i is the device weight coefficient, a i , b i are online learning parameters, dynamically updated through the LSTM network, s i is one of the load rate, health index, and real-time production capacity, and the sigmoid function is specifically

[0197] Parameter interaction and efficiency calculation example:

[0198] The state parameters of a certain CNC machine tool are as follows: load rate s1 = 0.85 (85% load), health index s2 = 0.7, real-time production capacity s3 = 120 pieces per hour;

[0199] Parameter configuration: weight coefficients w1 = 0.5, w2 = 0.3, w3 = 0.2

[0200] Learning parameters a1 = 4, b1 = -3; a2 = 5, b2 = -2; a3 = 2, b3 = 0

[0201] Calculation steps:

[0202] Calculate the sigmoid output of each state item:

[0203] Load rate item: sigmoid(4 * 0.85 - 3) = sigmoid(0.4) ≈ 0.598

[0204] Health index item: sigmoid(5 * 0.7 - 2) = sigmoid(1.5) ≈ 0.818

[0205] Real-time production capacity item: sigmoid(2 * 120 - 0) = sigmoid(240) ≈ 1.0

[0206] Weighted summation:

[0207] η p = 0.5 * 0.598 + 0.3 * 0.818 + 0.2 * 1.0 ≈ 0.299 + 0.245 + 0.2 = 0.744

[0208] Conclusion: The comprehensive efficiency of the device is 74.4%, indicating that it is in a high-efficiency operating state.

[0209] Define the process timing constraint matrix Tc, and the element t ij represents the minimum switching time from device i to device j;

[0210] S22: Energy management layer modeling: Among them is the comprehensive carbon emission intensity of the system at time t, is the carbon emission factor of the power grid power supply at time t, is the output power of the local microgrid at time t, C diesel is the carbon emission factor of the diesel generator, is the output power of the diesel generator at time t The total power consumption of the system at time t is

[0211] Calculation example: Energy management of a certain data center

[0212] (The carbon intensity of the power grid is relatively high);

[0213] (Photovoltaic full power generation);

[0214] (The diesel engine is not enabled);

[0215]

[0216] Calculate:

[0217] Strategy trigger:

[0218] The system automatically reduces non-critical loads (such as standby air conditioners), and will Reduce to 1000 kW,

[0219] Make Further reduce to 144 gCO2 / kWh.

[0220] Deploy an edge-cloud collaborative federated learning framework, and each production line edge node trains a local energy consumption model The central server aggregates model parameters Where θ g Is the global model parameter vector (representing the weights of the energy consumption prediction model aggregated by the central server), K is the number of edge nodes participating in federated learning (such as control units of different production lines in the factory), θ k Is the local model parameter of the kth edge node (obtained by training with local historical data);

[0221] S23: External power grid layer interaction: Subscribe to the real-time power grid data stream through the IEC 61850-7-420 standard interface, including: Regional power grid carbon emission intensity (Updated every 5 minutes), time-of-use electricity price curve λ t And demand response incentive signal DR t And renewable energy predicted output

[0222] Construct the power grid interaction revenue function: Where H is the optimization time domain (representing the time span for calculating revenue), Is the electricity sales volume, Is the electricity purchase volume, Is the predicted value of renewable energy output (representing the predicted renewable energy power generation at time τ), and β is the renewable energy consumption penalty coefficient;

[0223] Dynamic optimization scenario example: A factory energy storage system participates in the power grid demand response

[0224] Input parameters:

[0225] H = 6 hours, with a significant peak-valley difference in time-of-use electricity prices;

[0226] (Photovoltaic prediction): [200, 180, 150, 120, 50, 0], kW) (decreasing over time);

[0227] β = 0.05 yuan / kW 2 .

[0228] Optimization results:

[0229] Sell electricity during the peak electricity price period (14:00 - 16:00) Revenue: 1.2×100 = 120 yuan;

[0230] Purchase electricity when the photovoltaic output is insufficient (18:00) Actual predicted value Penalty cost: 0.05×80 2 = 320 yuan.

