Optimal dispatching method for incremental distribution network of source-grid-load-storage coordination in zero-carbon park

Abstract

This invention discloses an optimized scheduling method for incremental distribution networks in zero-carbon industrial parks, involving the field of power system optimization scheduling technology. It includes acquiring predicted data on renewable energy power and load power within the scheduling cycle, along with external grid interaction constraints. An improved multi-objective adaptive evolutionary algorithm is used to solve these constraints, generating operational strategy curves for distributed power sources, energy storage systems, and flexible loads. The optimized scheme is then distributed to the incremental distribution network control system for online scheduling. During the scheduling process, actual operating data is collected, and closed-loop feedback is used to continuously correct the predicted dataset for the next cycle. The algorithm's adaptive optimization improves the accuracy of multi-objective scheduling, while real-time data correction reduces prediction errors. This method achieves continuous and smooth control of all equipment, strengthens the synergy between the power source, grid, load, storage, and external grid interactions, and ensures that scheduling commands are highly adapted to actual operating conditions, guaranteeing stable, efficient, and low-carbon operation of the zero-carbon industrial park distribution network.

Description

Technical Field

[0001] This invention belongs to the field of power system optimization and dispatching technology, specifically a method for optimizing and dispatching incremental distribution networks with source-grid-load-storage coordination in zero-carbon industrial parks. Background Technology

[0002] In the zero-carbon industrial park, the incremental distribution network routine dispatching collects initial forecast data on renewable energy power and load power, and performs calculations in conjunction with fixed interactive constraints from the external power grid. Traditional solutions use conventional multi-objective optimization algorithms to solve the problem, employing independent control methods for distributed power sources, energy storage systems, and flexible loads to generate discrete output setpoints. The dispatching scheme is issued in an open-loop manner, without introducing operational data to interfere with subsequent processes.

[0003] Conventional multi-objective evolutionary algorithms have fixed parameters, weak adaptive adjustment capabilities, and limited multi-objective optimization accuracy, failing to generate continuous and smooth equipment operation strategy curves. Scheduling relies solely on initial prediction data, making it unable to correct deviations between actual operation and predictions. The coordinated regulation of power generation, grid, load, and storage is ineffective, and external grid interaction constraints are difficult to match with internal operating conditions. Flexible loads, energy storage regulation, and fluctuations in new energy output are disconnected, resulting in poor adaptability between commands and operating conditions, making it impossible to simultaneously achieve low-carbon and economical operation in the industrial park.

[0004] The problem of improving the algorithm's adaptive capability and the open-loop operation of scheduling needs to be addressed. This can be achieved by improving the algorithm to generate continuous operation strategies, establishing a feedback mechanism for actual operation data, and continuously correcting the predicted data to realize real-time collaborative online scheduling between the source network, load storage, and external network. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; To this end, this invention proposes a method for optimizing the scheduling of incremental distribution networks in zero-carbon industrial parks that integrates source-grid-load-storage coordination, including: Obtain the prediction data set of the incremental distribution network within the scheduling cycle. The prediction data set includes new energy power prediction data, load power prediction data, and external power grid interaction constraints. The predicted dataset is input into an improved multi-objective adaptive evolutionary algorithm to generate a collaborative optimization scheduling scheme, which defines the operation strategy curves of distributed power sources, energy storage systems and flexible loads within the park. The aforementioned collaborative optimization scheduling scheme is sent to the control system of the incremental distribution network to perform online scheduling of distributed power sources, energy storage systems and flexible loads; During the scheduling process, actual operating data of the incremental distribution network is collected, and the actual operating data is fed back to the predicted data set for the next scheduling cycle for rolling correction.

[0006] Furthermore, obtain the predicted data set of the incremental distribution network within the scheduling cycle, including: Collect historical power generation data of distributed power sources, historical electricity consumption data of loads, and historical meteorological data within the park; Based on the historical meteorological data, the new energy power prediction data is generated through a power prediction model that combines a physical model and a neural network. By analyzing the patterns in the historical electricity consumption data and combining it with the park's production plan information, the load power prediction data is generated using a time-series decomposition prediction model. Electricity price curves, power interaction limits, and carbon emission factors are obtained from the external power grid dispatch center as external power grid interaction constraints.

[0007] Furthermore, the working principle of the improved multi-objective adaptive evolutionary algorithm includes: Initialize and generate a population containing multiple candidate scheduling solutions, each candidate scheduling solution representing a complete source-grid-load-storage coordinated operation plan; Calculate the target value of carbon emission intensity and economic cost for each candidate scheduling scheme solution in the population, and perform non-dominated ranking of the population based on these two target values; Differentiated adaptive crossover and mutation probabilities are assigned to candidate scheduling schemes at different sorting levels, with the mutation probability of higher-level solutions being lower than that of lower-level solutions. Based on the differentiated crossover probability, crossover operations are performed between individuals in the population, and based on the differentiated mutation probability, mutation operations are performed on individuals in the population to generate offspring populations. From the merged parent and offspring populations, a new parent population is selected based on non-dominated sorting and crowding distance, and this process is repeated iteratively until the preset convergence condition is met. A Pareto front solution set is selected from the final population, and the final collaborative optimization scheduling scheme is determined from it based on the preset decision preferences.

