A distributed photovoltaic grid-connected regional power grid real-time monitoring and coordinated control system

Abstract

The present application relates to the field of real-time monitoring and coordinated control of power grid, in particular to a kind of distributed photovoltaic grid-connected regional power grid real-time monitoring and coordinated control system, system includes sensing acquisition, multi-time scale prediction, self-evolution calibration and distributed coordination control module;Sensing acquisition module is collected wide frequency domain electrical quantity and multidimensional meteorological quantity by double-layer sensing network, and the repair data is obtained by abnormal detection;Multi-time scale prediction module integrates ultra-short-term, short-term and medium and long-term prediction, realizes cross-scale cooperation based on joint state space model etc.;Self-evolution calibration module is based on deep reinforcement learning to compensate the residual error of multi-scale prediction online;Distributed coordination control module combines improved alternating direction multiplier method and non-dominated sorting genetic algorithm, obtains optimal power instruction that meets voltage, frequency and harmonic ternary constraint;The present application improves the stability of power grid operation under high penetration rate photovoltaic grid-connected scene.

Description

Technical Field

[0001] This invention relates to the field of real-time monitoring and coordinated control of power grids, specifically to a real-time monitoring and coordinated control system for distributed photovoltaic grid-connected regional power grids. Background Technology

[0002] As the penetration rate of distributed photovoltaic power continues to rise, the high proportion of photovoltaic grid connection poses a severe challenge to the safe and stable operation of regional power grids. Traditional monitoring systems mostly adopt a single time scale acquisition mode, which makes it difficult to take into account the millisecond-level dynamic changes of electrical quantities and the second-level slow fluctuations of meteorological quantities, and cannot capture the grid disturbances caused by the randomness of photovoltaic output in a timely manner.

[0003] Existing prediction models suffer from insufficient cross-scale coordination. Ultra-short-term, short-term, and medium-to-long-term predictions are independent, with asynchronous parameter updates, making it difficult to adapt to the rapid fluctuations in photovoltaic power. This results in large prediction errors and limits the accuracy of coordinated control. At the control level, traditional methods struggle to balance the conflicting objectives of power tracking, minimizing grid loss, and control costs. Furthermore, they fail to adequately address the coupling of voltage, frequency, and harmonic constraints, leading to issues such as voltage exceeding limits, excessive frequency fluctuations, and harmonic pollution in high-penetration scenarios. Therefore, a real-time monitoring and coordination control system for distributed photovoltaic grid-connected regional power grids is proposed to solve the above problems. Summary of the Invention

[0004] To address the technical problems mentioned in the background, this invention provides a real-time monitoring and coordination control system for distributed photovoltaic grid-connected regional power grids.

[0005] The objective of this invention can be achieved through the following technical solutions: This invention provides a real-time monitoring and coordinated control system for distributed photovoltaic grid-connected regional power grids, including a sensor acquisition module, a multi-timescale prediction module, a self-evolving calibration module, and a distributed coordinated control module.

[0006] The sensor acquisition module collects broadband electrical quantities and multi-dimensional meteorological quantities respectively based on a two-layer sensor network. The specific process is as follows: Each photovoltaic inverter's outlet is equipped with a set of bottom-level sensors, and upper-level sensors are installed at the highest points of the photovoltaic modules and power distribution area. The bottom-level sensors include voltage sensors, current sensors, phase angle sensors, and switch status sensors. The upper-level sensors include the photovoltaic array area sensor group and the regional observation point meteorological station. Each bottom-level sensor collects three-phase voltage, three-phase current, voltage phase angle, and switch status quantities in real time, and divides them into conventional electrical quantities and wideband components. The wideband components are subjected to Fourier transform in real time to extract harmonic components and generate composite data frames containing fundamental, harmonic, and wideband noise. The photovoltaic array area sensor group collects photovoltaic irradiance, backsheet temperature, and module surface temperature in real time, and the regional observation point meteorological station collects wind speed, wind direction, ambient temperature, cloud thickness, and cloud movement speed. Data collected by each bottom-level sensor and each top-level sensor is input into the anomaly detection model to construct a temporal feature matrix XR. An autoencoder E is then used to compress the multi-dimensional temporal feature matrix into a latent vector. Decoder D reconstructs the latent vectors to obtain the reconstructed data. Calculate the reconstruction error Its calculation logic is as follows: When reconstruction error If the value exceeds the dynamic error threshold, it is considered an outlier and replaced with reconstructed data to obtain repaired data.

