Multi-target dynamic optimization unit load intelligent distribution method

Abstract

The invention relates to the field of power equipment data processing, in particular to a multi-target dynamic optimization unit load intelligent distribution method, which adapts a micro-service cluster through a containerization protocol, adopts an OPC UA protocol and an MQTT protocol to respectively collect thermal power and new energy data at different frequencies, dynamically analyzes an IEC 61850 power grid instruction, converts and stores the IEC 61850 power grid instruction into a Kafka message queue, and sends the IEC 61850 power grid instruction to a network server. The problem of communication protocol heterogeneity is solved. Based on an IEEE 1588 protocol, sub-millisecond clock synchronization of edge nodes is realized, a cubic spline interpolation algorithm is adopted to align multi-frequency data, an interpolation window is dynamically adjusted in combination with a genetic algorithm, and a spatial-temporal characteristic matrix is generated and stored in a Redis database. And performing multi-objective optimization solution through an NSGA-II algorithm, and dynamically adjusting the weight of an objective function to adapt to the frequency modulation requirement of the power grid. A closed-loop feedback mechanism corrects model parameters through Kalman filtering, actual data are written back to a feature matrix, the robustness of the system is improved in combination with a hierarchical fault-tolerant strategy, the problem of space-time mismatch of data is effectively solved, and the economical efficiency and safety of load distribution are improved.

Description

Technical Field

[0001] The present invention relates to the field of power equipment data processing, and particularly to an intelligent load distribution method for units with multi-objective dynamic optimization. Background Art

[0002] In the power plant scenario, the intelligent load distribution of units is a technical method that dynamically calculates the optimal output combination of each unit by establishing a multi-objective optimization model, combining the operating characteristics of the units, the grid dispatching requirements, and the economic indicators. This process is based on the real-time collected unit efficiency curves, coal consumption characteristics, environmental protection emission data, and grid frequency fluctuation information, and uses mixed integer programming or heuristic algorithms. Under the constraints of power balance, ramp rate, minimum start-stop time, etc., the objective functions are to minimize the power supply coal consumption, control pollutant emissions, and balance the equipment life, and solve the global optimal solution set. Through the online data correction and rolling optimization mechanism, the system can adaptively adjust the unit load distribution strategy, thereby improving the overall operating economy of the power plant, effectively supporting the grid peak shaving and frequency modulation requirements, and maintaining the safe and stable operation of the power system.

[0003] In the data processing of the intelligent load distribution of units in the power plant scenario, there are problems in the real-time fusion and consistency maintenance of multi-source heterogeneous data. Specifically, the operating data of thermal power units (such as steam temperature, steam pressure, and flow parameters collected by the DCS system), the output fluctuation data of new energy units, the grid dispatching instructions, and the environmental monitoring data have data flow asynchrony problems due to sampling frequency differences, communication protocol heterogeneity, and timestamp asynchrony, resulting in spatio-temporal mismatch of the input parameters of the load distribution model, which in turn affects the dynamic calibration accuracy of the unit efficiency characteristic curve, making the solution result of the optimization algorithm deviate from the optimal solution of the actual working condition under multiple constraints, and restricting the real-time performance and economy of the load distribution strategy. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides an intelligent load distribution method for units with multi-objective dynamic optimization, which solves the problem of spatio-temporal mismatch of the input parameters of the load distribution model caused by the data flow asynchrony of multi-source heterogeneous data due to sampling frequency differences, communication protocol heterogeneity, and timestamp asynchrony in the power plant scenario.

[0005] To solve the above technical problems, the specific technical solution of the present invention is as follows: The intelligent load distribution method for units with multi-objective dynamic optimization provided by the present invention includes: Collect the operating parameters of thermal power units, the equipment status of new energy power stations, and the grid dispatching instructions through the protocol adaptation microservice cluster deployed on the edge side. After the protocol conversion engine, the multi-source heterogeneous data is unified into a spatio-temporal feature matrix including timestamps, unit status codes, and environmental monitoring values, and stored in the Redis time series database; When the Redis time series database finishes writing the spatio-temporal feature matrix, it triggers the online calibration microservice to extract the boiler efficiency correlation features from the spatio-temporal feature matrix, uses the non-linear least squares method to fit and generate the thermal power coal consumption characteristic curve. At the same time, it calculates the boundary parameters of the adjustable range of new energy output through the sliding window filtering algorithm, and writes the calibrated curve coefficients and interval parameters into the MongoDB parameter database; In response to the version update event of the MongoDB parameter database, start the hybrid optimization engine to build a rolling optimization model that includes unit output constraints and frequency modulation requirements. After generating the initial solution set through the genetic algorithm, use the non-dominated sorting multi-objective optimization algorithm to screen the Pareto front of the solution set, and output the load distribution plan to the MySQL industrial control database; According to the load distribution plan stored in the MySQL industrial control database, send control instructions to the thermal power DCS system and new energy power stations through the OPC protocol, and collect the coal consumption measurement values and output deviation amounts in the actual execution data. Correct the prediction parameters of the thermal power coal consumption characteristic curve through the Kalman filtering algorithm, and at the same time trigger the protocol adaptation microservice cluster to recalculate the interpolation weight coefficients of the spatio-temporal feature matrix to form a closed-loop feedback.

[0006] Furthermore, for the multi-objective dynamic optimization unit load intelligent allocation method of the present invention, the protocol adaptation microservice cluster performs the following operations: Collect the steam turbine speed and main steam pressure parameters of the thermal power DCS system through the containerized OPC UA protocol interface at a sampling rate of 1 Hz, and at the same time subscribe to the output current data of the photovoltaic inverter through the MQTT protocol interface at a sampling rate of 10 Hz; Call the dynamically loaded IEC 61850 protocol parsing module to convert the grid dispatching instructions into a JSON format data stream including timestamps and instruction codes; Write the converted thermal power parameters, new energy data, and grid instructions into independent partitions of the Kafka message queue according to the data source type. Among them, the thermal power data partition is set with a 10-minute life cycle, and the new energy data partition is set with a 5-minute life cycle.

[0007] Furthermore, for the multi-objective dynamic optimization unit load intelligent allocation method of the present invention, the generation process of the spatio-temporal feature matrix includes: Synchronize the clocks of each edge node based on the IEEE 1588 protocol, and control the timestamp error of the collected data within the range of ±1 ms; Use the cubic spline interpolation algorithm to downsample the 10 Hz current data of the new energy power station to 1 Hz to generate a time series synchronized with the thermal power data; When the irradiance change rate in the environmental monitoring data is detected to exceed 50 W / m² / s, trigger the genetic algorithm to adjust the sliding window length from 500 ms to 2 s, and recalculate the interpolation weight coefficient.

[0008] Furthermore, for the intelligent unit load distribution method with multi-objective dynamic optimization according to the present invention, the dynamic fitting algorithm specifically includes: Extract a set of feature points from the boiler efficiency-load relationship field of the spatio-temporal feature matrix, and use the Levenberg-Marquardt algorithm to fit a quadratic polynomial coal consumption curve; When the ash content deviation in the coal quality detection data exceeds 2% of the historical average, retrieve the curve coefficients of the same coal quality label in the MongoDB parameter library for the most recent 7 days and perform weighted averaging; Apply a moving average filter with a window length of 30 seconds to the new energy output data, and generate the upper and lower limits of the output interval after removing outliers based on the RANSAC algorithm.