[0008] Furthermore, the target values ​​for carbon emission intensity and economic cost are calculated for each candidate scheduling scheme solution in the population, including: Based on the operation strategy curves in the candidate scheduling scheme solutions, calculate the net interaction power between the park and the external power grid during the scheduling period; By combining the carbon emission factor in the external power grid interaction constraint, the carbon emissions generated by the interaction with the external power grid are calculated, and then summed with the carbon emissions of the distributed power sources within the park to obtain the total carbon emissions. The ratio of the total carbon emissions to the total energy supply of the park is used as the carbon emission intensity target value. The target economic cost is obtained by summing the fuel cost of distributed power sources, the cost of purchasing electricity from the external power grid, the cycle loss cost of the energy storage system, and the incentive compensation cost of flexible loads during the scheduling cycle.

[0009] Furthermore, the collaborative optimization scheduling scheme is distributed to the control system of the incremental distribution network to perform online scheduling of distributed power sources, energy storage systems, and flexible loads, including: The operation strategy curves in the aforementioned collaborative optimization scheduling scheme are analyzed and decomposed into active power setting instructions for photovoltaic inverters, charging and discharging power setting instructions for energy storage converters, and power regulation instructions for flexible load controllers. The active power setting command, charge / discharge power setting command, and power adjustment command are sent to the corresponding field device controllers via the communication network. The field equipment controller drives the photovoltaic inverter, energy storage converter and flexible load switch to execute corresponding command actions.

[0010] Furthermore, during the scheduling execution process, actual operational data of the incremental distribution network is collected, including: By deploying smart meters and sensors at key nodes of the distribution network, the actual output of distributed power sources, the actual power of loads, the actual state of charge of energy storage systems, and the actual interactive power of points of common coupling are periodically collected. The collected raw data is cleaned and time-aligned to form a real-world dataset with consistent time series.

[0011] Furthermore, the actual operational data is fed back to the predicted data set for rolling correction of the next scheduling cycle, including: By comparing the predicted power output of new energy sources with the actual output data of distributed power sources, a new energy power prediction error sequence is generated. Compare the predicted load power data with the actual load power data to generate a load power prediction error sequence; Using the aforementioned new energy power prediction error sequence and load power prediction error sequence, an autoregressive correction model for the prediction error is trained. When generating the prediction data set for the next scheduling cycle, the output of the autoregressive correction model is used as a correction term and superimposed on the preliminary prediction result to obtain the corrected prediction data set.

[0012] Furthermore, before distributing the collaborative optimization scheduling scheme to the control system of the incremental distribution network, a scheme security verification step is also included: The operation strategy curve in the aforementioned collaborative optimization scheduling scheme is input into the power flow calculation model of the distribution network to calculate the voltage of each node and the power of each branch in each time period within the scheduling cycle. Verify whether the calculated voltage of each node is within the allowable upper and lower voltage limits, and verify whether the power of each branch exceeds its thermal stability limit. If a limit violation occurs, the collaborative optimization scheduling scheme is adjusted to reduce the source load power that causes the limit violation, and the improved multi-objective adaptive evolutionary algorithm is called again for local optimization until all verifications pass.

[0013] Furthermore, the power flow calculation model of the distribution network is a three-phase unbalanced power flow model that takes into account the control characteristics of distributed generation inverters.

[0014] Furthermore, the training process of the autoregressive correction model includes: The new energy power prediction error sequence and the load power prediction error sequence are used as training samples; Construct a neural network model whose structure includes an input layer, a hidden layer, and an output layer. The input layer contains the prediction error values ​​for multiple consecutive historical time periods, and the output layer contains the prediction error correction values ​​for the next time period. Using the actual error value of the previous time step as a supervision signal, the neural network model is trained using the backpropagation algorithm until its loss function on the validation set converges and its prediction performance stabilizes.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The improved multi-objective adaptive evolutionary algorithm can dynamically adjust internal parameters based on scheduling objectives and distribution network operating status. It combines new energy sources, load forecasting, and external grid interaction constraints to perform global optimization, directly outputting continuous operation strategy curves for distributed power sources, energy storage systems, and flexible loads. The algorithm's optimization process closely matches the dynamic characteristics of incremental distribution networks, its multi-objective optimization dimensions highly match the operational needs of zero-carbon industrial parks, its equipment control commands are continuous and smooth, its distributed power sources, energy storage, and flexible load operating statuses are closely linked, its control rhythms are mutually adapted, and its external grid interaction power constraints correspond in real-time to the internal operating status of the distribution network.

[0016] During the dispatch execution phase, real-time operational data of the incremental distribution network is collected and fed back to the forecasting stage for rolling corrections of renewable energy power, load power, and external grid interaction constraints for the next cycle. The deviation between forecast data and actual operating conditions continues to narrow, and the dispatching scheme is updated iteratively with the real-time status of the power grid, ensuring that online dispatching instructions are highly consistent with actual operating scenarios. The timing of energy storage charging and discharging, the range of flexible load adjustment, and renewable energy output and load fluctuations are synchronously adapted, gradually offsetting dispatching deviations caused by initial forecast errors. The distribution network maintains a closed-loop control state, and all units of the power generation, grid, load, and storage systems maintain coordinated operation with the external grid. Attached Figure Description

