A coal power unit coupled multi-element energy storage active power balance control method

Abstract

The present application relates to the technical field of power system regulation, and specifically discloses a kind of coal power unit coupling multi-element energy storage's electric power active balance control method, for solving the problem that the existing static active safety margin evaluation method cannot accurately depict multi-stage frequency response, lacks online adaptive power distribution and continuous smooth scheduling under complex working conditions such as high wind power penetration, large disturbance, system inertia decline, leading to large frequency drop, slow recovery, output mutation, comprising building collaborative characteristic model, online calculation dominant feature, segmented optimization scheduling, closed-loop parameter updating;The present application realizes the rapid response, smooth connection and accurate optimization of frequency support and economic dispatch under high wind power penetration scenario by building hierarchical inertia and multi-stage frequency modulation collaborative model, using multi-scale online feature extraction and adaptive power distribution, segmented gradient optimization scheduling and closed-loop online parameter updating.

Description

Technical Field

[0001] This invention relates to the field of power system control technology, and in particular to a power active power balance control method for coal-fired power units coupled with multi-element energy storage. Background Technology

[0002] Chinese invention patent application CN119340999A discloses a method for assessing the active power safety margin of a power system considering the uncertainty of renewable energy output. It constructs a static active power safety domain model based on active power balance, line phase angle difference, and unit output constraints, solves for the safety domain boundary using the maximum volume optimization method, and assesses the static active power safety margin under risk and safety conditions using interval mathematics and the Lagrange multiplier method, respectively. However, this method primarily focuses on static margin quantification and fails to reflect the multi-stage frequency response characteristics of the system under large disturbances or net load fluctuations. It lacks characterization of inertial response, primary / secondary / tertiary frequency regulation, and ramp-up capability. Furthermore, the model parameters are based on fixed operating conditions and cannot be adaptively adjusted, making the assessment results difficult to adapt to real-world scenarios with rapidly changing wind and solar power output. Moreover, it fails to closely integrate the assessment results with the coordinated control and economic dispatch of units and energy storage, thus failing to guide coal-fired power units coupled with multi-element energy storage systems to achieve rapid active power balance and frequency stability under dynamic conditions. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a power active power balance control method for coal-fired power units coupled with multi-element energy storage. By constructing a hierarchical inertial and multi-stage frequency regulation collaborative model, adopting multi-scale online feature extraction and adaptive power allocation, segmented optimized scheduling and closed-loop online parameter updates, it can achieve rapid response, smooth connection and precise optimization of frequency support and economic dispatch in high wind power penetration scenarios.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A power active power balance control method for coal-fired power units coupled with multi-element energy storage includes the following steps:

[0006] Step 1: Construct a collaborative characteristic model: Based on the power dynamic characteristics of coal-fired power units and energy storage units, establish an equivalent mathematical model with graded inertia and multi-stage frequency regulation capabilities.

[0007] Step 2, Online calculation of dominant characteristics: Based on the power system frequency deviation and the real-time output measurement values ​​of coal-fired power units and energy storage units, extract the dominant dynamic characteristics and calculate the power distribution coefficient of each coordinating unit;

[0008] Step 3, segmented optimization scheduling: Construct a segmented optimization scheduling model that takes into account both power constraints and economic operation objectives, and solve the joint output scheme of coal-fired power units and energy storage units;

[0009] Step 4, Closed-loop parameter update: Based on real-time load fluctuations and frequency deviations, the equivalent inertia coefficient, frequency modulation slope, response time parameters, and power distribution coefficient in the equivalent mathematical model of Step 1 are adjusted online in a closed loop.

[0010] In step three, the segmented optimization scheduling model constructs a sub-model for each scheduling period, which includes a cost function and energy storage charging and discharging constraints. Within this sub-model, the output allocation of coal-fired power units and energy storage units is calculated iteratively through gradient descent. Based on the convergence criterion, the joint output scheme for the scheduling period is output, and the gradient step size is dynamically adjusted. Each scheduling period is executed sequentially and connected.

[0011] It should be noted that, firstly, this invention constructs an equivalent mathematical model with hierarchical inertia and multi-stage frequency regulation capabilities to achieve inertial response of generating units and energy storage devices; secondly, it constructs a segmented optimization scheduling model for each scheduling period, incorporating economic costs and energy storage charging and discharging constraints, and solves the joint output scheme through dynamic step size adjustment and accelerated gradient iteration; finally, based on recursive least squares and sliding window algorithms, it performs real-time closed-loop identification and updating of the equivalent inertia coefficient, frequency regulation slope, and allocation coefficient in the model to maintain consistency between parameters and operating conditions. This invention overcomes the shortcomings of existing technologies, which are limited to static active power safety domain evaluation and cannot characterize the multi-stage frequency response of the system.

[0012] As a further aspect of the present invention, in step three, the segmented optimization scheduling model uses the combined output result of the previous scheduling period as the initial value for hot start, and sets a sliding overlapping time window in the sub-model; the objective function of the sub-model introduces an economic cost function and a frequency deviation penalty term in parallel, and adaptively adjusts the weights of the economic cost function and the frequency deviation according to the real-time frequency difference ratio; an increasing / decreasing tightening strategy is adopted for the upper and lower limits of the charging and discharging of the energy storage unit, and the boundary change rate is set according to the current state of charge; and a smooth transition correction is performed between the combined output schemes of adjacent scheduling periods.

[0013] It should be noted that the present invention introduces the following technical features in step three: using the combined output result of the previous time period as the initial value for hot start, and setting a sliding overlap time window to ensure the continuity of the scheme; incorporating economic cost and frequency deviation penalty terms in parallel into the objective function of the sub-model, and dynamically adjusting the weights of the two according to the real-time frequency deviation ratio; adopting an increasing / decreasing tightening strategy based on the current state of charge for the upper and lower limits of the energy storage unit to match the actual constraints; and performing smooth transition correction between outputs of adjacent scheduling time periods to avoid sudden changes in output, thereby realizing the dynamic connection and optimized scheduling of combined output in multiple time periods, overcoming the shortcomings of the existing technology that only quantifies active power margin based on the static safety domain boundary and fails to consider the continuity and dynamic coupling between scheduling time periods.

