New energy grid-connected voltage control method based on multi-level cooperation and adaptive algorithm
The new energy grid-connected voltage control method using multi-level collaborative and adaptive algorithms solves the grid instability problem caused by wind and solar power output fluctuations in integrated wind-solar-storage power stations. It achieves accurate assessment and proactive response, reduces grid operating costs, and improves the capacity for new energy absorption and grid stability.
Patent Information
- Application Number
- CN202611260328.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-19
- Publication Date
- 2026-11-10
Smart Images

Figure CN123073833A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power generation technology, and more specifically, to a new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms. Background Technology
[0002] The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms is mainly used to solve the problem of unstable grid node voltage and frequency caused by the fluctuation of wind and solar power output during the grid connection of wind, solar and energy storage integrated power stations. It optimizes the peak-shaving function of energy storage and the accuracy of power prediction, reduces grid operating costs, improves the capacity of new energy consumption, ensures the safe and stable operation of the grid, and adapts to the actual operating needs of wind, solar and energy storage integrated power stations.
[0003] In existing technologies, voltage control for traditional renewable energy grid connection often uses a single parameter to assess the impact of wind and solar power output. Energy storage devices are adjusted through a passive response mode after the voltage exceeds the limit. Moreover, there is a lack of a multi-level collaborative mechanism between cloud, edge, and terminal. Data at each level cannot be effectively exchanged. Power prediction relies heavily on traditional models and has not been fully optimized by combining the spatiotemporal characteristics of wind and solar power output with the probability of strong impacts. It can only achieve basic voltage regulation and power prediction and cannot adapt to the operating needs of power plants under complex operating conditions.
[0004] However, in actual power plant operation, numerous shortcomings exist. Taking a wind-solar-storage integrated power plant located in a windy region in northern China as an example, the specific manifestations are as follows: sudden changes in wind speed in spring and rapid cloud cover obscuring the photovoltaic array in summer lead to severe fluctuations in wind and solar power output, impacting the power grid like ocean waves and causing voltage and frequency fluctuations at grid connection nodes. Traditional renewable energy grid connection voltage control cannot accurately assess the amplitude, rate, and probability of these fluctuations, and can only passively respond after the grid connection node voltage exceeds the limit. The power grid dispatch and control center needs to reserve a large amount of buffer resources, increasing operating costs and posing a risk of voltage exceeding the limit, which does not meet the requirements of wind-solar-storage integrated power plants for optimizing energy storage and peak shaving. At the same time, the spatiotemporal characteristics of wind and solar power output at power plants are complex, and different wind power sources... Due to differences in terrain and wind speed distribution, photovoltaic power plants experience varying output fluctuations caused by weather conditions. Traditional grid-connected voltage control for new energy lacks a multi-level collaborative mechanism involving the cloud, edge, and terminal, resulting in ineffective data exchange between different levels. This leads to unclear spatiotemporal characteristics and large power prediction errors in traditional prediction models. Previously, the power grid dispatch and control center needed to reserve 200MW of thermal power reserve capacity (accounting for 20% of the total installed capacity of the power plant) to compensate for prediction deviations, which significantly increased the operating costs of the power plant and squeezed the space for new energy consumption. This does not meet the development requirements of prioritizing the development of new energy and enhancing comprehensive competitiveness. Therefore, we provide a new energy grid-connected voltage control method based on multi-level collaboration and adaptive algorithms. Summary of the Invention
[0005] The purpose of this invention is to provide a new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, a new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms is provided, comprising the following steps:
[0007] S1. Collect operational data related to wind power generation and photovoltaic power generation, and after denoising and normalization processing, obtain a standardized dataset containing wind power output data and photovoltaic output data.
[0008] Based on the dynamic deviation between the real-time wind power output ratio and the benchmark wind power output ratio in the standardized dataset, the initial weighting coefficients of wind power and photovoltaic power are adaptively adjusted to obtain the final weighting coefficients of wind power and photovoltaic power.
[0009] Based on the wind power output change rate and photovoltaic output change rate in the standardized dataset, and combined with the final weighting coefficients of wind power and photovoltaic power, a comprehensive fluctuation impact factor is calculated to quantify the impact intensity of wind-solar coordinated fluctuations on the power grid; and the corresponding impact level is classified according to the magnitude of the comprehensive fluctuation impact factor.
[0010] S2. Based on the standardized dataset, the rolling prediction method is used to predict the wind power output and photovoltaic output at future times, and combined with the final weighting coefficients of wind power and photovoltaic power, the predicted value of the comprehensive fluctuation impact factor at future times is calculated.
[0011] The comprehensive fluctuation impact factor at the current moment and the comprehensive fluctuation impact factor of the previous acquisition period are obtained. Combined with the predicted value of the comprehensive fluctuation impact factor at the future moment, the impact assessment data used to characterize the impact characteristics are calculated. The impact assessment data includes at least the impact amplitude, impact rate and strong impact probability.
[0012] Based on the impact level, a corresponding adaptive adjustment coefficient is selected. The adaptive adjustment coefficient is an energy storage power adjustment coefficient independent of the weighting coefficient. Based on the comprehensive fluctuation impact factor, the adaptive adjustment coefficient, and the rated charging and discharging power of the energy storage device, the charging and discharging power of the energy storage device at the current moment is dynamically calculated, and the energy storage device is controlled to adjust its charging and discharging state according to the charging and discharging power.
[0013] S3. Obtain long-term historical data of the power plant, and perform time alignment and spatial aggregation of the long-term historical data with the impact assessment data and the standardized dataset to extract spatiotemporal-impact fusion features; based on a preset time window, construct the spatiotemporal-impact fusion features into a multi-dimensional coupled spatiotemporal feature matrix;
[0014] S4. Using the long-term historical data and the spatiotemporal-impact fusion features, generate an initial prediction coefficient matrix; combine the spatiotemporal feature matrix and the prediction coefficient matrix to calculate the initial prediction error, and superimpose the initial prediction error to form an ultra-short-term power prediction value; extract the peak value of the ultra-short-term prediction error and calculate the actual required thermal power reserve capacity, which can be quickly supplemented when a power output gap occurs.
[0015] Preferably, the method for determining the final weighting coefficients for wind power generation and photovoltaic power generation includes the following steps:
[0016] The initial weighting coefficients for wind power generation and photovoltaic power generation are set based on the proportion of total installed capacity of wind power generation and photovoltaic power generation in the power station.
[0017] The real-time wind power output ratio is calculated based on the wind power output and photovoltaic output in the standardized dataset, and the output deviation is calculated based on the real-time wind power output ratio and the benchmark wind power output ratio.
[0018] By calculating and adjusting the weighting coefficients based on the output deviation, and then dynamically adjusting the corresponding weighting coefficients in conjunction with the initial weighting coefficients, the final weighting coefficients for wind power generation and photovoltaic power generation are obtained.
[0019] Preferably, the comprehensive wave impact factor is calculated through the following steps:
[0020] Obtain the comprehensive volatility impact factor calculated in step S1;
[0021] The normalized wind power output change rate and the normalized photovoltaic output change rate are obtained by calculating the two types of output change rates separately and normalizing them.
[0022] The absolute values of the normalized wind power output change rate and the normalized photovoltaic output change rate are taken respectively. The two types of absolute output change rates are weighted and coupled together by the final weight coefficients of wind power and photovoltaic power to construct a comprehensive fluctuation impact factor.
[0023] Preferably, the calculation of the impact assessment data and the charging and discharging power of the energy storage device includes the following method steps:
[0024] The total number of prediction time points is obtained by predicting the time domain and the prediction step size, and the impact magnitude is obtained by iterating through the prediction value of the comprehensive fluctuation impact factor.
[0025] The impact rate is obtained by calculating the change in the impact factor between adjacent time points.
[0026] The probability of a strong impact is calculated based on the total number of predicted time points, and impact assessment data is formed by combining the impact amplitude and impact rate.
[0027] Based on the calculated comprehensive fluctuation impact factor coupled with the adaptive adjustment coefficient γ, a dynamic control model is constructed and the charging and discharging power of the energy storage device is calculated.
[0028] Preferably, the adaptive adjustment coefficient includes the following method steps:
[0029] The corresponding adaptive adjustment coefficient is selected according to the impact level:
[0030] Set strong impact threshold and low impact threshold. When the comprehensive fluctuation impact factor is greater than or equal to the strong impact threshold, record the corresponding adaptive adjustment coefficient.
[0031] When the comprehensive volatility impact factor is greater than or equal to the low impact threshold and less than the strong impact threshold, the corresponding adaptive adjustment coefficient is recorded.
[0032] When the comprehensive fluctuation impact factor is less than the low impact threshold, the corresponding adaptive adjustment coefficient is recorded.
[0033] Preferably, the extraction of spatiotemporal-impact fusion features includes the following method steps:
[0034] Obtain the comprehensive wave impact factor calculated in step S1 and the impact assessment data calculated in step S2;
[0035] The impact assessment data is time-aligned with the standardized dataset, and combined with the long-term historical data of the power plant, time-series alignment and spatial aggregation are performed based on a preset time window.
[0036] Spatial feature vectors are extracted based on wind speed, and temporal feature vectors are extracted based on wind power output and irradiance. The spatial feature vectors and temporal feature vectors are fused to obtain the spatiotemporal features of wind and solar power output.
[0037] The comprehensive wave impact factor, impact amplitude, impact rate and strong impact probability constitute the impact factor feature. The impact factor feature is then linked and fused with the spatiotemporal features of wind and solar power output to extract the spatiotemporal-impact fusion feature.
[0038] Preferably, the construction of the multidimensional coupled spatiotemporal feature matrix includes the following method steps:
[0039] The matrix is arranged vertically to display full-time samples, and each set of time-series samples simultaneously includes the main features of wind power generation and photovoltaic power generation.
