Power grid adaptive pre-adjustment frequency method, device and equipment and storage medium

Abstract

The application discloses a power grid adaptive pre-regulation frequency method and device, equipment and storage medium, and belongs to the technical field of power system automation and control. The method comprises the following steps: wavelet transform is performed on historical fluctuation data of a net load and initial prediction data of a next period, noise bands are suppressed and main characteristic bands are corrected, and a net load prediction sequence of a plurality of future pre-regulation frequency execution points is obtained after reconstruction; equivalent power-frequency sensitivity is obtained, estimated frequency offset and resource sensitivity decomposition results are obtained in combination with the net load prediction sequence; the estimated frequency offset, the resource sensitivity decomposition results and scenario prior characteristics are taken as states, candidate pre-regulation frequency schemes are simulated and constraint checking is performed, a deep Q network is trained to obtain a pre-regulation frequency comprehensive ranking strategy; a target pre-regulation frequency scheme is selected and is mapped into a pre-regulation frequency power trajectory and a pre-regulation frequency parameter adjustment amount of the future pre-regulation frequency execution points. The embodiment of the application can improve the foresight and robustness of new energy power grid frequency control.

Description

Technical Field

[0001] This application relates to the field of power system automation and control technology, and in particular to a power grid adaptive pre-frequency regulation method, device, equipment and storage medium. Background Technology

[0002] With the continuous increase in the installed capacity and penetration rate of new energy sources such as wind power and photovoltaics, the proportion of grid synchronous machines is decreasing, and the system inertia and adjustable active power reserves are reduced. The rapid random fluctuations of new energy sources and loads are directly transmitted to the system frequency, resulting in a contraction of the frequency safety margin. Traditional frequency regulation methods rely on passive adjustment of real-time frequency deviations, lacking forward-looking identification of future frequency offset risks, which easily leads to problems such as amplified frequency deviations, frequent frequency regulation actions, and inefficient use of reserves. Existing methods are mostly based on predictions of the original power sequence, failing to effectively distinguish between key fluctuation components that contribute to frequency behavior and high-frequency noise, resulting in inaccurate frequency risk assessments. Therefore, a new method is needed that can collaboratively perceive future frequency risks and perform adaptive forward-looking adjustments. Summary of the Invention

[0003] The purpose of this application is to provide a power grid adaptive pre-frequency control method, device, equipment, and storage medium that can effectively improve the foresight, accuracy, and robustness of frequency control in new energy power grids.

[0004] To achieve the above objectives, a first aspect of this application provides a power grid adaptive pre-frequency regulation method, comprising: Wavelet transform is performed on the historical fluctuation data of net load and the initial prediction data of the next period. Noise frequency bands are suppressed and main characteristic frequency bands are corrected based on prior features. After wavelet reconstruction, net load prediction sequences for multiple future pre-frequency adjustment execution points are obtained. Based on the power-frequency sensitivity mapping table and the prior characteristics of the current operating conditions, dynamic values ​​are assigned to obtain the equivalent power-frequency sensitivity. Combined with the net load prediction sequence, calculations are performed to obtain the estimated frequency offset and resource sensitivity decomposition results for each future pre-frequency adjustment execution point. Using the estimated frequency offset, resource sensitivity decomposition results, and scenario prior features as states, the candidate pre-frequency modulation schemes are simulated and constrained. A deep Q-network is trained through a multi-objective reward function to obtain a comprehensive pre-frequency modulation ranking strategy. The target pre-frequency modulation scheme is selected according to the pre-frequency modulation comprehensive sorting strategy, and the target pre-frequency modulation scheme is mapped to the pre-frequency modulation power trajectory and pre-frequency modulation parameter adjustment amount of the future pre-frequency modulation execution point.

[0005] Compared with existing technologies, the adaptive pre-frequency regulation method for power grids provided in this application has the following advantages: By performing wavelet transform on historical net load fluctuation data and initial prediction data for the next time period, noise is suppressed and the main characteristic frequency bands are corrected based on prior features, and then wavelet reconstruction is performed, which improves the accuracy of net load prediction at future pre-frequency regulation execution points and provides reliable support for subsequent frequency assessment; by combining the power-frequency sensitivity mapping table with the prior features of the current operating conditions to dynamically assign equivalent power-frequency sensitivity, and in conjunction with the net load prediction sequence, fine-grained assessment of frequency offset at multiple future execution points and resource sensitivity analysis can be accurately achieved. The solution provides a precise basis for formulating pre-frequency regulation schemes. Using relevant evaluation results and scenario prior features as the state, a comprehensive ranking strategy for pre-frequency regulation is obtained through simulation verification and training a deep Q-network with a multi-objective reward function, ensuring the comprehensive optimality of candidate pre-frequency regulation schemes. Finally, the target pre-frequency regulation scheme is selected and mapped to the pre-frequency regulation power trajectory and the pre-frequency regulation parameter adjustment amount, realizing adaptive tuning of pre-frequency regulation parameters. This can proactively identify frequency risks and actively pre-adjust, effectively avoiding problems such as frequency difference amplification and frequent frequency regulation actions, improving the foresight, accuracy and robustness of power grid pre-frequency regulation control, optimizing the utilization efficiency of reserve resources, and adapting to the needs of complex power grid operating conditions.

[0006] In some embodiments, the wavelet transform is performed on the historical fluctuation data of net load and the initial prediction data for the next time period. Noise frequency bands are suppressed based on prior features, and the main characteristic frequency bands are corrected. After wavelet reconstruction, a net load prediction sequence for multiple future pre-tuning execution points is obtained, including: The historical fluctuation data of net load is combined with the initial forecast data for the next period in chronological order to form a combined sequence; Wavelet transform decomposition is performed on the combined sequence to obtain multiple frequency band detail components; Based on the prior characteristics of the multiple frequency band detail components, noise frequency bands are distinguished from main characteristic frequency bands; Noise in the noise band is suppressed based on a level threshold, and the amplitude of the main characteristic frequency band is corrected based on a level correction coefficient to obtain the processed frequency band coefficient. Wavelet reconstruction is performed on the processed frequency band coefficients to obtain the purified time series, and the purified time series and the combined sequence have the same length. Based on the purification time series, time-series prediction is performed to obtain net load prediction sequences for multiple future pre-frequency adjustment execution points.

[0007] In some embodiments, the prior features include at least one of the following: new energy penetration rate, system inertia level, tidal current tension at key sections, and typical weather scenarios.

[0008] In some embodiments, the dynamic assignment based on the power-frequency sensitivity mapping table and prior characteristics of the current operating condition to obtain the equivalent power-frequency sensitivity, combined with the net load prediction sequence, yields the estimated frequency offset and resource sensitivity decomposition results for each future pre-frequency adjustment execution point, including: In the offline phase, small disturbance sensitivity analysis is performed on various adjustable resources under multiple typical operating conditions, and a power-frequency sensitivity mapping table indexed by prior operating condition characteristics is established. During the online phase, prior features of the current operating condition are extracted and matched and interpolated with typical operating conditions in the power-frequency sensitivity mapping table to dynamically assign the equivalent power-frequency sensitivity of various adjustable resources under the current operating condition. Based on the net load forecast sequence and the current planned value, calculate the power balance deviation at each future pre-frequency adjustment execution point; The estimated frequency offset for each future pre-frequency modulation execution point is calculated based on the equivalent power-frequency sensitivity and the power balance deviation. Based on the equivalent power-frequency sensitivity and the adjustable capacity of various adjustable resources, the sensitivity index of different adjustable resources to the estimated frequency offset at each future pre-frequency adjustment execution point is calculated, and the resource sensitivity decomposition result is obtained.