[0017] Figure 1 This is a state diagram of the zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method described in this invention. Figure 2A flowchart illustrating the process of obtaining the predicted data set for the incremental distribution network within the scheduling cycle; Figure 3 A flowchart illustrating the work on the improved multi-objective adaptive evolutionary algorithm. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention provides a method for optimizing the scheduling of incremental distribution networks with coordinated source-grid-load-storage systems in zero-carbon industrial parks. The specific method includes: A system architecture is established comprising a renewable energy power prediction module, a multi-objective optimization module, a scheduling execution module, and a feedback correction module. After system startup, the renewable energy power prediction module acquires renewable energy power prediction data, load power prediction data, and external grid interaction constraints within the scheduling cycle, forming a prediction data set. The multi-objective optimization module receives the prediction data set and uses an improved multi-objective adaptive evolutionary algorithm to solve for and generate a collaborative optimization scheduling scheme that defines the operating strategy curves of distributed power sources, energy storage systems, and flexible loads within the industrial park. The scheduling execution module parses the collaborative optimization scheduling scheme, issues instructions to the incremental distribution network control system, and drives distributed power sources, energy storage systems, and flexible loads to perform online scheduling according to the operating strategy curves. The feedback correction module collects actual operating data of the incremental distribution network during scheduling execution, transmits the actual operating data back to the prediction module, and continuously corrects the prediction data set for the next scheduling cycle, completing closed-loop optimization.

[0020] In one embodiment of the present invention, see [reference] Figure 2 Historical power generation data, historical load consumption data, and historical meteorological data of distributed power sources within the industrial park are collected and stored in a historical database. Based on the historical meteorological data, a power prediction model combining physical models and neural networks is used to generate time-period new energy power prediction data within the scheduling cycle. The model input consists of meteorological sequences of wind speed, irradiance, and temperature, and the output consists of predicted output sequences of photovoltaic and wind power. The daily and weekly seasonal patterns of historical electricity consumption data are analyzed, and combined with the industrial park's production plan scheduling information, a time-series decomposition prediction model is used to decompose the data into trend, periodic, and random terms. These terms are then superimposed and reconstructed to generate load power prediction data. Time-of-use electricity price curves, tie-line power interaction limits, and regional power grid average carbon emission factors are obtained from the external power grid dispatch center interface and organized into external power grid interaction constraints. These constraints, along with the aforementioned prediction data, constitute the prediction data set.

[0021] In its implementation, this embodiment relies on the historical database and real-time data acquisition interface deployed in the park's data center. The data collected covers second-level historical power generation data from rooftop photovoltaic arrays and distributed wind turbines within the park, 15-minute interval historical electricity consumption data from industrial and commercial loads, and historical meteorological data recorded by the weather station for concurrent wind speed, total horizontal radiation, and ambient temperature. All historical data are archived and stored using a unified time label, and missing values ​​are filled using linear interpolation between adjacent time periods. When processing historical meteorological data, the power prediction model combining the physical model and neural network converts the irradiance sequence for the next 24 hours into theoretical photovoltaic output under standard test conditions in the physical model branch. The neural network branch takes irradiance, temperature, and humidity as inputs and learns the nonlinear mapping between the conversion efficiency decay of the local photovoltaic array and shading. The weighted fusion of the two branches' outputs generates new energy power prediction data, predicting the photovoltaic output. The calculation is expressed as: in: This indicates the photoelectric conversion efficiency of a photovoltaic module under standard operating conditions. The effective light-receiving area of ​​the photovoltaic array. For time period The total radiation intensity on the horizontal plane, and Time periods Ambient temperature and relative humidity, This is the nonlinear correction term output by the neural network model. Wind power output prediction uses historical wind speed data to drive a similar-day clustering model, matching historical output patterns of wind turbines within the same season and wind speed range to generate an output sequence.

[0022] In some embodiments, when processing historical load data, the time-series decomposition prediction model first identifies the differences in daily load curves between weekdays and holidays, separating a trend term reflecting long-term growth trends, a daily periodic term with a 24-hour period, and a weekly periodic term with a 168-hour period. The random residual term is smoothed using a moving average method. Then, known start-up and shutdown events of high-power equipment in the park's production plan are superimposed to correct the predicted load values ​​for the corresponding time periods, generating load power prediction data. It can be understood that the acquisition of external grid interaction constraints is achieved through the secure access zone of the power dispatch data network. The system calls the next day's 96-point time-of-use electricity price curve published by the dispatch center, the upper and lower limits of the allowed exchange power at the park's grid connection point, and the annual average carbon emission factor of the provincial power grid, encapsulating them as structured constraint parameters and adding them to the prediction data set. Optionally, the output resolution of the new energy power prediction data and the load power prediction data is set to 15 minutes, consistent with the time step of the optimized dispatch. The prediction data set is transmitted to the optimization module in JSON format.

[0023] In a specific implementation example, for generating forecast data for a typical sunny summer day, historical meteorological data shows that the total horizontal radiation reaches 850 W / m² at 12:00 noon. 2 With an ambient temperature of 32℃, the neural network correction term output accounted for 12% of the theoretical output, and the predicted peak photovoltaic output was 5.8MW, with the deviation from the actual midday photovoltaic output of 5.72MW controlled within 1.4%. On the load side, due to the concentrated start-up of compressors in the park's cold storage facilities, the time-series decomposition model superimposed a 300kW event-type load correction at 14:00, resulting in a predicted load value of 12.5MW, close to the actual maximum load of 12.45MW for the day. The time-of-use electricity price obtained from external grid interaction constraints was 1.2 yuan / kWh for peak hours, 0.8 yuan / kWh for off-peak hours, and 0.4 yuan / kWh for valley hours. The upper limit of the interconnection power was set to 8MW, and the carbon emission factor was taken as 0.56kgCO2 / kWh. All parameters were written into the prediction data set for the optimization module to use.