[0231] Total revenue: The revenue and penalties for all time periods need to be integrated, and the optimal purchase / sale electricity strategy can be found through an optimization algorithm (such as dynamic programming).

[0232] S24: Cross-layer collaborative decision-making mechanism, defining three-layer interaction constraints;

[0233] S241: Hard production timing constraint: X p ·X c ≥D min , where X p is the production plan matrix (dimension m×n, m is the number of devices, n is the number of time slices), X c is the process timing constraint matrix (dimension n×k, k is the number of process stages), and D min is the lower limit of the delivery date;

[0234] S242: Energy flexibility constraint: where is the grid demand response capacity margin, is the baseline electricity purchase volume;

[0235] Example description: Cross-layer collaborative optimization of an automobile manufacturing plant

[0236] Input of the production layer:

[0237] X p : Production plan for 5 production lines in the next 6 hours (matrix dimension 5×6);

[0238] X c:Time sequence constraints for welding, painting, and general assembly processes (30 min / stage).

[0239] Energy layer input:

[0240] Baseline predicted power purchase value (based on historical data);

[0241] Grid demand response instruction (allowing ±200 kW fluctuation).

[0242] Optimization result:

[0243] ADMM converges after 3 iterations (residual < 0.01), generating a strategy:

[0244] Increase power purchase during low electricity price periods (λ t = 0.3 yuan / kWh) to complete high-energy-consuming processes in advance; sell the energy storage power during high electricity price periods (λ t == 1.5 yuan / kWh), and at the same time delay non-critical processes to low-carbon periods.

[0245] Effect: The total cost is reduced by 18%, and carbon emissions are reduced by 12%.

[0246] S243: Use the distributed ADMM algorithm to solve the cross-layer optimization problem, synchronize the Lagrange multipliers every 15 seconds, and the core iteration formula is: Where X p is the production plan variable (representing the decision variables of the production system, such as equipment start-stop plan, production capacity allocation), Z is the auxiliary variable (representing the decision variables of the energy management system or grid interaction, such as power purchase volume, energy storage charge and discharge plan), Y is the Lagrange multiplier vector, p is the penalty coefficient, k is the number of iterations, argmin is to take the minimum value of the function, and L p is the augmented Lagrangian function, in the form of: Where f(X p ) is the production layer objective function (such as maximizing equipment utilization rate), g(Z) is the energy layer objective function (such as minimizing power purchase cost), the linear term Y T (X p -Z) is to coordinate the cross-layer differences, and the quadratic penalty term forces X p to approach Z.

[0247] Example description: ADMM parameter configuration for a smart factory

[0248] Objective: Coordinate the production plans of 5 production lines and the output of the microgrid to minimize the total cost.

[0249] Parameter setting: ρ0 = 1.0, upper limit of adaptive adjustment; ρ max = 5.0; Synchronization frequency: 15 seconds / time; Termination condition: residual ∈ = 0.1 or k max = 50.

[0250] Result: Average number of iterations: Converged in 12 times;

[0251] Cost savings: 23% (compared with centralized optimization);

[0252] Communication load: Reduced by 65% (due to local computing).

[0253] In a preferred embodiment, in S3, using the improved NSGA-III algorithm to solve the Pareto front specifically includes:

[0254] S31: Adaptive reference point generation: According to the objective function dimension m (m≥3) and population size N, construct a dynamic reference point set on the hyperplane, adopt the polar coordinate hierarchical division strategy, generate H equally spaced reference points along each objective axis, and the number of spacing Introduce the device safety threshold constraint direction as an additional reference point, and the weight vector W s = [0, 0,..., α,..., 0], where α corresponds to the penalty coefficient of the safety threshold target item; Based on the historical optimization results, dynamically adjust the reference point distribution density through principal component analysis (PCA) to make the reference point cluster gather towards the high-potential solution area;

[0255] S32: Integration of constraint handling and deadlock detection: Device safety threshold constraint, process timing deadlock detection, feasibility priority selection mechanism;