[0009] In some embodiments, the method further includes: When calculating the estimated frequency offset of each future pre-frequency adjustment execution point, the measured power flow at the current time is compared with the planned value at the current time to obtain the reference error, and the estimated frequency offset of each future pre-frequency adjustment execution point is corrected based on the reference error.

[0010] In some embodiments, the process of using the estimated frequency offset, resource sensitivity decomposition results, and scenario prior features as states to simulate and constrain candidate pre-frequency modulation schemes, and training a deep Q-network through a multi-objective reward function to obtain a comprehensive pre-frequency modulation ranking strategy, includes: Based on the estimated frequency offset, resource sensitivity decomposition results, adjustable resource constraints, and scenario prior features, the state vector of the agent is constructed. The action space of the agent is constructed based on a set of candidate pre-frequency power allocation schemes that satisfy basic safety constraints; Design a multi-objective reward function that includes frequency security indicators, control cost indicators, power ramp smoothness indicators, and operational constraint violation penalties. Dynamically adjust the weights of each constraint indicator in the multi-objective reward function based on scenario prior features to construct the training objective of the agent. In the simulation environment, frequency trajectory simulation and operational constraint verification are performed on the candidate pre-frequency modulation schemes selected by the agent. The feedback value of the multi-objective reward function is calculated based on the verification results, and a deep Q-network is trained to obtain a pre-frequency modulation comprehensive ranking strategy for online ranking of the candidate pre-frequency modulation schemes.

[0011] In some embodiments, the operational constraint verification includes: frequency over-limit verification, unit output and ramp constraint verification, standby constraint verification, and line power flow constraint verification.

[0012] To achieve the above objectives, a second aspect of this application provides a power grid adaptive pre-frequency regulation device, the device comprising: The prediction module is used to perform wavelet transform on the historical fluctuation data of net load and the initial prediction data of the next period. Based on the prior features, the noise frequency band is suppressed and the main feature frequency band is corrected. After wavelet reconstruction, the net load prediction sequence of multiple future pre-frequency adjustment execution points is obtained. The calculation module is used to dynamically assign values ​​based on the power-frequency sensitivity mapping table and the prior characteristics of the current operating conditions to obtain the equivalent power-frequency sensitivity. Combined with the net load prediction sequence, the module calculates the estimated frequency offset and resource sensitivity decomposition results for each future pre-frequency adjustment execution point. The training module is used to simulate and constrain the candidate pre-frequency modulation schemes by taking the estimated frequency offset, resource sensitivity decomposition results and scenario prior features as states, and to train a deep Q network through a multi-objective reward function to obtain a comprehensive pre-frequency modulation ranking strategy. The mapping module is used to select a target pre-frequency modulation scheme according to the pre-frequency modulation comprehensive sorting strategy, and map the target pre-frequency modulation scheme to the pre-frequency modulation power trajectory and pre-frequency modulation parameter adjustment amount of the future pre-frequency modulation execution point.

[0013] To achieve the above objectives, a third aspect of this application provides an electronic device, the electronic device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the method described in the first aspect.

[0014] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when executed, controls the device containing the computer-readable storage medium to perform the method described in the first aspect. Attached Figure Description

[0015] Figure 1 This is a flowchart of a power grid adaptive pre-frequency regulation method provided in an embodiment of this application; Figure 2This is a schematic diagram of a power grid adaptive pre-frequency regulation device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0018] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0019] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0020] As the installed capacity and penetration rate of new energy sources such as wind power and photovoltaics continue to climb, the proportion of traditional synchronous generating units in the power grid is gradually declining. This leads to a significant decrease in system inertia and adjustable active power reserve levels. The rapid random fluctuations in new energy output and load power have an increasingly direct impact on system frequency, resulting in a continuously narrowing system frequency safety margin. Against this backdrop, the traditional frequency control system, centered on inertial support, primary frequency regulation, and automatic generation control (AGC), primarily relies on real-time or near-real-time frequency deviation feedback for passive adjustment. It lacks the ability to proactively identify and adjust frequency deviations at future execution points, which can easily lead to a series of problems such as amplified frequency deviations, frequent redundant frequency regulation actions, and low efficiency in the utilization of reserve resources.

[0021] In terms of new energy output and net load forecasting, existing technologies mostly employ time-domain statistical models or deep learning models to directly model and analyze the raw power time series. These methods fail to effectively distinguish and process prediction errors, measurement noise, and invalid high-frequency jitter, and lack frequency domain decomposition and noise suppression mechanisms that incorporate prior features of the scenario. Such methods struggle to accurately differentiate the contribution of different frequency band fluctuation components to the system's frequency behavior, resulting in a large number of high-frequency interference components in the predicted sequence that are unrelated or weakly related to frequency control. Ultimately, this leads to frequency risk assessments based on the predicted data that are either conservative or optimistic, severely limiting the accuracy and robustness of subsequent pre-frequency regulation decisions.

[0022] On the other hand, in the power-frequency sensitivity assessment stage, existing technologies are generally based on simplified static equivalent models or linearized analysis methods under a few typical operating conditions. They assume that the frequency response coefficient to active power disturbances remains constant within a specific range, making it difficult to fully characterize the dynamic evolution of prior characteristics of operating conditions such as renewable energy penetration rate, system inertia level, and power flow at key sections. At the same time, most existing methods only perform frequency sensitivity calculations for a single operating point or a short time period, failing to achieve fine-grained frequency offset assessment and adjustable resource sensitivity decomposition for future multi-operation points, and making it even more difficult to form a deep coupling with rolling forecast results.

[0023] Based on this, embodiments of this application provide a power grid adaptive pre-frequency control method, apparatus, device, and storage medium, which can effectively improve the foresight, accuracy, and robustness of frequency control in new energy power grids.

[0024] Please see Figure 1 , Figure 1 This is an optional flowchart of the power grid adaptive pre-frequency regulation method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S104.

[0025] Step S101: Perform wavelet transform on the historical fluctuation data of net load and the initial prediction data of the next period. Based on the prior features, suppress the noise frequency band and correct the main feature frequency band. After wavelet reconstruction, obtain the net load prediction sequence of multiple future pre-frequency adjustment execution points. Step S102: Dynamically assign values ​​based on the power-frequency sensitivity mapping table and the prior characteristics of the current operating conditions to obtain the equivalent power-frequency sensitivity. Combine this with the net load prediction sequence to calculate the estimated frequency offset and resource sensitivity decomposition results for each future pre-frequency adjustment execution point. Step S103: Using the estimated frequency offset, resource sensitivity decomposition results and scenario prior features as the state, simulate and constrain the candidate pre-frequency modulation scheme, and train a deep Q network through a multi-objective reward function to obtain the pre-frequency modulation comprehensive ranking strategy. Step S104: Select the target pre-frequency modulation scheme according to the pre-frequency modulation comprehensive sorting strategy, and map the target pre-frequency modulation scheme to the pre-frequency modulation power trajectory and pre-frequency modulation parameter adjustment amount of the future pre-frequency modulation execution point.