[0007] The multi-timescale prediction module integrates ultra-short-term, short-term, and medium-to-long-term predictions, and then achieves cross-scale coordination and dynamic parameter updates through joint state-space model, sensitivity matrix, and stochastic model prediction and control. The specific process is as follows: The multi-timescale prediction module includes an ultra-short-term power prediction unit, a short-term voltage sensitivity estimation unit, and a medium-to-long-term cluster schedulable capacity assessment unit. Convert the repaired data into a state vector. The control state space model is set based on the state vector. The control state space model includes the control input vector, the disturbance vector, and the output vector. The calculation logic is as follows: ,in For active power command increment, The reactive power command increment is k, where k is the current time point; the disturbance vector is... It is a set of uncontrollable disturbance variables, reflecting random changes in the external environment and load; output vector The parameters are output by combining the control input vector and the disturbance vector; the time-state equation is obtained by embedding the control input vector and the disturbance vector into the state equation. This time-state equation describes the evolution of the system state over time, and its calculation logic is as follows: ,in This is the state transition matrix, describing the effect of temperature on power or the coupling relationship between voltage and frequency. To control the input matrix, the effect of the power command on the state is quantized. As the perturbation matrix, the output vector is embedded into the output equation to obtain the relational output equation. Its calculation logic is as follows: ,in For the output matrix, For direct transmission matrix; The parameters of the control state-space model are updated in real time using a recursive least squares algorithm with a forgetting factor. These parameters include the state transition matrix, control input matrix, disturbance matrix, output matrix, and direct transmission matrix. The regression vector is set as follows: The parameter vector to be estimated is The update is performed recursively based on the regression vector and the vector of parameters to be estimated. The update logic is as follows: ,in This is the vector of parameter estimates for the next time step. The gain matrix represents the magnitude of the parameter update, and its calculation logic is as follows: ,in This is an abnormal factor, and its value range is... , The covariance matrix is ​​used; rolling optimization is used to generate the predicted output sequence within the first preset time period. The predicted output sequence includes the active and reactive power of each photovoltaic node, the voltage amplitude and voltage phase angle of key nodes, and the first preset time period is set to a time of ten seconds. The short-term voltage sensitivity estimation unit obtains the mapping relationship between photovoltaic active power, inverter reactive power and node voltage within a second preset time period by linearizing the sensitivity matrix. The second preset time period is between 1 min and 15 min. Using the real-time voltage of each photovoltaic node, the real-time reactive power output of the inverter, and the active power of the load as inputs, photovoltaic active power-voltage sensitivity matrices, inverter reactive power-voltage sensitivity matrices, and load active power-voltage sensitivity matrices are established respectively; the difference in photovoltaic active power at adjacent time points is calculated. and voltage difference The logic for calculating the difference is as follows: , ,in The current photovoltaic active power is... The photovoltaic active power at the previous moment. This is the current voltage value. Number the inverter. Number the voltage monitoring points, select time periods when the inverter reactive power and load changes are less than the corresponding preset change thresholds, and calculate using the least squares method. Obtain the photovoltaic active power-voltage sensitivity matrix ,in The first in the matrix Line number The elements of the column are used to calculate the reactive power difference of the inverter at adjacent time points. Solve using the least squares method Obtain the inverter reactive power-voltage sensitivity matrix Where QH is the inverter reactive power; the load active power difference is obtained by subtracting the load active power at adjacent times. ,pass Obtain the load active power-voltage sensitivity matrix The photovoltaic active-voltage sensitivity matrix, the inverter reactive-voltage sensitivity matrix, and the load active-voltage sensitivity matrix are integrated to obtain the voltage sensitivity matrix TK. The voltage stability margin is then calculated based on this sensitivity matrix, and the calculation logic is as follows: , This represents the maximum active power output vector of the photovoltaic cluster. The medium- and long-term cluster dispatchable capacity assessment unit probabilistically evaluates the dispatch potential of photovoltaic clusters, fitting the irradiance probability density function for the third preset time period using a two-parameter Beta distribution. The third preset time period is from 15 minutes to 1 hour, and its function is expressed as: Where G represents the real-time photovoltaic irradiance. For maximum irradiance, For Beta functions, Let be the shape parameter of the Beta distribution, when The time distribution degenerates into a uniform distribution. The distribution is biased towards areas with high irradiance. The time distribution is biased towards low irradiance regions; a stochastic predictive control model is established based on the irradiance probability density function, and the schedulable capacity is obtained by solving the stochastic predictive control model; the calculation logic of the stochastic predictive control model is as follows: Where K is the total number of time points. In order to exchange power with the main network, This represents the current real-time state of charge of the energy storage. For energy storage target state of charge, and With fixed weighting coefficients of 0.59 and 0.41 respectively, the data are combined with ultra-short-term, short-term, and medium-to-long-term forecast data to form a three-dimensional forecast vector.

[0008] The self-evolutionary calibration module is based on a deep reinforcement learning algorithm, combining a shared feature extractor and an independent decision head for online compensation of multi-scale prediction residuals. The specific process is as follows: The three-dimensional prediction vector, real-time measured data of the two-layer sensor network, historical calibration error sequence and environmental disturbance are combined as input data. The local spatiotemporal features of the input data and the temporal dependency of the historical calibration error sequence are extracted by the shared feature extractor. The two are concatenated into a feature vector and input into the scaled decision head. Each calibration quantity is output as a stepwise correction value of the corresponding ultra-short-term, short-term and medium-to-long-term state quantity. The input data is set as the state space, and the calibration values ​​corresponding to the ultra-short-term, short-term, and medium-to-long-term time scales are used as the action space. The calibration errors of the ultra-short-term, short-term, and medium-to-long-term time scales are squared respectively, and then accumulated according to different weights, with the negative value taken to obtain the accuracy reward. Then, the calibration value of each time scale is detected. If the absolute value of a certain calibration value exceeds the preset maximum safety threshold, the square value of the excess part is accumulated, multiplied by a penalty coefficient, and the negative value is taken to obtain the constraint penalty. The accuracy reward and constraint penalty are set as the reward function. Historical sample data is obtained. Each sample contains the state space, action space, and reward function. Samples are stored in an experience replay pool. Several samples are randomly selected each time. This method minimizes the error between the predicted Q-value and the target Q-value by updating the network through maximizing the action space using the Critic gradient until a preset number of iterations are reached to complete training and obtain a reinforced calibration model. Ultra-short-term historical predicted values ​​and measured values ​​are input into the reinforced calibration model, which outputs an ultra-short-term calibration matrix and merges it with the predicted output sequence to obtain ultra-short-term updated values. Short-term historical average errors are input into the reinforced calibration model, which outputs a corrected voltage sensitivity matrix. The corrected voltage sensitivity matrix and the voltage sensitivity matrix are merged to obtain the complete voltage sensitivity matrix. The latest medium- and long-term meteorological data are input into the reinforced calibration model, which outputs a schedulable capacity correction value. This correction value is then embedded into the schedulable capacity to obtain the complete schedulable capacity.