[0009] Furthermore, for the intelligent unit load distribution method with multi-objective dynamic optimization according to the present invention, the hybrid optimization engine performs the following operations: Generate an initial solution set containing the output values of thermal power units (in 0.1 MW steps) and the photovoltaic frequency regulation reserve capacity (in 0.01 MW steps) through the genetic algorithm; Use the NSGA-II algorithm to perform non-dominated sorting on the solution set, and calculate the three-dimensional objective function values of the thermal power coal consumption, NOx emissions, and gearbox loss coefficient of each solution; When the frequency regulation demand issued by the grid EMS system exceeds 100 MW, adjust the coal consumption target weight from 0.5 to 0.3, and increase the frequency regulation reserve weight from 0.2 to 0.4.

[0010] Furthermore, for the intelligent unit load distribution method with multi-objective dynamic optimization according to the present invention, the closed-loop feedback process includes: sending the load distribution plan in the MySQL industrial control database to the thermal power unit DCS system through the OPC DA protocol, and caching the latest three versions of the plan including timestamps at the edge network node; Collect the deviation between the actual coal consumption value and the predicted value of the plan. When the deviation exceeds 2 times the standard deviation of the thermal power coal consumption characteristic curve, trigger the protocol adaptation microservice cluster to recalculate the interpolation weight of the spatio-temporal feature matrix; Based on the coal consumption curve coefficients stored in the MongoDB parameter database, use the Kalman filter algorithm to update the process noise covariance matrix Q, and correct the state transition parameters in the boiler efficiency prediction model.

[0011] Furthermore, for the intelligent load distribution method of multi-objective dynamic optimization of units in the present invention, the processing of the multi-constraint conditions includes: constructing a constraint set including the maximum ramp rate of 2% per minute of thermal power units, the upper and lower limits of photovoltaic output intervals, and the demand for grid frequency regulation reserve capacity; For individuals violating the ramp rate constraint, a penalty function value of (actual ramp - 2%)²×10^4 is imposed; through the gRPC call to the frequency regulation capacity query interface of the grid EMS system, verify whether the frequency regulation reserve values of each plan in the solution set meet the real-time requirements.

[0012] Furthermore, for the intelligent load distribution method of multi-objective dynamic optimization of units in the present invention, the cooperation between the protocol adaptation microservice cluster and the hybrid optimization engine includes: When a message verification error occurs in the IEC 61850 protocol parsing module, trigger the MongoDB parameter database to roll back to the calibration parameters of the previous valid version; During the iteration process of the genetic algorithm, adopt a distributed transaction lock mechanism to ensure the consistency of the instruction data in the Kafka message queue and the plan records in the MySQL industrial control database.

[0013] Furthermore, for the intelligent load distribution method of multi-objective dynamic optimization of units in the present invention, the process of reverse correction of online monitoring data includes: Write the actual coal consumption data as a new column into the spatio-temporal feature matrix table of the Redis time series database; Trigger the incremental learning thread of the dynamic fitting algorithm, and only update the constant term coefficient in the quadratic polynomial of the coal consumption characteristic; When the volatility of the wind speed monitoring data exceeds 30% of the historical average, increase the crossover probability of the genetic algorithm from 0.8 to 0.9 to accelerate the population evolution.

[0014] Furthermore, for the intelligent load distribution method of multi-objective dynamic optimization of units in the present invention, the system fault tolerance mechanism includes: detecting the heartbeat signal of the hybrid optimization engine through the Consul service registry, and if there is no response for 5 consecutive seconds, switch to the greedy algorithm based on priority to distribute the load; When a timeout occurs in the detection of the MySQL database connection, read the last successfully issued load distribution plan from the edge network node and continue to execute; For the calibration parameters missing in the MongoDB parameter library, automatically trigger the online calibration microservice to re-execute the coal consumption curve fitting process.

[0015] Advantages of the present invention; The present invention realizes the standardized access and asynchronous processing of multi-source heterogeneous data by constructing a protocol adaptation microservice cluster, and uses dynamic clock synchronization and interpolation downsampling technologies to eliminate the sampling frequency differences, generating a spatio-temporal aligned feature matrix, thus solving the problem of input parameter mismatch caused by asynchronous data streams in traditional methods; based on the dynamic weight adjustment mechanism of a hybrid optimization engine and the NSGA-II non-dominated sorting algorithm, it realizes multi-dimensional objective optimization while satisfying the physical constraints of the units and the grid frequency regulation requirements, improving the global optimality and real-time performance of the load distribution strategy; through a closed-loop feedback link, the actual operation data is injected back into the model parameter correction process, and the prediction accuracy of the coal consumption characteristic curve is continuously optimized by combining Kalman filtering and incremental learning technologies, forming a data-driven self-correction system. At the same time, relying on a hierarchical fault tolerance mechanism, the system continues to operate continuously in abnormal scenarios, comprehensively improving the economy, safety and environmental friendliness of power plant load distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the drawings.

[0017] Figure 1 It is a flowchart of an intelligent load distribution method for multi-objective dynamic optimization of units provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention will be described in detail below in conjunction with the accompanying drawings. To better understand the objectives of the present invention, the present invention will be further described in detail below.

[0019] Please refer to Figure 1 , the present invention provides an intelligent load distribution method for multi-objective dynamic optimization of units, including: Step S101, collecting the operating parameters of thermal power units, the equipment status of new energy power stations and the grid dispatching instructions through a protocol adaptation microservice cluster deployed on the edge side, and unifying the multi-source heterogeneous data into a spatio-temporal feature matrix including a timestamp, a unit status code and an environmental monitoring value through a protocol conversion engine, and storing it in a Redis time series database; Step S102: When the Redis time-series database finishes writing the spatio-temporal feature matrix, trigger the online calibration microservice to extract the boiler efficiency-related features from the spatio-temporal feature matrix, generate the thermal power coal consumption characteristic curve by fitting with the non-linear least squares method. At the same time, calculate the boundary parameters of the adjustable interval of new energy output through the sliding window filtering algorithm, and write the calibrated curve coefficients and interval parameters into the MongoDB parameter database; Step S103: In response to the version update event of the MongoDB parameter database, start the hybrid optimization engine to build a rolling optimization model that includes unit output constraints and frequency regulation requirements. After generating the initial solution set through the genetic algorithm, use the non-dominated sorting multi-objective optimization algorithm to screen the Pareto front of the solution set, and output the load distribution plan to the MySQL industrial control database; Step S104: According to the load distribution plan stored in the MySQL industrial control database, send control instructions to the thermal power DCS system and the new energy power station through the OPC protocol, and collect the coal consumption measurement values and output deviation amounts in the actual execution data. Correct the prediction parameters of the thermal power coal consumption characteristic curve through the Kalman filtering algorithm, and at the same time trigger the protocol adaptation microservice cluster to recalculate the interpolation weight coefficients of the spatio-temporal feature matrix to form a closed-loop feedback.