[0256] S321: Device safety threshold constraint: Define the constraint violation degree where v j is the device state parameter (such as temperature, current harmonic distortion rate), is the safety threshold; Embed a dynamic penalty function in the fitness calculation: where f(x) is the original objective function, the core index to be optimized (such as energy consumption, production efficiency, etc.), γ j = η·e βt 、Exponentially grows with the iteration number t, η is the initial penalty coefficient, the basic penalty weight for constraint violation, and β is the penalty growth rate, the exponential factor that controls the growth of the penalty coefficient with the iteration number;

[0257] Industrial scenario application example: Scheduling optimization of a rolling mill production line in a steel plant

[0258] Objective function: f(x) = Energy consumption (kW) + Equipment loss cost (yuan / hour)

[0259] Constraints: Roll temperature ≤ 120 °C (c1 = max(0, T - 120));

[0260] Current harmonic distortion rate ≤ 7% (c2 = max(0, THD - 7%)).

[0261] Parameter setting: η = 200 yuan / °C, β = 0.03, t max = 80;

[0262] η1 for the key constraint (temperature) is 300, and η2 for the secondary constraint (harmonics) is 100.

[0263] Optimization results:

[0264] Final solution temperature violation degree c1 = 0 °C, harmonic violation degree c2 = 0.2% (acceptable);

[0265] The total cost is reduced by 18% compared to the static penalty, and there is no equipment overheating failure.

[0266] S322: Process sequence deadlock detection: Construct a resource allocation matrix A, where the element a ij = 1 indicates that process i occupies equipment j; after each crossover and mutation, detect whether the process sequence of the offspring individual forms a circular wait chain: traverse the process dependency graph based on depth - first search (DFS) and mark the loops; if a deadlock is detected, trigger the repair operator, randomly select a process in the loop, and insert it into a conflict - free timing position;

[0267] S323: Feasibility - first selection mechanism: In non - dominated sorting, preferentially select individuals with a constraint violation degree ∑c j = 0; if the number of feasible solutions is insufficient, retain some low - violation - degree solutions in ascending order of ∑c j to maintain population diversity;

[0268] S33: Elite retention and diversity enhancement: Adaptive crossover and mutation, environmental selection, external archive update;

[0269] S331: Adaptive crossover and mutation:

[0270] Crossover probability decreases with iteration to improve late - stage convergence, where t max is the maximum number of iterations;

[0271] Mutation probability n is the variable dimension, and the perturbation is increased when the constraint violation degree is high;

[0272] Industrial application case: Optimization of the reactor scheduling in a chemical plant

[0273] Problem dimension: n = 30 (10 devices × 3 control parameters);

[0274] Population size: N = 150, maximum number of iterations t max = 200;

[0275] Constraints: temperature ≤ 150 °C, pressure ≤ 10 MPa.

[0276] Optimization results:

[0277] Change in crossover probability: p c Decreases from 0.9 to 0.4 to avoid ineffective crossovers in the later stage;

[0278] Mutation probability response: When the sum of c j suddenly increases due to abnormal raw materials in a certain batch, p m Increases from 0.03 to 0.15 to quickly repair infeasible solutions;

[0279] Effect: The proportion of feasible solutions increases from 45% in the initial stage to 98% in the later stage, and the optimization time is reduced by 22%.

[0280] S332: Environmental selection: Combine reference point association and crowding distance to select N individuals from the combined population (parent + offspring). Prioritize non-dominated solutions with a high degree of association with the nearest reference point. Within the same reference point region, select solutions with a large crowding distance to maintain the distribution breadth;

[0281] S333: External archive update: Update the global Pareto front archive every 5 generations using the ε-domination screening mechanism, and eliminate historical solutions that are ε-dominated by new solutions (ε = 0.05).