[0026] Steps S101 to S104 as illustrated in this application embodiment improve the accuracy of net load prediction for future pre-frequency adjustment execution points by performing wavelet transform on historical net load fluctuation data and initial prediction data for the next time period, suppressing noise and correcting main characteristic frequency bands based on prior features, and then performing wavelet reconstruction. This provides reliable support for subsequent frequency assessment. By combining the power-frequency sensitivity mapping table with the dynamic assignment of prior features of the current operating condition to obtain the equivalent power-frequency sensitivity, and using it in conjunction with the net load prediction sequence, fine-grained assessment of frequency offsets at multiple future execution points and decomposition of resource sensitivity can be accurately achieved, providing a basis for pre-frequency adjustment. The scheme provides precise basis for formulation; based on relevant evaluation results and scenario prior characteristics, a pre-frequency modulation comprehensive ranking strategy is obtained through simulation verification and training a deep Q-network with a multi-objective reward function to ensure the comprehensive optimality of candidate pre-frequency modulation schemes; finally, the target pre-frequency modulation scheme is selected and mapped to the pre-frequency modulation power trajectory and pre-frequency modulation parameter adjustment amount, realizing adaptive tuning of pre-frequency modulation parameters. It can proactively identify frequency risks and actively pre-adjust, effectively avoiding problems such as frequency difference amplification and frequent frequency modulation actions, improving the foresight, accuracy and robustness of power grid pre-frequency modulation control, optimizing the utilization efficiency of reserve resources, and adapting to the needs of complex power grid operating conditions.

[0027] In step S101 of some embodiments, net load can be the difference between the output of new energy sources and the load power in the power grid, which is an indicator reflecting the active power balance of the power grid; historical fluctuation data can be past time series data covering typical fluctuation characteristics of net load; initial forecast data for the next period can be the preliminary forecast value of future net load obtained through day-ahead / short-term forecasts or scheduling plans; wavelet transform can be a frequency domain analysis method that decomposes time series data into approximate and detail components of different frequency bands; prior features can be features obtained based on power grid operation scenarios and historical statistics, including new energy penetration rate, system inertia, weather scenarios, correlation between each frequency band and frequency behavior, etc.; noise frequency band can be a frequency band component with weak correlation to the frequency behavior of the power grid system and small amplitude; main characteristic frequency band can be a frequency band component that is highly correlated with the frequency behavior of the power grid system and plays a dominant role in frequency fluctuation; wavelet reconstruction can be the inverse transformation process of restoring the processed wavelet decomposed components to the time series; future pre-frequency regulation execution point can be the specific time node for the power grid to carry out pre-frequency regulation control; net load prediction sequence can be a sequence formed by matching the net load prediction values ​​of each future pre-frequency regulation execution point in chronological order.

[0028] In some embodiments, the prior features include at least one of the following: new energy penetration rate, system inertia level, tidal current tension at key sections, and typical weather scenarios.

[0029] In some embodiments, wavelet transform is performed on historical fluctuation data of net load and initial forecast data for the next time period. Noise frequency bands are suppressed based on prior features, and the main characteristic frequency bands are corrected. After wavelet reconstruction, a net load forecast sequence for multiple future pre-tuning execution points is obtained, including: The historical fluctuation data of net load is combined with the initial forecast data for the next period in chronological order to form a combined sequence; Wavelet transform decomposition of the combined sequence yields multiple frequency band detail components; Based on the prior characteristics of multiple frequency band detail components, noise frequency bands are distinguished from main characteristic frequency bands; Noise in the noise band is suppressed based on the level threshold, and the amplitude of the main characteristic frequency band is corrected based on the level correction coefficient to obtain the processed frequency band coefficient. Wavelet reconstruction is performed on the processed frequency band coefficients to obtain the purified time series. The purified time series and the combined sequence have the same length. Based on the purification time series, time-series prediction is performed to obtain net load prediction sequences for multiple future pre-frequency adjustment execution points.

[0030] Specifically, net load is the difference between the output of new energy sources and the power of the load in the power grid. To analyze its fluctuation characteristics, a segment of length is first selected. The historical fluctuation data window contains two time series: the new energy output series, denoted as a continuous time form. The load power sequence is denoted in continuous-time form. .

[0031] Furthermore, sampling can be performed at a uniform interval. Discretize the above continuous time series (e.g., 1s, 2s, or 4s) to construct historical fluctuation data (sequence) of net load. :

[0032] in, For discrete-time indexing, Representing the current moment, therefore this discrete sequence It exactly covers the length of The historical time period corresponds perfectly to the time range of the historical fluctuation data window.

[0033] Initial forecast data for the next period This is obtained through existing day-ahead / short-term forecasts or scheduling plans, covering Each time step corresponds to the time period in which the future pre-frequency adjustment execution point will be located:

[0034] Furthermore, historical fluctuation data is concatenated with initial forecast data for the next time period on the time axis to form a combined sequence for frequency domain analysis. :

[0035] in, Representing the current moment, Combined sequences Historical data section: From arrive Historical fluctuation data of net load at any given time, total Each time step reflects the actual net load fluctuation characteristics over a period of time. Combined sequences Future predictions section: From arrive Initial net load forecast at time , total Each time step contains forecast information for future periods.

[0036] It is evident that this combined sequence not only retains the actual fluctuation characteristics of the past period but also includes the initial prediction of the future period, providing a complete input for subsequent wavelet analysis.

[0037] Subsequently, the combined sequence Wavelet transform decomposition is performed to obtain multiple frequency band detail components. Specifically, it is necessary to first select wavelet basis functions suitable for power time series analysis (such as the Daubechies, Symlets, or Coiflets family), and then determine the number of decomposition levels based on the sampling interval and the frequency band of interest. Then, for the combined sequence Perform Discrete Wavelet Transform (DWT):

[0038] in, It can be the first The layer approximate component corresponds to a low-frequency, slowly varying trend; It can be the first Layer detail components, corresponding to the first The fluctuation components of each frequency band.

[0039] In addition to obtaining multiple frequency band detail components, the Lth-level approximate component corresponding to the slow-varying trend at low frequencies will also be obtained, along with the sampling interval. With the number of decomposition layers Establish each wavelet detail component The correspondence between the actual time scale / frequency range prepares for subsequent frequency band differentiation.

[0040] Next, based on the prior features of multiple frequency band detail components, noise bands and main characteristic bands are distinguished. These prior features include at least three categories: scenario-based features, statistical features, and operational constraint features. Scenario-based features include: current typical weather scenarios (strong convection, typhoons, cold waves, long periods of overcast skies, etc.), renewable energy penetration rate ranges, system inertia levels, and tidal current tension at key sections. Statistical features include: the correlation between the energy proportion of each frequency band and system frequency offset indicators (maximum frequency difference, RoCoF, etc.) calculated based on historical large samples. Operational constraint features include: unit ramp-up constraints and frequency regulation response time scales. Through these prior features, the main characteristic bands highly correlated with frequency behavior and the noise bands with weak correlation and small amplitude can be accurately identified.

[0041] Subsequently, noise in the noise band is suppressed based on a level threshold, and the amplitude of the main characteristic frequency band is corrected based on a level correction coefficient to obtain the processed frequency band coefficients; where the level threshold is... With level correction factor In the offline phase, based on the aforementioned prior features, the coefficient distribution and contribution to frequency behavior of each frequency band are statistically analyzed and configured. The two together constitute the "prior feature-frequency band parameter mapping table". During online operation, the table is selected or interpolated according to the current scenario.

[0042] It should be noted that the level threshold Used to identify noise components with "very small amplitude and weak correlation with frequency behavior"; level correction coefficient Used to adjust the amplitude of key characteristic frequency bands to avoid excessive smoothing or amplification of critical fluctuations during reconstruction.