[0020] The intelligent load distribution method for multi-objective dynamic optimization of units provided by the present invention realizes the real-time fusion and dynamic optimization of multi-source heterogeneous data based on the edge computing framework. By deploying a containerized protocol adaptation microservice cluster, the OPCUA protocol is used to collect the steam turbine speed and main steam pressure parameters of thermal power units at a sampling rate of 1 Hz. At the same time, the output current data of photovoltaic inverters is subscribed through the MQTT protocol at a sampling rate of 10 Hz to solve the problem of heterogeneous communication protocols between thermal power and new energy devices. For the IEC 61850 format data of grid dispatching instructions, it is converted into a JSON format data stream containing timestamps and instruction codes by dynamically loading the protocol parsing module to unify the data coding specifications. The converted multi-source data is divided into independent partitions of the Kafka message queue according to the types of thermal power and new energy. The thermal power data partition is set with a 10-minute life cycle, and the new energy data partition is set with a 5-minute life cycle to realize the buffering and asynchronous processing of high-throughput data streams.

[0021] During the generation process of the spatio-temporal feature matrix, sub-millisecond clock synchronization is performed on each edge node based on the IEEE 1588 protocol to eliminate the timestamp deviation at the data acquisition end. The 10Hz high-frequency current data of the new energy power station is downsampled to 1Hz using the cubic spline interpolation algorithm to align its time series with that of the thermal power data. When the irradiance change rate in the environmental monitoring data exceeds 50W / m² / s, the genetic algorithm is triggered to dynamically extend the sliding window length to 2 seconds and optimize the interpolation weight coefficient to solve the data distortion problem in the case of sudden light changes. The generated spatio-temporal feature matrix is stored in the Redis time series database, and the fields include unit status codes, environmental parameters, and a unified time reference, providing a standardized input for subsequent calibration.

[0022] The online calibration microservice extracts the correlation features between boiler efficiency and load from the Redis database, fits a quadratic polynomial coal consumption curve using the Levenberg-Marquardt algorithm, and establishes a quantitative relationship model between load and coal consumption. When it is detected that the deviation of the coal quality ash content exceeds 2% of the historical average, the curve coefficients of the same coal quality label in the MongoDB parameter library in the recent 7 days are automatically retrieved, and temporary calibration parameters are generated through weighted averaging to avoid model inaccuracy caused by coal quality fluctuations. After applying a sliding average filter with a 30-second window length to the new energy output data and removing abnormal outliers using the RANSAC algorithm, the upper and lower limits of the adjustable output range are calculated, and the dynamic constraint boundary parameters are written into MongoDB.

[0023] The hybrid optimization engine responds to the version update event of the MongoDB parameter library, constructs a multi-objective optimization model including the ramp rate of thermal power units, the new energy output range, and the grid frequency regulation requirements. An initial solution set including thermal power output values (in 0.1MW steps) and photovoltaic frequency regulation capacities (in 0.01MW steps) is generated through the genetic algorithm, and the NSGA-II algorithm is used for non-dominated sorting to calculate the three-dimensional objective function values of the coal consumption, NOx emissions, and equipment loss coefficients corresponding to each solution. When the frequency regulation requirement issued by the grid EMS exceeds 100MW, the objective function weights are dynamically adjusted, the coal consumption weight is reduced from 0.5 to 0.3, and the frequency regulation reserve weight is increased from 0.2 to 0.4. The Pareto optimal solution set that meets the real-time requirements is generated and stored in the MySQL industrial control database.

[0024] The closed-loop feedback link distributes the load distribution plan in MySQL to the unit DCS system through the OPC DA protocol and caches the last three versions of the plan at the edge node. The deviation data between the actual coal consumption value and the predicted value is collected. When the deviation exceeds twice the standard deviation of the coal consumption characteristic curve, the protocol adaptation microservice is triggered to recalculate the interpolation weights of the spatio-temporal feature matrix. The Kalman filter algorithm is used to update the process noise covariance matrix, correct the state transition parameters of the boiler efficiency prediction model, and synchronously write the actual coal consumption data as a new dimension into the spatio-temporal feature matrix table of the Redis database, forming a data-driven parameter self-correction mechanism. When the environmental parameters change suddenly, the crossover probability of the adaptive genetic algorithm is adjusted to 0.9 to accelerate the convergence speed of the optimization engine and complete the full-process closed-loop control from data acquisition, model optimization to execution feedback.

[0025] Specifically, for the intelligent load distribution method of multi-objective dynamic optimization of the unit described in the present invention, the protocol adaptation microservice cluster performs the following operations: Collect the steam turbine speed and main steam pressure parameters of the thermal power DCS system through the containerized OPC UA protocol interface at a sampling rate of 1 Hz, and subscribe to the output current data of the photovoltaic inverter through the MQTT protocol interface at a sampling rate of 10 Hz; Call the dynamically loaded IEC 61850 protocol parsing module to convert the grid dispatching instruction into a JSON format data stream containing a timestamp and an instruction code; Write the converted thermal power parameters, new energy data, and grid instructions into independent partitions of the Kafka message queue according to the data source type, where the thermal power data partition is set with a 10-minute life cycle, and the new energy data partition is set with a 5-minute life cycle.

[0026] The protocol adaptation microservice cluster realizes the multi-protocol access capability through containerization technology. It uses the OPC UA protocol interface to collect the real-time operation parameters of the thermal power unit DCS system at a fixed sampling frequency of 1 Hz, including key indicators such as steam turbine speed and main steam pressure, to solve the problem of data access standardization in the thermal power control system. For the high-frequency data characteristics of new energy field station equipment, it subscribes to the output current data of the photovoltaic inverter through the MQTT protocol interface at a sampling rate of 10 Hz to achieve millisecond-level data collection and meet the monitoring requirements of rapid fluctuations in new energy output. The differential data collection strategies for the two types of data effectively balance the data timeliness differences between thermal power and new energy equipment.

[0027] The protocol conversion module adopts a dynamic loading mechanism and automatically calls the corresponding parsing components according to the data type. For the IEC 61850 format instruction data issued by the power grid dispatching center, a dedicated protocol parsing module is loaded to extract the instruction code and the effective timestamp, and convert them into a structured JSON data stream containing the operation type, target value, and execution time window. The semantic information of the original protocol is retained during the conversion process to ensure the integrity and traceability of the dispatching instructions, providing an accurate power grid demand input for subsequent optimization calculations.

[0028] The converted multi-source data stream is divided into storage paths according to the data category. The thermal power operation parameters are written into the "thermal_data" topic partition of the Kafka message queue, the new energy data is written into the "renewable_data" topic partition, and the power grid instructions are stored separately in the "grid_cmd" partition. Different life cycle parameters are set for each partition. The data in the thermal power partition is retained for 10 minutes to adapt to its relatively stable data update rhythm, and the data in the new energy partition is retained for 5 minutes to match its high-frequency update characteristics. The partition strategy combines the time window management mechanism, which not only maintains the timeliness of data processing but also avoids system overload caused by sudden data peaks.

[0029] The data buffer layer adopts a multi-level storage architecture. The thermal power parameters in the Kafka message queue are batch-written into the Redis time series database through batch processing, while the new energy data is updated in real time using stream processing. The different storage strategies for the two types of data match the front-end acquisition frequency, providing a stable and reliable data supply for the spatio-temporal feature matrix generation module. The power grid dispatching instruction data is directly passed through to the optimization calculation module after format conversion, forming a complete preprocessing link from data acquisition, conversion to storage.

[0030] Specifically, for the intelligent allocation method of unit load with multi-objective dynamic optimization described in the present invention, the generation process of the spatio-temporal feature matrix includes: Based on the IEEE 1588 protocol, clock synchronization is performed on each edge node, and the timestamp error of the collected data is controlled within the range of ±1ms. The 10Hz current data of the new energy power station is downsampled to 1Hz using the cubic spline interpolation algorithm to generate a time series synchronized with the thermal power data. When it is detected that the irradiance change rate in the environmental monitoring data exceeds 50W / m² / s, the genetic algorithm is triggered to adjust the sliding window length from 500ms to 2s, and the interpolation weight coefficient is recalculated.