[0282] In a preferred embodiment, the cross-domain strategy execution and feedback in S4 specifically include:

[0283] S41: Instruction conversion and protocol adaptation:

[0284] Encode and convert the optimized generated strategy instruction set according to the target system type:

[0285] Production line control instructions: Convert to IEC 61131-3 structured text (ST code) executable by PLC, and embed device start / stop timing sequences, motor speed set values, and safety interlock conditions;

[0286] Energy storage system instructions: Convert to charge / discharge power curves (SOC-P curves), and attach battery temperature protection thresholds;

[0287] Power grid interaction instructions: Encapsulate demand response signals (such as load reduction amounts, interruptible time periods) through the IEC 61850 protocol;

[0288] Deploy a multi-protocol conversion gateway to support real-time interoperability of OPC UA, Modbus TCP, and MQTT protocols, ensuring that instructions are synchronously sent to each terminal within 200 ms;

[0289] Description of the embodiment: Apply this method in a semiconductor packaging workshop:

[0290] Objective function: 4 dimensions (energy consumption, equipment health, production capacity, carbon emissions);

[0291] Constraints: Equipment temperature ≤ 85°C, current harmonic distortion rate ≤ 8%;

[0292] The 12 processes need to meet the timing constraints of coating → cutting → welding → testing;

[0293] Optimization results: The coverage rate of the Pareto front solution set is increased to 92% (the benchmark NSGA-III is 68%); the success rate of deadlock repair reaches 100%, and the production interruption time is reduced by 89%; the number of safety threshold violations is reduced from an average of 5.2 times per hour to 0.1 times per hour.

[0294] S42: Execution status monitoring and data feedback: Collect the feedback data after the strategy execution through the device layer sensors and the SCADA system, including the production system (actual output, equipment failure code, process delay time), the energy system (real-time energy consumption, energy storage SOC status, photovoltaic output fluctuation), and the grid interaction (demand response execution rate, real-time electricity price change, carbon emission factor update value);

[0295] Add spatio-temporal consistency tags (timestamp accuracy ±10 ms, device location code) to the feedback data to construct a traceable closed-loop data chain;

[0296] S43: Dynamic parameter adjustment and model update: Dynamically correct the weight coefficients and perform online learning of the model;

[0297] S431: Dynamically correct the weight coefficients: According to the latest carbon emission factor of the power grid (such as the regional power grid carbon intensity gCO2 / kWh updated every 15 minutes) and the production urgency index, adjust the weights of the optimization function according to the following formula:

[0298]

[0299] where α t is the dynamic energy consumption weight (the weight of the energy consumption minimization target at time t, dynamically adjusted according to the power grid carbon emission intensity), α0 is the initial energy consumption weight (the preset basic weight of the energy consumption target), is the real-time power grid carbon intensity (carbon emission per unit power of the regional power grid at time t), k is the adjustment coefficient, γ tis the dynamic production efficiency weight (the production efficiency target weight at time t, adjusted according to the difference between the planned and actual production capacity), γ 0为 The initial production efficiency weight (the preset basic weight of the production target), is the actual production capacity in the previous period (the actual output at time t - 1), is the production plan throughput in the current period;

[0300] Engineering application example:

[0301] When the power grid is in a high-carbon period If α0 = 0.6 and k = 0.3, then: α t = 0.6 * (1 + 0.3 * 900 / 1000) = 0.6 * 1.27 = 0.76

[0302] The energy consumption weight increases by 27%, and the system preferentially schedules energy-saving equipment or reduces non-critical loads.

[0303] S432: Model online learning: Adopt an incremental federated learning framework. Each edge node updates the sub-model parameters based on local feedback data. The central server aggregates the gradients and distributes the global model. The update period ≤ 5 minutes;

[0304] S44: Exception handling and policy rollback:

[0305] Real-time detection of cross-system instruction conflicts (such as the contradiction between the production line acceleration demand and the power grid load reduction instruction), triggering dynamic priority arbitration: If the power grid is in a peak electricity price period and the carbon emission factor > the preset threshold, force the green energy-saving mode to be enabled and limit the power of non-critical equipment; If the production delay exceeds the tolerance limit (such as > 15%), temporarily switch to the production guarantee mode and allow the standby diesel generator to be called for power supply;

[0306] When the abnormal health status of the equipment is detected (such as the harmonic component of the vibration spectrum exceeding the standard), automatically generate a preventive maintenance work order and adjust the scheduling strategy.