[0014] As a further aspect of the present invention, when wind power installed capacity accounts for 50% to 70% of the total power generation capacity, energy storage unit installed capacity accounts for 10% to 20% of the total power generation capacity, and wind power output fluctuation rate exceeds ±15% / min, the sliding overlap time window length of the segmented optimization scheduling model is 60 seconds, the initial value for hot start is the average value of joint output in the last 10 seconds of the previous period, the frequency deviation penalty weight is calculated based on the real-time maximum rate of decline of 0.2Hz / s, and is doubled when the rate of decline exceeds 0.2Hz / s, the tightening rate of the upper and lower limits of energy storage unit charging and discharging is 10% of the current state of charge, and the smooth transition correction of the joint output scheme in adjacent scheduling periods adopts three-point linear interpolation.

[0015] It should be noted that when wind power accounts for more than 50%, the inertia of traditional synchronous machines is largely replaced, and the initial frequency drop of the system is faster. Although 10% to 20% of energy storage capacity can provide frequency regulation support, its capacity and power release rate need to be finely constrained, otherwise it will be difficult to cope with short-term drastic fluctuations exceeding ±15% / min. In high-speed fluctuation environments, it is necessary to quickly suppress frequency drops while avoiding sudden changes in dispatching instructions that could affect equipment safety and market economy. In high-fluctuation operating conditions, dispatching models switch frequently, and without overlapping windows and smooth transitions, output jumps or "frame drops" can easily occur. A 60-second sliding overlapping time window eliminates transient jumps during the switching of different sub-models, ensuring continuous output under high fluctuations. Using the average of the last 10 seconds of the previous period as the initial value enables a rapid response to the current operating conditions without introducing excessive initial errors. When the frequency drop rate exceeds 0.2Hz / s, the penalty weight is doubled to enhance the frequency regulation response during severe fluctuations. A 10% tightening rate ensures that the energy storage constraint dynamically matches the SOC change, avoiding out-of-bounds operation. Three-point linear interpolation enables seamless connection of output schemes between adjacent time periods.

[0016] As a further aspect of the present invention, in step one, the collaborative characteristic model forms an equivalent mathematical model through the following interconnected sub-modules:

[0017] Cascaded inertial submodule: It consists of three to five inertial links connected in series, each with independently adjustable inertia and damping, to simulate the multi-level inertial response of coal-fired power units and energy storage units;

[0018] Parallel frequency modulation submodule: Three gain channels are arranged in parallel after each inertial element to handle primary, secondary and tertiary frequency modulation at different time scales, and each channel is configured with response amplitude and time delay separately;

[0019] Derivative feedback submodule: A frequency change rate feedback path is set between the inertial cascade chain and the parallel frequency modulation channel to separate the inertial support from the frequency modulation action;

[0020] Energy storage constraint mapping submodule: The upper and lower limits of the state of charge and the maximum charging and discharging power of each energy storage unit are used as model input constraints and mapped to the corresponding state variables;

[0021] Parameter identification submodule: Based on the recursive least squares optimal estimation algorithm, frequency deviation, power response and state of charge data are collected periodically, and the parameters of each inertial element and frequency modulation channel are updated online.

[0022] It should be noted that in step one of this invention, the cooperative characteristic model is refined into five interconnected sub-modules: the cascaded inertia sub-module simulates the multi-level inertial support of the generating unit and energy storage through three to five series inertial links; the parallel frequency modulation sub-module arranges primary, secondary, and tertiary gain channels after each inertial link to cover frequency modulation actions at different time scales; the derivative feedback sub-module sets up a frequency change rate path to decouple the inertial and frequency modulation functions; the energy storage constraint mapping sub-module maps the upper and lower limits of the state of charge and the charging and discharging power as model input constraints; and the parameter identification sub-module updates the parameters of each sub-module online based on the recursive least squares algorithm. This modular structure can accurately characterize the multi-stage frequency support process and keep the model parameters synchronized with the actual operating conditions, making up for the shortcomings of existing static evaluation methods in terms of dynamic response and parameter adaptation.

[0023] As a further aspect of the present invention, step two, which involves extracting the dominant dynamic features and calculating the power allocation coefficients of each cooperative unit, includes:

[0024] Step 21: Perform multi-scale decomposition processing on the power system frequency and output signal to obtain the dominant response components representing inertial support, primary frequency regulation and secondary frequency regulation respectively.

[0025] Step 22: Identify the corresponding modal energy weights based on the amplitude time series and energy spectrum characteristics of each dominant response component;

[0026] Step 23: The modal energy weights are fused with the real-time adjustable power capacity of the coal-fired power unit and the energy storage unit according to the adaptive mapping rule to generate preliminary power allocation coefficients;

[0027] Step 24: Dynamically normalize the sub-allocation coefficients and perform boundary corrections based on the upper and lower limits of the remaining capacity of each unit to output the final power allocation coefficients.

[0028] As a further aspect of the present invention, step 21 employs continuous wavelet transform to perform multi-scale decomposition of frequency deviation and output signal, using Morlet wavelet as the mother wavelet, and setting three scales of 0.5 seconds, 2 seconds, and 8 seconds corresponding to different time constants; step 22 applies Hilbert transform to each scale component to obtain instantaneous amplitude envelopes for subsequent energy analysis; step 23 calculates spectral entropy based on the amplitude envelopes of each scale to characterize the energy distribution weights at different adjustment stages; step 24 utilizes Mamdani-type fuzzy inference to fuse the spectral entropy weights of each scale with the real-time adjustable capacity of the unit and energy storage unit to generate preliminary allocation coefficients, applies a five-second window Savitzky-Golay filter to the preliminary allocation coefficients, and outputs them as the final power allocation coefficients after smoothing.