[0031] The generation of the spatio-temporal feature matrix is based on precise time synchronization and dynamic data processing mechanisms. By deploying the PTP (Precision Time Protocol) master-slave clock architecture of the IEEE 1588 protocol, a sub-millisecond-level synchronization network is established among edge nodes. The master clock node periodically broadcasts time synchronization messages, and the slave nodes calculate the clock offset based on the message transmission delay, converging the data acquisition timestamp error to the range of ±1 ms. This synchronization mechanism provides a unified time reference for the temporal alignment of multi-source heterogeneous data, eliminating data misalignment caused by device clock drift.

[0032] After the timestamp correction of the 10Hz high-frequency current data of the new energy power station, the cubic spline interpolation algorithm is used to construct a continuous curve function, and the interpolation points are calculated at each 1-second time node to generate a temporal data sequence matching the 1Hz sampling rate of thermal power. The interpolation process retains the fluctuation characteristics of the original data, and realizes the smooth transition of the data through piecewise polynomial function fitting, avoiding the stepped distortion caused by traditional linear interpolation. The synchronized temporal data is reorganized according to the unified time axis to form a multi-dimensional matrix structure including unit status codes, environmental parameters, and grid commands.

[0033] The dynamic window adjustment module monitors the change gradient of environmental parameters in real time. When the meteorological monitoring unit detects that the irradiance change rate exceeds 50 W / m² / s, it triggers the initialization of the genetic algorithm population. The initial population includes parameter combinations such as window length and interpolation weight coefficient, and the fitness function takes the weighted sum of data reconstruction error and calculation overhead as the optimization goal. After 20 generations of iteration, the optimal parameter combination is selected, the sliding window is extended from the benchmark 500 ms to 2 s, and the interpolation weight coefficient allocation strategy is updated. This mechanism enhances the data fidelity ability in the photovoltaic sudden change scenario while maintaining the timeliness of the data, providing a stable input for the subsequent optimization model.

[0034] Specifically, for the intelligent allocation method of unit load with multi-objective dynamic optimization described in the present invention, the dynamic fitting algorithm specifically includes: Extract the feature point set from the boiler efficiency-load relationship field of the spatio-temporal feature matrix, and use the Levenberg-Marquardt algorithm to fit the quadratic polynomial coal consumption curve; When the ash content deviation in the coal quality detection data exceeds 2% of the historical average, retrieve the curve coefficients of the same coal quality label in the past 7 days from the MongoDB parameter library for weighted average; Apply a sliding average filter with a window length of 30 seconds to the new energy output data, and generate the upper and lower limits of the output interval after removing outliers based on the RANSAC algorithm.

[0035] The dynamic fitting algorithm realizes the accurate modeling of the unit operation characteristics through multi-stage data processing. Extract the data field of the boiler efficiency-load relationship in the spatio-temporal feature matrix from the Redis time series database, perform outlier removal and normalization preprocessing on the discrete feature point set to form a data sample set that meets the requirements of curve fitting. Use the Levenberg-Marquardt algorithm to perform non-linear least squares optimization on the sample set, solve the quadratic polynomial coefficients to characterize the quantitative relationship between coal consumption and load, and establish a dynamic mathematical model of the efficiency characteristics of thermal power units.

[0036] The coal quality anomaly processing module monitors the ash content index in the fuel detection data in real time. When the deviation between the detected value and the historical average exceeds 2%, it triggers the fuzzy query function of the MongoDB parameter library. Construct a composite retrieval condition based on the coal quality label, sulfur content and volatile matter characteristics, and obtain the set of coal consumption curve coefficients under the same coal quality classification in the recent 7 days. Assign a decreasing weight to the retrieval results according to the time distance, and generate temporary calibration parameters through weighted average to maintain the prediction accuracy of the coal consumption characteristic curve under the fuel fluctuation scenario.

[0037] The new energy output data processing link adopts a time window alignment mechanism, applies a 30-second fixed-length moving average filter to the photovoltaic / wind power data stream to smooth short-term power fluctuations. Based on the RANSAC algorithm, construct an output value distribution model, iteratively calculate the distance residuals between the data points and the model, and mark the sampling points outside the 3-fold standard deviation range as outliers and remove them. According to the data distribution characteristics after filtering, calculate the mean plus or minus 2 times the standard deviation as the upper and lower boundary values of the output adjustable interval, providing dynamic constraint parameters for the optimization model. The processed interval parameters and the thermal power coal consumption curve are jointly written into MongoDB to form a complete calibration result of the unit efficiency characteristics.

[0038] Specifically, for the intelligent load distribution method of multi-objective dynamic optimization of the unit described in the present invention, the hybrid optimization engine performs the following operations: Generate an initial solution set containing the output values of thermal power units (step size of 0.1 MW) and the photovoltaic frequency modulation reserve capacity (step size of 0.01 MW) through the genetic algorithm; Use the NSGA-II algorithm to perform non-dominated sorting on the solution set, and calculate the three-dimensional objective function values of the thermal power coal consumption, NOx emissions and gearbox loss coefficient of each solution; When the frequency modulation demand issued by the grid EMS system exceeds 100 MW, adjust the coal consumption target weight from 0.5 to 0.3, and increase the frequency modulation reserve weight from 0.2 to 0.4.

[0039] The hybrid optimization engine realizes the dynamic generation of the load distribution strategy based on the multi-objective collaborative optimization mechanism. The initial population is constructed by real number coding. The output value of the thermal power unit is discretely coded with a minimum adjustment step of 0.1 MW, and the frequency regulation reserve capacity of the photovoltaic is quantified with an accuracy of 0.01 MW. Each solution individual includes the output combination of the thermal power and new energy units and the corresponding frequency regulation capacity parameters. The constraint verification logic is embedded in the coding process to exclude invalid solutions that violate the minimum output or ramp rate limit of the unit, improving the quality of the initial solution set.

[0040] In the non-dominated sorting stage, the NSGA-II algorithm framework is adopted to evaluate the three-dimensional objective function of the population individuals. The coal consumption of the thermal power is calculated based on the currently calibrated quadratic polynomial curve for real-time coal consumption, the NOx emissions are converted through the emission factor model, and the gearbox loss coefficient is derived from the cumulative damage model of historical operation data. After the calculation, the solution set is sorted hierarchically according to the Pareto dominance relationship of the objective function values, and the candidate solution set with the best diversity is selected by combining the crowding degree calculation, providing a high-quality solution space for subsequent decision-making.

[0041] The dynamic weight adjustment module monitors the frequency regulation demand data published by the grid EMS system in real time. When the demand value exceeds the 100 MW threshold, the online weight optimization mechanism is triggered. The adaptability of the current population is evaluated through the selection operator of the genetic algorithm. The weight of the coal consumption target is gradually attenuated from the reference value of 0.5 to 0.3, while the weight of the frequency regulation reserve is increased from 0.2 to 0.4. After the weight vector is adjusted, the fitness value of each solution individual is recalculated to guide the optimization direction to tilt towards the grid frequency regulation demand, generating a load distribution plan that takes into account both economy and grid security. The adjusted weight parameters are synchronously updated to the MongoDB parameter library, providing a new set of reference parameters for subsequent rolling optimization.