[0307] From the above description, it can be seen that the above embodiments of the present invention achieve the following technical effects:

[0308] 1. Multi-modal data fusion-driven dynamic scheduling: Through the spatio-temporal alignment and confidence-weighted fusion of heterogeneous sensor networks, combined with an adaptive acquisition strategy, a millisecond-level dynamic response is achieved; Compared with traditional static scheduling, the equipment abnormal detection accuracy is increased to 93%, and the process switching delay is reduced by 62%, significantly improving the scheduling flexibility under complex disturbances.

[0309] 2. Cross - domain collaborative optimization of production - energy - power grid: Based on the improved NSGA - III algorithm and three - layer interaction model, solve the Pareto front in the target space of more than 5 dimensions, and support the collaborative optimization of multiple indicators such as energy consumption, carbon emissions, and equipment life. In practical applications, automatically reduce the load by 20% - 35% during high - carbon periods, achieve a comprehensive carbon emission reduction of 12% - 18%, and ensure that the production efficiency volatility is <5%.

[0310] 3. Edge - cloud federated learning architecture: Adopt lightweight federated learning and OPC UA multi - protocol instruction synchronization to achieve the dual goals of data privacy protection and real - time optimization. In practical applications, the global model prediction error is reduced from 8% to 3%, the communication bandwidth occupancy is reduced by 37%, and the daily average power consumption of edge nodes is reduced by 18%.

[0311] 4. Dynamic - weight carbon - production game mechanism: Dynamically adjust the weight coefficient of the objective function through the real - time power grid carbon intensity and production demand, and combine ADMM distributed optimization to solve the conflict between energy efficiency and production capacity during high - carbon periods. In practical applications, the time taken to adjust the scheduling strategy during peak - time electricity prices is compressed from the minute level to 15 seconds, the demand response revenue is increased by 48,000 yuan per time, and the carbon emission intensity is decreased by 19% year - on - year.

[0312] Those of ordinary skill in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present invention (including the claims) is limited to these examples. Under the concept of the present invention, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of the present invention as described above, which are not provided in detail for the sake of brevity.

[0313] The present invention aims to cover all such substitutions, modifications, and variations that fall within the broad scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent scheduling and optimization method for industrial electrical automation systems, characterized in that It includes the following specific steps: S1: Real-time acquisition and fusion of multi-modal data: Synchronously collect multi-modal data at the device layer through distributed sensing units, including electrical parameters, environmental parameters, and production task data, perform spatio-temporal alignment on heterogeneous data streams, and construct a unified feature vector using a confidence-weighted algorithm; S2: Cross-system collaborative modeling: Establish a three-layer interaction model of the production system, energy management system, and external power grid; S3: Green energy-saving oriented dynamic optimization: Construct a multi-objective optimization function: min(α·E total +β·C carbon -γ·P throughput ), and use the improved NSGA-III algorithm to solve the Pareto front. The constraint conditions include the device safety threshold and the process timing deadlock detection; S4: Cross-domain policy execution and feedback: Obtain the optimal solution after data fusion from S1 to cross-domain modeling to dynamic optimization, convert the optimal solution set into executable instructions, and synchronously send them to the production line PLC, energy storage system, and grid dispatching terminal through the OPCUA protocol; Based on the device execution status and external power grid feedback data, dynamically update the model parameters and weight coefficients in S2 to form a closed-loop optimization.