[0029] It should be noted that in step two of this invention, the following steps are proposed: First, Morlet wavelet is used to perform multi-scale decomposition of the frequency deviation and output signal at three scales: 0.5s, 2s, and 8s, to obtain the dominant response components of each stage. Then, Hilbert transform is used to extract the instantaneous amplitude envelope and calculate the spectral entropy to characterize the energy distribution weights under different time constants. Next, Mamdani-type fuzzy inference is used to fuse the aforementioned spectral entropy weights with the real-time adjustable capacity of the generating unit and energy storage unit according to an adaptive mapping rule, generating preliminary power allocation coefficients. Finally, a five-second window Savitzky-Golay filter is applied to the preliminary coefficients for smoothing, outputting the final power allocation coefficients, thereby achieving real-time quantification and dynamic allocation of the multi-stage frequency support process. This invention overcomes the shortcomings of existing technologies, which only assess the steady-state active power margin of the system through static safety domain boundaries but fail to quantify the staged response characteristics of wind power and energy storage during frequency disturbances, neglect the time-varying influence of power allocation at each stage of inertial support, primary frequency regulation, and secondary frequency regulation, and cannot achieve real-time adaptive allocation for random load mutations.

[0030] As a further aspect of the present invention, in the segmented optimization scheduling model of step three, an accelerated gradient algorithm is used for the gradient descent iteration process of each sub-model, specifically including:

[0031] In the first iteration, only the current gradient information is used as the update direction;

[0032] Starting from the second iteration, the gradient of this iteration is linearly superimposed with the direction of the previous iteration at a ratio of 0.8, and used as the new search direction;

[0033] Before each iteration, the momentum coefficient is fine-tuned within a range of ±0.05 based on the difference between the current objective function value of the sub-model and the objective function value of the previous iteration.

[0034] When the improvement of the objective function is less than 0.001% for three consecutive iterations or the number of iterations exceeds 50, the iteration of this sub-model is terminated and the result is output.

[0035] It should be noted that this invention introduces an accelerated gradient algorithm in the piecewise optimization scheduling model in step three: the first iteration uses the current gradient, and subsequent iterations linearly superimpose the current gradient with the previous update direction by 0.8. Before each iteration, the momentum coefficient is finely adjusted by ±0.05 based on the difference between the current and previous objective function values ​​to smooth the search path and accelerate convergence. Simultaneously, a termination criterion is set for three consecutive improvements below 0.001% or more than 50 iterations to ensure that each sub-model terminates efficiently while meeting accuracy requirements. This invention overcomes the shortcomings of existing technologies where piecewise optimization only uses conventional gradient descent, failing to balance iterative convergence speed and directional stability, and prone to slow convergence or frequent oscillations in highly volatile scenarios.

[0036] As a further aspect of the present invention, an economic cost term and a frequency deviation penalty term are set in parallel in the objective function of the sub-model in step three, and the weight of the frequency deviation penalty is dynamically adjusted in the following manner:

[0037] The maximum frequency drop of the system in the last 30 seconds is sampled and averaged at a period of 5 seconds to generate the real-time maximum drop rate value.

[0038] Multiply the rate of decline value by a preset scaling factor as the frequency penalty weight;

[0039] During the same period, the weight of the economic cost item was set at 1.2 times the proportion of wind power installed capacity;

[0040] The frequency penalty weight is reset every 40 seconds.

[0041] As a further aspect of the present invention, in the segmented optimization scheduling model in step three, when using the accelerated gradient algorithm for the gradient descent iteration process of each sub-model, the initial value of the gradient step size is set to 0.1. After each iteration, the step size is adaptively decayed based on the local second derivative / inverse of the gradient norm of the objective function at the current iteration point. When the iteration step size decreases to below 0.01 or the improvement of the objective function value in continuous iterations is below 0.0001, the convergence condition is reached, and the iteration process is terminated immediately.

[0042] It should be noted that, in the objective function of the sub-model in step three of this invention, an economic cost term and a frequency deviation penalty term are set in parallel. A moving average of the maximum rate drop within the last 30 seconds is calculated using a 5-second period, and this real-time rate drop value is multiplied by a preset proportional coefficient to generate a dynamic frequency penalty weight. Simultaneously, the weight of the economic cost term is set to 1.2 times the wind power installed capacity ratio, and is reset every 40 seconds. Furthermore, in the piecewise gradient optimization iteration, the initial step size is set to 0.1, and after each iteration, it adaptively decays based on the local second derivative or the inverse of the gradient norm of the current objective function. The process terminates immediately when the step size falls below 0.01 or the continuous iteration improvement is less than 0.0001, ensuring that the optimization process is both efficient and stable in highly volatile scenarios. This invention overcomes the shortcomings of existing technologies that only quantify active power margin within a static evaluation framework, fail to tightly couple economic dispatch with the frequency secondary response, and lack an adaptive acceleration mechanism for the gradient optimization process.

[0043] As a further aspect of the present invention, the closed-loop parameter update process in step four includes:

[0044] The real-time frequency deviation and power response data are sampled in a sliding window of 10 seconds.

[0045] The equivalent inertia coefficient, first / second / third-order frequency modulation slope, and power allocation coefficient in the equivalent mathematical model are identified online using a recursive least squares algorithm with a forgetting factor of 0.98.

[0046] An update is triggered every 5 seconds, and the parameters obtained in this identification are exponentially weighted and fused with the results of the previous update at a ratio of 70%:30%.

[0047] The parameters obtained after fusion are limited to within ±10% of the previous updated value;

[0048] The above fusion results are then subjected to a first-order exponential smoothing filter with a time constant of 0.5 seconds, and then loaded into the equivalent mathematical model for the next control cycle.