[0024] In one embodiment of the present invention, when the improved multi-objective adaptive evolutionary algorithm is run, see [reference]. Figure 3 An initial population containing several candidate scheduling solutions is generated. Each solution is encoded as a complete operational plan for distributed power output, energy storage charging and discharging power, and flexible load adjustment within a scheduling cycle. The population is traversed to calculate the dual objectives of carbon emission intensity and economic cost for each solution: The net interaction power between the park and the external power grid is calculated based on the operational strategy curve in the solution. Combined with the carbon emission factor in the external power grid interaction constraints, the carbon emissions caused by the interaction are calculated. The total carbon emissions are obtained by summing the carbon emissions from distributed power sources such as gas turbines within the park. The ratio of total carbon emissions to the total energy supply of the park is the carbon emission intensity target value. The economic cost target value is obtained by accumulating the fuel cost of distributed power sources, the cost of purchasing electricity from the external grid, the aging cost of energy storage cycles, and the incentive compensation cost of flexible loads. The population is non-dominated and sorted according to the dual objective values, dividing it into Pareto levels. Differentiated operators are configured for individuals at different levels: high-level solutions retain high-quality structures and are assigned lower crossover and mutation probabilities; low-level solutions enhance exploration and are assigned higher mutation probabilities. Differential crossover and mutation operations are performed to generate offspring. After merging parent and offspring populations, new parents are selected based on non-dominated sorting and crowding distance, iterating until a convergence threshold is reached. The Pareto front solution set is extracted from the final generation population, and the final collaborative optimization scheduling scheme is selected based on park operation preferences using a weighted average.

[0025] In practice, the improved multi-objective adaptive evolutionary algorithm runs on the optimization engine of the park's energy management server. The scheduling cycle is set to 24 hours and the time resolution to 15 minutes. The population is initialized to a size of 100. Each candidate scheduling scheme is encoded as a 96-dimensional vector. The dimensions correspond to the photovoltaic output setpoint, the wind power output setpoint, the energy storage charging and discharging power, and the three types of flexible load adjustment. The solution space boundary is set based on the maximum power of the photovoltaic inverter, the rated power of the energy storage, and the adjustable range of the load. When calculating the carbon emission intensity target value for each candidate scheduling scheme solution, the hourly power generation and load demand within the park are obtained by analyzing the operation strategy curve. The difference is used as the net interaction power between the park and the external power grid. Combined with the time-period carbon emission factor of 0.58 kgCO2 / kWh in the external power grid interaction constraint, the carbon emission introduced by electricity purchase is calculated. The carbon emission of the two gas turbine units in the park is linearly converted according to the output to obtain the total carbon emission. Dividing the total power supply of the park gives the carbon emission intensity target value. The economic cost target value is calculated by summing the cumulative photovoltaic operation and maintenance cost of 0.05 yuan / kWh, the gas turbine unit fuel cost of 0.35 yuan / kWh, the cost of purchasing electricity from the external grid according to the time-of-use price, the energy storage cycle aging cost of 0.08 yuan / kWh, and the flexible load interruption compensation unit price of 0.20 yuan / kWh. The economic cost target value is obtained by summing the cumulative photovoltaic operation and maintenance cost of 0.05 yuan / kWh, the gas turbine unit fuel cost of 0.35 yuan / kWh, the cost of purchasing electricity from the external grid according to the time-of-use price, the energy storage cycle aging cost of 0.08 yuan / kWh, and the flexible load interruption compensation unit price of 0.20 yuan / kWh. The dual target values ​​are stored in the evaluation vector of the solution.

[0026] In some embodiments, during the non-dominated sorting stage, the population solutions are sorted in ascending order based on the target values ​​of carbon emission intensity and economic cost. Solutions with a carbon emission intensity below 0.25 kg / kWh and an economic cost below 80,000 yuan are initially assigned to the first non-dominated layer. The remaining solutions are then sorted by dominance relationship, with the highest-level solutions accounting for 15% of the total. The differentiated crossover and mutation probabilities are configured as follows: 0.7 for the first-level solutions and 0.02 for the mutation; 0.8 for the second-level solutions and 0.03 for the mutation; and 0.9 for the crossover and 0.06 for solutions at the third level and below, maintaining stability of high-level solutions and diversity of low-level solutions. Simulated binary crossover is used for crossover, and polynomial mutation is used for mutation. The offspring population size is maintained at 100. After merging the parent and offspring populations, the non-dominated sorting is repeated. Duplicate solutions are eliminated by calculating crowding distance, and new parents are selected for the next iteration. The maximum number of iterations is set to 200, and the convergence condition is that the Pareto front movement distance is less than 0.001 for 20 consecutive generations.