[0009] The distributed coordination control module obtains the optimal power command based on the improved alternating direction multiplier method and the non-dominated sorting genetic algorithm, taking into account the three-element constraints of voltage, frequency, and harmonics. The specific process is as follows: The power allocation of photovoltaic and energy storage devices is used as a decision variable and encoded using chromosome real numbers to obtain the photovoltaic power chromosome. and energy storage power chromosome 'm' represents the energy storage device number. N initial candidate solutions are randomly generated based on the photovoltaic power chromosome and the energy storage power chromosome, and integrated to obtain a non-dominated sort. Each solution is represented as a chromosome. A multi-objective fitness function is calculated using the initial candidate solutions. The multi-objective fitness function includes power tracking objective, grid loss objective, and control cost objective. The power tracking objective... The calculation logic is as follows: ,in Active power target value; network loss target The calculation logic is as follows: , and Number the different voltage monitoring points. For the electrical conductance between monitoring points, The voltage phase angle difference; the calculation logic for the control cost target is as follows: The system checks whether each chromosome meets voltage, frequency, and harmonic constraints. If not, its sorting priority is lowered. Specifically, the verification voltage value is calculated using a sensitivity matrix. Its calculation logic is as follows: ,in Based on the base voltage value, This is the power allocation vector; if the verification voltage values ​​of all monitoring points satisfy the first voltage threshold ≤ If the voltage is less than or equal to the second voltage threshold, the initial candidate solution is a voltage-feasible solution and retains its original priority; otherwise, it is marked as a voltage constraint violation solution and downgraded by one level in the non-dominated ranking. For frequency constraints, the sum of all active power and the current measured local load in the initial candidate solutions are extracted. The difference between the absolute values ​​of the sum of all active power and the current measured local load is used to calculate the active power imbalance. The active power imbalance is multiplied by a preset frequency deviation coefficient to obtain the frequency deviation. If the frequency deviation is less than or equal to the first preset deviation threshold, the initial candidate solution is a frequency-feasible solution with no penalty; otherwise, it is marked as a severely frequency-violated solution and removed from the non-dominated ranking. The harmonic constraint verification process involves obtaining a composite data frame of fundamental and harmonic frequencies and calculating the total harmonic distortion rate of each device. If the total harmonic distortion rate of the device is less than or equal to the preset distortion threshold, the candidate solution is a harmonic-feasible solution and retains its original ranking; otherwise, it is marked as a harmonic-exceeding solution and removed from the non-dominated ranking. Genetic operations and solution set generation: Divide several chromosomes into Pareto levels, with the first level being the global optimum. Calculate and sort the distance crowding between solutions of the same level. Select the first level and the solution with the largest distance crowding as the parent generation. Randomly perturb the offspring to obtain new candidate solutions. After completing the preset number of iterations, output the first-level Pareto solution. The main controller filters the first-level Pareto solution through real-time scenarios. During peak load periods, it selects the solution with the minimum network loss target; when photovoltaic fluctuations are large, it selects the solution with the minimum power tracking target; and when the voltage is critical, it selects the voltage-feasible solution. For each device's selected solution, Lagrange multipliers are added, and a global optimization update is performed to output the optimal global adjustment instruction for the corresponding device. The global optimization update formula is as follows: ,in The solution selected for the device, the first The total number of devices. For equipment number, For the h-th iteration, update the global optimization variables. The Lagrange multiplier for the h-th iteration sends the optimal global adjustment command to the server to adjust the active power of the photovoltaic and energy storage devices.

[0010] Compared with the prior art, the beneficial effects of the present invention are: The distributed coordination control module integrates the improved alternating direction multiplier method and the non-dominated sorting genetic algorithm to generate the optimal power command under the premise of satisfying the three-element constraints of voltage, frequency and harmonics. It balances the requirements of power tracking, grid loss control and cost optimization, improves the stability and economy of grid operation, and provides core technical support for the safe and efficient grid connection of high-penetration photovoltaics. The multi-timescale prediction module dynamically updates model parameters through a cross-scale collaborative mechanism of joint state-space model, sensitivity matrix and stochastic model prediction and control, so as to achieve accurate connection of ultra-short-term, short-term and medium-to-long-term predictions and provide a reliable basis for scheduling decisions at different time scales. The self-evolutionary calibration module is based on a deep reinforcement learning algorithm. It uses a shared feature extractor and a scaled decision head to perform online compensation for prediction residuals at multiple time scales, effectively reducing prediction errors, optimizing the voltage sensitivity matrix and schedulable capacity, and improving the adaptability of prediction results to actual operating conditions. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to illustrate the main idea of ​​the present invention.

[0012] Figure 1 This is a schematic diagram of the principle of the present invention. Detailed Implementation

[0013] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 also within the scope of protection of the present invention.