[0040] It is divided into three core feature columns, which respectively carry wind power output and photovoltaic output, wind speed and irradiance, and comprehensive fluctuation impact factor at various times.
[0041] Preferably, the generation of the prediction coefficient matrix and the ultra-short-term power prediction values includes the following method steps:
[0042] By combining long-term historical data with spatiotemporal-impact fusion characteristics, a prediction coefficient matrix coupled with multi-level collaborative data is generated;
[0043] After obtaining the basic prediction value by coupling the calculation with the prediction coefficient matrix as the link, the initial prediction error is then calculated as the state of the reinforcement learning model.
[0044] The ultra-short-term power prediction is finally formed by superimposing the basic prediction value with the initial prediction error.
[0045] Preferably, the reinforcement learning model includes the following method steps:
[0046] By summing up the squared terms of the ultra-short-term prediction error, a correction coefficient is introduced to correlate the strong shock probability with the ultra-short-term prediction error, and a reinforcement learning reward function is constructed.
[0047] The gradient is obtained by taking the partial derivative of the reinforcement learning reward function with respect to the prediction coefficient matrix. Then, the dynamic adjustment of the prediction coefficient matrix is calculated based on the gradient, and used as the action to build the reinforcement learning model.
[0048] Using state as input, action as optimization output, and reinforcement learning reward function as iterative feedback, the reinforcement learning model is constructed by combining gradient and learning rate.
[0049] Preferably, calculating the actual configured thermal power reserve capacity includes the following steps:
[0050] The prediction coefficient matrix, spatiotemporal-impact fusion features, and long-term historical data of the power plant are used for iterative optimization. The prediction error data is recorded after the prediction coefficient matrix is iteratively optimized.
[0051] The optimized power output prediction value is output by combining the optimized prediction coefficient matrix, spatiotemporal feature matrix and prediction error data;
[0052] The maximum prediction error is obtained, and the actual configured thermal power reserve capacity is calculated by combining it with the optimized power output prediction value.
[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0054] 1. This new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms uses a standardized dataset and an input sample window to predict the wind power output for the next minute. It then calculates the predicted value of the comprehensive fluctuation impact factor using the final weighting coefficient, calculates the impact assessment data, and selects an adaptive adjustment coefficient based on the impact level. This adaptive adjustment coefficient is used to dynamically calculate the charging and discharging power of the energy storage device, and the charging and discharging is adjusted according to the energy storage device's charging and discharging power. Active prediction is achieved through an MPC objective function incorporating the comprehensive fluctuation impact factor. Combined with real-time fluctuation data of wind and solar power in the standardized dataset, it realizes active prediction of impact parameters and early activation of energy storage buffers. This completely solves the core defects of traditional methods, which cannot accurately assess the impact amplitude, rate, and probability, and can only passively respond after the grid-connected node voltage exceeds the limit. Simultaneously, the comprehensive fluctuation impact factor dynamically adjusts the output ratio, adapting to both the strong wind power impact caused by sudden wind speed changes in spring and the sudden changes in solar power output caused by cloud cover in summer. It perfectly adapts to the characteristics of wind and solar power output fluctuations in different seasons and time periods, improving the overall assessment accuracy and adaptability compared to traditional single-parameter impact assessment methods.
[0055] 2. This new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms acquires long-term historical data, combines spatiotemporal-impact fusion features to generate a prediction coefficient matrix, calculates the initial prediction error using the spatiotemporal feature matrix, and then superimposes the initial prediction error to form an ultra-short-term power prediction value. It obtains the peak value of the ultra-short-term prediction error for the next 60 minutes and calculates the actual configured thermal power reserve capacity. When a short-term power output gap occurs in the grid, it directly calls upon the actual configured thermal power reserve capacity, commanding thermal power units to quickly increase output to fill the gap. Through the spatiotemporal feature matrix and prediction coefficient matrix, the multi-level collaborative architecture of the reinforcement learning model dynamically corrects the prediction error, adapting to the complex spatiotemporal characteristics and frequent impacts of power plants. Furthermore, it optimizes the reinforcement learning reward function with a strong impact probability, solving the shortcomings of traditional prediction models such as unclear spatiotemporal features and large errors. This makes the reinforcement learning model more closely aligned with the actual operating scenarios of power plants, thereby improving prediction accuracy. Attached Figure Description
[0056] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0058] Example 1
[0059] This invention provides a new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms. Please refer to [link to relevant documentation]. Figure 1 The method includes the following steps:
[0060] The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms is mainly used to solve the problem of unstable voltage and frequency at the grid connection node caused by the fluctuation of wind and solar power output during the grid connection of integrated wind, solar and energy storage power stations. It optimizes the peak-shaving function of energy storage and the accuracy of power prediction, reduces grid operating costs, improves the capacity for new energy absorption, ensures the safe and stable operation of the grid, and adapts to the actual operating needs of integrated wind, solar and energy storage power stations. In the existing technology, for traditional new energy grid-connected voltage control, a single parameter is often used to evaluate the impact of wind and solar power output. The energy storage device is adjusted through the passive response mode after the voltage exceeds the limit. Moreover, there is a lack of a multi-level collaborative mechanism between cloud, edge and terminal. Data at each level cannot be effectively communicated. Power prediction relies on traditional models and does not fully combine the spatiotemporal characteristics of wind and solar power output and the probability of strong impact for optimization. It can only achieve basic voltage regulation and power prediction and cannot adapt to the operating needs of power stations under complex operating conditions. Therefore, in step S1, the relevant operating data of wind power generation and photovoltaic power generation are collected, and after denoising and normalization processing, a standardized dataset containing wind power output data and photovoltaic output data is obtained.
[0061] Based on the dynamic deviation between the real-time wind power output ratio and the benchmark wind power output ratio in the standardized dataset, the initial weighting coefficients of wind power and photovoltaic power are adaptively adjusted to obtain the final weighting coefficients of wind power and photovoltaic power.
[0062] Based on the wind power output change rate and photovoltaic output change rate in the standardized dataset, and combined with the final weighting coefficients of wind power and photovoltaic power, a comprehensive fluctuation impact factor is calculated to quantify the impact intensity of wind-solar coordinated fluctuations on the power grid. The corresponding impact levels are then classified according to the magnitude of the comprehensive fluctuation impact factor. The specific steps include the following:
[0063] S1.1 Since integrated wind-solar-storage power stations are generally located in windy areas of northern China, and wind speeds change drastically in spring and clouds rapidly obscure photovoltaic arrays in summer, wind and solar power output fluctuates dramatically, impacting the power grid like ocean waves. This causes voltage and frequency fluctuations at grid connection points, making it impossible to accurately assess the amplitude, rate, and probability of these fluctuations. The power grid dispatch and control center needs to reserve significant buffer resources, increasing operating costs and posing a risk of voltage exceeding limits. To accurately capture and address this fluctuation issue and ensure stable grid operation, smart sensors are deployed at each wind turbine nacelle in the wind farm, the photovoltaic array in the photovoltaic power station, the energy storage compartment of the new energy storage equipment, and key nodes where the power station connects to the grid. These sensors collect wind power generation data with a 10ms acquisition cycle. The wind power generation data includes the wind speed of each wind turbine in the wind farm. Wind power output (The actual active power generated and connected to the grid by the wind turbine at a certain moment), where wind speed data is collected at the turbine hub to avoid measurement errors caused by terrain obstruction and ensure the accuracy of wind power-related data; simultaneously, intelligent sensors are used to collect photovoltaic and energy storage-related data, including the irradiance of photovoltaic arrays in each photovoltaic power station. Photovoltaic power output (Active power generated and fed into the grid by photovoltaic power generation equipment at a certain moment), and the charging and discharging power of energy storage equipment. In addition to the remaining capacity of energy storage devices, it also utilizes smart sensors to collect grid-related data, including the voltage at the power plant's grid connection nodes. ,frequency Because sensor malfunctions and extreme weather events (such as severe sandstorms and heavy rain) can cause abrupt data changes, the collected data is denoised and normalized to remove the abrupt data caused by sensor malfunctions and extreme weather events, resulting in a standardized dataset applicable to the entire process. , This refers to the corresponding timestamp. A standardized dataset is a universally applicable, well-organized dataset formed by denoising, removing anomalous jumps, and normalizing the raw 10ms data. This eliminates dimensional differences and interference from anomalous data. It is primarily used as a unified input data source for weight calculation, impact factor construction, power prediction, and energy storage regulation. It can eliminate interference from different physical dimensions, sensor anomalies, and differences in installed capacity, achieving 100% data consistency and cross-platform compatibility. Power plant transplantation and reuse; if the original collected data is used directly, the dimensions are chaotic, outlier interference is serious, matrix operations are completely ineffective, the prediction accuracy and control accuracy drop sharply, and multi-level data collaboration cannot be achieved.