[0043] Noise band suppression follows specific rules for detail components that are labeled as "noise bands" by prior features. Using a level threshold The process is handled using a graded noise reduction rule: When the magnitude of the detail component coefficients satisfies At that time, The discrete coefficients are set to zero or subjected to strong suppression to completely filter out such weakly correlated minute noise; When the magnitude of the detail component coefficients satisfies However, if the correlation between the component and the frequency behavior is still low based on prior features, then the amplitude is weakened by soft thresholding or proportional reduction in order to reduce the interference of such components on the wavelet reconstruction results.

[0044] The amplitude correction of the main characteristic frequency band is achieved through the formula Implementation, in which Determined by prior features under the corresponding working conditions. A value slightly greater than 1 can enhance low- and mid-frequency fluctuations, or a value slightly less than 1 can suppress short-term spikes. Simultaneously, for the low-frequency approximate component... Generally, strong thresholding is not performed. Only when there is a significant measurement offset or abnormal trend is a slow correction based on offset detection used to maintain the integrity of the system's long-term trend information.

[0045] After processing the frequency band coefficients, wavelet reconstruction is performed on the processed frequency band coefficients to obtain the purified time series. Purification time series and combined sequence The lengths are the same; specifically, the inverse wavelet transform (IDWT) is used to reconstruct the processed frequency band coefficients (i.e., the low-frequency approximate components and the detail components of each frequency band) to obtain the purified time series. :

[0046] Specifically, the processed frequency band coefficients refer to the Lth-level low-frequency approximation components after threshold suppression and level correction. and various details The reconstructed purified time series can effectively suppress high-frequency noise and invalid jitter, while retaining and moderately enhancing mid- and low-frequency fluctuations that are closely related to frequency behavior, providing a more physically reliable input for subsequent time series prediction and pre-tuning decisions.

[0047] Finally, based on the cleanup time series, time-series forecasting is performed to obtain net load forecast sequences for multiple future pre-frequency adjustment execution points; specifically, the cleanup time series... The historical data is used as input to the time series forecasting model, and the corresponding location in future time periods is used as the forecast target. The net load / output forecasts for multiple future time periods are obtained through a short-term forecasting model (which can be a statistical model or a deep learning model):

[0048] Based on the pre-frequency modulation strategy design, these future prediction values ​​are aligned in time with future pre-frequency modulation execution points (such as AGC adjustment cycle, primary frequency modulation evaluation cycle, etc.) to finally form the net load prediction sequence corresponding to the future pre-frequency modulation execution points. This sequence serves as a key input for the subsequent future execution point frequency evaluation and prior perception multi-objective deep Q-learning pre-frequency modulation comprehensive ranking strategy, which can significantly reduce the interference of measurement noise and invalid high-frequency jitter on subsequent decisions and provide high-quality prediction basis data.

[0049] This embodiment achieves frequency domain noise reduction for net load prediction, improves the accuracy of prediction data, and provides reliable data support for subsequent frequency assessment.

[0050] In step S102 of some embodiments, the power-frequency sensitivity mapping table can be a table that organizes and stores the power-frequency sensitivity coefficients obtained through small disturbance analysis under various typical operating conditions offline, according to the prior feature dimension; the current operating condition can be the real-time / near-real-time operating state of the power grid, specifically characterized by various prior features; the equivalent power-frequency sensitivity can be the power-frequency sensitivity coefficient adapted to the current operating condition, dynamically assigned by the sensitivity mapping table; the estimated frequency offset can be the estimated deviation of the system frequency relative to the reference value at the future pre-frequency adjustment execution point under the assumption that the pre-frequency adjustment parameters are not adjusted for the time being; the resource sensitivity decomposition result can be the quantitative decomposition data of the influence degree and potential regulation contribution of each adjustable resource on the frequency offset of each future pre-frequency adjustment execution point.

[0051] In some embodiments, an equivalent power-frequency sensitivity is obtained by dynamically assigning values ​​based on a power-frequency sensitivity mapping table and prior characteristics of the current operating condition. This equivalent power-frequency sensitivity is then calculated using the net load forecast sequence to obtain the estimated frequency offset and resource sensitivity decomposition results for each future pre-frequency adjustment execution point, including: During the offline phase, small disturbance sensitivity analysis was performed on various adjustable resources under multiple typical operating conditions, and a power-frequency sensitivity mapping table indexed by prior operating condition characteristics was established. During the online phase, prior features of the current operating conditions are extracted and matched and interpolated with typical operating conditions in the power-frequency sensitivity mapping table to dynamically assign the equivalent power-frequency sensitivity of various adjustable resources under the current operating conditions. Based on the net load forecast sequence and the current planned value, calculate the power balance deviation at each future pre-frequency adjustment execution point; The estimated frequency offset for each future pre-frequency modulation execution point is obtained by calculating the equivalent power-frequency sensitivity and power balance deviation. Based on the equivalent power-frequency sensitivity and the adjustable capacity of various adjustable resources, the sensitivity index of different adjustable resources to the estimated frequency offset at each future pre-frequency adjustment execution point is calculated, and the resource sensitivity decomposition results are obtained.

[0052] Specifically, firstly, in the offline phase, small disturbance sensitivity analysis is performed on various adjustable resources under multiple typical operating conditions to establish a power-frequency sensitivity mapping table indexed by prior operating condition characteristics. Specifically, a set of operating conditions covering typical operating states needs to be selected, including different load levels (low load, off-peak, peak, etc.), different renewable energy penetration ranges (e.g., 30%-50%, 50%-70%, above 70%), different system inertia levels (high inertia, medium inertia, low inertia), and different power flow levels and directions at key sections (light load, heavy load, reverse power flow, etc.). Each combination of operating conditions constitutes a "typical operating condition point." Subsequently, at each typical operating condition point, active small disturbances are applied to various resources that can participate in pre-frequency regulation (including conventional thermal / hydropower units, wind / photovoltaic units with frequency regulation capabilities, energy storage devices, load-side adjustable resources, etc.). The system frequency change is observed through dynamic simulation or linearization analysis, and the power-frequency sensitivity coefficients of various resources under that operating condition are calculated. For example, for the first... Applying class resources The active power change is observed through dynamic simulation or linearization analysis, and the corresponding system frequency change is monitored. The power-frequency sensitivity coefficient was calculated. :

[0053] Among them, superscript Indicates the first A typical operating condition point.

[0054] Finally, the sensitivity results under each typical operating condition were organized into a success rate-frequency sensitivity mapping table according to "prior features + resource type". The operating condition index is uniquely identified by prior features such as the new energy penetration rate range, inertia level, and power flow level at key sections. The resource dimension records the sensitivity coefficient of each type of adjustable resource. It can also include optional attributes such as frequency stiffness and reserve level, providing basic data for dynamic assignment during the online phase.

[0055] Next, in the online phase, prior features of the current operating condition are extracted and matched and interpolated with typical operating conditions in the power-frequency sensitivity mapping table to dynamically assign the equivalent power-frequency sensitivity of various adjustable resources under the current operating condition. Specifically, prior features are extracted from real-time and near-real-time data, including but not limited to the current renewable energy penetration rate (the proportion of renewable energy output in the entire network or region), the current equivalent inertia level of the system (which can be estimated online or obtained through offline calibration curves), the current power flow magnitude and direction at key sections, the current reserve level, frequency stiffness, etc., and this information is organized into a prior feature vector. Then, based on this vector, several similar typical operating condition points are selected in the power-frequency sensitivity mapping table, and the similarity weight is determined by the operating condition distance metric (such as the weighted sum of the normalized differences of each prior feature). Operating condition points with higher similarity are assigned greater weights; operating condition points with similarity below a preset threshold are not included in the interpolation.