[0049] It should be noted that the present invention proposes a closed-loop parameter update process in step four: Frequency deviation and power response data are sampled using a 10-second sliding window; a recursive least squares algorithm with a forgetting factor of 0.98 is used to identify the equivalent inertia coefficient, frequency modulation slopes at each level, and power allocation coefficients online; an update is triggered every 5 seconds, and the results are exponentially weighted and fused with the previous results at a 70%:30% ratio, limiting the fused parameters to within ±10%; finally, a first-order exponential smoothing filter with a 0.5-second time constant is applied, and the data is reloaded into the equivalent mathematical model to ensure that the model parameters are synchronized with actual operating conditions. This invention overcomes the shortcomings of existing technologies that use offline fixed parameters for safety margin assessment, which cannot adapt to the dynamic characteristic changes of coal-fired power units and energy storage units caused by load and wind power output variations in high-fluctuation, high-permeability scenarios.

[0050] The technical advantages of the power active power balance control method for coal-fired power units coupled with multi-element energy storage in this invention are as follows:

[0051] This invention constructs a cooperative characteristic equivalent model with hierarchical inertia and multi-stage frequency regulation capabilities. It employs multi-scale dynamic feature extraction and adaptive power allocation to achieve rapid response at each stage of inertial support and primary / secondary / tertiary frequency regulation. Through segmented optimized scheduling and dynamic adjustment of frequency / economic dual weights, it outputs a continuous and smooth joint power output scheme. Online recursive least squares identification and closed-loop parameter updates are used to correct model parameters in real time to match operating conditions. Furthermore, by utilizing sliding overlapping windows and multi-point interpolation transitions, it eliminates output abrupt changes during sub-model switching. This results in a significant reduction in power system frequency drops and a substantial shortening of recovery time in scenarios with high wind power penetration and high volatility, while simultaneously considering operational economy and energy storage safety. This achieves multiple improvements in frequency stability, economic dispatch efficiency, and output continuity. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method proposed in this invention;

[0053] Figure 2 This is a diagram illustrating the process of extracting the dominant dynamic features and calculating the power allocation coefficients of each cooperative unit in step two of this invention.

[0054] Figure 3 A diagram showing the control interface for applying the method proposed in this invention. Detailed Implementation

[0055] The technical solutions of 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 within the scope of protection of the present invention.

[0056] Example 1. As... Figure 1 As shown, the present invention proposes a power active power balance control method for coal-fired power units coupled with multi-element energy storage, comprising the following steps:

[0057] Step 1: Construct a collaborative characteristic model: Based on the power dynamic characteristics of coal-fired power units and energy storage units, establish an equivalent mathematical model with graded inertia and multi-stage frequency regulation capabilities.

[0058] Step 2, Online calculation of dominant characteristics: Based on the power system frequency deviation and the real-time output measurement values ​​of coal-fired power units and energy storage units, extract the dominant dynamic characteristics and calculate the power distribution coefficient of each coordinating unit;

[0059] Step 3, segmented optimization scheduling: Construct a segmented optimization scheduling model that takes into account both power constraints and economic operation objectives, and solve the joint output scheme of coal-fired power units and energy storage units;

[0060] Step 4, Closed-loop parameter update: Based on real-time load fluctuations and frequency deviations, the equivalent inertia coefficient, frequency modulation slope, response time parameters, and power distribution coefficient in the equivalent mathematical model of Step 1 are adjusted online in a closed loop.

[0061] In step three, the segmented optimization scheduling model constructs a sub-model for each scheduling period, which includes a cost function and energy storage charging and discharging constraints. Within this sub-model, the output allocation of coal-fired power units and energy storage units is calculated iteratively through gradient descent. Based on the convergence criterion, the joint output scheme for the scheduling period is output, and the gradient step size is dynamically adjusted. Each scheduling period is executed sequentially and connected.

[0062] It should be specifically noted that in step three, the segmented optimization scheduling model uses the combined output result of the previous time period of each scheduling period as the initial value for hot start, and sets a sliding overlapping time window in the sub-model; the objective function of the sub-model introduces the economic cost function and the frequency deviation penalty term in parallel, and adaptively adjusts the weights of the economic cost function and the frequency deviation according to the real-time frequency difference ratio; the upper and lower limits of the charging and discharging of the energy storage unit adopt an increasing / decreasing tightening strategy, and sets the boundary change rate according to the current state of charge; and performs smooth transition correction between the combined output schemes of adjacent scheduling periods.

[0063] In addition, when wind power capacity accounts for 50% to 70% of the total power generation capacity, energy storage capacity accounts for 10% to 20% of the total power generation capacity, and wind power output fluctuation exceeds ±15% / min, the sliding overlap time window length of the segmented optimization scheduling model is 60 seconds, the initial value for hot start is the average value of the combined output in the last 10 seconds of the previous period, the frequency deviation penalty weight is calculated based on the real-time maximum rate of decline of 0.2Hz / s, and is doubled when the rate of decline exceeds 0.2Hz / s, the tightening rate of the upper and lower limits of energy storage unit charging and discharging is 10% of the current state of charge, and the smooth transition correction of the combined output scheme in adjacent scheduling periods adopts three-point linear interpolation.

[0064] It should be noted that in the piecewise optimization scheduling model in step three, the gradient descent iteration process of each sub-model uses an accelerated gradient algorithm, specifically including:

[0065] In the first iteration, only the current gradient information is used as the update direction;

[0066] Starting from the second iteration, the gradient of this iteration is linearly superimposed with the direction of the previous iteration at a ratio of 0.8, and used as the new search direction;

[0067] Before each iteration, the momentum coefficient is fine-tuned within a range of ±0.05 based on the difference between the current objective function value of the sub-model and the objective function value of the previous iteration.