[0064] Synchronously collect and record basic capacity parameters of the power station, including the capacity of a single wind turbine. The single-unit capacity of photovoltaic modules is The total installed capacity of the power station is (Wind power is) and photovoltaics The typical range for the rate of change of wind power output is: The typical range for the photovoltaic power output change rate is: This provides a basis for the subsequent normalization of the rate of change of output and the setting of the impact threshold;
[0065] S1.2 Based on the wind and solar installed capacity structure of the power station, wind power accounts for 60% of the total installed capacity and photovoltaic accounts for 40%. Furthermore, the power station is located in a windy northern region, where wind power is the core power source. Therefore, an initial weighting coefficient is set. , The real-time wind power output ratio is calculated by using wind power output and photovoltaic power output from a standardized dataset. The real-time wind power output ratio refers to the real-time ratio of the total active power output of renewable energy sources to that of wind power. This is achieved by obtaining the rated total installed capacity of wind turbines and the rated total installed capacity of photovoltaic power, and then calculating the benchmark wind power output ratio. ,in, The moving average (time-varying reference value) of the real-time wind power output ratio over the past 10 minutes represents the typical wind power output ratio during the current meteorological period, rather than a fixed installed capacity ratio. Physically, it represents the typical wind and solar power output ratio under current meteorological conditions. The benchmark wind power output ratio refers to the ratio benchmark of the station's rated installed capacity, a fixed design value, not an instantaneous ratio of real-time active power. The comparison between the real-time and benchmark wind power output ratios is used to determine whether to adjust the weighting coefficient. When the real-time wind power output ratio is greater than the benchmark wind power output ratio for three consecutive 10ms data collection cycles, it indicates that wind power is the absolute dominant output, and the weighting coefficient is adjusted accordingly. The output deviation is calculated based on the real-time and benchmark wind power output ratios. Then, calculate and adjust the weighting coefficients according to the output deviation. , The α value is slightly increased only to match the dominant output characteristics, and then the dynamic adjustment is stopped immediately. When the real-time wind power output ratio is less than the benchmark wind power output ratio for three consecutive 10ms data collection cycles, it indicates that the wind power output ratio is abnormally low and the real-time photovoltaic output ratio is high. The adjustment weight coefficient is then determined, and the adjustment weight coefficient is calculated separately according to the output deviation. , ,Will Upgraded to 0.8 If the weighting is lowered to 0.2, dynamic adjustment will immediately cease. If the real-time wind power output ratio equals the benchmark wind power output ratio, it is determined that the weighting coefficient adjustment will not be triggered, and the initial weighting coefficient will be directly used as the final weighting coefficient. To avoid imbalance in the output ratio, accurately match the output disturbance caused by a sudden increase in photovoltaic power generation equipment, and ensure the quantification accuracy of the wind and solar power output fluctuation factor, dynamic adjustments are made using the corrected weighting coefficient and the initial weighting coefficient respectively to obtain the final weighting coefficient. , Otherwise, the initial weight coefficients will be used as the final weight coefficients, and the weight adjustment results will be applied synchronously to the power prediction scenario to ensure the consistency of parameters throughout the process.
[0066] The final weighting coefficient refers to a weighted coefficient that is dynamically adjusted based on the deviation between the installed capacity ratio of wind and solar power plants and the real-time wind and solar power output ratio. Corresponding wind power, final weighting coefficient Corresponding to photovoltaics, the constraints are The range is mainly used for weighted fusion of wind and solar power output change rates to accurately allocate the contribution of wind power and photovoltaic power to the comprehensive fluctuation impact factor. This allows for dynamic matching of real-time wind and solar power output ratios, perfectly adapting to the power output fluctuation characteristics under different operating conditions such as strong winds in spring and cloudy weather in summer, so that the quantification error of the impact factor is less than or equal to 1.3%. If a fixed initial weight is used, it cannot adapt to the real-time changes in the proportion of wind and solar power output, the impact factor is completely distorted, and energy storage regulation and voltage control lose their accuracy.
[0067] S1.3 Calculate the rate of change of wind power output using wind power output. This is used to reflect the impact rate caused by sudden changes in wind speed, such as the rapid rise and fall of wind turbine output in a short period of time during strong winds in spring. The rate of change of photovoltaic output is calculated based on the photovoltaic output. This is used to reflect sudden changes in photovoltaic output caused by cloud cover in summer; to eliminate the impact of differences in installed capacity of different power plants on the impact assessment, the rate of change of wind power and photovoltaic output is normalized to obtain the normalized rate of change of wind power output. and normalized photovoltaic power output change rate ,in, The total installed capacity of the power plant, and its normalized range. Dimensionless and portable across power stations; it takes the absolute values of the normalized wind power output change rate and the normalized photovoltaic output change rate, and uses an absolute value algorithm to shield the interference of wind and photovoltaic output rise and fall directions, purely quantifying the severity of disturbances. Combined with the final weight coefficients, it couples the absolute output change rates of wind power and photovoltaic power respectively and weights and fuses them to construct a comprehensive fluctuation impact factor in real time. The specific algorithm formula for the comprehensive volatility impact factor is as follows: ;in, The phase coordination coefficient is incremented by 1 for in-direction fluctuations to enhance the superposition effect, and decremented by 1 for reverse fluctuations to reflect the cancellation effect, truly representing the net power impact on the grid side. The comprehensive fluctuation impact factor is a shared feature quantity throughout the entire process: ① used for real-time control of energy storage charging and discharging; ② directly embedded in the spatiotemporal feature matrix; ③ used for optimization of the reinforcement learning reward function. It is the same physical quantity throughout the process, only performing time-series backtracking or data alignment, without any mathematical transformations. The value range of the comprehensive fluctuation impact factor is strictly defined. The upper limit of the threshold is 1 (corresponding to the theoretical limit impact of a sudden change in 100% of the total installed capacity); the larger the comprehensive fluctuation impact factor, the greater the impact intensity on the power grid caused by the combined power fluctuation rate and fluctuation amplitude. Based on the comprehensive fluctuation impact factor, strong, medium and weak impact levels are divided, and the strong impact threshold range is set as follows. The medium impact threshold range is The low impact threshold range is The comprehensive fluctuation impact factor and the strong impact threshold are used to determine whether a strong impact exists on the power grid. When the comprehensive fluctuation impact factor is greater than or equal to... When the comprehensive fluctuation impact factor is greater than or equal to 0.3 and less than 0.8, the power grid is judged to have a strong impact; when the comprehensive fluctuation impact factor is less than 0.3, the power grid is judged to have a moderate impact; when the comprehensive fluctuation impact factor is less than 0.3, the power grid is judged to have a weak impact. Multi-power station simulation verification shows that the thresholds for strong, moderate, and weak impacts are adaptable. It is fully portable and reasonable for power plants of different sizes;
[0068] The comprehensive fluctuation impact factor refers to a dimensionless quantitative index obtained by weighting and fusing the normalized wind power output change rate and the normalized photovoltaic power output change rate with a final weighting coefficient. Its value range is strictly defined as follows: This system is used to uniformly characterize the impact intensity of wind and solar power output fluctuations on the power grid. It is mainly used to classify strong, medium, and weak impact levels, calculate the charging and discharging power of energy storage devices, embed a spatiotemporal feature matrix, and serve as the core optimization basis for the reinforcement learning reward function. It also couples the fluctuation characteristics of wind and solar power, and can dynamically adjust the weights according to the real-time output ratio of wind and solar power. The impact quantification accuracy is improved by more than 40% compared to a single indicator. The same feature quantity is shared throughout the process, and there is no data transformation distortion. If the comprehensive fluctuation impact factor is replaced by a single wind power output change rate or a single solar power output change rate, it cannot take into account the synergistic fluctuation characteristics of wind and solar power. The impact assessment deviation is greater than or equal to 35%, which leads to frequent malfunctions in energy storage regulation and increases the risk of grid connection voltage exceeding the limit by 80%. It cannot be adapted to the synergistic operation scenario of wind, solar, and energy storage.
[0069] The determination of the impact threshold range is based on the following methods and steps:
[0070] The range of values for the comprehensive volatility impact factor is strictly defined as follows: (Normalized dimensionless), the upper limit of the impact threshold is 1 (corresponding to the theoretical limit impact of a sudden change in 100% of the total installed capacity); the total installed capacity of this power station is (Wind power is) and photovoltaics When the comprehensive fluctuation impact factor is 0.8, it corresponds to a sudden power output change of 80% of the rated installed capacity; when the comprehensive fluctuation impact factor is 0.3, it corresponds to a sudden power output change of 30% of the rated installed capacity, which is close to the upper limit of the actual power output fluctuation of the power station. According to actual measurement statistics: when the comprehensive fluctuation impact factor is greater than or equal to 0.8, the risk of grid connection voltage exceeding the limit is greater than or equal to 90%; when the comprehensive fluctuation impact factor is within... Within the specified range, voltage fluctuations are within a safe range; when the comprehensive fluctuation impact factor is less than 0.3, there is no risk of voltage fluctuation, and the impact threshold perfectly matches the grid safety boundary; the rated charging and discharging power of the energy storage equipment in this power station is 200MW, and the maximum charging and discharging rate is 50MW / min: when the comprehensive fluctuation impact factor is greater than or equal to 0.8, the sudden change in output exceeds the maximum buffering capacity of the energy storage equipment, and must be judged as a strong impact; when the comprehensive fluctuation impact factor is within... Within the specified range, the energy storage device can buffer normally, classifying it as a moderate impact. When the comprehensive fluctuation impact factor is less than 0.3, no significant action from the energy storage device is required, classifying it as a weak impact. Further analysis using the power plant's historical data from the past year to count the number of strong, moderate, and weak impact events revealed that impact thresholds of 0.3 and 0.8 achieve an impact level classification accuracy of 98.7%. Therefore, three candidate impact thresholds (0.2 / 0.7, 0.3 / 0.8, and 0.4 / 0.9) were set for closed-loop simulation. The simulation results verified that when the candidate impact thresholds... At thresholds of 0.2 and 0.7, the false alarm rate is high, energy storage devices operate frequently, and lifespan is significantly reduced. When the candidate impact thresholds are 0.3 and 0.8, the voltage over-limit rate is lowest (0.3%), the energy storage device operation frequency is optimal, and the early warning accuracy reaches 99.1%. When the candidate impact thresholds are 0.4 and 0.9, the false alarm rate is high, and the grid impact risk exceeds the limit. Further field verification under extreme conditions of strong winds in spring and cloudy weather in northern windy regions shows that impact thresholds of 0.3 and 0.8 balance grid safety, energy storage lifespan, and early warning reliability, making them the optimal values. Therefore, [the following is a continuation of the previous sentence, but the context is unclear]. Set as the strong impact threshold range Set as medium impact threshold range Set as a low impact threshold range, and verified through transplantation to multiple power plants of the same scale, the impact threshold is adaptable. Wind-solar-storage power stations are both portable and feasible.