[0014] Please refer to Figure 1 As shown, the present invention provides a real-time monitoring and coordinated control system for distributed photovoltaic grid-connected regional power grids, including a sensor acquisition module, a multi-timescale prediction module, a self-evolving calibration module, and a distributed coordinated control module.

[0015] The sensor acquisition module collects broadband electrical quantities and multi-dimensional meteorological quantities respectively based on a two-layer sensor network. The specific process is as follows: Each photovoltaic inverter's outlet is equipped with bottom-level sensors, and upper-level sensors are installed at high points in the photovoltaic modules and power distribution area. The bottom-level sensors include voltage sensors, current sensors, phase angle sensors, and switch status sensors. The upper-level sensors include the photovoltaic array sensor group and the regional observation point meteorological station. Each bottom-level sensor collects three-phase voltage, three-phase current, voltage phase angle, and switch status quantities in real time, and divides them into conventional electrical quantities and wide-frequency components. Conventional electrical quantities refer to grid signals with a frequency range of 0 to 2 kHz, including the fundamental and low-order harmonic components traditionally monitored in the power system. Wide-frequency components refer to high-frequency signals with a frequency range of 2 kHz to 150 kHz, including high-frequency harmonics generated by inverter switching operations, grid resonance signals, and electromagnetic interference. The wide-frequency components are subjected to Fourier transform in real time to extract harmonic components and generate composite data frames containing the fundamental, harmonics, and wide-frequency noise. The photovoltaic array sensor group collects photovoltaic irradiance, backsheet temperature, and module surface temperature in real time, and the regional observation point meteorological station collects wind speed, wind direction, ambient temperature, cloud thickness, and cloud movement speed. Data collected by each bottom-level sensor and each top-level sensor is input into the anomaly detection model to construct a temporal feature matrix XR. An autoencoder E is then used to compress the multi-dimensional temporal feature matrix into a latent vector. Decoder D reconstructs the latent vectors to obtain the reconstructed data. Calculate the reconstruction error Its calculation logic is as follows: When reconstruction error If the value exceeds the dynamic error threshold, it is considered an outlier and replaced with reconstructed data to obtain repaired data.

[0016] The multi-timescale prediction module integrates ultra-short-term, short-term, and medium-to-long-term predictions, and then achieves cross-scale coordination and dynamic parameter updates through joint state-space model, sensitivity matrix, and stochastic model prediction and control. The specific process is as follows: The multi-timescale prediction module includes an ultra-short-term power prediction unit, a short-term voltage sensitivity estimation unit, and a medium-to-long-term cluster schedulable capacity assessment unit. It should be noted that the ultra-short-term power prediction unit, in conjunction with the state-space model, predicts outputs based on the rapid fluctuations in photovoltaic power output and grid conditions, providing an instantaneous state reference for real-time control. Convert the repaired data into a state vector. The control state space model is set based on the state vector. The control state space model includes the control input vector, the disturbance vector, and the output vector. The calculation logic is as follows: ,in For active power command increment, The input vector represents the reactive power command increment, where k is the current time point. It should be noted that this input vector is the power adjustment command issued by the photovoltaic inverter, instructing the inverter to increase or decrease the amount of active and reactive power output; the disturbance vector... This is a set of uncontrollable disturbance variables, reflecting random changes in the external environment and load, such as changes in ambient temperature and humidity, sudden changes in light intensity, or fluctuations in wind speed; the output vector... The parameters output by the combination of the control input vector and the disturbance vector, such as actual output power, grid connection point voltage, ambient temperature, and solar irradiance, are obtained by embedding the control input vector and the disturbance vector into the state equation to obtain the time state equation. This time-state equation describes the evolution of the system state over time, and its calculation logic is as follows: ,in This is the state transition matrix, describing the effect of temperature on power or the coupling relationship between voltage and frequency. To control the input matrix, the effect of the power command on the state is quantized. Let be the perturbation matrix, describing the propagation characteristics of the perturbation; embedding the output vector into the output equation yields the relational output equation. The observable mapping relationship between output and state is calculated as follows: ,in To produce the output matrix, select the key observations from the state vector. For direct transmission matrix, characterize the effect of control commands on the output; To address the strong randomness of photovoltaic power output, a recursive least squares algorithm with a forgetting factor is used to update the parameters of the control state-space model in real time. These parameters include the state transition matrix, control input matrix, disturbance matrix, output matrix, and direct transmission matrix. The regression vector is set as follows: The parameter vector to be estimated is The update is performed recursively based on the regression vector and the vector of parameters to be estimated. The update logic is as follows: ,in This is the vector of parameter estimates for the next time step. The gain matrix represents the magnitude of the parameter update, and its calculation logic is as follows: ,in This is an abnormal factor, and its value range is... , The larger the covariance matrix, the greater the error in parameter estimation. Rolling optimization is used to generate the predicted output sequence within the first preset time period. The predicted output sequence includes the active and reactive power of each photovoltaic node, the voltage amplitude and voltage phase angle of key nodes. The first preset time period is set to a time of ten seconds. The short-term voltage sensitivity estimation unit obtains the mapping relationship between photovoltaic active power, inverter reactive power and node voltage within a second preset time period by linearizing the sensitivity matrix. The second preset time period is between 1 min and 15 min. Using the real-time voltage of each photovoltaic node, the real-time reactive power output of the inverter, and the active power of the load as inputs, photovoltaic active power-voltage sensitivity matrices, inverter reactive power-voltage sensitivity matrices, and load active power-voltage sensitivity matrices are established respectively; the difference in photovoltaic active power at adjacent time points is calculated. and voltage difference The logic for calculating the difference is as follows: , ,in The current photovoltaic active power is... The photovoltaic active power at the previous moment. This is the current voltage value. Number the inverter. Number the voltage monitoring points, select time periods when the inverter reactive power and load changes are less than the corresponding preset change thresholds, and calculate using the least squares method. Obtain the photovoltaic active power-voltage sensitivity matrix ,in The first in the matrix Line number The elements of the column are used to calculate the reactive power difference of the inverter at adjacent time points. Solve using the least squares method Obtain the inverter reactive power-voltage sensitivity matrix Where QH is the inverter reactive power; the load active power difference is obtained by subtracting the load active power at adjacent times. ,pass Obtain the load active power-voltage sensitivity matrix The photovoltaic active-voltage sensitivity matrix, the inverter reactive-voltage sensitivity matrix, and the load active-voltage sensitivity matrix are integrated to obtain the voltage sensitivity matrix TK. The voltage stability margin is then calculated based on this sensitivity matrix, and the calculation logic is as follows: , This represents the maximum active power output vector of the photovoltaic cluster. The medium- and long-term cluster dispatchable capacity assessment unit probabilistically evaluates the dispatch potential of photovoltaic clusters, fitting the irradiance probability density function for the third preset time period using a two-parameter Beta distribution. The third preset time period is from 15 minutes to 1 hour, and its function is expressed as: Where G represents the real-time photovoltaic irradiance. For maximum irradiance, For Beta functions, Let be the shape parameter of the Beta distribution, when The time distribution degenerates into a uniform distribution. The distribution is biased towards areas with high irradiance. The time distribution is biased towards low irradiance regions; a stochastic predictive control model is established based on the irradiance probability density function, and the schedulable capacity is obtained by solving the stochastic predictive control model; the calculation logic of the stochastic predictive control model is as follows: Where K is the total number of time points. In order to exchange power with the main network, This represents the current real-time state of charge of the energy storage. For energy storage target state of charge, and With fixed weighting coefficients of 0.59 and 0.41 respectively, the data are combined with ultra-short-term, short-term, and medium-to-long-term forecast data to form a three-dimensional forecast vector.