2. The intelligent scheduling and optimization method of the industrial electrical automation system according to claim 1, characterized in that, The specific content of S1 includes: S11: Synchronous acquisition of multi-modal data, using a hierarchical sensing architecture, including: S111: Hard real-time data layer: Synchronously collect electrical parameters through the FPGA embedded module at a sampling rate of ≥1kHz, align the trigger signal by the IEEE 1588PTP protocol, and the clock deviation ≤0.5ms; S112: Intermediate frequency data layer: Based on the time-driven mode, collect environmental parameters once every 100ms, and embed the device location code in the data packet; S113: Event-driven layer: Respond to production system events and capture associated process constraint data and device health status codes in real time; S12: Spatio-temporal alignment and confidence evaluation: S121: Time-axis alignment: Add a unified time reference label to multi-source data streams, and use a sliding window mechanism to compensate for transmission delays. The window size W is: W = max(T trans-max - T trans-min , 2 ms); S122: Spatial Topological Binding: Map the device location encoding to the three-dimensional factory coordinate system to construct a spatio-temporal correlation matrix M st , where the matrix element m ij represents the multi-modal data set of the i-th device within the time slice j; S13: Confidence-weighted feature fusion: Calculate the dynamic confidence weight w for each data source k , and generate a unified input based on the weighted feature vector ; S131: Data quality factor: Calculate the initial weight by the entropy weight method based on the signal-to-noise ratio, data loss rate, and acquisition frequency stability S132: Context-related correction: When the device is in a high-load or abnormal state, increase the weight of vibration spectrum data by 30%-50%; S133: Time decay coefficient: weight for historical work order data According to Decay.

3. The intelligent scheduling and optimization method for the industrial electrical automation system according to claim 2, wherein In S131, the initial weights are calculated by the entropy weight method Specifically, it includes: S131a: Calculate the index proportion: For each index j, calculate the proportion p of each i data source ij : S131b: Calculate the information entropy e j : When p ij = 0, define p ij lnp ij = 0; S131c: Calculate the difference coefficient and weight: Coefficient of difference d j : d j = 1 - e j ; Initial weight 4. The intelligent scheduling and optimization method of the industrial electrical automation system according to claim 1, wherein, In S2, the three-layer interaction model specifically includes: The production system layer defines the mapping relationship between device status and production efficiency; The energy management layer embeds a carbon emission intensity calculation module, associates the real-time electricity price of the power grid with the proportion of renewable energy; The external power grid layer obtains load forecasting and demand response signals through the API interface.

5. The intelligent scheduling and optimization method of the industrial electrical automation system according to claim 4, characterized in that The specific content of S2 includes: S21: Modeling of the production system layer: Based on the device state vector S p = [load rate, health index, real-time production capacity], construct the device efficiency mapping function: Define the process timing constraint matrix Tc, and the element t ij represents the minimum switching time from device i to device j; S22: Energy management layer modeling: Deploy a federated learning framework for edge-cloud collaboration, and each edge node on the production line trains a local energy consumption model The central server aggregates model parameters S23: Interaction with the external power grid layer: Subscribe to the real-time power grid data stream through the IEC 61850-7-420 standard interface, including: the carbon emission intensity of the regional power grid Time-of-use electricity price curve λ t and the demand response incentive signal DR t , the predicted output of renewable energy Construct the grid interaction revenue function: S24: Cross-layer collaborative decision-making mechanism, defining three-layer interaction constraints; S241: Hard Constraint on Production Timing: X p ·X c ≥D min ; S242: Energy Flexibility Constraint: S243: Use the distributed ADMM algorithm to solve the cross-layer optimization problem, synchronize the Lagrange multipliers every 15 seconds, and the core iteration formula is:

6. The intelligent scheduling and optimization method for the industrial electrical automation system according to claim 1, wherein, In S3, using the improved NSGA-III algorithm to solve the Pareto front specifically includes: S31: Adaptive reference point generation: According to the objective function dimension m and the population size N, construct a dynamic reference point set on the hyperplane, adopt a polar coordinate hierarchical partitioning strategy, and generate H equally spaced reference points along each objective axis, with the number of intervals Introduce the device safety threshold constraint direction as an additional reference point, and the weight vector W s = [0, 0,..., α,..., 0]; Based on the historical optimization results, dynamically adjust the reference point distribution density through principal component analysis to make the reference point cluster gather towards the high-potential solution region; S32: Integration of constraint handling and deadlock detection: Device safety threshold constraints, process timing deadlock detection, feasibility priority selection mechanism; S33: Elite retention and diversity enhancement: Adaptive crossover and mutation, environmental selection, external archive update.