[0071] Step S2: Based on the standardized dataset, use the rolling prediction method to predict the wind power output and photovoltaic output at future times, and combine the final weighting coefficients of wind power and photovoltaic power to calculate the predicted value of the comprehensive fluctuation impact factor at future times.
[0072] The comprehensive fluctuation impact factor at the current moment and the comprehensive fluctuation impact factor of the previous acquisition period are obtained. Combined with the predicted value of the comprehensive fluctuation impact factor at the future moment, the impact assessment data used to characterize the impact characteristics are calculated. The impact assessment data includes at least the impact amplitude, impact rate and strong impact probability.
[0073] Based on the impact level, a corresponding adaptive adjustment coefficient is selected, which is an energy storage power adjustment coefficient independent of the weighting coefficient. Based on the comprehensive fluctuation impact factor, the adaptive adjustment coefficient, and the rated charging and discharging power of the energy storage device, the charging and discharging power of the energy storage device at the current moment is dynamically calculated, and the energy storage device is controlled to adjust its charging and discharging state according to the charging and discharging power. Specifically, the method includes the following steps:
[0074] S2.1 When a strong shock occurs, based on the standardized dataset, sample data from the standardized dataset is extracted for the consecutive 10 minutes and 10ms sampling frequencies preceding the current moment. The sample data includes standardized wind speed, wind power output, irradiance, photovoltaic power output, grid-connected node voltage, and energy storage device charging and discharging power, forming the input sample window for rolling prediction. The latest 10-minute sample is retained in a sliding window manner. Based on the strong temporal continuity of short-term wind and solar fluctuations, autoregressive rolling extrapolation is used to predict the trend of standardized wind speed. Standardized wind speed is used as the time series input, and the short-term change gradient is fitted. The prediction step size for each prediction step within the next minute is calculated point by point through extrapolation. Wind speed forecast Irradiance prediction value Predicted voltage values at grid connection nodes Predicted charging and discharging power of energy storage devices By utilizing the mapping relationship between standardized wind speed and standardized wind power output in the input sample window, the predicted wind speed value is substituted into this mapping relationship to convert it into the predicted wind power output value for each moment in the next minute. Simultaneously, by utilizing the mapping relationship between irradiance and standardized photovoltaic output in the input sample window, the predicted irradiance value is substituted into this mapping relationship to convert it into the predicted photovoltaic output value for each moment within the next minute. To fully adapt to the rapid fluctuations in wind and solar power output and the need for rapid on-site adjustments, the predicted value of the comprehensive fluctuation impact factor for future moments is calculated by combining the predicted wind and solar power output values with the final weighting coefficients. Then, obtain the grid-connected node voltage reference value from the database. Upper limit of allowable deviation of grid-connected node voltage (Based on the reference to GB / T12325-2008 "Power Quality - Supply Voltage Deviation" document, 4.1, the sum of the absolute values of the positive and negative deviations of the supply voltage of 35kV and above shall not exceed 10% of the nominal voltage; 4.2, the deviation of the three-phase supply voltage of 20kV and below shall be ±7% of the nominal voltage; 4.3, the deviation of the single-phase supply voltage of 220V shall be +7% and -10% of the nominal voltage; therefore, the value is taken as ±5% of the rated voltage, which is a fixed standard value.)
[0075] To ensure that all physical quantities involved in the calculations of the subsequently constructed MPC objective function are dimensionless pure numbers, conforming to the engineering mathematics principle of linear weighted summation, the voltage deviation is calculated using the predicted and reference values of the grid-connected node voltage. The voltage deviation is then normalized to obtain the dimensionless voltage deviation factor. ,in, For the predicted voltage value of the grid-connected node, Reference value for grid-connected node voltage, The upper limit of the allowable voltage deviation at the grid connection node; the dimensionless voltage deviation factor eliminates the dimension of voltage. Impact, range of values is The dimensions are completely consistent with the predicted value of the comprehensive fluctuation impact factor at future moments, allowing for legitimate linear weighted calculations. Then, with minimizing the impact of the impact on the grid voltage as the core objective, voltage stability weighting coefficients and impact suppression weighting coefficients are introduced to construct an MPC objective function that incorporates the predicted value of the comprehensive fluctuation impact factor and the dimensionless voltage deviation factor. Specific algorithm formula:
[0076] ;in, This refers to the prediction time domain, which is 1 minute, an ultra-short-term prediction to ensure real-time performance and adapt to the characteristics of rapid fluctuations in wind and solar power output on site. This refers to the prediction step size, which is 1 step every 10 seconds, matching the terminal's 10ms-level sampling frequency. and These refer to the voltage stability weighting coefficient and the surge suppression weighting coefficient, respectively, to enable proactive prediction of voltage fluctuations and surges, replacing the traditional passive response mode after voltage exceeds the limit.
[0077] The basis for determining the voltage stability weighting factor of 0.6 and the surge suppression weighting factor of 0.4 is as follows:
[0078] According to the safety specifications for grid-connected operation of power systems, the control priority of voltage stability at grid-connected nodes is higher than that of fluctuation and impact suppression. Following the principle that the weighting coefficient for voltage stability is greater than that for impact suppression, two sets of candidate weighting combinations are randomly selected. Substituting the values into the MPC objective function, simulation iterations were conducted. The voltage exceedance rate, energy storage regulation frequency, and impulse suppression effect were statistically analyzed under the two sets of candidate weight combinations. Simulation comparisons revealed that when the voltage stability weight coefficient was... The shock suppression weighting coefficient is At that time, the impact suppression effect is insufficient, and energy storage regulation is too frequent; when the voltage stability weighting coefficient is The shock suppression weighting coefficient is At that time, the voltage over-limit rate was the lowest, the impact suppression effect was the best, and the energy storage operation frequency was moderate. Combined with on-site verification of typical extreme operating conditions in windy northern regions during spring strong winds and summer cloud cover, candidate weight combinations were identified. It can quickly smooth out voltage fluctuations and accurately respond to impact changes, adapting to actual power plant operating scenarios, therefore... These are respectively used as the voltage stability weighting coefficient and the surge suppression weighting coefficient;
[0079] During the implementation of the objective function, multiple constraints are set simultaneously to ensure overall feasibility and security. These constraints include:
[0080] First, grid-connected node voltage constraints are implemented. This involves acquiring grid-connected node voltage and historical grid-connected node voltage data, statistically analyzing the long-term grid-connected node voltage fluctuation range of the power plant, and combining this with a sample database of operating conditions used in determining the strong impact threshold. Dangerous voltage limits for grid-connected nodes are then eliminated to ultimately obtain the lower limit of the grid-connected node voltage. and upper limit To ensure that the predicted voltage value of the grid-connected node always remains within the safe operating range, the specific formula is as follows: In the formula, This refers to the predicted voltage value at the grid-connected node.
[0081] Second, the charging and discharging power of energy storage devices is constrained. Relevant energy storage data is obtained, and the actual capacity of the power station's energy storage devices is used as the basis: the lower limit is set to 0, limiting the operating range of the energy storage devices; the rated maximum charging and discharging power of the energy storage devices... That is, the upper limit, which takes the value of The rated maximum charge and discharge power of energy storage equipment is mainly determined by the parameters on the energy storage compartment nameplate and the power station design documents. This is to avoid exceeding the hardware capabilities of the energy storage equipment and causing overload failures. The specific formula is as follows: In the formula, Predicted charging and discharging power values for energy storage devices;
[0082] Third, constrain the charge and discharge rates of energy storage devices and obtain constraint values. Historical charge and discharge data of energy storage devices, constraint values Based on the safety operation specifications and long-term field operation experience of energy storage equipment manufacturers, and combined with historical charge and discharge data of energy storage equipment, the maximum power change rate that will not cause damage to the battery and inverter of the energy storage equipment is calculated. This limits the speed of energy storage operation and prevents sudden power changes from causing equipment overheating, lifespan degradation, or secondary grid impact. The specific formula is as follows: ;
[0083] S2.2, targeting Millisecond-level acquisition cycle and The second-level prediction step size has a 1000-fold scale difference. To avoid information loss and quantization errors during the compression of high-resolution original data, we first... The raw data undergoes multi-scale adaptation preprocessing: A 10-second single-step aggregation window is used, and then a sliding weighted average aggregation is applied to the 1000 points of raw data collected in 10ms within the single-step aggregation window, retaining the effective values of subsequent impact amplitude, impact rate, and strong impact probability. Simultaneously, a second-order Butterworth low-pass anti-aliasing filter is used to remove millisecond-level random noise and anomalous jumps. Through a feature anchoring strategy, impact assessment data of impact amplitude, impact rate, and strong impact probability are retained, strictly controlling the multi-scale conversion quantization error to within 0.5%, achieving lossless mapping from millisecond-level collected data to second-level prediction input, completely solving the problem of information attenuation and error amplification caused by thousand-fold scale differences. Based on this, since the prediction time domain is 1 minute and the single-step prediction step size is 10 seconds / step, the prediction time point is calculated according to the prediction time domain and prediction step size. ,correspond to There are a total of 6 prediction time points, which is the total number of prediction time points. The predicted values of the comprehensive volatility impact factor are iterated through in the prediction time domain. The predicted values of the comprehensive volatility impact factor corresponding to the 6 prediction time points are extracted and sorted from largest to smallest. The largest predicted impact factor value is selected, which is the value of the impact magnitude. The time difference between two adjacent timestamps is calculated based on a 10ms acquisition period. It takes 10ms. Then, based on the time difference, calculate the normalized wind power output change rate of the previous data collection period. and normalized photovoltaic power output change rate The predicted value of the comprehensive fluctuation impact factor for the previous data collection period is calculated by combining the normalized wind power and photovoltaic output change rates from the previous data collection period with the final weighting coefficients. The change in the impact factor at adjacent times is calculated using the predicted value of the comprehensive fluctuation impact factor from the previous data collection period and the comprehensive fluctuation impact factor itself. Then, divide the change in the impact factor between adjacent moments by the time difference to obtain the rate of change of the impact factor over time, which is the impact rate. ;
[0084] Impact assessment data refers to a three-dimensional assessment index derived from comprehensive fluctuation impact factors: impact amplitude equals the maximum impact intensity, impact rate equals the rate of impact change, and the probability of a strong impact equals the confidence level of impact occurrence. It is primarily used to support adaptive energy storage regulation, serve as feature input for reinforcement learning models, and enable proactive prediction of grid impact risks. It comprehensively characterizes impact features from three dimensions: intensity, rate, and risk probability, enabling prediction followed by regulation, replacing the traditional passive response mode after voltage exceedance. If only a single impact factor is used, it is impossible to distinguish between impact intensity, rate of change, and probability of occurrence; energy storage regulation lags significantly, voltage fluctuations cannot be mitigated, and grid impact risks cannot be prevented in advance.