[0017] The self-evolutionary calibration module is based on a deep reinforcement learning algorithm, combining a shared feature extractor and an independent decision head for online compensation of multi-scale prediction residuals. The specific process is as follows: The three-dimensional prediction vector, real-time measured data from the two-layer sensor network, historical calibration error sequences, and environmental disturbances are combined as input data. The local spatiotemporal features of the input data and the temporal dependencies of the historical calibration error sequences are extracted by a shared feature extractor. The spatiotemporal features include high-frequency components of power fluctuations, and the temporal dependencies include error accumulation trends. The two are concatenated into a feature vector and input to the scaled decision head. Each calibration quantity is output as a progressive correction value for the corresponding ultra-short-term, short-term, and medium-to-long-term state quantities. For example, the ultra-short-term includes progressive correction values ​​for state quantities such as photovoltaic active or reactive power, node voltage, and frequency. The input data is set as the state space, and the calibration values ​​corresponding to the ultra-short-term, short-term, and medium-to-long-term time scales are used as the action space. The calibration errors of the ultra-short-term, short-term, and medium-to-long-term time scales are squared respectively, and then accumulated according to different weights, with the negative value taken to obtain the accuracy reward. Then, the calibration value of each time scale is detected. If the absolute value of a certain calibration value exceeds the preset maximum safety threshold, the square value of the excess part is accumulated, multiplied by a penalty coefficient, and the negative value is taken to obtain the constraint penalty. The accuracy reward and constraint penalty are set as the reward function. Historical sample data is obtained. Each sample contains the state space, action space, and reward function. Samples are stored in an experience replay pool. Several samples are randomly selected each time. This method minimizes the error between the predicted Q-value and the target Q-value by updating the network through maximizing the action space using the Critic gradient until a preset number of iterations are reached to complete training and obtain a reinforced calibration model. Ultra-short-term historical predicted values ​​and measured values ​​are input into the reinforced calibration model, which outputs an ultra-short-term calibration matrix and merges it with the predicted output sequence to obtain ultra-short-term updated values. Short-term historical average errors are input into the reinforced calibration model, which outputs a corrected voltage sensitivity matrix. The corrected voltage sensitivity matrix and the voltage sensitivity matrix are merged to obtain the complete voltage sensitivity matrix. The latest medium- and long-term meteorological data are input into the reinforced calibration model, which outputs a schedulable capacity correction value. This correction value is then embedded into the schedulable capacity to obtain the complete schedulable capacity.