[0056] Weighted interpolation is performed on the selected operating points to transform the "discrete typical operating condition sensitivity" into a "continuous sensitivity estimate of the current actual operating condition," thus obtaining the equivalent power-frequency sensitivity under the current operating condition.

[0057] in, For the first Power-frequency sensitivity coefficient at a typical operating condition point; The weights are normalized based on similarity.

[0058] The above steps realize the equivalent power-frequency sensitivity of various adjustable resources under the current operating conditions based on the dynamic assignment of prior features. Compared with static or single-condition fixed sensitivity, this mechanism can more accurately reflect the sensitivity of frequency to active power disturbances in high-proportion renewable energy power grids under different operating conditions.

[0059] Then, based on the net load forecast sequence and the current planned value, the power balance deviation at each future pre-regulation execution point is calculated. Taking a certain region as an example, the net load forecast sequence is used as a reference. Compared with the current scheduling plan or baseline operating point, through the formula Calculate the power balance deviation at the future pre-frequency modulation execution point, where, For future pre-frequency modulation execution points The predicted power generation output For future pre-frequency modulation execution points Forecasted load, To correspond to the future pre-frequency adjustment execution point The planned power or the current day plan.

[0060] Subsequently, the estimated frequency offset for each future pre-frequency modulation execution point is calculated based on the equivalent power-frequency sensitivity and power balance deviation. Assuming the current pre-frequency modulation parameters remain unchanged or are executed according to the baseline scheme, the power balance deviation of the future pre-frequency modulation execution point is mapped to the dynamically assigned sensitivity coefficient using the formula... The predicted frequency shift is obtained, where This is a frequency reference value. For the j-th type of resource, the future pre-frequency adjustment execution point The active power deviation or adjustable space, Let be the equivalent power-frequency sensitivity of resource j under the current operating conditions.

[0061] In some embodiments, the method further includes: When calculating the estimated frequency offset of each future pre-frequency adjustment execution point, the measured power flow at the current time is compared with the planned value at the current time to obtain the reference error, and the estimated frequency offset of each future pre-frequency adjustment execution point is corrected based on the reference error.

[0062] Specifically, when calculating the estimated frequency offset of each future pre-frequency modulation execution point, a comparison mechanism between the measured power flow at the current moment and the planned value at the corresponding moment is introduced. If the measured power flow deviates significantly from the planned value, the frequency assessment of the future pre-frequency modulation execution point can be corrected by superimposing the deviation trend, thus avoiding idealized results caused by relying solely on the planned value and enhancing the reliability of the assessment.

[0063] Finally, based on equivalent power-frequency sensitivity In addition to the adjustable capacity of various adjustable resources, the sensitivity index of different adjustable resources to the estimated frequency offset at each future pre-frequency adjustment execution point is calculated to obtain the resource sensitivity decomposition result. Specifically, two types of sensitivity indexes can be defined: one is frequency influence, i.e. The larger the value, the more significant the impact of unit active power adjustment on frequency; secondly, the potential adjustment contribution, i.e. ,in The available adjustment capacity of resource j is given. These indicators are calculated for each future pre-frequency tuning execution point, forming a two-dimensional sensitivity distribution of "execution point-resource". This is the complete resource sensitivity decomposition result. The resource sensitivity decomposition result can be used to determine which resources to prioritize at a specific execution point. It also provides state features for subsequent prior perception multi-objective deep Q-learning pre-frequency tuning comprehensive ranking strategies, helping the agent identify "low-cost, high-efficiency" pre-frequency tuning resource combinations at different execution points.

[0064] Through the above steps, this embodiment organically combines the power output prediction with the power-frequency sensitivity mapping table, realizing the adaptive and precise assessment of the frequency offset and resource sensitivity of the future pre-frequency regulation execution point in a high-proportion new energy power grid, providing a solid frequency layer foundation for subsequent pre-frequency regulation comprehensive sorting and pre-correction.

[0065] In step S103 of some embodiments, the scenario prior features are prior features characterizing the current operation scenario of the power grid, including the penetration rate of new energy sources, the level of system inertia, the power flow of key sections, and the type of risk scenario; the candidate pre-frequency regulation scheme can be a set of schemes that meet the basic security constraints of the power grid and include the pre-frequency regulation output trajectories of different adjustable resources at multiple future pre-frequency regulation execution points; simulation and constraint verification can be to conduct power grid frequency response simulation on the candidate pre-frequency regulation schemes and verify operational constraints such as frequency exceeding limits, unit output / ramp constraints, and line power flow constraints; the multi-objective reward function can be a quantitative reward calculation function that takes into account frequency security, frequency regulation cost, power ramp smoothness, and operational constraint violation penalties, and the weight of each objective can be dynamically adjusted according to the prior features; the deep Q network can be a deep learning network that realizes the calculation of the comprehensive utility value of the pre-frequency regulation scheme and is the main model of deep Q learning; the pre-frequency regulation comprehensive ranking strategy can be a strategy that sorts the candidate pre-frequency regulation schemes from high to low according to the comprehensive utility value, and the comprehensive utility value is output by the deep Q network, reflecting the overall merits of the schemes.

[0066] In some embodiments, the estimated frequency offset, resource sensitivity decomposition results, and scenario prior features are used as states to simulate and constrain candidate pre-frequency modulation schemes. A deep Q-network is trained through a multi-objective reward function to obtain a comprehensive pre-frequency modulation ranking strategy, including: Based on the estimated frequency offset, resource sensitivity decomposition results, adjustable resource constraints, and scenario prior features, the state vector of the agent is constructed. The action space of the agent is constructed based on a set of candidate pre-frequency power allocation schemes that satisfy basic safety constraints; Design a multi-objective reward function that includes frequency security indicators, control cost indicators, power ramp smoothness indicators, and operational constraint violation penalties. Dynamically adjust the weights of each constraint indicator in the multi-objective reward function based on scenario prior features to construct the training objectives of the agent. In the simulation environment, frequency trajectory simulation and operation constraint verification are performed on the candidate pre-frequency modulation schemes selected by the agent. The feedback value of the multi-objective reward function is calculated based on the verification results, and a deep Q network is trained to obtain a pre-frequency modulation comprehensive ranking strategy for online ranking of candidate pre-frequency modulation schemes.

[0067] Specifically, firstly, based on the estimated frequency offset, resource sensitivity decomposition results, adjustable resource constraints, and prior scenario features, a state vector of the agent is constructed. At each pre-frequency adjustment decision moment... State vector Its composition includes multi-dimensional information: one is the estimated frequency offset sequence of the future pre-frequency modulation execution point. The first is the summative indicators derived from it, such as the maximum frequency difference and frequency difference integral; the second is the equivalent power-frequency sensitivity of various adjustable resources under the current operating conditions. The information includes: 1) resource characteristics such as adjustable capacity and ramping constraints; 2) preliminary assessment characteristics of each candidate pre-frequency modulation scheme in terms of frequency, security, and cost; and 3) scenario-specific prior characteristics such as renewable energy penetration rate, system inertia level, power flow at key sections, and types of risk scenarios. This information is encoded in a fixed order to form a priori perception state. As input to the deep Q-network, it provides the agent with complete environmental information needed for decision-making.