[0085] The comprehensive fluctuation impact factor prediction value and the strong impact threshold are used to determine whether there is a strong impact risk on the power grid at six predicted time points within the next minute. When the comprehensive fluctuation impact factor prediction value is greater than or equal to 0.8, it is determined that there is a strong impact risk on the power grid at the six predicted time points within the next minute. The number of time points with strong impact risk is recorded. Based on the temporal distribution characteristics of equally spaced prediction time points, a calculation is performed based on the number of time points with strong impact risk and the total number of prediction time points. Furthermore, by combining the statistical patterns of the correlation between fluctuation shocks and strong shock events in the long-term historical operating data of the power plant over the past year, Mapped to strong impact probability The strong impact probability is defined by an engineering probability standard, referring to the confidence level of the occurrence of a strong impact event on the power grid caused by wind and solar power output fluctuations within the predicted time domain of one minute in the future, rather than the instantaneous probability of occurrence at a single moment. This engineering probability definition, based on the statistical characteristics of sampling at equal intervals, can intuitively quantify the overall risk level of a strong impact. When the strong impact probability is closer to 1, it indicates a wider time range covered by the strong impact within the next minute and a higher overall impact risk, which can be legally used as probabilistic feature input data for subsequent reinforcement learning models, avoiding semantic confusion and feature misuse. When impact assessment data of impact amplitude, impact rate, and strong impact probability are obtained, to achieve accurate and adaptive buffering of wind and solar power fluctuations, a dynamic control model for the charging and discharging power of energy storage equipment is constructed based on a comprehensive fluctuation impact factor. This model directly couples the comprehensive fluctuation impact factor with the remaining capacity of the energy storage equipment, using the real-time comprehensive fluctuation impact factor as the core of control, combined with the rated maximum charging and discharging power of the energy storage equipment. Introducing an adaptive adjustment coefficient To achieve flexible matching of buffering force, the charging and discharging power of the energy storage device at the current moment is calculated by linking the comprehensive fluctuation impact factor, the rated maximum charging and discharging power of the energy storage device, and the adaptive adjustment coefficient. ,in, This is the SOC closed-loop correction coefficient. hour ; Do not discharge. Charging is prohibited. The energy storage device adjusts its charging and discharging power at the current moment, down to the millisecond level; the adaptive adjustment coefficient is set to a value. This formula breaks the limitation of fixed power output, enabling the energy storage output to be dynamically adjusted in real time according to the intensity of the grid impact. The more severe the impact, the closer the output power of the energy storage device is to the rated maximum charging and discharging power of the energy storage device. While quickly smoothing out voltage fluctuations at the grid connection node, it also takes into account the operating life of the energy storage device.
[0086] The adaptive adjustment coefficient refers to the flexible adjustment coefficient of energy storage set according to the impact level: 0.95 for strong impact, 0.6 for medium impact, and 0.3 for weak impact. It is mainly used to dynamically match the impact intensity and calculate the real-time charging and discharging power of energy storage equipment. It can follow the principle that the stronger the impact, the greater the buffering force, taking into account the voltage smoothing effect and energy storage life, and reserving a 5% safety margin to avoid energy storage overload. If a fixed adjustment coefficient is used, insufficient adjustment during strong impact will lead to voltage exceeding the limit, while excessive adjustment during weak impact will aggravate the loss of energy storage life and increase the equipment failure rate by 50%.
[0087] The core principles for determining the adaptive adjustment coefficient must simultaneously satisfy: ① the principle that the stronger the impact, the greater the buffering force; ② the principle that it does not exceed the rated charging and discharging power and rate constraints of the energy storage; ③ the principle of smoothing voltage fluctuations while minimizing energy storage lifespan loss; ④ the principle of reserving a safety margin to avoid overloading the energy storage equipment. The adaptive adjustment coefficient is then quantified and derived based on the impact level: In a strong impact scenario, the risk of grid voltage exceeding the limit is greater than or equal to 90%, requiring the energy storage equipment to buffer rapidly at near full power. Considering the hardware limit of the energy storage equipment's rated maximum charging and discharging power of 200MW, a 5% safety margin is reserved to prevent overloading, thus deriving an adaptive adjustment coefficient of 0.95, which can output 190MW of buffering power to quickly smooth out voltage exceeding the limit caused by a strong impact; In a moderate impact scenario, the voltage fluctuation is within a safe range, and full-power buffering is not required. The goal is to balance the buffering effect with the lifespan of the energy storage equipment, thus deriving an adaptive adjustment coefficient of 0. 6. It can output 120MW of buffer power, balancing the impact mitigation effect with the economic efficiency of energy storage operation. In the case of a weak impact scenario, there is no risk of voltage fluctuation, and only minor adjustments to the energy storage equipment are required to minimize energy storage operation losses. Therefore, the adaptive adjustment coefficient is derived to be 0.3, which can output 60MW of buffer power, avoiding excessive adjustment that would damage the energy storage life. Multiple candidate adaptive adjustment coefficient combinations (0.9 / 0.5 / 0.2, 0.95 / 0.6 / 0.3, 1.0 / 0.7 / 0.4) were set up for closed-loop simulation. The simulation results show that the candidate adaptive adjustment coefficient combination (0.95 / 0.6 / 0.3) has the lowest voltage over-limit rate (0.3%), the best energy storage operation frequency, and a 40% reduction in life loss. Other candidate adaptive adjustment coefficient combinations have the problems of insufficient buffering leading to voltage over-limit and excessive adjustment damaging the energy storage equipment. Therefore, (0.95 / 0.6 / 0.3) was determined as the adaptive adjustment coefficient for different impact scenarios.
[0088] As described above, after achieving dynamic regulation of the charging and discharging power of energy storage equipment and precise buffering of different levels of impact, this project proactively predicts impacts by dynamically adjusting the final weighting coefficient, using ultra-short-term forecast data (including wind speed, irradiance, grid connection node voltage, and predicted wind and solar power output), and incorporating a comprehensive fluctuation impact factor into the MPC objective function. Combined with real-time fluctuation data of wind and solar power in a standardized dataset, this achieves proactive prediction of impact parameters and early activation of energy storage buffering. This completely solves the core shortcomings of traditional methods, which cannot accurately assess the magnitude, rate, and probability of impacts and can only passively respond after voltage exceeds limits. Furthermore, the comprehensive fluctuation impact factor is not a single parameter but incorporates the characteristics of sudden changes in wind power output in windy northern regions, and its final weighting coefficient... and It can be dynamically adjusted according to the ratio of wind and solar power output of the power station. It can adapt to the strong impact of wind power caused by sudden changes in wind speed in spring, as well as the sudden changes in photovoltaic power output caused by cloud cover in summer. It perfectly adapts to the characteristics of wind and solar power output fluctuations in different seasons and at different times. Compared with the traditional single parameter (such as only using the rate of change of wind power output) to assess the impact, it improves the overall assessment accuracy and adaptability, and can solve the problem of wind power impact in windy areas in the north.
[0089] This step fully utilizes the comprehensive fluctuation impact factor defined in S1, without adding or changing any physical quantities, to achieve direct linkage between impact quantification and energy storage device charging and discharging regulation; at the same time, the impact assessment data and comprehensive fluctuation impact factor time series data generated in this step will be completely input into S3 and used as the core embedded quantity of the spatiotemporal feature matrix.
[0090] The spatiotemporal feature matrix refers to a multi-dimensional coupling matrix constructed with a 72-hour time window. It includes three types of features horizontally: power output, meteorology, and impact, and vertically, it contains full-time-series samples. It is mainly used to integrate multi-level data from terminals, edge devices, and the cloud as the core feature carrier for power prediction and reinforcement learning. This can clarify the spatiotemporal distribution pattern of wind and solar power output, break down the isolation barriers of multi-level data, and solve the core defect of fuzzy spatiotemporal features in traditional prediction. Without the spatiotemporal feature matrix, if only a single time-series or spatial feature is used, the prediction error is greater than or equal to 15%, which cannot adapt to the power output fluctuation pattern caused by terrain differences and weather changes.