[0042] Specifically, for the intelligent load distribution method of multi-objective dynamic optimization described in the present invention, the closed-loop feedback process includes: sending the load distribution plan in the MySQL industrial control database to the DCS system of the thermal power unit through the OPC DA protocol, and caching the latest three versions of the plan including timestamps at the edge network node; Collecting the deviation between the actual coal consumption value and the predicted value of the plan. When the deviation exceeds 2 times the standard deviation of the thermal power coal consumption characteristic curve, triggering the protocol adaptation microservice cluster to recalculate the interpolation weight of the spatio-temporal feature matrix; Based on the coal consumption curve coefficients stored in the MongoDB parameter database, the Kalman filter algorithm is used to update the process noise covariance matrix Q, and the state transition parameters in the boiler efficiency prediction model are corrected.

[0043] The closed-loop feedback process realizes the self-optimization of the model through a data-driven parameter correction mechanism. The OPC DA industrial communication protocol is used to transmit the load distribution plan stored in the MySQL industrial control database to the DCS control system of the thermal power unit. The protocol message encapsulation contains a standardized instruction structure including the unit number, output value, and execution time window. The edge network node synchronously caches the plan versions of the last three timestamps, and each version is attached with a verification hash value. When the main link communication is interrupted, it automatically switches to the latest valid cached instruction to maintain the continuous execution of the control instruction.

[0044] The actual operation data acquisition module continuously monitors the measured coal consumption value of the thermal power unit and calculates the absolute deviation between it and the predicted value of the plan. The deviation detection unit correlates the standard deviation data of the coal consumption characteristic curve stored in the MongoDB parameter library. When the deviation for three consecutive sampling periods exceeds 200% of the standard deviation value, it triggers the interpolation weight recalculation process of the protocol adaptation microservice cluster. The recalculation instruction is transmitted to the spatio-temporal feature matrix generation module through the message queue, driving the genetic algorithm to re-optimize the sliding window parameters and interpolation coefficients.

[0045] The model parameter correction link is based on the Kalman filter framework, and extracts the quadratic polynomial coefficients of the current coal consumption curve from MongoDB as the initial value of the state vector. The update of the process noise covariance matrix Q uses the recursive least squares algorithm, and combines the real-time coal consumption deviation data to dynamically adjust the diagonal element values of the matrix, optimizing the state transition equation parameters of the boiler efficiency prediction model. The corrected coefficient set is written back to the new version record in the MongoDB parameter library, triggering the incremental model reconstruction of the hybrid optimization engine, forming a closed-loop learning link from execution feedback to parameter update.

[0046] Specifically, for the intelligent load distribution method of the multi-objective dynamic optimization of the unit described in the present invention, the multi-constraint condition processing includes: constructing a constraint set including the maximum ramp rate of 2% per minute of the thermal power unit, the upper and lower limits of the photovoltaic output interval, and the grid frequency modulation reserve capacity requirement; For individuals violating the ramp rate constraint, a penalty function value of (actual ramp amount - 2%)²×10^4 is imposed; the frequency modulation capacity query interface of the grid EMS system is called through gRPC to verify whether the frequency modulation reserve values of each plan in the solution set meet the real-time requirements.

[0047] The feasibility of the optimization model is ensured by the hierarchical verification mechanism for multi-constraint processing. The constraint set construction module extracts the historical operation data of the thermal power unit from the MongoDB parameter library, and statistically generates a dynamic constraint threshold of 2% of the maximum ramp rate per minute, which is updated in conjunction with the thermal characteristic parameters of the unit. The upper and lower limits of the photovoltaic output range are calculated based on the distribution characteristics of the output data after sliding window filtering, and the mean plus or minus two standard deviations are used to form a dynamic boundary constraint. The grid frequency regulation reserve capacity demand is dynamically obtained by subscribing to the real-time frequency deviation signal released by the EMS system, forming a multi-dimensional constraint system that includes equipment physical limitations and grid operation requirements.

[0048] The penalty function processing module intervenes in the fitness calculation stage of the genetic algorithm to perform constraint verification on the climbing rate parameters of the individuals in the population. When it is detected that the output change in adjacent time periods exceeds the 2% threshold, the quadratic penalty function calculation process is triggered. The penalty function value is amplified by a factor of 10,000 based on the square difference between the actual climbing amount and the threshold, and the calculation result is superimposed on the fitness value of the objective function, so that individuals that violate the constraints are naturally eliminated in the population selection stage. This mechanism guides the optimization direction to converge to the feasible solution space while maintaining population diversity.

[0049] The frequency regulation capacity verification phase is implemented through the gRPC remote call framework. After the NSGA-II algorithm generates a Pareto solution set, a two-way communication link is established with the power grid EMS system. The call process encapsulates the frequency regulation reserve capacity query request message, carrying the unit output value and time window parameters in the load distribution plan, and the EMS system returns the real-time frequency regulation capacity margin assessment result. Plans that fail the verification are automatically marked as invalid solutions, triggering the optimization engine to re-execute local search until a feasible solution set that meets the real-time needs of the power grid is generated. The verification results are synchronously written to the audit log of the MySQL industrial control database to provide data support for strategy backtracking.

[0050] Specifically, in the multi-objective dynamic optimization unit load intelligent allocation method of the present invention, the collaboration between the protocol adaptation microservice cluster and the hybrid optimization engine includes: When a message verification error occurs in the IEC 61850 protocol parsing module, the MongoDB parameter database is triggered to roll back to the previous valid version calibration parameter; During the iterative process of the genetic algorithm, a distributed transaction lock mechanism is used to ensure the consistency of the instruction data in the Kafka message queue and the plan records in the MySQL industrial control database.

[0051] The collaboration between the protocol adaptation microservice cluster and the hybrid optimization engine realizes the improvement of system robustness through exception handling and data consistency mechanisms. When the IEC 61850 protocol parsing module detects an abnormal message checksum or semantic integrity error, it triggers a rollback operation of the parameter database. The error detection unit traverses and validates the syntax tree of the control command field in the message based on the ASN.1 data structure verification rules, and generates an exception event when an invalid opcode or out-of-bounds parameter is found. This event notifies the MongoDB parameter database through the message bus to perform a version rollback, retrieves the last calibrated parameter version that passed the consistency check from the operation log, overwrites the data record in the current abnormal state, and maintains the reliability of the input parameters of the optimization model.

[0052] The distributed transaction lock mechanism intervenes during the iterative process of the genetic algorithm and adopts an optimistic lock strategy to manage the interaction between the Kafka message queue and the MySQL database. When the fitness of the population individuals is calculated, a unique transaction ID is assigned to each solution set. Before writing the pre-record of the solution set into the MySQL industrial control database, the instruction partition of the Kafka message queue is locked first. The transaction manager coordinates data writing through a two-phase commit protocol. In the first phase of pre-commit, the solution set metadata is written into the temporary topic of Kafka. In the second phase, it is committed to permanent storage after successful insertion into MySQL. If any phase fails, a transaction rollback is triggered. This mechanism effectively prevents data state inconsistency problems caused by network jitter during the optimization process.