[0068] Secondly, based on a set of candidate pre-frequency modulation power allocation schemes that satisfy basic safety constraints, the action space of the agent is constructed. (Candidate pre-frequency modulation power allocation schemes) It can be generated through empirical rules (such as a sensitivity-priority allocation scheme), preliminary solution results from heuristic algorithms, or perturbations of existing scheduling schemes. Each scheme contains the pre-frequency modulation output trajectory of different adjustable resources at multiple future pre-frequency modulation execution points. Action Space It is defined as selecting from this set of candidate solutions:

[0069] That is, each action corresponds to a candidate pre-frequency modulation scheme, and the deep Q network will analyze each action... Output the corresponding overall utility value: When running online, for Sort the candidate schemes from largest to smallest to obtain the pre-frequency tuning comprehensive ranking result, which provides a basis for subsequent scheme ranking.

[0070] Next, a multi-objective reward function is designed, which includes frequency security indicators, control cost indicators, power ramp smoothness indicators, and operational constraint violation penalties. The weights of each constraint indicator in the multi-objective reward function are dynamically adjusted based on the scenario prior features to construct the training objectives of the agent.

[0071] The multi-objective reward function can be expressed by the following formula:

[0072] in, The frequency performance indicator is obtained by weighting the maximum frequency difference, frequency difference integral, recovery time, etc. at the future pre-frequency tuning execution point. This is a frequency regulation cost indicator, covering frequency regulation power consumption, start-up and shutdown costs, energy storage lifespan loss, etc. This is an indicator of power ramp smoothness, such as the sum of squares of the rate of change of output of each resource; As a penalty for violations of operational constraints, this penalty will be significantly increased when frequency limits are exceeded, unit / energy storage output or ramping constraints are violated, standby constraints are violated, or line power flow constraints are violated.

[0073] Meanwhile, the multi-objective reward function introduces weight coefficients that are dynamically adjusted based on scene prior features. In high-risk scenarios such as high penetration of new energy sources, low system inertia, and tight power flow at critical sections, the weight of frequency safety indicators will be increased, such as... In stable scenarios, the weight of economic indicators should be appropriately increased, such as... This achieves an adaptive trade-off between frequency security, economy, and power ramp-up smoothness.

[0074] Finally, in the simulation environment, frequency trajectory simulation and operational constraint verification are performed on the candidate pre-frequency modulation schemes selected by the agent. Based on the verification results, the feedback value of the multi-objective reward function is calculated, and a deep Q-network is trained to obtain a comprehensive pre-frequency modulation ranking strategy for online ranking of candidate pre-frequency modulation schemes. This simulation environment integrates functional modules such as frequency domain noise reduction prediction, future pre-frequency modulation execution point frequency evaluation, grid frequency response, and operational constraint verification. When the agent is in a certain state... Select a certain action When a candidate pre-frequency modulation scheme is selected, the simulation environment will execute that scheme and generate the frequency trajectory for the future pre-frequency modulation execution point. And the output trajectory of each resource.

[0075] In some embodiments, operational constraint verification includes: frequency over-limit verification, unit output and ramping constraint verification, standby constraint verification, and line power flow constraint verification.

[0076] Subsequently, the frequency was checked to see if it exceeded the limit, whether the unit / energy storage output and ramping constraints were met, and whether the standby constraints and line power flow constraints were compliant. The degree of violation was then included in the constraint violation penalty item of the reward function. Furthermore, calculation Based on these results, the feedback value of the multi-objective reward function is calculated. Furthermore, by combining deep Q-learning mechanisms such as experience replay and target networks, deep Q-networks are improved. Iterative updates are performed. Through multiple rounds of simulation and constraint verification training, the agent gradually learns to select the pre-frequency modulation scheme with the best overall utility under different prior scenarios and resource constraints, and finally forms a stable pre-frequency modulation comprehensive ranking strategy. This pre-frequency modulation comprehensive ranking strategy can sort candidate pre-frequency modulation schemes from largest to smallest according to their comprehensive utility value during online operation, providing a direct basis for the selection of pre-frequency modulation schemes.

[0077] This embodiment realizes a multi-objective comprehensive evaluation of the pre-frequency tuning scheme, ensuring the optimality of the sorting strategy and the adaptability to the scenario.

[0078] In step S104 of some embodiments, the target pre-frequency regulation scheme can be the pre-frequency regulation scheme with the best comprehensive utility value selected from the candidate pre-frequency regulation schemes according to the pre-frequency regulation comprehensive ranking strategy; the pre-frequency regulation power trajectory can be the curve of the pre-frequency regulation active power output of different adjustable resource allocations changing with time at each future pre-frequency regulation execution point; the pre-frequency regulation parameter adjustment amount can be the specific correction value of the control parameters such as the pre-frequency regulation base value, primary frequency regulation participation coefficient, and adjustment level of each frequency regulation participating unit of the power grid.

[0079] The target pre-frequency modulation scheme is selected based on the pre-frequency modulation comprehensive sorting strategy, and the target pre-frequency modulation scheme is mapped to the pre-frequency modulation power trajectory and pre-frequency modulation parameter adjustment amount at the future pre-frequency modulation execution point.

[0080] Specifically, during actual online operation, as the net load forecast and the frequency assessment results of the future pre-frequency adjustment execution point are continuously updated, the system will continuously utilize the output data of steps S101 and S102 to construct a current state vector that includes the estimated frequency offset of the future pre-frequency adjustment execution point, resource sensitivity decomposition results, adjustable resource constraints, and scenario prior features. The state vector is then input into the trained deep Q-network. The deep Q-network outputs a comprehensive utility value for each candidate pre-modulation scheme. This comprehensive utility value fully reflects the overall performance of each candidate scheme across multiple objective dimensions, including frequency security, control cost, power ramp smoothness, and operational constraint satisfaction. The pre-tuning comprehensive ranking strategy is based on this comprehensive utility value. All candidate pre-frequency regulation schemes are sorted from largest to smallest, and then the scheme with the highest ranking is selected as the target pre-frequency regulation scheme. This target pre-frequency regulation scheme can be used as the main scheme for direct implementation or as a backup scheme to deal with sudden changes in power grid operating conditions, ensuring that the selected target pre-frequency regulation scheme is the comprehensive optimal solution after multi-objective trade-offs under the current power grid operating conditions.

[0081] After determining the target pre-frequency regulation scheme, this abstract scheme strategy needs to be mapped into specific execution instructions that can be directly applied to real-time power grid control, namely, the pre-frequency regulation power trajectory and pre-frequency regulation parameter adjustment amount at future pre-frequency regulation execution points. The pre-frequency regulation power trajectory is generated by extracting the pre-frequency regulation output values ​​of various adjustable resources at each future pre-frequency regulation execution point from the target pre-frequency regulation scheme, thus forming pre-frequency regulation output curves for various resources over time. This trajectory clearly defines the active power adjustment amount that each type of adjustable resource needs to provide at different future pre-frequency regulation execution points, and serves as a direct basis for pre-correcting the predicted frequency deviation at future pre-frequency regulation execution points.

[0082] Meanwhile, the generation of pre-frequency regulation parameter adjustments further transforms the aforementioned pre-frequency regulation power trajectory into specific control parameter correction values ​​for each participating frequency regulation unit. This includes adjusting the pre-frequency regulation base value of each frequency regulation resource, correcting the primary frequency regulation participation coefficient, and setting the regulation level of the generating unit or energy storage device. By adjusting these control parameters, online adaptive tuning of the pre-frequency regulation parameters of the renewable energy power grid is achieved. The entire process relies on a "prior perception multi-objective deep Q-learning pre-frequency regulation comprehensive ranking strategy." Based on frequency domain noise reduction prediction and dynamic sensitivity assessment, it completes the intelligent ranking and pre-correction of pre-frequency regulation schemes, ultimately significantly improving the foresight and robustness of frequency control in high-proportion renewable energy power grids under strong fluctuations and complex operating conditions.