[0091] Step S3: Obtain long-term historical data of the power plant, and perform time alignment and spatial aggregation of the long-term historical data with the impact assessment data and the standardized dataset to extract spatiotemporal-impact fusion features; based on a preset time window, construct a multi-dimensional coupled spatiotemporal feature matrix from the spatiotemporal-impact fusion features; specifically including the following method steps:
[0092] S3.1. The dynamic regulation of energy storage device charging and discharging power, enabling precise buffering of different levels of impact, indicates that the real-time response to wind and solar power fluctuations has been initially resolved. However, considering the actual operating scenarios of power plants and the scheduling needs of the grid dispatch control center, there are still issues such as different wind speed distributions due to terrain variations in different wind farms; and inconsistent power output fluctuations when photovoltaic power plants are affected by weather. Furthermore, the previous lack of a multi-level collaborative mechanism between the cloud, edge, and terminal levels prevented effective data exchange, leading to significant power prediction errors. This directly resulted in the grid dispatch control center needing to reserve 200MW of thermal power reserve capacity (accounting for 20% of the total installed capacity of the power plant) to compensate for prediction deviations, significantly increasing the overall operating costs of the power plant and squeezing the space for renewable energy consumption. To address these issues, the grid dispatch control center first automatically receives impact assessment data from each edge terminal, and simultaneously receives standardized datasets synchronously transmitted from the edge terminals. The cloud automatically retrieves its own stored long-term historical data, including wind power output, wind speed, and photovoltaic irradiance under different seasons and weather conditions over the past year, with a 72-hour time window. The wind power output transmitted from the edge within 72 hours Wind speed Irradiance and comprehensive volatility shock factor Furthermore, the impact assessment data and the long-term historical data in the cloud are uniformly aligned to the minute-level time dimension to eliminate data misalignment caused by time difference. The aligned operational data strictly retains the physical uniqueness of the comprehensive fluctuation impact factor within 72 hours during the alignment process. Then, the average wind speed and average irradiance of the single wind turbine and single photovoltaic power generation equipment at the edge are spatially summarized to obtain the spatially summarized operational data, which is consistent with the regional or power station-level dimensions in the long-term historical data in the cloud. The regional or power station-level refers to the spatial scale after data summarization.
[0093] The spatial locations of wind turbines and photovoltaic equipment are obtained, and the space is divided into grids according to the physical installation coordinates of the wind turbines and photovoltaic arrays, based on the wind speed over a 72-hour period. Using this as the data source, spatial statistical analysis and correlation calculations are used to quantify the regional distribution patterns. The mean and variance of wind speed for each wind turbine are statistically analyzed grid by grid over 72 hours to obtain the wind speed statistics for each grid, characterizing the differences in wind speed amplitude across different terrain regions. The Pearson correlation coefficient is used to calculate the correlation coefficient of wind speed between adjacent wind turbines, quantifying the regional distribution differences caused by terrain shading and wind turbulence. The spatial location of wind turbine equipment, the wind speed statistics for each grid, and the correlation coefficients of wind speed between adjacent wind turbines are concatenated in a fixed-dimensional order to form a spatial feature vector, comprehensively representing the regional distribution patterns and spatial correlations of wind speed. Furthermore, combined with the spatial locations of wind turbines and photovoltaic equipment, the spatial correlation of wind turbine and photovoltaic equipment output is extracted; based on the wind power output over 72 hours... and irradiance This study analyzes the fluctuation patterns of wind and solar power output in different seasons and intraday periods, extracts time feature vectors, combines comprehensive fluctuation impact factors, explores the temporal correlation between wind and solar power output fluctuations and impact intensity, quantifies the wind and solar power generation patterns using spatial and temporal feature vectors, and derives the spatiotemporal characteristics of wind and solar power output. The comprehensive fluctuation impact factor, impact amplitude, impact rate, and strong impact probability constitute the impact factor features.
[0094] The spatiotemporal characteristics of wind and solar power output and the characteristics of impact factors are aligned to the same timestamp and subjected to maximum and minimum normalization to ensure that the numerical ranges of the two types of characteristics are consistent. The Pearson correlation coefficient is used to calculate the correlation strength between the spatiotemporal characteristics of wind and solar power output and the characteristics of impact factors. The fusion weights are assigned according to the correlation magnitude to strengthen the coupling effect of highly correlated features. The weighted spatiotemporal characteristics of wind and solar power output and the characteristics of impact factors are obtained and then concatenated in an orderly manner according to the dimensions. Redundant and weakly correlated features are removed to finally form spatiotemporal-impact fusion features. Through the spatiotemporal-impact fusion features, the strong correlation between the spatiotemporal distribution of wind speed and the probability of strong impacts, and the strong correlation between the temporal pattern of photovoltaic power output and the impact rate are obtained.
[0095] Spatiotemporal-impact fusion features refer to the feature vector obtained by aligning the spatiotemporal features of wind and solar power output with the features of impact factors in time series and weighting them by correlation. It is mainly used as the core input for training the prediction coefficient matrix and reinforcement learning model to improve the adaptability of prediction to impact. This can deeply couple the spatiotemporal distribution patterns of wind and solar power with the impact risk of the power grid, improving the prediction accuracy by 30% and reducing the prediction error in strong impact scenarios by 60%. If the spatiotemporal features and impact features are used separately, the prediction model cannot associate the spatiotemporal patterns of wind and solar power with the impact risk, the prediction deviation will continue to expand, and it will not be able to adapt to the frequent impact conditions.
[0096] S3.2 To transform the correlation or relevance into standardized input that can be directly used for model training and prediction optimization, a time frame of 72 hours is constructed. The multidimensional coupled spatiotemporal feature matrix F (dimension: ), will measure wind power output within 72 hours Photovoltaic power output, wind speed Irradiance Combined volatility shock factor Time-series deep fusion specifically includes the following methodological steps:
[0097] The matrix is arranged vertically with full-time samples from the most recent moment to within 72 hours. Each set of time-series samples simultaneously includes the main characteristics of wind power and photovoltaic power. The full-time samples refer to samples taken over a 72-hour period. The time window is 1 hour and the sampling step is 1 hour. There are 72 rows in total, corresponding to 72 consecutive sampling times within 72 hours. Each row represents a time series sample at one time.
[0098] From the previous time point Begin by continuously backtracking to moments within a 72-hour period. This matrix forms a complete and continuous time-series data set. The main characteristics of wind power refer to the wind power output, wind speed, and the corresponding comprehensive fluctuation impact factor at each moment. The main characteristics of photovoltaic power refer to the photovoltaic power output, irradiance, and the corresponding comprehensive fluctuation impact factor at each moment. It is horizontally divided into three core feature columns, totaling three columns. The first column carries the wind power output and photovoltaic power output at each moment, completely depicting the temporal fluctuation trajectory of wind power and photovoltaic power output. The second column corresponds to and matches the wind speed and irradiance at each moment, restoring the core meteorological elements that drive the changes in power output of wind turbines and photovoltaics. The third column synchronously embeds the comprehensive fluctuation impact factor at each moment, directly linking the power output fluctuation characteristics with the grid impact intensity. Through this matrix construction process, the scattered spatiotemporal correlation and impact correlation features are integrated into a unified and standardized temporal feature structure, thereby breaking the defects of multi-level data isolation and spatiotemporal feature ambiguity, and providing a regular and high-quality data foundation for subsequent power prediction model optimization.
[0099] Build The specific formula for the multidimensional coupled spatiotemporal feature matrix is as follows:
[0100] The spatiotemporal feature matrix integrates three types of data, originally scattered across terminals, edges, and the cloud, into a single matrix carrier, enabling end-to-end data interoperability between terminal data, edge aggregation, and cloud fusion. The specific integration includes the following methods and steps:
[0101] For the first time, a spatiotemporal feature matrix is used to deeply bind spatial feature vectors, output spatial correlation, and temporal feature vectors, while incorporating the linkage characteristics of comprehensive fluctuation impact factors. This accurately depicts the complex spatiotemporal characteristics of wind and solar power output in power plants, thus achieving deep coupling of the three types of features. This solves the defects of traditional prediction models, such as unclear spatiotemporal features, isolated multi-level data, and ambiguous spatiotemporal features. Specifically, terminal data is obtained through standardized datasets collected by sensors and transmitted via edge devices; edge data includes impact assessment data and operational data spatially summarized at the edge; and cloud data includes long-term historical data. The three types of data from the terminal, edge, and cloud are time-aligned and spatially summarized again to obtain aligned and summarized data. Noise is then removed from the aligned and summarized data, and residual abnormal data (such as jump values caused by extreme weather) is eliminated. The effective data subset is retained and integrated into the spatiotemporal feature matrix, completely breaking down the barriers of previously isolated multi-level data.
[0102] Step S4: Using the long-term historical data and the spatiotemporal-impact fusion features, generate an initial prediction coefficient matrix; combine the spatiotemporal feature matrix and the prediction coefficient matrix to calculate the initial prediction error, and superimpose the initial prediction error to form an ultra-short-term power prediction value; extract the peak value of the ultra-short-term prediction error and calculate the actual required thermal power reserve capacity for rapid replenishment when a power output gap occurs. Specifically, this includes the following steps:
[0103] By acquiring nearly one year of long-term historical data from the power station via the cloud, comprehensively covering the power station's operating patterns under all conditions, a sufficient data foundation is provided for training and fitting the prediction coefficient matrix. Using spatiotemporal-impact fusion features as the training input feature set, a multiple linear regression model based on the least squares method is employed to generate the prediction coefficient matrix. Specifically, the following steps are implemented: Obtaining the power station's historical actual total power generation output. (Including wind power output and photovoltaic power output) are used as the dependent variable, and combined with the independent variables of spatiotemporal-impact fusion characteristics, a multiple linear regression model is constructed: ;in, The prediction coefficient matrix, i.e., the parameter vector to be solved, represents the contribution weight of historical features at each moment in the spatiotemporal feature matrix to the current predicted value. It is a spatiotemporal characteristic matrix, consisting of wind power output, wind speed, irradiance, photovoltaic output, and comprehensive fluctuation impact factor arranged in chronological order within a 72-hour time window. The dataset is a standardized dataset for the current moment, containing normalized operational data such as real-time wind speed and irradiance; it uses the prediction coefficient matrix to be solved as a link, combined with the spatiotemporal feature matrix. Coupled with the standardized dataset, the basic predicted value θ is obtained. , This represents the theoretical prediction of the model under parameter θ.