[0053] The exception recovery module forms a linkage with the optimization engine. After the parameter rollback operation is completed, it sends a model reconstruction instruction to the hybrid optimization engine. The engine reloads the valid version of the coal consumption curve coefficient and output interval parameters from MongoDB, extracts the corresponding spatio-temporal feature matrix of the corresponding period from the Redis time series database based on the rollback timestamp, and re-initializes the genetic algorithm population. During the reconstruction process, the read offset of the Kafka message queue is synchronized and rolled back to avoid dirty data flowing into the subsequent processing flow, realizing the full-link state recovery in abnormal scenarios.

[0054] Specifically, for the intelligent allocation method of unit load with multi-objective dynamic optimization described in the present invention, the online monitoring data reverse correction process includes: Writing the actual coal consumption data as a new column into the spatio-temporal feature matrix table of the Redis time series database; Triggering the incremental learning thread of the dynamic fitting algorithm and only updating the constant term coefficient in the quadratic polynomial of the coal consumption characteristic; When the volatility of the wind speed monitoring data exceeds 30% of the historical average, increasing the crossover probability of the genetic algorithm from 0.8 to 0.9 to accelerate the population evolution.

[0055] The reverse correction process of online monitoring data realizes the dynamic optimization of model parameters through a hierarchical data fusion mechanism. After the actual coal consumption data completes quality verification, it is written as a new field into the spatio-temporal feature matrix table structure of the Redis time series database. The new columns include the measured coal consumption value, the acquisition timestamp, and the data version identifier. The write operation adopts columnar storage extension technology, which forms a spatio-temporal association between real-time operation data and historical feature data without affecting the query efficiency of the original matrix, providing a complete data basis for incremental learning.

[0056] After the incremental learning module detects a Redis table structure change event, it starts an independent thread to perform parameter fine-tuning. The thread extracts the measured coal consumption data of the latest three cycles from Redis, applies the local optimization mode of the Levenberg-Marquardt algorithm, fixes the quadratic and linear coefficients of the quadratic polynomial coal consumption curve, and only performs gradient descent optimization on the constant term. The optimization process adopts a read-write lock mechanism to freeze the model call interface during parameter update to maintain the continuous availability of the online calibration service.

[0057] The environmental parameter mutation response mechanism calculates the rolling volatility index of wind speed monitoring data in real time and establishes a dynamic threshold based on the mean and standard deviation of the past 30 sampling cycles. When the real-time volatility breaks through 30% of the historical mean, it sends an algorithm parameter adjustment instruction to the hybrid optimization engine, increasing the crossover probability of the genetic algorithm from the benchmark value of 0.8 to 0.9. The parameter adjustment signal is broadcast to all computing nodes through the message queue, triggering the re-initialization of the population, accelerating the spread of high-quality gene combinations using the increased gene exchange probability, and shortening the convergence time of the optimization solution. The adjusted crossover probability parameter is synchronously recorded in the operation log of MongoDB to provide a traceability basis for subsequent optimization processes.

[0058] Specifically, for the intelligent load distribution method of multi-objective dynamic optimization of the unit load described in the present invention, the system fault tolerance mechanism includes: detecting the heartbeat signal of the hybrid optimization engine through the Consul service registry, and switching to the greedy algorithm based on priority to distribute the load if there is no response for 5 consecutive seconds; When a MySQL database connection timeout is detected, read the last successfully issued load distribution plan from the edge network node and continue to execute; For the calibration parameters missing in the MongoDB parameter library, automatically trigger the online calibration microservice to re-execute the coal consumption curve fitting process.

[0059] The system fault tolerance mechanism improves the system reliability through a hierarchical fault detection and redundancy processing strategy. The Consul service registration center configures a health check probe to periodically receive the heartbeat signals sent by the hybrid optimization engine, and the probe interval is set to 1 second. When the heartbeat detection fails continuously for 5 times, the service status of the optimization engine is marked as abnormal, and the service discovery module immediately switches the traffic to the standby algorithm module. The standby algorithm adopts a priority greedy strategy, allocates the load in ascending order of the coal consumption rate of thermal power units, preferentially meets the basic power demand of the power grid, and maintains the minimum operation guarantee of the power plant.

[0060] The database connection fault tolerance module monitors the TCP connection status of the MySQL industrial control database in real time. When the three-way handshake request times out continuously and the total delay exceeds 3 seconds, it is determined that the database service is unavailable. The edge network node maintains the local cache of the load distribution plans successfully issued in the last three times, and each cached plan is attached with an SHA-256 checksum and a generation timestamp. During the fault switch, the control instruction issuing module automatically selects the valid plan with the latest timestamp, and after verifying the data integrity through the checksum, it bypasses the database and directly sends instructions to the unit control system.

[0061] The parameter integrity verification process scans the calibration parameter records in the MongoDB parameter library through a scheduled task. When it detects that the digital field of the coal consumption curve system is missing or the version numbers are discontinuous, it triggers the compensation execution mechanism of the online calibration microservice. The compensation process extracts the historical spatio-temporal feature matrix of the last 24 hours from the Redis time series database, re-executes the entire process of data cleaning, feature extraction, and curve fitting, generates a new calibration parameter version, and writes it into the isolation storage area. During the parameter reconstruction, the load distribution system automatically enables the default parameter set to maintain the basic optimization function, and performs a hot switch after the new parameters pass the consistency check.

[0062] The specific implementation mode of the present invention constructs a four-layer architecture system including a data acquisition layer, a feature processing layer, an optimization calculation layer, and an execution feedback layer based on the actual operation scenario of the power plant. The OPC UA protocol adaptation microservice is deployed on the edge server in the thermal power unit control room to collect the steam turbine speed and main steam pressure parameters at a sampling period of 1 Hz. At the same time, the inverter output current data is obtained through the MQTT microservice of the photovoltaic power station edge node at a frequency of 10 Hz. The IEC 61850 format instructions issued by the power grid dispatching center are converted by a dynamically loaded protocol parsing module to generate a JSON instruction stream with a millisecond-level timestamp. The converted multi-source data is written into independent partitions of the Kafka message queue according to three categories: thermal power, new energy, and power grid. The thermal power partition is set with a 10-minute life cycle, and the new energy partition is set with a 5-minute life cycle to achieve the buffering and asynchronous processing of high-frequency data.

[0063] In the spatio-temporal feature matrix generation stage, the IEEE 1588 protocol is used to construct a precise clock synchronization network. The master clock node sends synchronization messages every 2 seconds, and the edge nodes calculate the clock offset compensation amount to control the timestamp error within ±1ms. Cubic spline interpolation downsampling is performed on the 10Hz current data of new energy sources, and a time series aligned with the thermal power data is generated at every whole second. When the meteorological monitoring unit detects that the irradiance change rate exceeds 50W / m² / s, the genetic algorithm optimization process is triggered. The individual encoding of the population includes the sliding window length from 500ms to 2s and the interpolation weight combination, and the optimal parameter configuration is obtained through 20 generations of iteration to improve the data fidelity under mutation scenarios.

[0064] The dynamic calibration module extracts the boiler efficiency-load relationship data in the spatio-temporal feature matrix from the Redis database, and uses the Levenberg-Marquardt algorithm to fit the quadratic polynomial coal consumption curve. When the coal quality ash deviation exceeds 2%, the MongoDB historical database is retrieved for parameter fusion. After the new energy output data is filtered by a sliding average with a window length of 30 seconds, the RANSAC algorithm is used to remove the outliers outside ±3σ, and the upper and lower limits of the output interval are calculated. The hybrid optimization engine responds to the parameter library version change event, generates an initial solution set with a thermal power output step of 0.1MW and a PV reserve step of 0.01MW using real number encoding, performs three-dimensional objective non-dominated sorting through the NSGA-II algorithm, and reduces the coal consumption weight from 0.5 to 0.3 when the grid frequency modulation demand exceeds 100MW. The optimization results are written into the MySQL industrial control library.