[0018] The distributed coordination control module obtains the optimal power command based on the improved alternating direction multiplier method and the non-dominated sorting genetic algorithm, taking into account the three-element constraints of voltage, frequency, and harmonics. The specific process is as follows: The power allocation of photovoltaic and energy storage devices is used as a decision variable and encoded using chromosome real numbers to obtain the photovoltaic power chromosome. and energy storage power chromosome 'm' represents the energy storage device number. N initial candidate solutions are randomly generated based on the photovoltaic power chromosome and the energy storage power chromosome, and integrated to obtain a non-dominated sort. Each solution is represented as a chromosome. A multi-objective fitness function is calculated using the initial candidate solutions. The multi-objective fitness function includes power tracking objective, grid loss objective, and control cost objective. The power tracking objective... The calculation logic is as follows: ,in Active power target value; network loss target The calculation logic is as follows: , and Number the different voltage monitoring points. For the electrical conductance between monitoring points, The voltage phase angle difference; the calculation logic for the control cost target is as follows: The system checks whether each chromosome meets voltage, frequency, and harmonic constraints. If not, its sorting priority is lowered. Specifically, the verification voltage value is calculated using a sensitivity matrix. Its calculation logic is as follows: ,in Based on the base voltage value, This is the power allocation vector; if the verification voltage values ​​of all monitoring points satisfy the first voltage threshold ≤ If the voltage is less than or equal to the second voltage threshold, the initial candidate solution is a voltage-feasible solution and retains its original priority; otherwise, it is marked as a voltage constraint violation solution and downgraded by one level in the non-dominated ranking. For frequency constraints, the sum of all active power and the current measured local load in the initial candidate solutions are extracted. The difference between the absolute values ​​of the sum of all active power and the current measured local load is used to calculate the active power imbalance. The active power imbalance is multiplied by a preset frequency deviation coefficient to obtain the frequency deviation. If the frequency deviation is less than or equal to the first preset deviation threshold, the initial candidate solution is a frequency-feasible solution with no penalty; otherwise, it is marked as a severely frequency-violated solution and removed from the non-dominated ranking. The harmonic constraint verification process involves obtaining a composite data frame of fundamental and harmonic frequencies and calculating the total harmonic distortion rate of each device. If the total harmonic distortion rate of the device is less than or equal to the preset distortion threshold, the candidate solution is a harmonic-feasible solution and retains its original ranking; otherwise, it is marked as a harmonic-exceeding solution and removed from the non-dominated ranking. Genetic operations and solution set generation: Several chromosomes are divided into Pareto levels, with the first level being the global optimum. The distance crowding between solutions in the same level is calculated and sorted. The solution with the largest distance crowding in the first level is selected as the parent. The offspring are randomly perturbed to obtain new candidate solutions. After completing a preset number of iterations, the first-level Pareto solution is output. It should be noted that the first-level Pareto solution includes five internally selected candidate solutions for scenarios such as coverage tracking priority or network loss priority. The main controller filters the first-level Pareto solution through real-time scenarios. During peak load periods, it selects the solution with the minimum network loss target; when photovoltaic fluctuations are large, it selects the solution with the minimum power tracking target; and when the voltage is critical, it selects the voltage-feasible solution. For each device's selected solution, Lagrange multipliers are added, and a global optimization update is performed to output the optimal global adjustment instruction for the corresponding device. The global optimization update formula is as follows: ,in The solution selected for the device, the first The total number of devices. For equipment number, For the h-th iteration, update the global optimization variables. The Lagrange multiplier for the h-th iteration sends the optimal global adjustment command to the server to adjust the active power of the photovoltaic and energy storage devices.

[0019] All the above formulas are performed numerically after dimensionless processing. Dimensionless processing can be achieved through conventional methods such as standardization and normalization, which will not be elaborated here. The formula parameters are obtained through software simulation and fitting based on a large amount of measured data, and can reflect the actual operating rules. The preset parameters can be flexibly configured by those skilled in the art according to specific application scenarios.

[0020] The embodiments of the present invention can be implemented by software, hardware, firmware, or any combination thereof. When implemented in software, it can be presented in whole or in part as a computer program product, which includes one or more computer instructions or computer programs. When the instructions or programs are loaded or executed on a computer, all or part of the processes and functions described in the embodiments of the present invention will be implemented. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0021] Computer instructions can be stored in a computer-readable storage medium or transmitted between computer-readable storage media via wired or wireless means (such as infrared, wireless, microwave, etc.). The transmission path can involve websites, computers, servers, or data centers. The computer-readable storage medium can be any usable medium that can be accessed by a computer, or a data storage device such as a server or data center that contains multiple usable media. The media types include magnetic media (floppy disks, hard disks, magnetic tapes), optical media (DVDs), and semiconductor media (solid-state drives), etc.

[0022] It should be understood that the program numbers in the embodiments of the present invention do not represent the execution order. The execution order is determined by the function and internal logic and should not constitute a limitation on the implementation process. Those skilled in the art will recognize that the units and algorithm steps of the various examples in the embodiments can be implemented by a combination of electronic hardware, computer software and hardware. The specific implementation method depends on the application scenario and design constraints of the technical solution, but its implementation should not exceed the protection scope of the present invention.

[0023] It should be noted that the systems, apparatuses, and methods disclosed in this invention can be implemented in other forms. For example, the unit division in the apparatus embodiment is only a logical functional division, and other division methods can be used in actual implementation. For instance, multiple units or components can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling or communication connection between the components can be achieved through interfaces, which can be electrical, mechanical, or other forms.

[0024] The separation components may or may not be physically separated, and the displayed units may or may not physically exist; they may be concentrated in one location or distributed across multiple network units. Some or all of the units can be selected to implement the solution of this embodiment according to actual needs.

[0025] Furthermore, each functional unit can be integrated into a processing unit, or exist as a separate physical entity, or two or more units can be integrated into one unit. If the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this, the contribution or function of the technical solution of this invention to the prior art can be embodied in a software product, stored in a storage medium, and includes several instructions to cause a computer device (personal computer, server, network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, and optical disks.