[0104] Then, based on the actual total power generation in history Compared with model predictions The optimization objective is to minimize the sum of squared errors between the two sides, and a loss function is constructed accordingly. ; This represents the total deviation between the model's theoretical predictions and historical true values for any set of θ values. By taking the partial derivative of the loss function with respect to θ and setting it to zero, the optimal prediction coefficient matrix that minimizes the sum of squared global errors can be obtained through a closed-form solution. ,in, This refers to the transpose operation. This closed-form solution guarantees that the prediction coefficient matrix θ is the globally optimal parameter, and the best-fit coefficients in the least squares sense can be obtained without iteration.
[0105] Specifically, the dimension definition of a standardized dataset is: to standardize the dataset into... Feature column vectors (dimension: The optimal prediction coefficient matrix θ, which is strictly matched with the number of columns (3 columns) of the spatiotemporal feature matrix F, is substituted into the hypothesis function based on the dimension matching rule to calculate... The initial prediction error is calculated by comparing the total power output with the baseline forecast. The initial prediction error directly quantifies the degree of deviation between the model's preliminary prediction results and the actual operating conditions of the power plant. Moreover, the initial prediction error is controlled within 8%, which will directly serve as the basis for correcting the reinforcement learning model. Through iterative optimization, the deviation will be continuously reduced, thereby achieving an adaptive improvement in prediction accuracy.
[0106] The prediction coefficient matrix refers to the coefficient matrix obtained by closed-form calculation based on the least squares method of multiple linear regression, with dimensions... It serves as a core bridge connecting spatiotemporal features and power prediction; it is mainly used to calculate the basic power prediction value, which serves as the optimization object of the reinforcement learning model and dynamically corrects the prediction error; this allows for data-driven closed-loop solutions, without hyperparameters or iterative oscillations, providing a stable and accurate initial starting point for reinforcement learning, thus improving the convergence speed by 50%; if empirical coefficients are used, and the basic prediction error is greater than or equal to 12%, the reinforcement learning iteration will continue to diverge and fail to converge, and the power prediction will completely fail.
[0107] By adding the initial prediction error to the base prediction value, the final ultra-short-term power prediction value is formed. Formula for ultra-short-term power prediction algorithm: , where k represents the discrete time point number (step index) in the prediction time domain, and t+k represents the kth prediction time after the current time t. This corresponds to the prediction time points in 1-minute increments over the next 60 minutes; the ultra-short-term prediction period of the next 60 minutes is the optimization range. Prediction deviation control and power grid impact risk are coupled into an integrated reward and punishment criterion to obtain the ultra-short-term prediction error. First, the ultra-short-term prediction error is normalized to eliminate the influence of power dimensions, thus obtaining the normalized ultra-short-term prediction error. Then, by accumulating the squared terms of the ultra-short-term prediction error point by point... The overall prediction deviation is quantified to represent the degree of deviation at the k-th prediction time after the current time t, and a correction coefficient is introduced. The probability of a strong short-term shock By correlating the normalized ultra-short-term prediction error with the data, a dimensionless reinforcement learning reward function is constructed. In particular, under the scenario of moderate to strong impact of spring gales, the reinforcement learning reward function will automatically increase the penalty for prediction deviation, guide the algorithm to iterate towards higher accuracy, rely on the reward mechanism to continuously guide and dynamically optimize the prediction coefficient matrix, and finally bring the prediction error to within 5%, accurately matching the accuracy requirements of the power grid dispatch and control center.
[0108] Obtain the correction coefficient The basis for the value of 0.5 is:
[0109] The correction coefficient is determined automatically through multiple simulation iterations based on the long-term historical data of the power plant over the past year, the correlation between the probability distribution of strong impacts and prediction errors, and the correction weights of prediction errors and strong impact probabilities. A coefficient of 0.5 is used as the equilibrium coefficient to avoid a single factor from dominating the iteration and to ensure the stable convergence of the reinforcement learning model. By matching the strong impact characteristics of wind turbines and photovoltaic equipment in windy areas of northern China and combining the impact level, the error penalty under strong impact scenarios is made moderate. After on-site simulation verification of the power plant, when the correction coefficient is 0.5, the prediction error convergence speed is the fastest and there is no parameter oscillation. It is most suitable for the power plant output fluctuations and the accuracy requirements of the power grid dispatch and control center, and fits the dynamic buffering characteristics of energy storage equipment, so as to achieve the synergistic optimization of prediction correction and impact mitigation.
[0110] Once the reinforcement learning reward function is known, the gradient is obtained by taking the partial derivative of the reward function with respect to the prediction coefficient matrix. Then, the dynamic adjustment amount of the prediction coefficient matrix is calculated based on the gradient. And as an action for subsequently building a reinforcement learning model, where, This refers to the prediction coefficient matrix from the previous round. During the iteration process, the initial prediction error is used to correct the prediction accuracy, thereby improving the time-series prediction accuracy of the comprehensive volatility impact factor. This refers to the learning rate. The learning rate is obtained through the following steps: The long-term historical data of the power station over the past year is used as the basic dataset for testing and screening the learning rate. Combining the spatiotemporal-impact fusion characteristics and the power output characteristics of sudden wind and solar changes in windy areas in the north, multiple candidate learning rates are set to conduct pre-iterative simulations. The update stability of the prediction coefficient matrix and the convergence speed of the prediction error are observed under different values. Subsequently, using the impact assessment data and the constraints of the charging and discharging power and the charging and discharging rate of the energy storage equipment, values that would cause the model to oscillate and diverge or converge too slowly are eliminated, and the optimal value range is locked. Finally, combined with the needs of the power grid dispatch and control center in the hourly prediction time domain, after multiple rounds of simulation verification, the learning rate that balances iterative stability and convergence efficiency is determined to be 0.01.
[0111] The initial prediction error is normalized to obtain a normalized initial prediction error. This normalized initial prediction error is then used as the state input for constructing the reinforcement learning model. The dynamic adjustment is used as the action output for constructing the reinforcement learning model, and the reinforcement learning reward function serves as the iterative feedback. The gradient guides the iteration direction, and the action magnitude is calculated using the learning rate to construct the reinforcement learning model. The state space is a two-dimensional state space with 2×1 column vector dimensions, and the action space refers to the prediction coefficient matrix... The dimensional row vectors are therefore the action space is A two-dimensional matrix, where each dimension corresponds to a single-parameter adjustment of the prediction coefficients, is used to achieve fine-grained optimization of the coefficient matrix; the two-dimensional state is used as input to the reinforcement learning model. The dimensionless action is used as the output of the reinforcement learning model; the dimensionless reward function is used as the iterative feedback, the gradient is calculated through the policy function to guide the iterative direction, and the action magnitude is determined by the learning rate; the convergence criterion is: when the ultra-short-term prediction error is stably converged to within 5% for 5 consecutive rounds, the iteration is determined to be complete and the optimization is stopped.
[0112] By conducting full-process iterative optimization using the prediction coefficient matrix, spatiotemporal-impact fusion features, and nearly one year of historical data from the power plant, the reinforcement learning model continuously corrects its output actions using a reinforcement learning reward function during the iteration process. The parameters of the prediction coefficient matrix are dynamically adjusted to gradually compress the prediction error amplitude. After multiple rounds of adaptive iteration, the reinforcement learning model dynamically optimizes the prediction coefficient matrix to its optimal state, resulting in an optimized prediction coefficient matrix. The prediction error data after iteration is then recorded, and the prediction error stabilizes and converges to within 5%, accurately matching the precision requirements of the power grid dispatch and control center. Using the optimized prediction coefficient matrix, the constructed spatiotemporal feature matrix, and an effective data subset, combined with the prediction error data after the reinforcement learning model iteration, optimized power output prediction values are output (for medium to long term: outputting the power grid dispatch plan for the next 24 hours to determine the power plant's grid connection power output plan for the next day; for ultra-short term: outputting the power output allocation plan for the next hour). The optimized power output prediction values include wind power output prediction values and photovoltaic power output prediction values, with the prediction error controlled within 5%, meeting the precision requirements of the power grid dispatch and control center.
[0113] The peak value of the ultra-short-term forecast error over the next 60 minutes is obtained as a dynamic driving factor for increasing reserve capacity. The maximum forecast error is then extracted from the peak value. The actual configured thermal power reserve capacity is calculated by using the maximum prediction error. The system calculates the wind power output difference and photovoltaic (PV) output difference using the wind power output and PV output forecast values for the next hour, respectively, and sets wind power output deviation ranges and PV output deviation ranges. It then compares the wind power output difference and PV output difference values for the next hour with these respective deviation ranges to determine if any deviation exists. If the wind power output difference and PV output difference exceed their respective deviation ranges, it is determined that there is a deviation between the wind power output and PV output forecast values for the next hour and the optimized output forecast values, indicating a short-term power output gap in the grid. The grid dispatch and control center then directly calls upon the actual configured thermal power reserve capacity, ordering thermal power units to quickly increase output to fill the gap, avoiding frequency and voltage fluctuations at grid connection nodes and ensuring grid power supply security. This refers to the minimum reserve capacity of thermal power plants. The minimum reserve capacity of thermal power plants is... The basis includes the following methods and steps: calling up the long-term historical data of the power station for the past year, statistically analyzing the minimum basic fluctuation amplitude of wind power output and photovoltaic output under different seasons and weather conditions and the minimum power gap under extreme conditions, quantifying the normal lower limit of wind and solar power output fluctuations, and providing actual operational data support for the reserve capacity threshold.