[0027] In a specific implementation, for an optimization case on a cloudy winter day, the economic cost of the initial solutions ranged from 75,000 to 120,000 yuan, and the carbon emission intensity ranged from 0.22 to 0.42 kg / kWh. After iterative optimization, the economic cost of the Pareto front solution set was compressed to 72,000-78,000 yuan, and the carbon emission intensity decreased to 0.18-0.21 kg / kWh. Finally, based on the park's priority preference for carbon emission reduction, an equilibrium solution was selected with a carbon emission intensity of 0.19 kg / kWh and an economic cost of 74,500 yuan. This corresponds to the full photovoltaic power consumption, midday charging and evening discharging of energy storage, and flexible load peak-shaving and valley-shifting in the operation strategy curve. It can be understood that the quantitative calculation of the carbon emission intensity target value relies on the precise correlation between net interactive power and the carbon emission factor. During the period The calculation expression is: in: For time period The total load demand of the park For time period The photovoltaic power output is predicted. For time period Wind power forecast output, For time period Output of distributed gas turbine units. Negative value. This indicates that the park sells electricity to the grid, and carbon emissions are counted on the park's side; (This is a partial sentence and doesn't need a direct translation.) This indicates that the park purchases electricity from the power grid, and carbon emissions are calculated based on the amount of electricity purchased.

[0028] Optionally, the energy storage cycle loss cost coefficient in the calculation of the economic cost target value is adjusted in segments according to the state of charge (SOC). When the SOC is below 30%, the coefficient increases to 0.10 yuan / kWh cycle, and when it is above 80%, it decreases to 0.07 yuan / kWh cycle, reflecting the nonlinear characteristics of battery aging. It can be understood that the Pareto front solution set is selected using a normalized weighted method, with carbon emission intensity weight set to 0.7 and economic cost weight set to 0.3, meeting the priority orientation of zero-carbon operation in the park. In some embodiments, the population initialization introduces a historical excellent solution injection mechanism, inserting 20% ​​of the gene fragments of the previous day's best solution into the initial population to accelerate the convergence process. Optionally, in the mutation operation, a continuity constraint is added to the mutation of energy storage charging and discharging power over continuous periods to avoid power jumps in adjacent periods exceeding the converter's response capability.

[0029] In one embodiment of the present invention, the scheduling execution subsystem analyzes the operation strategy curve of the collaborative optimization scheduling scheme and decomposes it into three types of executable instructions: an active power setting instruction for the photovoltaic inverter specifies the output power value of the photovoltaic array at each time period; a charging and discharging power setting instruction for the energy storage converter specifies the charging and discharging power and direction of the energy storage unit; and a power adjustment instruction for the flexible load controller determines the switching time and adjustment amount of interruptible loads. The instruction set is sent to the field device controller via industrial Ethernet. The photovoltaic inverter adjusts the PWM modulation wave to achieve active power output tracking, the energy storage converter switches the charging and discharging mode and adjusts the DC side current, and the flexible load controller drives the contactor to perform load switching. During the execution process, smart meters deployed at distribution nodes and common connection points collect the actual output of distributed power sources, the actual power of loads, the actual state of charge of energy storage, and the actual interactive power of tie lines at fixed sampling intervals. The data acquisition unit performs anomaly filtering and timestamp alignment on the raw sampled values ​​to generate a time-series consistent actual operation dataset.

[0030] In practical implementation, the scheduling execution subsystem analyzes the operation strategy curves defined in the collaborative optimization scheduling scheme, and breaks down the power instructions for photovoltaics, energy storage, and flexible loads into three types of equipment-level instructions in 15-minute time periods: the active power setting instruction for photovoltaic inverters specifies the output percentage of the rooftop photovoltaic array in Zone A and the ground photovoltaic array in Zone B of the park, such as setting the photovoltaic output to 60% of the available capacity at 09:00; the charging and discharging power setting instruction for energy storage converters specifies the power magnitude and direction of the lithium battery energy storage units in Zone C, with positive values ​​for discharging and negative values ​​for charging, such as setting the discharge power to 500kW at 11:30; and the power adjustment instruction for flexible load controllers determines the switching time and adjustment amount of the load that can be moved in Workshop D, such as cutting off the refrigeration unit load by 300kW at 13:00. The instruction set is sent via the Modbus-TCP protocol through the industrial Ethernet communication network. The photovoltaic inverter adjusts the pulse width modulation duty cycle to track the active power setting instruction. The energy storage converter switches the bidirectional DC / AC conversion mode and adjusts the DC side inductor current to achieve power control. The flexible load controller drives the magnetic latching relay to perform load switching.

[0031] In some embodiments, the field device controller monitors the command execution status at a 50ms sampling period. The output voltage and current waveforms of the photovoltaic inverter verify that the active power output deviation from the set value is less than 3%. The DC-side current fluctuation of the energy storage converter is within ±5A. The contact status of the flexible load switch is confirmed by the IO module to indicate that the action is in place. During execution, smart meters and sensors deployed at key nodes of the distribution network synchronously collect actual operating data: the distributed power source collects the three-phase voltage and current of the photovoltaic array and the output power of the combiner box; the load end collects the total active power of the production line motors and lighting circuits; the energy storage end collects the battery cluster voltage and current and the state of charge reported by the management system; and the point of common coupling collects the grid interaction power measured by the gate meter. The data acquisition unit performs the Laida criterion on the raw sampled values ​​to remove gross errors, synchronizes the timestamps with the GPS clock, and aligns to generate a consistent time series of actual operating datasets. The recording interval and optimization step size are kept consistent (see Table 1).