[0026] The above description is merely a specific embodiment of the present invention, but its scope of protection is not limited thereto. Any variations or substitutions readily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A real-time monitoring and coordinated control system for distributed photovoltaic grid-connected regional power grids, comprising a sensor acquisition module, a multi-timescale prediction module, a self-evolving calibration module, and a distributed coordinated control module, characterized in that: The sensor acquisition module collects wideband electrical quantities and multi-dimensional meteorological quantities based on a two-layer sensor network. The multi-timescale prediction module integrates ultra-short-term, short-term, and medium-to-long-term predictions, and then achieves cross-scale coordination and dynamic parameter updates through joint state-space model, sensitivity matrix, and stochastic model prediction control. The self-evolutionary calibration module is based on a deep reinforcement learning algorithm and combines a shared feature extractor and an independent decision head to perform online compensation for multi-scale prediction residuals. The distributed coordination control module obtains the optimal power command based on the improved alternating direction multiplier method and the non-dominated sorting genetic algorithm, which are constrained by voltage, frequency and harmonics. The distributed coordination control module checks whether each chromosome satisfies voltage, frequency, and harmonic constraints using a non-dominated sorting genetic algorithm, specifically as follows: Power allocation for photovoltaic (PV) and energy storage devices is used as decision variables and encoded using chromosome real numbers to obtain PV power chromosomes and energy storage power chromosomes. Based on these chromosomes, N initial candidate solutions are randomly generated and integrated to obtain a non-dominated ranking. Each solution is represented as a chromosome. A multi-objective fitness function is calculated using the initial candidate solutions. This function includes power tracking objectives, network loss objectives, and control cost objectives. Each chromosome is checked to see if it meets voltage, frequency, and harmonic constraints. If not, its ranking priority is reduced. Specifically, the verification voltage value is calculated using the sensitivity matrix. If the verification voltage values ​​of all monitoring points satisfy the condition: first voltage threshold ≤ verification voltage value ≤ second voltage threshold, then the initial candidate solution is a voltage-feasible solution and retains its original priority; otherwise, it is marked as... Solutions violating voltage constraints are downgraded one level in the non-dominated ranking. For frequency constraints, the sum of all active power and the current measured local load are extracted from the initial candidate solutions. The absolute difference between the sum of all active power and the current measured local load is used to calculate the active power imbalance. This imbalance is multiplied by a preset frequency deviation coefficient to obtain the frequency deviation. If the frequency deviation is less than or equal to the first preset deviation threshold, the initial candidate solution is considered frequency-feasible and receives no penalty; otherwise, it is marked as a severely frequency-violated solution and removed from the non-dominated ranking. The harmonic constraint verification process involves acquiring composite data frames of the fundamental and harmonic frequencies and calculating the total harmonic distortion rate (THD) of each device. If the THD of the device is less than or equal to the preset distortion threshold, the initial candidate solution is considered harmonic-feasible and retains its original ranking; otherwise, it is marked as a harmonic-exceeding solution and removed from the non-dominated ranking.

2. The distributed photovoltaic grid-connected regional power grid real-time monitoring and coordination control system according to claim 1, characterized in that, The distributed coordination and control module's genetic operations and solution set generation are as follows: Several chromosomes are divided into Pareto levels, with the first level being the global optimum. The distance crowding between solutions of the same level is calculated and sorted. The solution with the largest distance crowding in the first level is selected as the parent. The offspring are randomly perturbed to obtain new candidate solutions. After completing a preset number of iterations, the Pareto solution of the first level is output.

3. The distributed photovoltaic grid-connected regional power grid real-time monitoring and coordination control system according to claim 2, characterized in that, The specific steps for the distributed coordination and control module to output the optimal global adjustment instruction are as follows: The main controller filters the first-level Pareto solution through real-time scenarios. During peak load periods, it selects the solution with the minimum network loss target; when photovoltaic fluctuations are large, it selects the solution with the minimum power tracking target; and when the voltage is critical, it selects the voltage-feasible solution. For each device, a Lagrange multiplier is added to the selected solution, and a global optimization update is performed to output the optimal global adjustment instruction for the corresponding device. The optimal global adjustment instruction is then sent to the server to adjust the active power of the photovoltaic and energy storage devices.

4. The real-time monitoring and coordinated control system for distributed photovoltaic grid-connected regional power grids according to claim 1, characterized in that, The process by which the self-evolutionary calibration module outputs correction values ​​for ultra-short-term, short-term, and medium-to-long-term state variables: The three-dimensional prediction vector, real-time measured data from the two-layer sensor network, historical calibration error sequences, and environmental disturbances are combined as input data. The local spatiotemporal features of the input data and the temporal dependencies of the historical calibration error sequences are extracted by a shared feature extractor. The two are concatenated into a feature vector and input into the scaled decision head to output various calibration quantities.