7. The intelligent scheduling and optimization method of the industrial electrical automation system according to claim 6, characterized in that, In S32, the integration of constraint handling and deadlock detection specifically includes: S321: Device safety threshold constraint: Define the degree of constraint violation Embed a dynamic penalty function in fitness calculation: S322: Process Timing Deadlock Detection: Construct a resource allocation matrix A, where the element a ij = 1 indicates that process i occupies device j; after each crossover mutation, detect whether the process sequence of the offspring individual forms a circular wait chain: traverse the process dependency graph based on depth-first search and mark the loops; if a deadlock is detected, trigger a repair operator, randomly select a process in the loop, and insert it into a conflict-free timing position; S323: Feasibility Priority Selection Mechanism: In non-dominated sorting, individuals with a constraint violation degree of ∑c j = 0 are preferentially selected; if there are insufficient feasible solutions, a partial number of solutions with low violation degrees are retained in ascending order of ∑c j to maintain the diversity of the population.

8. The intelligent scheduling and optimization method of the industrial electrical automation system according to claim 7, characterized in that, In S33, the elite retention and diversity enhancement specifically includes: S331: Adaptive crossover and mutation: Crossover probability Decrease with iteration to improve late convergence; Mutation probability S332: Environment Selection: Combine reference point association and crowding distance to select N individuals from the merged population. Preferentially retain non-dominated solutions with a high degree of association with the nearest reference point. Within the same reference point region, select solutions with a large crowding distance to maintain the distribution breadth. S333: External Archive Update: Every 5 generations, use the ε-domination screening mechanism to update the global Pareto front archive and remove historical solutions that are ε-dominated by new solutions.

9. The intelligent scheduling and optimization method for an industrial electrical automation system according to claim 1, wherein The cross-domain strategy execution and feedback in S4 specifically include: S41: Instruction Conversion and Protocol Adaptation: Encode and convert the optimized generated strategy instruction set according to the target system type: Production line control instructions: Convert to IEC 61131-3 structured text executable by PLC, and embed device start / stop timing sequences, motor speed set values, and safety interlock conditions. Energy storage system instructions: Convert to charge / discharge power curves and append battery temperature protection thresholds. Grid interaction instructions: Package demand response signals through the IEC 61850 protocol. Deploy a multi-protocol conversion gateway to support real-time interoperability of OPC UA, Modbus TCP, and MQTT protocols, and ensure that instructions are synchronously sent to each terminal within 200 ms. S42: Execution Status Monitoring and Data Return: Collect feedback data after strategy execution through device-level sensors and the SCADA system, including the production system, energy system, and grid interaction. Add spatio-temporal consistency tags to the feedback data to construct a traceable closed-loop data chain. S43: Dynamic Parameter Adjustment and Model Update: Dynamically correct weight coefficients and perform online model learning. S44: Exception Handling and Strategy Rollback: Real-time detect cross-system instruction conflicts and trigger dynamic priority arbitration: If the grid is in a peak electricity price period and the carbon emission factor > the preset threshold, forcibly enable the green energy-saving mode and limit the power of non-critical equipment; if the production delay exceeds the tolerance limit, temporarily switch to the production guarantee mode and allow the use of standby diesel generators for power supply. When an abnormality is detected in the device health status, automatically generate a preventive maintenance work order and adjust the scheduling strategy.

10. The intelligent scheduling and optimization method of the industrial electrical automation system according to claim 9, characterized in that In S43, the dynamic parameter adjustment and model update specifically include: S431: Dynamically Correct Weight Coefficients: According to the latest carbon emission factor of the grid and the production urgency index, adjust the weights of the optimization function according to the following formula: S432: Online Model Learning: Adopt an incremental federated learning framework. Each edge node updates the parameters of the sub-model based on local feedback data. The central server aggregates the gradients and distributes the global model, and the update period ≤ 5 minutes.

Images