[0083] In one specific embodiment, a power grid in a high-proportion inland area with new energy sources is taken as the application object. The installed capacity of new power sources (centralized wind farms and photovoltaic power stations) in this area accounts for 55% to 65%, with supporting electrochemical energy storage power stations. The number of conventional thermal power and hydropower units is limited, and they often face operating conditions such as cloud cover and rapid wind and solar power uphill. The net load exhibits random high-frequency fluctuations and medium-to-low-frequency drift characteristics, and the system inertia changes greatly, making it difficult to adapt to traditional frequency regulation methods.

[0084] Under conditions such as rapid wind and solar power uphill, cloud cover, and sudden load changes, the dispatch center initiates this method for pre-frequency regulation control. First, the dispatch center acquires recent measured data on wind power, solar power output, and regional load at fixed sampling intervals, such as 2 seconds, forming a historical fluctuation data window. Simultaneously, it obtains the initial forecast curve for the next time period from the day-ahead forecasting system, and splices them together on the time axis to form a combined sequence. Wavelet decomposition is performed on the combined sequence to obtain approximate components and several detail components at different time scales. Based on a priori feature library formed through offline statistics, combined with information such as current renewable energy penetration rate, system inertia level, key section power flow, and weather scenarios, level thresholds and level correction coefficients are dynamically assigned to each frequency band: detail coefficients identified as noise bands with small amplitudes are suppressed or set to zero, while coefficients of mid-to-low frequency characteristic bands highly correlated with frequency behavior are appropriately amplified or corrected. Subsequently, wavelet reconstruction is performed using the processed frequency band coefficients to obtain a purified time series, which is input into a short-term time series prediction model to output net load prediction sequences for multiple future pre-frequency regulation execution points.

[0085] Next, relying on the power-frequency sensitivity mapping table built in the offline phase (covering typical operating conditions with different new energy penetration rates, inertia levels, and cross-sectional power flow combinations), the dispatch center extracts the prior features of the current operating condition (equivalent inertia, cross-sectional power flow under light load). It matches three similar operating condition points in the power-frequency sensitivity mapping table and obtains the equivalent power-frequency sensitivity of conventional units, energy storage, and frequency-regulating photovoltaic under the current operating condition through similarity-weighted interpolation. Combining the net load forecast sequence with the corresponding current planned value, the power balance deviation of each pre-frequency regulation execution point is calculated. The baseline error between the current measured power flow and the planned value is added for correction, and the predicted frequency offset of each execution point in the future is estimated. At the same time, the frequency influence and potential regulation contribution of various resources are decomposed to form the "execution point - frequency offset - resource sensitivity" evaluation result.

[0086] Subsequently, the dispatch center initiated a priori perception multi-target deep Q-learning ranking strategy, encoding the estimated frequency offset, resource sensitivity, and scenario prior features into state vectors, which were then input into the trained deep Q-network. The network's action space consisted of 12 candidate pre-frequency modulation schemes generated based on empirical rules. The reward function included four key indicators: frequency safety (maximum frequency difference, recovery time), frequency modulation cost (energy storage lifetime loss, frequency modulation power consumption), power ramp smoothness, and constraint violation penalties (frequency exceeding limits, unit ramp constraints). Furthermore, the weights of frequency safety and constraint penalties were dynamically increased based on the current low-inertia, high-risk scenario. Frequency trajectory simulation and constraint verification were performed on the candidate schemes in a simulation environment, and reward values ​​were calculated and fed back to train the network. Finally, the optimal target pre-frequency modulation scheme was selected based on its comprehensive utility value.

[0087] Finally, the dispatch center maps the target pre-frequency regulation scheme into specific execution instructions: extracting the output values ​​of each adjustable resource at the future pre-frequency regulation execution point to form a pre-frequency regulation power trajectory; converting the output allocation results into the adjustment amount of the energy storage pre-frequency regulation base value, the correction value of the primary frequency regulation participation coefficient of conventional units, and the setting of the photovoltaic frequency regulation adjustment level, realizing online adaptive tuning of pre-frequency regulation parameters and correcting the predicted frequency deviation in advance. After the operating condition ends, the dispatch center updates the actual frequency response effect, frequency regulation cost, and constraint satisfaction status to the prior feature statistics library and the power-frequency sensitivity mapping table, further optimizing the accuracy and adaptability of subsequent pre-frequency regulation control.

[0088] This method achieves frequency domain noise reduction for net load forecasting by combining wavelet transform with prior features, thus improving the accuracy of forecast data. Relying on the dynamic assignment of the power-frequency sensitivity mapping table, it realizes fine-grained assessment of frequency offset at multiple future pre-frequency regulation execution points and precise decomposition of resource sensitivity. A comprehensive ranking strategy obtained by training a deep Q-network with a multi-objective reward function ensures the overall optimality of the pre-frequency regulation scheme. Finally, through scheme mapping, it achieves adaptive tuning of the pre-frequency regulation power trajectory and pre-frequency regulation parameters, realizing the forward-looking identification and proactive pre-adjustment of grid frequency risks. This effectively avoids problems such as frequency deviation amplification and frequent frequency regulation actions, improving the forward-looking nature, accuracy, and robustness of grid pre-frequency regulation control. Simultaneously, it optimizes the utilization efficiency of adjustable resources and can well adapt to the complex and variable operating conditions of high-proportion renewable energy grids.

[0089] Please see Figure 2 This application also provides a power grid adaptive pre-frequency regulation device, which can implement the above-mentioned power grid adaptive pre-frequency regulation method. The device includes: Prediction module 201 is used to perform wavelet transform on historical fluctuation data of net load and initial prediction data for the next period, suppress noise frequency bands based on prior features and correct main feature frequency bands, and obtain net load prediction sequences for multiple future pre-frequency adjustment execution points after wavelet reconstruction. The calculation module 202 is used to dynamically assign values ​​based on the power-frequency sensitivity mapping table and the prior characteristics of the current operating conditions to obtain the equivalent power-frequency sensitivity. Combined with the net load prediction sequence, it is used to calculate the estimated frequency offset and resource sensitivity decomposition results for each future pre-frequency adjustment execution point. Training module 203 is used to simulate and constrain candidate pre-frequency modulation schemes by taking the estimated frequency offset, resource sensitivity decomposition results and scenario prior features as states, and to train a deep Q network through a multi-objective reward function to obtain a comprehensive ranking strategy for pre-frequency modulation. The mapping module 204 is used to select the target pre-frequency modulation scheme according to the pre-frequency modulation comprehensive sorting strategy, and map the target pre-frequency modulation scheme to the pre-frequency modulation power trajectory and pre-frequency modulation parameter adjustment amount of the future pre-frequency modulation execution point.

[0090] The specific implementation of the power grid adaptive pre-frequency regulation device is basically the same as the specific implementation of the power grid adaptive pre-frequency regulation method described above, and will not be repeated here.

[0091] Thirdly, embodiments of this application provide an electronic device, see [link to relevant documentation]. Figure 3 The diagram shown is a structural schematic of an electronic device provided in this application.

[0092] like Figure 3 As shown, the device includes: Memory 31 is used to store computer programs; Processor 32 is used to execute computer programs; When the processor 32 executes the computer program, it implements the power grid adaptive pre-frequency regulation method as described in any of the above embodiments.

[0093] For example, a computer program may be divided into one or more modules / units, one or more of which are stored in memory 31 and executed by processor 32 to complete this application. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in an electronic device.