[0068] When the improvement of the objective function is less than 0.001% for three consecutive iterations or the number of iterations exceeds 50, the iteration of this sub-model is terminated and the result is output.

[0069] It should be noted that the economic cost term and the frequency deviation penalty term are set in parallel in the sub-model objective function in step three, and the weight of the frequency deviation penalty is dynamically adjusted as follows:

[0070] The maximum frequency drop of the system in the last 30 seconds is sampled and averaged at a period of 5 seconds to generate the real-time maximum drop rate value.

[0071] Multiply the rate of decline value by a preset scaling factor as the frequency penalty weight;

[0072] During the same period, the weight of the economic cost item was set at 1.2 times the proportion of wind power installed capacity;

[0073] The frequency penalty weight is reset every 40 seconds.

[0074] It should be noted that in the segmented optimization scheduling model in step three, when using the accelerated gradient algorithm for the gradient descent iteration process of each sub-model, the initial value of the gradient step size is set to 0.1. After each iteration, the step size is adaptively decayed based on the local second derivative of the objective function at the current iteration point / the inverse of the gradient norm. When the iteration step size decreases to below 0.001 or the improvement of the objective function value in continuous iterations is below 0.0001, the convergence condition is reached, and the iteration process is terminated immediately.

[0075] To clearly describe the implementation scheme of the above technical solution, the following embodiments illustrate its technical effects compared with the prior art.

[0076] In a 1000MW power generation system, comprising 250MW of coal-fired power units, 600MW of wind power units, and 150MW of energy storage units, the power system load experiences a step increase of 80MW within 5 seconds after 10 seconds. MATLAB simulations are used to analyze two proposed solutions. Solution 1 employs the baseline approach, a traditional segmented optimization scheduling method with no hot start, no sliding window, a fixed penalty weight of 0.3, no tightening strategy, and no interpolation. Solution 2 utilizes the proposed solution, employing hot start, a 60-second sliding overlapping window, doubled real-time rate reduction threshold, and a 10% tightening rate. Three-point linear interpolation is used. In step three, the initial value of the instruction iteration for each scheduling period during hot start is the average value of the joint output in the last 10 seconds of the previous period. For the sliding overlapping window, the sub-model time window length is 60 seconds, and the window overlap is 30 seconds to maintain the continuity of the scheme. For the dynamic penalty weight, the maximum rate of decrease in the last 30 seconds is calculated with a 5-second cycle. When the rate of decrease is >0.2Hz / s, the weight value is doubled. For the tightening strategy, when the energy storage unit's state of charge is 50%, the upper and lower limits of charging and discharging are tightened by 10%. For smooth interpolation, three points are taken at the beginning and end of each of the three adjacent periods, and linear interpolation is used for seamless transition. The simulation results of the two technical schemes are compared, as shown in Table 1.

[0077] Table 1. Comparison of Simulation Results for Scheme 1 and Scheme 2

[0078]

[0079] As can be seen from the table, Scheme 2 is significantly better than the baseline in terms of frequency drop, RoCoF, and recovery speed. Output mutation and SoC lower limit are improved, and the economic cost increases only slightly. The above embodiments fully verify the synergistic effect of the various technical features in step three. Using the average output value of the last 10 seconds of the previous period as the initial value of the next period's iteration significantly reduces the optimization search space, makes the gradient iteration converge faster, and reduces the initial power jump. The 60-second long and 30-second overlapping time window design ensures the output continuity between adjacent sub-models and avoids power command gaps or mutations caused by model switching. The penalty coefficient is dynamically adjusted based on the real-time maximum drop rate. When the drop rate exceeds 0.2Hz / s, the frequency compensation intensity is automatically strengthened, which can improve the frequency regulation response intensity in a timely manner under large disturbance conditions. The 10% state of charge is tightened on the upper and lower limits of energy storage charging and discharging to ensure that the energy storage device does not run out of bounds during rapid fluctuations, while reserving a certain margin for subsequent frequency support. Seamless connection is achieved between output schemes in adjacent scheduling periods through three-point linear interpolation, which effectively suppresses command mutations at the time boundary and improves the smoothness of the overall output curve.

[0080] Example 2. The difference between Example 2 and Example 1 is that this example describes steps one, two, and four of a power active power balance control method for a coal-fired power unit coupled with multi-element energy storage.

[0081] In step one of the technical solution of this invention, the collaborative characteristic model forms an equivalent mathematical model through the following interconnected sub-modules:

[0082] Cascaded inertial submodule: It consists of three to five inertial links connected in series, each with independently adjustable inertia and damping, to simulate the multi-level inertial response of coal-fired power units and energy storage units;

[0083] Parallel frequency modulation submodule: Three gain channels are arranged in parallel after each inertial element to handle primary, secondary and tertiary frequency modulation at different time scales, and each channel is configured with response amplitude and time delay separately;

[0084] Derivative feedback submodule: A frequency change rate feedback path is set between the inertial cascade chain and the parallel frequency modulation channel to separate the inertial support from the frequency modulation action;

[0085] Energy storage constraint mapping submodule: The upper and lower limits of the state of charge and the maximum charging and discharging power of each energy storage unit are used as model input constraints and mapped to the corresponding state variables;

[0086] Parameter identification submodule: Based on the recursive least squares optimal estimation algorithm, frequency deviation, power response and state of charge data are collected periodically, and the parameters of each inertial element and frequency modulation channel are updated online.