[0065] In the closed-loop execution stage, the load plan is sent to the DCS system through the OPC DA protocol, and the edge gateway caches the instruction versions of the last three timestamps. When the actual coal consumption deviation exceeds 2 times the standard deviation, the interpolation weight recalculation process is triggered, and the Kalman filter module updates the process noise covariance matrix Q value. The Consul service monitoring center detects the heartbeat of the optimization engine every 5 seconds and switches to the greedy algorithm to allocate loads in case of failure. When the MySQL connection times out for 3 seconds, the edge cache plan is automatically enabled. When the MongoDB parameters are missing, the online calibration service is triggered to reconstruct the coal consumption curve, forming a complete data-driven optimization closed-loop.

[0066] The technical features involved in the present invention are explained as follows: Protocol adaptation microservice cluster: OPC UA protocol (Open Platform Communications Unified Architecture): An industrial automation communication standard protocol used for data acquisition in the thermal power DCS system, providing a secure and reliable data transmission channel and supporting the encapsulation of complex data structures.

[0067] MQTT protocol (Message Queuing Telemetry Transport): A lightweight publish / subscribe messaging protocol used for high-frequency collection of photovoltaic inverter data, suitable for low-bandwidth and high-concurrency scenarios of new energy devices.

[0068] IEC 61850: An international standard protocol for power system automation that defines the semantic model and communication specification of grid dispatching instructions, which needs to be parsed into JSON format for unified data processing.

[0069] Data processing and storage module: Kafka message queue: A distributed stream processing platform that divides independent partitions according to data source types (such as thermal_data / renewable_data), sets different lifecycle periods (10 minutes for thermal power and 5 minutes for new energy), and realizes multi-frequency data buffering and asynchronous decoupling.

[0070] Redis time series database: An in-memory database that stores spatio-temporal feature matrices containing timestamps, unit codes, and environmental parameters, supporting high-speed read / write and time range queries.

[0071] Optimization algorithm module: NSGA-II algorithm (Non-dominated Sorting Genetic Algorithm II): A multi-objective genetic algorithm that screens the Pareto optimal solution set through non-dominated sorting and crowding degree calculation to solve the multi-objective trade-off problem of coal consumption, emissions, and equipment losses.

[0072] RANSAC algorithm (Random Sample Consensus): A robust regression algorithm used for detecting outliers in new energy output data, which iteratively eliminates outliers deviating from the main distribution based on random sampling.

[0073] gRPC (Google Remote Procedure Call): A high-performance RPC framework used to call the grid EMS system interface to verify the demand for frequency modulation reserve capacity.

[0074] Consul: A distributed service mesh solution that realizes microservice registration discovery and health check. The heartbeat detection interval is 1 second, and 5 consecutive failures trigger service degradation.

[0075] IEEE 1588 PTP (Precision Time Protocol): A precision clock synchronization protocol that compensates for the clock deviation between the master and slave nodes, with the timestamp error controlled within ±1ms.

[0076] Levenberg - Marquardt algorithm: A non - linear least - squares optimization algorithm used for fitting coal consumption characteristic curves, balancing the convergence of gradient descent and the Gauss - Newton method.

[0077] SHA - 256: A cryptographic hash algorithm that generates a checksum for the load distribution plan to ensure the integrity of edge - cached data.

[0078] TCP handshake timeout: A database connection detection mechanism. If the handshake fails three times in a row (total delay > 3 seconds), it is determined that the connection is abnormal, triggering local cache reading.

[0079] Explanation of the collaborative relationship of the technical features of the present invention: Data acquisition and processing chain: Protocol adaptation microservice (OPC UA / MQTT) → Kafka partition buffering → Clock synchronization (IEEE 1588) → Interpolation and downsampling → Redis spatio - temporal matrix generation.

[0080] Example: After interpolating and aligning the thermal power 1Hz data and new - energy 10Hz data, a matrix input with a unified time reference is formed.

[0081] Optimization and execution closed - loop: MongoDB parameter update → NSGA - II multi - objective optimization → MySQL plan storage → OPC DA instruction issuance → Actual data feedback → Kalman filter correction.

[0082] Example: The deviation of coal consumption prediction triggers the recalculation of the interpolation weight genetic algorithm, forming data - driven parameter self - optimization.

[0083] Abnormal handling link: IEC 61850 parsing exception → MongoDB version rollback → Optimization engine reconstruction → Kafka offset rollback.

[0084] Example: When the grid command verification fails, it rolls back to the previous valid parameter version to maintain the stability of the optimization model.

[0085] Spatio - temporal alignment mechanism: Cubic spline interpolation combined with dynamic window adjustment (500ms → 2s) to improve the data fidelity in the new - energy sudden - change scenario.

[0086] Hybrid optimization strategy: Collaboration between genetic algorithm (global search) and NSGA - II (local optimization), reducing the solution time and improving the quality of the solution set.

[0087] Closed - loop self - correction: Kalman filter real - time corrects the prediction model, making the coal consumption error stable in the long term and superior to traditional open - loop systems.

[0088] The present invention solves the spatio-temporal mismatch problem of multi-source heterogeneous data in power plants by constructing a multi-level data processing architecture. First, a containerized protocol adaptation microservice cluster is deployed. The OPC UA protocol is used to collect the parameters of the thermal power DCS system at a fixed frequency of 1 Hz, and the new energy device data is subscribed at 10 Hz through the MQTT protocol to eliminate communication protocol heterogeneity. For the grid IEC 61850 command data, the protocol parsing module is dynamically loaded and converted into a unified JSON format. The Kafka message queue partitions are divided according to the data source type and different lifecycles are set to achieve asynchronous buffering and standardized encapsulation of multi-frequency data streams.

[0089] Based on the IEEE 1588 precision clock protocol, sub-millisecond time synchronization is performed on each edge node to construct a unified time reference system. The cubic spline interpolation algorithm is used to downsample the 10 Hz high-frequency data of the new energy station to 1 Hz to generate a time series synchronized with the thermal power data. When the environmental parameters mutate, the genetic algorithm is triggered to dynamically adjust the sliding window length and recalculate the interpolation weight coefficients to solve the timing misalignment problem caused by the data sampling rate difference. After the processed multi-source data is aligned by timestamp, it is written into the Redis time series database to form a spatio-temporal feature matrix containing unit status codes, environmental parameters, and grid commands.

[0090] A data-driven dynamic calibration mechanism is constructed through a hybrid optimization engine. When abnormal coal quality parameters are detected, similar working condition curves in the historical database are retrieved for weighted fusion to maintain the accuracy of the coal consumption model. The NSGA-II algorithm is used to perform non-dominated sorting on the multi-objective optimization solution set, and the objective function weights are dynamically adjusted in combination with the real-time frequency modulation requirements. In the closed-loop feedback link, the actual execution data is written back to the spatio-temporal feature matrix, the model noise parameters are corrected through the Kalman filter algorithm, and the interpolation weight recalculation process is triggered to form a full-link self-correcting system from data collection, model optimization to execution feedback.