[0094] The processor 32 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0095] The memory 31 can be used to store computer programs and / or modules. The processor 32 implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory 31 and calling the data stored in the memory 31. The memory 31 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 31 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0096] It should be noted that the aforementioned electronic devices include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 3The structural diagram is merely an example of the electronic device described above and does not constitute a limitation on the electronic device. It may include more components than shown in the diagram, or combine certain components, or use different components.

[0097] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed, implements the power grid adaptive pre-frequency regulation method of any of the above embodiments.

[0098] It should be understood that the implementation of all or part of the above-described adaptive pre-frequency regulation method for power grids can also be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described adaptive pre-frequency regulation method for power grids. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium can be appropriately added to or subtracted according to the requirements of legislation and patent practice in the relevant jurisdiction. For example, in some relevant jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0099] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0100] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A power grid adaptive pre-frequency regulation method, characterized in that, include: Wavelet transform is performed on the historical fluctuation data of net load and the initial prediction data of the next period. Noise frequency bands are suppressed and main characteristic frequency bands are corrected based on prior features. After wavelet reconstruction, net load prediction sequences for multiple future pre-frequency adjustment execution points are obtained. Based on the power-frequency sensitivity mapping table and the prior characteristics of the current operating conditions, dynamic values ​​are assigned to obtain the equivalent power-frequency sensitivity. Combined with the net load prediction sequence, the estimated frequency offset and resource sensitivity decomposition results for each future pre-frequency adjustment execution point are obtained. Using the estimated frequency offset, resource sensitivity decomposition results, and scenario prior features as states, the candidate pre-frequency modulation schemes are simulated and constrained. A deep Q-network is trained through a multi-objective reward function to obtain a comprehensive pre-frequency modulation ranking strategy. The target pre-frequency modulation scheme is selected according to the pre-frequency modulation comprehensive sorting strategy, and the target pre-frequency modulation scheme is mapped to the pre-frequency modulation power trajectory and pre-frequency modulation parameter adjustment amount of the future pre-frequency modulation execution point.

2. The method according to claim 1, characterized in that, The process involves performing wavelet transform on historical fluctuation data of net load and initial prediction data for the next time period. Noise frequency bands are suppressed based on prior features, and the main characteristic frequency bands are corrected. After wavelet reconstruction, multiple net load prediction sequences for future pre-tuning execution points are obtained, including: The historical fluctuation data of net load is combined with the initial forecast data for the next period in chronological order to form a combined sequence; Wavelet transform decomposition is performed on the combined sequence to obtain multiple frequency band detail components; Based on the prior characteristics of the multiple frequency band detail components, noise frequency bands are distinguished from main characteristic frequency bands; Noise in the noise band is suppressed based on a level threshold, and the amplitude of the main characteristic frequency band is corrected based on a level correction coefficient to obtain the processed frequency band coefficient. Wavelet reconstruction is performed on the processed frequency band coefficients to obtain the purified time series, and the purified time series and the combined sequence have the same length. Based on the purification time series, time-series prediction is performed to obtain net load prediction sequences for multiple future pre-frequency adjustment execution points.

3. The method according to claim 2, characterized in that, The prior features include at least one of the following: new energy penetration rate, system inertia level, tidal current tension at key sections, and typical weather scenarios.

4. The method according to claim 1, characterized in that, The equivalent power-frequency sensitivity is obtained by dynamically assigning values ​​based on the power-frequency sensitivity mapping table and prior characteristics of the current operating conditions. This is then combined with the net load prediction sequence to calculate the estimated frequency offset and resource sensitivity decomposition results for each future pre-frequency adjustment execution point, including: In the offline phase, small disturbance sensitivity analysis is performed on various adjustable resources under multiple typical operating conditions, and a power-frequency sensitivity mapping table indexed by prior operating condition characteristics is established. During the online phase, prior features of the current operating condition are extracted and matched and interpolated with typical operating conditions in the power-frequency sensitivity mapping table to dynamically assign the equivalent power-frequency sensitivity of various adjustable resources under the current operating condition. Based on the net load forecast sequence and the current planned value, calculate the power balance deviation at each future pre-frequency adjustment execution point; The estimated frequency offset for each future pre-frequency modulation execution point is calculated based on the equivalent power-frequency sensitivity and the power balance deviation. Based on the equivalent power-frequency sensitivity and the adjustable capacity of various adjustable resources, the sensitivity index of different adjustable resources to the estimated frequency offset at each future pre-frequency adjustment execution point is calculated, and the resource sensitivity decomposition result is obtained.

5. The method according to claim 4, characterized in that, The method further includes: When calculating the estimated frequency offset of each future pre-frequency adjustment execution point, the measured power flow at the current time is compared with the planned value at the current time to obtain the reference error, and the estimated frequency offset of each future pre-frequency adjustment execution point is corrected based on the reference error.

6. The method according to claim 1, characterized in that, The process involves using the estimated frequency shift, resource sensitivity decomposition results, and scenario prior features as states to simulate and constrain candidate pre-frequency modulation schemes. A deep Q-network is trained using a multi-objective reward function to obtain a comprehensive pre-frequency modulation ranking strategy, including: Based on the estimated frequency offset, resource sensitivity decomposition results, adjustable resource constraints, and scenario prior features, the state vector of the agent is constructed. The action space of the agent is constructed based on a set of candidate pre-frequency power allocation schemes that satisfy basic safety constraints; Design a multi-objective reward function that includes frequency security indicators, control cost indicators, power ramp smoothness indicators, and operational constraint violation penalties. Dynamically adjust the weights of each constraint indicator in the multi-objective reward function based on scenario prior features to construct the training objective of the agent. In the simulation environment, frequency trajectory simulation and operational constraint verification are performed on the candidate pre-frequency modulation schemes selected by the agent. The feedback value of the multi-objective reward function is calculated based on the verification results, and a deep Q-network is trained to obtain a pre-frequency modulation comprehensive ranking strategy for online ranking of the candidate pre-frequency modulation schemes.

7. The method according to claim 6, characterized in that, The operational constraint verification includes: frequency over-limit verification, unit output and ramping constraint verification, standby constraint verification, and line power flow constraint verification.

8. A power grid adaptive pre-frequency regulation device, characterized in that, include: The prediction module is used to perform wavelet transform on the historical fluctuation data of net load and the initial prediction data of the next period. Based on the prior features, the noise frequency band is suppressed and the main feature frequency band is corrected. After wavelet reconstruction, the net load prediction sequence of multiple future pre-frequency adjustment execution points is obtained. The calculation module is used to dynamically assign values ​​based on the power-frequency sensitivity mapping table and the prior characteristics of the current operating conditions to obtain the equivalent power-frequency sensitivity. Combined with the net load prediction sequence, the module calculates the estimated frequency offset and resource sensitivity decomposition results for each future pre-frequency adjustment execution point. The training module is used to simulate and constrain the candidate pre-frequency modulation schemes by taking the estimated frequency offset, resource sensitivity decomposition results and scenario prior features as states, and to train a deep Q network through a multi-objective reward function to obtain a comprehensive pre-frequency modulation ranking strategy. The mapping module is used to select a target pre-frequency modulation scheme according to the pre-frequency modulation comprehensive sorting strategy, and map the target pre-frequency modulation scheme to the pre-frequency modulation power trajectory and pre-frequency modulation parameter adjustment amount of the future pre-frequency modulation execution point.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the power grid adaptive pre-frequency regulation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the power grid adaptive pre-frequency regulation method as described in any one of claims 1 to 7.

Images