[0087] In step one of this invention, the construction of the cooperative characteristic model follows this process: five interconnected sub-modules are superimposed and coupled layer by layer to form a complete equivalent mathematical model and solve the technical problems of insufficient system inertia, scarce frequency regulation resources, and rapidly changing constraints under high wind and solar penetration: The cascaded inertial sub-module first receives the frequency deviation signal of the power system, and passes it through three to five series inertial links in sequence. Each link uses independently set inertia coefficients and damping coefficients to filter and amplify the frequency deviation, generating a layer-by-layer decaying short-time inertial support output to compensate for the system inertia loss caused by wind power replacing synchronous machines; The parallel frequency regulation sub-module follows each inertial link with three parallel gain channels, corresponding to primary, secondary, and tertiary frequency regulation, respectively, and performs active power regulation with different amplitudes and delays on the primary inertial support signal at short, medium, and long time scales to ensure continuous and orderly power compensation after the initial inertial response, in the middle of the frequency drop, and at the end of the recovery; The derivative feedback sub-module is connected in the cascaded inertial chain and parallel A frequency change rate measurement loop is arranged in parallel between the frequency modulation channels to feed back the frequency derivative signal to the corresponding inertial link and frequency modulation gain channel. This is used to quickly distinguish between the two support mechanisms of "inertial" and "frequency modulation" and dynamically adjust the response intensity and timing of each channel. The energy storage constraint mapping submodule reads the upper and lower limits of the state of charge and the maximum charging / discharging power of each energy storage unit in real time. It maps these constraints to the input boundaries of the model and associates them with the available output of each inertial and frequency modulation channel. It automatically limits or relaxes the output of each channel to cope with the fluctuation of available capacity caused by rapid changes in the state of charge. The parameter identification submodule periodically collects system frequency deviation, unit and energy storage output and state of charge data with a 10-second sliding window. It uses a recursive least squares optimal estimation algorithm with a forgetting factor of 0.98 to identify the inertia coefficient, the frequency modulation slope of each level and the channel delay online. Every 5 seconds, the new and old parameters are weighted and merged in a 70%:30% ratio and then updated to the above submodules to ensure that the model parameters are synchronized with the actual operating conditions.

[0088] like Figure 2 As shown, the process of extracting the dominant dynamic features and calculating the power allocation coefficients of each cooperative unit in step two of the technical solution of this invention includes:

[0089] Step 21: Perform multi-scale decomposition processing on the power system frequency and output signal to obtain the dominant response components representing inertial support, primary frequency regulation and secondary frequency regulation respectively.

[0090] Step 22: Identify the corresponding modal energy weights based on the amplitude time series and energy spectrum characteristics of each dominant response component;

[0091] Step 23: The modal energy weights are fused with the real-time adjustable power capacity of the coal-fired power unit and the energy storage unit according to the adaptive mapping rule to generate preliminary power allocation coefficients;

[0092] Step 24: Dynamically normalize the sub-allocation coefficients and perform boundary corrections based on the upper and lower limits of the remaining capacity of each unit to output the final power allocation coefficients.

[0093] By accurately extracting the dominant response components of each stage of inertial support, primary frequency regulation, and secondary frequency regulation through multi-scale decomposition, and identifying modal energy weights based on amplitude time series and energy spectrum characteristics, the power demand of the entire frequency disturbance process can be quantitatively layered. These weights are then adaptively fused with the real-time adjustable capacity of the generating units and energy storage to generate preliminary allocation coefficients. Dynamic normalization and boundary correction are then performed in conjunction with the upper and lower limits of the remaining capacity, enabling real-time matching of the available output of each coordinating unit and avoiding excessive or insufficient resource allocation. The final output power allocation coefficients have fast response, high accuracy, and constraint compliance, effectively improving the accuracy and stability of the system's frequency support and reducing scheduling errors and energy storage overrun risks caused by static allocation.

[0094] Specifically, step 21 uses continuous wavelet transform to decompose the frequency deviation and output signal into multiple scales. The mother wavelet used is the Morlet wavelet, and three scales of 0.5 seconds, 2 seconds, and 8 seconds are set to correspond to different time constants. Step 22 applies Hilbert transform to each scale component to obtain the instantaneous amplitude envelope for subsequent energy analysis. Step 23 calculates the spectral entropy based on the amplitude envelope of each scale to characterize the energy distribution weights at different adjustment stages. Step 24 uses Mamdani-type fuzzy inference to fuse the spectral entropy weights of each scale with the real-time adjustable capacity of the unit and energy storage unit to generate preliminary allocation coefficients. The preliminary allocation coefficients are then smoothed using a five-second window Savitzky-Golay filter and output as the final power allocation coefficients.

[0095] By performing Morlet wavelet decomposition on the frequency deviation and output signal at three scales (0.5s, 2s, and 8s), the dominant response components corresponding to the inertial support, primary frequency modulation, and secondary frequency modulation stages are extracted respectively. Then, the instantaneous amplitude envelope of each component is obtained by Hilbert transform, and its spectral entropy is quantified, which can clearly characterize the dynamic energy distribution of different adjustment stages. Subsequently, the energy weights of each stage are adaptively fused with the adjustable capacity of the unit and energy storage through Madani-type fuzzy inference to generate accurate preliminary allocation coefficients. After smoothing by a five-second window Savitzky-Golay filter, active power support can be allocated in real time and hierarchically, which significantly improves the response speed and allocation accuracy to large disturbances and random fluctuations, while suppressing command mutations and noise interference, ensuring the continuity, stability, and robustness of active power balance control.

[0096] The closed-loop parameter update process in step four of the technical solution of this invention includes:

[0097] The real-time frequency deviation and power response data are sampled in a sliding window of 10 seconds.

[0098] The equivalent inertia coefficient, first / second / third-order frequency modulation slope, and power allocation coefficient in the equivalent mathematical model are identified online using a recursive least squares algorithm with a forgetting factor of 0.98.

[0099] An update is triggered every 5 seconds, and the parameters obtained in this identification are exponentially weighted and fused with the results of the previous update at a ratio of 70%:30%.

[0100] The parameters obtained after fusion are limited to within ±10% of the previous updated value;

[0101] The above fusion results are then subjected to a first-order exponential smoothing filter with a time constant of 0.5 seconds, and then loaded into the equivalent mathematical model for the next control cycle.