5. A real-time monitoring and coordinated control system for distributed photovoltaic grid-connected regional power grids according to claim 4, characterized in that, The self-evolutionary calibration module, based on a deep reinforcement learning algorithm combined with a shared feature extractor and an independent decision head, performs online compensation for multi-scale prediction residuals. The specific process is as follows: The input data is set as the state space, and the calibration values ​​corresponding to the ultra-short-term, short-term, and medium-to-long-term time scales are used as the action space. The calibration errors of the ultra-short-term, short-term, and medium-to-long-term time scales are squared respectively, and then accumulated according to different weights, with the negative value taken to obtain the accuracy reward. Then, the calibration value of each time scale is detected. If the absolute value of a certain calibration value exceeds the preset maximum safety threshold, the square value of the excess part is accumulated, multiplied by a penalty coefficient, and the negative value is taken to obtain the constraint penalty. The accuracy reward and constraint penalty are set as the reward function. Historical sample data is obtained. Each sample contains the state space, action space, and reward function. Samples are stored in an experience replay pool. Several samples are randomly selected each time. This method minimizes the error between the predicted Q-value and the target Q-value by updating the network through maximizing the action space using the Critic gradient until a preset number of iterations are reached to complete training and obtain a reinforced calibration model. Ultra-short-term historical predicted values ​​and measured values ​​are input into the reinforced calibration model, which outputs an ultra-short-term calibration matrix and merges it with the predicted output sequence to obtain ultra-short-term updated values. Short-term historical average errors are input into the reinforced calibration model, which outputs a corrected voltage sensitivity matrix. The corrected voltage sensitivity matrix and the voltage sensitivity matrix are merged to obtain the complete voltage sensitivity matrix. The latest medium- and long-term meteorological data are input into the reinforced calibration model, which outputs a schedulable capacity correction value. This correction value is then embedded into the schedulable capacity to obtain the complete schedulable capacity.

6. The real-time monitoring and coordinated control system for distributed photovoltaic grid-connected regional power grids according to claim 1, characterized in that, The multi-timescale prediction module integrates ultra-short-term, short-term, and medium-to-long-term predictions, and then achieves cross-scale coordination and dynamic parameter updates through joint state-space model, sensitivity matrix, and stochastic model prediction and control. The specific process is as follows: The repaired data is converted into state vectors. A control state space model is set based on the state vectors. The control state space model includes control input vectors, disturbance vectors, and output vectors. The control input vectors and disturbance vectors are embedded into the state equations to obtain the time state equations. The output vectors are embedded into the output equations to obtain the relational output equations. The parameters of the control state-space model are updated in real time using a recursive least squares algorithm with a forgetting factor. The parameters include the state transition matrix, control input matrix, disturbance matrix, output matrix, and direct transmission matrix. A regression vector and a parameter vector to be estimated are set, and the parameters are updated recursively based on the regression vector and the parameter vector to be estimated. A rolling optimization is used to generate the predicted output sequence for the first preset time period. The predicted output sequence includes the active and reactive power of each photovoltaic node, the voltage amplitude of key nodes, and the voltage phase angle. Using the real-time voltage of each photovoltaic node, the real-time reactive power output of the inverter, and the active power of the load as inputs, photovoltaic active power-voltage sensitivity matrices, inverter reactive power-voltage sensitivity matrices, and load active power-voltage sensitivity matrices are established respectively. Calculate the difference in photovoltaic active power and voltage between adjacent time points, screen for periods when the inverter reactive power and load changes are less than the corresponding preset change thresholds, and obtain the photovoltaic active power-voltage sensitivity matrix by solving the least squares method. Calculate the reactive power difference of the inverter between adjacent time points and obtain the inverter reactive power-voltage sensitivity matrix by solving the least squares method. Subtract the load active power between adjacent time points to obtain the load active power difference and obtain the load active power-voltage sensitivity matrix by solving the least squares method. Integrate the photovoltaic active power-voltage sensitivity matrix, the inverter reactive power-voltage sensitivity matrix, and the load active power-voltage sensitivity matrix to obtain the voltage sensitivity matrix. The voltage stability margin is calculated based on the sensitivity matrix. The medium- and long-term cluster dispatchable capacity assessment unit probabilistically assesses the dispatch potential of photovoltaic clusters. It fits the irradiance probability density function for the third preset period using a two-parameter Beta distribution, establishes a stochastic predictive control model based on the irradiance probability density function, solves the stochastic predictive control model to obtain the dispatchable capacity, and integrates ultra-short-term, short-term, and medium- and long-term prediction data into a three-dimensional prediction vector.

7. The real-time monitoring and coordinated control system for distributed photovoltaic grid-connected regional power grids according to claim 1, characterized in that, The sensor acquisition module collects broadband electrical quantities and multi-dimensional meteorological quantities respectively based on a two-layer sensor network. The specific process is as follows: Each photovoltaic inverter's outlet is equipped with a bottom-level sensor, and upper-level sensors are installed at the highest points of the photovoltaic modules and power distribution area. The bottom-level sensors include voltage sensors, current sensors, phase angle sensors, and switch status sensors. The upper-level sensors include the photovoltaic array area sensor group and the regional observation point meteorological station. Each bottom-level sensor collects three-phase voltage, three-phase current, voltage phase angle, and switch status quantities in real time, and divides them into conventional electrical quantities and wideband components. The wideband components are subjected to Fourier transform in real time to extract harmonic components and generate composite data frames containing fundamental, harmonic, and wideband noise. The photovoltaic array area sensor group collects photovoltaic irradiance, backsheet temperature, and module surface temperature in real time, and the regional observation point meteorological station collects wind speed, wind direction, ambient temperature, cloud thickness, and cloud movement speed.

8. A real-time monitoring and coordination control system for distributed photovoltaic grid-connected regional power grids according to claim 7, characterized in that, The sensing and acquisition module inputs the data collected by each bottom-level sensor and each top-level sensor into the anomaly detection model to construct a time-series feature matrix. An autoencoder is used to compress the multi-dimensional time-series feature matrix into a latent vector. The decoder reconstructs the latent vector to obtain reconstructed data and calculates the reconstruction error. When the reconstruction error is greater than the dynamic error threshold, it is determined to be an anomaly and the anomaly is replaced with the reconstructed data to obtain the repaired data.

Images