[0032] Table 1: Comparison of Command and Measured Power During Typical Time Periods It is understandable that the active power setting command for the photovoltaic inverter needs to be converted into a setting value that the inverter can recognize before it is issued, based on the rated power on the inverter's nameplate. Maximum allowable output under current lighting conditions Set command value The formula for calculation is: in: The active power setpoint specified for the operating strategy curve. The inverter overload protection factor is set to 0.95. Optionally, a soft constraint on the state of charge (SOC) is added when the energy storage charging and discharging power setting command is issued. When the SOC is below 15%, a zero-discharge command is forced; when it is above 95%, the charging command is blocked. In some embodiments, the flexible load power adjustment command adds a rising edge buffer for impulsive loads, and inserts a 2-second delay before the command is issued to prevent inrush current. Optionally, after the actual operating data is collected, an abnormal flag is marked. Data points with a voltage drop exceeding 10% or a power change exceeding 20% ​​are marked as invalid and do not participate in subsequent feedback correction. It is understood that the communication network adopts a redundant ring network topology, and the data transmission packet loss rate is less than 0.1%, ensuring the timeliness and reliability of command issuance.

[0033] In one embodiment of the present invention, the feedback correction module receives the actual operating dataset, compares the new energy power prediction data with the actual output of distributed power sources, and generates a new energy power prediction error sequence indexed by time; it also compares the load power prediction data with the actual load power to generate a load power prediction error sequence. The two error sequences are used as training samples input to the autoregressive correction model, which is a three-layer feedforward neural network. The input layer receives prediction error values ​​from multiple consecutive historical time periods, the hidden layer uses the ReLU activation function to fit a nonlinear relationship, and the output layer generates the prediction error correction value for the next time period. During training, the first N-1 points of the historical error sequence are used as input, and the Nth point is used as the supervision signal. The mean squared error loss function and the Adam optimizer are used to perform error backpropagation. After the validation set loss converges, the model parameters are saved. During the prediction of the next scheduling cycle, the current error sequence is input into the trained model, and the output correction term is superimposed on the preliminary prediction result to obtain the corrected new energy power prediction data and load power prediction data.

[0034] In practice, the feedback correction module runs on the edge server of the park's data center. It receives the actual operating dataset after the scheduling execution cycle ends, aligns the new energy power prediction data with the actual output data of distributed power sources according to time labels, and calculates the difference between the two in 15-minute intervals to generate a new energy power prediction error sequence. For example, the difference between the photovoltaic power prediction sequence and the actual photovoltaic power output sequence is denoted as the error sequence EPV(t). The load power prediction data is aligned with the actual load power data, and the load power prediction error sequence ELoad(t) is generated in the same way. The length of the two error sequences is consistent with the 96 time intervals of the scheduling cycle. Taking a certain spring equinox as an example, at 11:45 am, the photovoltaic power prediction output is 4120kW, the actual output is 3985kW, and the error value is 135kW; at 2:00 pm, the load prediction value is 10500kW, the actual value is 10728kW, and the error value is -228kW. The error sequence is stored in a circular buffer in chronological order for model training, as shown in Table 2.

[0035] Table 2: Statistical Analysis of Errors Between Forecast and Actual Data for Typical Time Periods The autoregressive correction model is trained using error sequences as samples. The model structure is a three-layer feedforward neural network. The input layer has 8 nodes, corresponding to the prediction error values ​​of 8 consecutive historical time periods. The hidden layer contains 16 neurons using the ReLU activation function to capture nonlinear fluctuations. The output layer has 1 node, corresponding to the prediction error correction value for the next time period. During training, the error sequence is divided into a sliding window of length 9. The first 8 error values ​​are used as input features, and the 9th error value is used as the supervision label. The mean squared error loss function is used, and the Adam algorithm is used to update the weights. The learning rate is set to 0.01, the batch size is set to 32, the number of training epochs is set to 500, and the validation set accounts for 20% to monitor overfitting. The model parameters are frozen after the loss convergence. In essence, after training, the model has the ability to extrapolate error trends, and the output correction value is used to adjust the prediction accuracy for the next period.

[0036] In practice, when generating the new energy power forecast and load power forecast for the next scheduling cycle, a trained autoregressive correction model is invoked to process the current error sequence, outputting the error correction value for the first time period of the next cycle, which is then superimposed on the preliminary forecast result to complete the correction: Photovoltaic power output after correction. The calculation is as follows: in: To contribute to the preliminary photovoltaic forecast for time period k, This is the photovoltaic error correction value output by the autoregressive correction model; the load-corrected predicted value is similarly superimposed with the load error correction value. For the above post-Spring Equinox prediction, the model outputs a photovoltaic error correction value of -122kW at 11:45 the next day based on historical errors. After superposition, the photovoltaic predicted value is adjusted from 4150kW to 4028kW, reducing the deviation from the actual value of 4015kW to 13kW; the load error correction value of +195kW corrects the predicted value from 10890kW to 11085kW, closer to the actual value of 11123kW. Optionally, the error sequence buffer length is set to the data of the most recent 7 days. New data added daily triggers incremental training to fine-tune the model parameters, adapting to seasonal changes. In some embodiments, the autoregressive correction model constructs independent networks for photovoltaic and load errors respectively. The photovoltaic error model input includes irradiance mutation identification features, and the load error model input includes holiday type features, improving the correction accuracy for specific scenarios. It can be understood that the corrected predicted data set is written to a JSON file and delivered to the optimization module, with the data structure retaining the original predicted value and the correction offset field. Optionally, before model training, the error sequence can be Z-score standardized to eliminate the influence of dimensions, and the output correction value can be de-standardized to restore the physical units.