[0102] It should be noted that the closed-loop update process described above is used in step four to ensure that the model parameters can track the rapid changes in system operating conditions in real time and suppress measurement noise. First, the 10-second sliding window rolling sampling can capture the dynamic characteristics of frequency deviation and output response within a short time scale. Second, the recursive least squares algorithm with a forgetting factor of 0.98 can both improve the identification accuracy by utilizing historical data and quickly "forget" outdated information, achieving online adaptation of parameters such as system inertia and frequency modulation slope. Third, triggering an update every 5 seconds and fusing the old and new parameters in a 70%:30% ratio can balance the identification convergence speed and the stability of the results. Subsequently, the parameters are limited to ±10% to avoid drastic parameter jumps caused by instantaneous disturbances. Finally, a 0.5-second exponential smoothing filter is applied to further filter out high-frequency noise, ensuring that the updated parameters transition smoothly and converge stably, so as to maintain the high-precision description of the actual operating state and continuous and reliable active power balance control of the equivalent mathematical model.

[0103] like Figure 3 As shown, applying the scheme proposed in this invention, the control interface, driven by this method, centrally displays the core components and operating status of the closed-loop online parameter update module of this invention through a visual interface: the "Optimize Parameter Configuration" area on the left allows setting the initial value of the gradient step size, the convergence threshold, the parameter fusion ratio, and the limited range; the "Parameter Evolution Curve" in the middle plots the dynamic changes of the equivalent inertia coefficient and the first / second / third-order frequency modulation slope in real time; the "Parameter Monitoring" panel at the bottom displays the 5-second update cycle, the 98% forgetting factor, the 10-second window, and the ±10% limited amplitude; and through the iteration progress bar and status indicators such as "normal," "updating," and "warning," online parameter debugging and early warning monitoring are realized.

[0104] In summary, combining Examples 1 and 2, this invention, through constructing a hierarchical inertia and multi-stage frequency regulation collaborative model, introducing multi-scale online feature extraction, segmented optimized scheduling, and closed-loop parameter updates, achieves rapid multi-stage frequency support and smooth active power allocation for wind power and energy storage in high wind power (50%–70% penetration) and high fluctuation (>±15% / min) scenarios. In Example 1, compared with the traditional baseline scheme, this invention increases the lowest system frequency point from 49.48Hz to 49.62Hz, reduces the maximum RoCoF from 0.35Hz / s to 0.22Hz / s, shortens the frequency recovery time from 18.7s to 11.3s, reduces the peak output surge by more than 60%, and increases the lowest SoC of energy storage from 42.1% to 45.3%, with only a slight increase in economic cost of 0.003 million yuan·s. Example 2 further demonstrates that online recursive least squares identification and parameter fusion can maintain the above performance under different system scales and load disturbance conditions. Therefore, this invention effectively overcomes the shortcomings of existing static evaluation methods in terms of dynamic response, continuity and parameter adaptation, and achieves a comprehensive improvement in frequency stability, output smoothness and economic dispatch efficiency.

[0105] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0106] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power active power balance control method for coal-fired power units coupled with multi-element energy storage, characterized in that, Includes the following steps: Step 1: Construct a collaborative characteristic model: Based on the power dynamic characteristics of coal-fired power units and energy storage units, establish an equivalent mathematical model with graded inertia and multi-stage frequency regulation capabilities. Step 2, Online calculation of dominant characteristics: Based on the power system frequency deviation and the real-time output measurement values ​​of coal-fired power units and energy storage units, extract the dominant dynamic characteristics and calculate the power distribution coefficient of each coordinating unit; Step 3, segmented optimization scheduling: Construct a segmented optimization scheduling model that takes into account both power constraints and economic operation objectives, and solve the joint output scheme of coal-fired power units and energy storage units; Step 4, Closed-loop parameter update: Based on real-time load fluctuations and frequency deviations, the equivalent inertia coefficient, frequency modulation slope, response time parameters, and power distribution coefficient in the equivalent mathematical model of Step 1 are adjusted online in a closed loop. In step three, the segmented optimization scheduling model constructs a sub-model for each scheduling period, which includes a cost function and energy storage charging and discharging constraints. Within this sub-model, the output allocation of coal-fired power units and energy storage units is calculated iteratively through gradient descent. Based on the convergence criterion, the joint output scheme for the scheduling period is output, and the gradient step size is dynamically adjusted. Each scheduling period is executed sequentially and connected.

2. The active power balance control method for a coal-fired power unit coupled with multi-element energy storage according to claim 1, characterized in that, In step three, the segmented optimization scheduling model uses the combined output result of the previous time period for each scheduling period as the initial value for hot start, and sets a sliding overlapping time window in the sub-model; the economic cost function and frequency deviation penalty term are introduced in parallel into the objective function of the sub-model, and the weights of the economic cost function and frequency deviation are adaptively adjusted according to the real-time frequency difference ratio; an increasing / decreasing tightening strategy is adopted for the upper and lower limits of charging and discharging of energy storage units, and the boundary change rate is set according to the current state of charge; and a smooth transition correction is performed between the combined output schemes of adjacent scheduling periods.

3. The active power balance control method for a coal-fired power unit coupled with multi-element energy storage according to claim 2, characterized in that, When wind power capacity accounts for 50%–70% of the total power generation capacity, energy storage capacity accounts for 10%–20% of the total power generation capacity, and wind power output fluctuation exceeds ±15% / min, the sliding overlap time window length of the segmented optimization scheduling model is 60 seconds, the initial value for hot start is the average value of the combined output in the last 10 seconds of the previous period, the frequency deviation penalty weight is calculated based on the real-time maximum rate of decline of 0.2Hz / s, and is doubled when the rate of decline exceeds 0.2Hz / s, the tightening rate of the upper and lower limits of energy storage unit charging and discharging is 10% of the current state of charge, and the smooth transition correction of the combined output scheme in adjacent scheduling periods adopts three-point linear interpolation.