Claims

1. A multi-objective dynamic optimization method for intelligent distribution of unit loads, characterized in that: include: The protocol adaptation microservice cluster deployed on the edge side collects the operating parameters of thermal power units, the status of new energy station equipment, and grid dispatch instructions. The protocol conversion engine unifies the multi-source heterogeneous data into a spatiotemporal feature matrix containing timestamps, unit status codes, and environmental monitoring values, and stores them in the Redis time series database. When the Redis time series database completes writing the spatiotemporal feature matrix, the online calibration microservice is triggered to extract boiler efficiency correlation features from the spatiotemporal feature matrix, and a thermal power coal consumption characteristic curve is generated by fitting using the nonlinear least squares method. At the same time, the adjustable interval boundary parameters of the new energy output are calculated by the sliding window filtering algorithm, and the calibrated curve coefficients and interval parameters are written into the MongoDB parameter database; In response to the version update event of the MongoDB parameter database, a hybrid optimization engine is started to build a rolling optimization model including unit output constraints and frequency regulation requirements, after an initial solution set is generated by a genetic algorithm, a non-dominated sorting multi-objective optimization algorithm is used to perform Pareto frontier screening on the solution set, and a load distribution plan is output to a MySQL industrial control database; According to the load distribution plan stored in the MySQL industrial control database, control instructions are issued to the thermal power DCS system and the new energy station through the OPC protocol, and the coal consumption measurement values ​​and output deviations in the actual execution data are collected. The predicted parameters of the thermal power coal consumption characteristic curve are corrected through the Kalman filter algorithm, and at the same time, the protocol adaptation microservice cluster is triggered to recalculate the interpolation weight coefficients of the spatiotemporal feature matrix to form a closed-loop feedback.

2. The multi-objective dynamic optimization unit load intelligent allocation method according to claim 1 is characterized in that: The protocol adaptation microservice cluster performs the following operations: The turbine speed and main steam pressure parameters of the thermal power DCS system are collected through the container-based OPC UA protocol interface at a sampling rate of 1 Hz, and the output current data of the photovoltaic inverter is subscribed through the MQTT protocol interface at a sampling rate of 10 Hz; Call the dynamically loaded IEC 61850 protocol parsing module to convert the power grid dispatching instruction into a JSON format data stream containing a timestamp and instruction code; The converted thermal power parameters, new energy data and power grid instructions are written into independent partitions of the Kafka message queue according to the data source type. The thermal power data partition is set to a 10-minute life cycle, and the new energy data partition is set to a 5-minute life cycle.

3. The multi-objective dynamic optimization unit load intelligent allocation method according to claim 2 is characterized in that: The generation process of the spatiotemporal feature matrix includes: Based on the IEEE 1588 protocol, the clock of each edge node is synchronized to control the timestamp error of the collected data within the range of ±1ms; The 10Hz current data of the new energy station is downsampled to 1Hz using the cubic spline interpolation algorithm to generate a time series synchronized with the thermal power data; When it is detected that the irradiance change rate in the environmental monitoring data exceeds 50W / m² / s, the genetic algorithm is triggered to adjust the sliding window length from 500ms to 2s and recalculate the interpolation weight coefficient.

4. The multi-objective dynamic optimization unit load intelligent allocation method according to claim 1 is characterized in that: The dynamic fitting algorithm specifically includes: Extracting a feature point set from the boiler efficiency-load relationship field of the spatiotemporal feature matrix, and fitting a quadratic polynomial coal consumption curve using a Levenberg-Marquardt algorithm; When the ash content deviation in the coal quality test data exceeds the historical average by 2%, the curve coefficients of the same coal quality number in the last 7 days are retrieved from the MongoDB parameter library for weighted averaging; A sliding average filter with a window length of 30 seconds is applied to the renewable energy output data, and the upper and lower limits of the output range are generated after removing abnormal points based on the RANSAC algorithm.

5. The multi-objective dynamic optimization unit load intelligent allocation method according to claim 4 is characterized in that: The hybrid optimization engine performs the following operations: Generate an initial solution set including the output value of thermal power units and photovoltaic frequency regulation reserve capacity through genetic algorithm; The NSGA-II algorithm is used to perform non-dominated sorting on the solution set, and the three-dimensional objective function values ​​of thermal power coal consumption, NOx emissions and gearbox loss coefficient of each solution are calculated; When the frequency regulation demand released by the power grid EMS system exceeds 100MW, the coal consumption target weight will be adjusted from 0.5 to 0.3, and the frequency regulation reserve weight will be increased from 0.2 to 0.

4.

6. The multi-objective dynamic optimization unit load intelligent allocation method according to claim 5 is characterized in that: The closed-loop feedback process includes: sending the load distribution plan in the MySQL industrial control database to the DCS system of the thermal power unit through the OPC DA protocol, and caching the latest three versions of the plan including the timestamp at the edge gateway node; Collect the deviation between the actual coal consumption value and the predicted value of the plan, and when the deviation exceeds twice the standard deviation of the thermal power coal consumption characteristic curve, trigger the protocol adaptation microservice cluster to recalculate the interpolation weight of the spatiotemporal feature matrix; Based on the coal consumption curve coefficients stored in the MongoDB parameter database, the Kalman filter algorithm is used to update the process noise covariance matrix Q, and the state transition parameters in the boiler efficiency prediction model are corrected.

7. The multi-objective dynamic optimization unit load intelligent allocation method according to claim 6 is characterized in that: The multi-constraint processing includes: constructing a constraint set including a maximum ramp rate of 2% per minute for thermal power units, upper and lower limits of photovoltaic output range, and grid frequency regulation reserve capacity requirements; For individuals that violate the climbing rate constraint, a penalty function value of (actual climbing amount - 2%)²×10^4 is imposed; the frequency regulation capacity query interface of the power grid EMS system is called through gRPC to verify whether the frequency regulation reserve values ​​of each plan in the solution meet the real-time demand.

8. The multi-objective dynamic optimization unit load intelligent allocation method according to claim 7 is characterized in that: The collaboration between the protocol adaptation microservice cluster and the hybrid optimization engine includes: When a message verification error occurs in the IEC 61850 protocol parsing module, the MongoDB parameter database is triggered to roll back to the previous valid version calibration parameter; During the iterative process of the genetic algorithm, a distributed transaction lock mechanism is used to ensure the consistency of the instruction data in the Kafka message queue and the plan records in the MySQL industrial control database.

9. The multi-objective dynamic optimization unit load intelligent allocation method according to claim 8 is characterized in that: The online monitoring data reverse correction process includes: Writing the actual coal consumption data as a newly added column into the spatiotemporal feature matrix table of the Redis time series database; Triggering the incremental learning thread of the dynamic fitting algorithm to update only the constant term coefficient in the quadratic polynomial of coal consumption characteristics; When the volatility of wind speed monitoring data exceeds the historical mean by 30%, the crossover probability of the genetic algorithm is increased from 0.8 to 0.9 to accelerate population evolution.

10. The multi-objective dynamic optimization unit load intelligent allocation method according to claim 9, characterized in that: The system fault tolerance mechanism includes: detecting the heartbeat signal of the hybrid optimization engine through the Consul service registration center, and switching to a priority-based greedy algorithm to distribute the load if there is no response for 5 consecutive seconds; When the MySQL database connection timeout is detected, the last successfully issued load distribution plan is read from the edge gateway node to continue execution; For calibration parameters missing in the MongoDB parameter library, the online calibration microservice is automatically triggered to re-execute the coal consumption curve fitting process.

Images