[0037] In one embodiment of the present invention, during the scheme safety verification stage, the operation strategy curve of the collaborative optimization scheduling scheme is input into the power flow calculation model of the distribution network. This model is a three-phase unbalanced power flow model that takes into account the PQ and VF control characteristics of distributed power inverters, and includes line impedance imbalance parameters and single-phase load distribution during modeling. The model calculates the three-phase voltage amplitude and phase angle of each node in each time period of the scheduling cycle, as well as the active and reactive power of each branch. The verification program scans the calculation results to verify whether the voltage of each node is within the range of 0.95~1.05 per unit, and whether the power of each branch exceeds the conductor thermal stability limit and the transformer capacity limit. If it is detected that the node voltage exceeds the lower limit or the branch is overloaded in a certain time period, it is determined to be a safety violation. The adjustment module reduces the output of distributed power sources that cause the exceedance of limits or increases the local energy storage charging power. If necessary, some non-critical flexible loads are cut off, a corrected operation strategy curve is generated, and the improved multi-objective adaptive evolutionary algorithm is called again for local optimization under the corrected constraints. The power flow is recalculated until all verifications pass.

[0038] In practical implementation, the safety verification step is performed before the collaborative optimization scheduling scheme is issued. The operation strategy curve is input into the power flow calculation model of the distribution network. This model is a three-phase unbalanced power flow model that considers the control characteristics of distributed generation inverters. The modeling parameters include the unbalanced values ​​of line resistance, reactance, and ground capacitance, as well as the distribution ratio of single-phase loads in each phase. The control mode of the distributed generation inverter is set to PQ control for constant power output or VF control for voltage stabilization support. The model calculates the three-phase voltage amplitude and phase angle of each node in each time period within the scheduling cycle, as well as the active and reactive power of each branch, with a step size of 15 minutes. The node voltage reference value is set to 10kV, and the voltage upper and lower limits range from 0.95pu to 1.05pu. The branch thermal stability limit is set according to the conductor type and transformer capacity. The verification program traverses the calculation results to identify voltage overruns and branch overload risks. For example, if the C-phase voltage of node N7 drops to 0.93pu during a certain time period, and the power transmitted by branch L12 reaches 4.8MW, exceeding the 4.5MW limit, it is judged as a safety violation.

[0039] In specific implementation, to address the aforementioned over-limit situations, the adjustment module reduced the source load power causing the over-limit. The low voltage at node N7 was attributed to excessive downstream distributed photovoltaic output combined with concentrated single-phase loads. Reduction measures included reducing the photovoltaic inverter output power by 300kW and reducing non-critical flexible loads by 200kW. The overload of branch L12 was caused by the upstream gas turbine operating at full capacity; the turbine output was reduced by 500kW to alleviate power flow pressure. The corrected operating strategy curve was re-input into the improved multi-objective adaptive evolutionary algorithm, performing local optimization under the added power constraints to generate alternative solutions that meet the safety boundaries. Power flow verification was performed, and the node voltage recovered to 0.97 pu, with branch power reduced to 4.2MW. After all verifications passed, the system was approved for deployment. It is understood that the calculation of the three-phase unbalanced power flow model needs to accurately reflect the impact of the neutral grounding method. For systems with ungrounded neutral points, the zero-sequence impedance parameter needs to be included, and the voltage amplitude at node i in the voltage correction equation... Calculate the relationship between the dependent admittance matrix and the injected current: in: This is the impedance matrix of a three-phase bus. The three-phase injected current vector for node i encompasses the current contributions of distributed power sources and loads. In some embodiments, the power flow calculation model supports the simulation of inverter reactive power regulation capabilities; when voltage exceeds limits, reactive power output can be adjusted first to replace active power reduction, reducing economic costs. Optionally, branch power verification incorporates N-1 anticipated fault scanning to simulate power flow transfer after any single branch is disconnected, verifying the safety of the scheme in fault states. It can be understood that the local optimization process retains the original dual-objective framework of carbon emission intensity and economic cost, adds safety constraints to limit the feasible region, and the Pareto solution set automatically avoids the risk of exceeding limits. In some embodiments, the power flow calculation model integrates a dynamic cable temperature rise model, estimating conductor temperature based on current and duration, replacing the static thermal stability limit to achieve more refined branch verification. Optionally, the voltage exceedance threshold is dynamically adjusted in segments, with the lower limit relaxed to 0.92 pu during peak load periods to adapt to short-term overload tolerance characteristics.

[0040] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for optimized scheduling of incremental distribution networks with coordinated generation, grid, load, and storage in zero-carbon industrial parks, characterized in that: The method includes: Obtain the prediction data set of the incremental distribution network within the scheduling cycle. The prediction data set includes new energy power prediction data, load power prediction data, and external power grid interaction constraints. The predicted dataset is input into an improved multi-objective adaptive evolutionary algorithm to generate a collaborative optimization scheduling scheme, which defines the operation strategy curves of distributed power sources, energy storage systems and flexible loads within the park. The aforementioned collaborative optimization scheduling scheme is sent to the control system of the incremental distribution network to perform online scheduling of distributed power sources, energy storage systems and flexible loads; During the scheduling process, actual operating data of the incremental distribution network is collected, and the actual operating data is fed back to the predicted data set for the next scheduling cycle for rolling correction.