4. The power active power balance control method for a coal-fired power unit coupled with multi-element energy storage according to claim 1, characterized in that, In step one, the collaborative characteristic model forms an equivalent mathematical model through the following interconnected sub-modules: Cascaded inertial submodule: It consists of three to five inertial links connected in series, each with independently adjustable inertia and damping, to simulate the multi-level inertial response of coal-fired power units and energy storage units; Parallel frequency modulation submodule: Three gain channels are arranged in parallel after each inertial element to handle primary, secondary and tertiary frequency modulation at different time scales, and each channel is configured with response amplitude and time delay separately; Derivative feedback submodule: A frequency change rate feedback path is set between the inertial cascade chain and the parallel frequency modulation channel to separate the inertial support from the frequency modulation action; Energy storage constraint mapping submodule: The upper and lower limits of the state of charge and the maximum charging and discharging power of each energy storage unit are used as model input constraints and mapped to the corresponding state variables; Parameter identification submodule: Based on the recursive least squares optimal estimation algorithm, frequency deviation, power response and state of charge data are collected periodically, and the parameters of each inertial element and frequency modulation channel are updated online.

5. The active power balance control method for a coal-fired power unit coupled with multi-element energy storage according to claim 1, characterized in that, Step two involves extracting the dominant dynamic features and calculating the power allocation coefficients of each cooperative unit, including: Step 21: Perform multi-scale decomposition processing on the power system frequency and output signal to obtain the dominant response components representing inertial support, primary frequency regulation and secondary frequency regulation respectively. Step 22: Identify the corresponding modal energy weights based on the amplitude time series and energy spectrum characteristics of each dominant response component; Step 23: The modal energy weights are fused with the real-time adjustable power capacity of the coal-fired power unit and the energy storage unit according to the adaptive mapping rule to generate preliminary power allocation coefficients; Step 24: Dynamically normalize the sub-allocation coefficients and perform boundary corrections based on the upper and lower limits of the remaining capacity of each unit to output the final power allocation coefficients.

6. The power active power balance control method for a coal-fired power unit coupled with multi-element energy storage according to claim 5, characterized in that, Step 21 uses continuous wavelet transform to perform multi-scale decomposition of frequency deviation and output signal. The mother wavelet used is Morlet wavelet, and three scales of 0.5 seconds, 2 seconds, and 8 seconds are set to correspond to different time constants. Step 22 applies Hilbert transform to each scale component to obtain the instantaneous amplitude envelope for subsequent energy analysis. Step 23 calculates spectral entropy based on the amplitude envelope of each scale to characterize the energy distribution weights of different adjustment stages. Step 24 uses Mamdani-type fuzzy inference to fuse the spectral entropy weights of each scale with the real-time adjustable capacity of the unit and energy storage unit to generate preliminary allocation coefficients. The preliminary allocation coefficients are then smoothed by applying a five-second window Savitzky-Golay filter and output as the final power allocation coefficients.

7. The active power balance control method for a coal-fired power unit coupled with multi-element energy storage according to claim 2, characterized in that, In the piecewise optimization scheduling model in step three, an accelerated gradient algorithm is used for the gradient descent iteration process of each sub-model, specifically including: In the first iteration, only the current gradient information is used as the update direction; Starting from the second iteration, the gradient of this iteration is linearly superimposed with the direction of the previous iteration at a ratio of 0.8, and used as the new search direction; Before each iteration, the momentum coefficient is fine-tuned within a range of ±0.05 based on the difference between the current objective function value of the sub-model and the objective function value of the previous iteration. When the improvement of the objective function is less than 0.001% for three consecutive iterations or the number of iterations exceeds 50, the iteration of this sub-model is terminated and the result is output.

8. The power active power balance control method for a coal-fired power unit coupled with multi-element energy storage according to claim 7, characterized in that, In the objective function of the sub-model in step three, an economic cost term and a frequency deviation penalty term are set in parallel, and the weight of the frequency deviation penalty is dynamically adjusted as follows: The maximum frequency drop of the system in the last 30 seconds is sampled and averaged at a period of 5 seconds to generate the real-time maximum drop rate value. Multiply the rate of decline value by a preset scaling factor as the frequency penalty weight; During the same period, the weight of the economic cost item was set at 1.2 times the proportion of wind power installed capacity; The frequency penalty weight is reset every 40 seconds.

9. The active power balance control method for a coal-fired power unit coupled with multi-element energy storage according to claim 7, characterized in that, In the segmented optimization scheduling model in step three, when using the accelerated gradient algorithm for the gradient descent iteration process of each sub-model, the initial value of the gradient step size is set to 0.

1. After each iteration, the step size is adaptively decayed based on the local second derivative of the objective function at the current iteration point / the inverse of the gradient norm. When the iteration step size decreases to below 0.001 or the improvement of the objective function value in continuous iterations is below 0.0001, the convergence condition is reached, and the iteration process is terminated immediately.

10. The power active power balance control method for a coal-fired power unit coupled with multi-element energy storage according to claim 1, characterized in that, Step four, the closed-loop parameter update process, includes: The real-time frequency deviation and power response data are sampled in a sliding window of 10 seconds. The equivalent inertia coefficient, first / second / third-order frequency modulation slope, and power allocation coefficient in the equivalent mathematical model are identified online using a recursive least squares algorithm with a forgetting factor of 0.

98. An update is triggered every 5 seconds, and the parameters obtained in this identification are exponentially weighted and fused with the results of the previous update at a ratio of 70%:30%. The parameters obtained after fusion are limited to within ±10% of the previous updated value; The fusion result is then subjected to a first-order exponential smoothing filter with a time constant of 0.5 seconds, and then loaded into the equivalent mathematical model for the next control cycle.

Images