2. The zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method as described in claim 1, characterized in that, Obtain the predicted data set of the incremental distribution network within the scheduling cycle, including: Collect historical power generation data of distributed power sources, historical electricity consumption data of loads, and historical meteorological data within the park; Based on the historical meteorological data, the new energy power prediction data is generated through a power prediction model that combines a physical model and a neural network. By analyzing the patterns in the historical electricity consumption data and combining it with the park's production plan information, the load power prediction data is generated using a time-series decomposition prediction model. Electricity price curves, power interaction limits, and carbon emission factors are obtained from the external power grid dispatch center as external power grid interaction constraints.

3. The zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method as described in claim 1, characterized in that, The working principle of the improved multi-objective adaptive evolutionary algorithm includes: Initialize and generate a population containing multiple candidate scheduling solutions, each candidate scheduling solution representing a complete source-grid-load-storage coordinated operation plan; Calculate the target value of carbon emission intensity and economic cost for each candidate scheduling scheme solution in the population, and perform non-dominated ranking of the population based on these two target values; Differentiated adaptive crossover and mutation probabilities are assigned to candidate scheduling schemes at different sorting levels, with the mutation probability of higher-level solutions being lower than that of lower-level solutions. Based on the differentiated crossover probability, crossover operations are performed between individuals in the population, and based on the differentiated mutation probability, mutation operations are performed on individuals in the population to generate offspring populations. From the merged parent and offspring populations, a new parent population is selected based on non-dominated sorting and crowding distance, and this process is repeated iteratively until the preset convergence condition is met. A Pareto front solution set is selected from the final population, and the final collaborative optimization scheduling scheme is determined from it based on the preset decision preferences.

4. The zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method as described in claim 3, characterized in that, Calculate the target carbon emission intensity and target economic cost for each candidate scheduling solution in the population, including: Based on the operation strategy curves in the candidate scheduling scheme solutions, calculate the net interaction power between the park and the external power grid during the scheduling period; By combining the carbon emission factor in the external power grid interaction constraint, the carbon emissions generated by the interaction with the external power grid are calculated, and then summed with the carbon emissions of the distributed power sources within the park to obtain the total carbon emissions. The ratio of the total carbon emissions to the total energy supply of the park is used as the carbon emission intensity target value. The target economic cost is obtained by summing the fuel cost of distributed power sources, the cost of purchasing electricity from the external power grid, the cycle loss cost of the energy storage system, and the incentive compensation cost of flexible loads during the scheduling cycle.

5. The zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method as described in claim 1, characterized in that, The collaborative optimization scheduling scheme is distributed to the control system of the incremental distribution network to perform online scheduling of distributed power sources, energy storage systems, and flexible loads, including: The operation strategy curves in the aforementioned collaborative optimization scheduling scheme are analyzed and decomposed into active power setting instructions for photovoltaic inverters, charging and discharging power setting instructions for energy storage converters, and power regulation instructions for flexible load controllers. The active power setting command, charge / discharge power setting command, and power adjustment command are sent to the corresponding field device controllers via the communication network. The field equipment controller drives the photovoltaic inverter, energy storage converter and flexible load switch to execute corresponding command actions.

6. The zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method as described in claim 5, characterized in that, During the scheduling execution process, actual operational data of the incremental distribution network is collected, including: By deploying smart meters and sensors at key nodes of the distribution network, the actual output of distributed power sources, the actual power of loads, the actual state of charge of energy storage systems, and the actual interactive power of points of common coupling are periodically collected. The collected raw data is cleaned and time-aligned to form a real-world dataset with consistent time series.

7. The zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method as described in claim 6, characterized in that, The actual operating data is fed back to the predicted data set for rolling correction of the next scheduling cycle, including: By comparing the predicted power output of new energy sources with the actual output data of distributed power sources, a new energy power prediction error sequence is generated. Compare the predicted load power data with the actual load power data to generate a load power prediction error sequence; Using the aforementioned new energy power prediction error sequence and load power prediction error sequence, an autoregressive correction model for the prediction error is trained. When generating the prediction data set for the next scheduling cycle, the output of the autoregressive correction model is used as a correction term and superimposed on the preliminary prediction result to obtain the corrected prediction data set.

8. The zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method as described in claim 1, characterized in that, Before distributing the collaborative optimization scheduling scheme to the control system of the incremental distribution network, a scheme security verification step is also included: The operation strategy curve in the aforementioned collaborative optimization scheduling scheme is input into the power flow calculation model of the distribution network to calculate the voltage of each node and the power of each branch in each time period within the scheduling cycle. Verify whether the calculated voltage of each node is within the allowable upper and lower voltage limits, and verify whether the power of each branch exceeds its thermal stability limit. If a limit violation occurs, the collaborative optimization scheduling scheme is adjusted to reduce the source load power that causes the limit violation, and the improved multi-objective adaptive evolutionary algorithm is called again for local optimization until all verifications pass.

9. The zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method as described in claim 8, characterized in that, The power flow calculation model for the distribution network is a three-phase unbalanced power flow model that takes into account the control characteristics of distributed generation inverters.

10. The zero-carbon industrial park source-grid-load-storage coordinated incremental distribution network optimization scheduling method as described in claim 7, characterized in that, The training process of the autoregressive correction model includes: The new energy power prediction error sequence and the load power prediction error sequence are used as training samples; Construct a neural network model whose structure includes an input layer, a hidden layer, and an output layer. The input layer contains the prediction error values ​​for multiple consecutive historical time periods, and the output layer contains the prediction error correction values ​​for the next time period. Using the actual error value of the previous time step as a supervision signal, the neural network model is trained using the backpropagation algorithm until its loss function on the validation set converges and its prediction performance stabilizes.

Images