[0114] Error coefficient for Basis for obtaining:
[0115] Based on the actual patterns of maximum prediction error and large-capacity thermal power reserve, and combined with the initial prediction error, a preliminary mathematical relationship is established for the linear response of reserve capacity with prediction error. This yields a candidate range for the error coefficient, completing the initial locking of the coefficient range. Utilizing the charging and discharging power constraints and charging and discharging rate constraints of energy storage devices, the output fluctuation range that energy storage can mitigate in real time is calculated. The output fluctuation compensation that energy storage can bear is deducted from the thermal power reserve demand, thereby compressing the candidate range of the error coefficient and eliminating unreasonable values that are too large or too small. Through spatiotemporal-impact fusion characteristics, targeting the moderate-to-strong impact probability of spring strong winds in windy areas of northern China, the additional safety margin required for reserve capacity under moderate impact scenarios is calculated. The error coefficient range is further fine-tuned to ensure that the error coefficient value meets safety requirements without causing reserve redundancy. Finally, using the last 1 years of long-term historical data, multiple rounds of closed-loop simulations are conducted on the candidate coefficients. Simulation verification shows that when… When the prediction error converges to within 5%, the thermal power reserve capacity can be stabilized at around 60MW, taking into account both the grid safety bottom line and the demand for new energy consumption. At the same time, the reinforcement learning model iterates without oscillation and adjusts the response speed to match the scheduling rhythm, thus automatically locking the error coefficient to 2.
[0116] This case utilizes the spatiotemporal feature matrix of S3 and the prediction coefficient matrix of S4 to dynamically correct prediction errors through a multi-level collaborative architecture of the reinforcement learning model. This adapts to the complex spatiotemporal characteristics and frequent impacts of power plants. Furthermore, the reinforcement learning reward function is optimized with the probability of strong impacts. This solves the shortcomings of traditional prediction models, such as unclear spatiotemporal features and large errors. By using a multi-level collaborative mechanism, the automatic interoperability and fusion of standardized terminal datasets, impact assessment data from the edge, and long-term historical data from the cloud are achieved. This makes the reinforcement learning model more closely aligned with the actual operating scenarios of power plants, thereby improving prediction accuracy.
[0117] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms, characterized in that: The methods and steps include the following: S1. Collect operational data related to wind power generation and photovoltaic power generation, and after denoising and normalization processing, obtain a standardized dataset containing wind power output data and photovoltaic output data. Based on the dynamic deviation between the real-time wind power output ratio and the benchmark wind power output ratio in the standardized dataset, the initial weighting coefficients of wind power and photovoltaic power are adaptively adjusted to obtain the final weighting coefficients of wind power and photovoltaic power. Based on the wind power output change rate and photovoltaic output change rate in the standardized dataset, and combined with the final weighting coefficients of wind power and photovoltaic power, a comprehensive fluctuation impact factor is calculated to quantify the impact intensity of wind-solar coordinated fluctuations on the power grid; and the corresponding impact level is classified according to the magnitude of the comprehensive fluctuation impact factor. S2. Based on the standardized dataset, the rolling prediction method is used to predict the wind power output and photovoltaic output at future times, and combined with the final weighting coefficients of wind power and photovoltaic power, the predicted value of the comprehensive fluctuation impact factor at future times is calculated. The comprehensive fluctuation impact factor at the current moment and the comprehensive fluctuation impact factor of the previous acquisition period are obtained. Combined with the predicted value of the comprehensive fluctuation impact factor at the future moment, impact assessment data for characterizing the impact characteristics are calculated. The impact assessment data includes at least the impact amplitude, impact rate and strong impact probability. Based on the impact level, a corresponding adaptive adjustment coefficient is selected. The adaptive adjustment coefficient is an energy storage power adjustment coefficient independent of the weighting coefficient. Based on the comprehensive fluctuation impact factor, the adaptive adjustment coefficient, and the rated charging and discharging power of the energy storage device, the charging and discharging power of the energy storage device at the current moment is dynamically calculated, and the energy storage device is controlled to adjust its charging and discharging state according to the charging and discharging power. S3. Obtain long-term historical data of the power plant, and perform time alignment and spatial aggregation of the long-term historical data with the impact assessment data and the standardized dataset to extract spatiotemporal-impact fusion features; Based on a preset time window, the spatiotemporal-impact fusion features are constructed into a multidimensional coupled spatiotemporal feature matrix; S4. Using the long-term historical data and the spatiotemporal-impact fusion features, generate an initial prediction coefficient matrix; combine the spatiotemporal feature matrix and the prediction coefficient matrix to calculate the initial prediction error, and superimpose the initial prediction error to form an ultra-short-term power prediction value; extract the peak value of the ultra-short-term prediction error and calculate the actual thermal power reserve capacity that needs to be configured, so as to quickly make up for the power output gap.
2. The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms according to claim 1, characterized in that: The method for determining the final weighting coefficients for wind power and photovoltaic power includes the following steps: The initial weighting coefficients for wind power generation and photovoltaic power generation are set based on the proportion of total installed capacity of wind power generation and photovoltaic power generation in the power station. The real-time wind power output ratio is calculated based on the wind power output and photovoltaic output in the standardized dataset, and the output deviation is calculated based on the real-time wind power output ratio and the benchmark wind power output ratio. By calculating and adjusting the weighting coefficients based on the output deviation, and then dynamically adjusting the corresponding weighting coefficients in conjunction with the initial weighting coefficients, the final weighting coefficients for wind power generation and photovoltaic power generation are obtained.
3. The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms according to claim 1, characterized in that: The comprehensive volatility impact factor is calculated through the following steps: Obtain the comprehensive volatility impact factor calculated in step S1; The normalized wind power output change rate and the normalized photovoltaic output change rate are obtained by calculating the two types of output change rates separately and normalizing them. The absolute values of the normalized wind power output change rate and the normalized photovoltaic output change rate are taken respectively. The two types of absolute output change rates are weighted and coupled together by the final weight coefficients of wind power and photovoltaic power to construct a comprehensive fluctuation impact factor.
4. The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms according to claim 1, characterized in that: The calculation of the impact assessment data and the charging and discharging power of the energy storage device includes the following methods and steps: The total number of prediction time points is obtained by predicting the time domain and the prediction step size, and the impact magnitude is obtained by iterating through the prediction value of the comprehensive fluctuation impact factor. The impact rate is obtained by calculating the change in the impact factor between adjacent time points. The probability of a strong impact is calculated based on the total number of predicted time points, and impact assessment data is formed by combining the impact amplitude and impact rate. Based on the calculated comprehensive fluctuation impact factor coupled with the adaptive adjustment coefficient γ, a dynamic control model is constructed and the charging and discharging power of the energy storage device is calculated.
5. The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms according to claim 4, characterized in that: The adaptive adjustment coefficient includes the following steps: The corresponding adaptive adjustment coefficient is selected according to the impact level: Set strong impact threshold and low impact threshold. When the comprehensive fluctuation impact factor is greater than or equal to the strong impact threshold, record the corresponding adaptive adjustment coefficient. When the comprehensive volatility impact factor is greater than or equal to the low impact threshold and less than the strong impact threshold, the corresponding adaptive adjustment coefficient is recorded. When the comprehensive fluctuation impact factor is less than the low impact threshold, the corresponding adaptive adjustment coefficient is recorded.
6. The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms according to claim 1, characterized in that: The extraction of spatiotemporal-impact fusion features includes the following method steps: Obtain the comprehensive wave impact factor calculated in step S1 and the impact assessment data calculated in step S2; The impact assessment data is time-aligned with the standardized dataset, and combined with the long-term historical data of the power plant, time-series alignment and spatial aggregation are performed based on a preset time window. Spatial feature vectors are extracted based on wind speed, and temporal feature vectors are extracted based on wind power output and irradiance. The spatial feature vectors and temporal feature vectors are fused to obtain the spatiotemporal features of wind and solar power output. The comprehensive wave impact factor, impact amplitude, impact rate and strong impact probability constitute the impact factor feature. The impact factor feature is then linked and fused with the spatiotemporal features of wind and solar power output to extract the spatiotemporal-impact fusion feature.
7. The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms according to claim 6, characterized in that: The construction of the multidimensional coupled spatiotemporal feature matrix includes the following steps: The matrix is arranged vertically to display full-time samples, and each set of time-series samples simultaneously includes the main features of wind power generation and photovoltaic power generation. It is divided into three core feature columns, which respectively carry wind power output and photovoltaic output, wind speed and irradiance, and comprehensive fluctuation impact factor at various times.
8. The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms according to claim 1, characterized in that: The generation of the prediction coefficient matrix and ultra-short-term power prediction values includes the following method steps: By combining long-term historical data with spatiotemporal-impact fusion characteristics, a prediction coefficient matrix coupled with multi-level collaborative data is generated; After obtaining the basic prediction value by coupling the calculation with the prediction coefficient matrix as the link, the initial prediction error is then calculated as the state of the reinforcement learning model. The ultra-short-term power prediction is finally formed by superimposing the basic prediction value with the initial prediction error.
9. The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms according to claim 8, characterized in that: The reinforcement learning model includes the following steps: By summing up the squared terms of the ultra-short-term prediction error, a correction coefficient is introduced to correlate the strong shock probability with the ultra-short-term prediction error, and a reinforcement learning reward function is constructed. The gradient is obtained by taking the partial derivative of the reinforcement learning reward function with respect to the prediction coefficient matrix. Then, the dynamic adjustment of the prediction coefficient matrix is calculated based on the gradient, and used as the action to build the reinforcement learning model. Using state as input, action as optimization output, and reinforcement learning reward function as iterative feedback, the reinforcement learning model is constructed by combining gradient and learning rate.
10. The new energy grid-connected voltage control method based on multi-level collaborative and adaptive algorithms according to claim 1, characterized in that: The calculation of the actual required thermal power reserve capacity includes the following steps: The prediction coefficient matrix, spatiotemporal-impact fusion features, and long-term historical data of the power plant are used for iterative optimization. The prediction error data is recorded after the prediction coefficient matrix is iteratively optimized. The optimized power output prediction value is output by combining the optimized prediction coefficient matrix, spatiotemporal feature matrix and prediction error data; The maximum prediction error is obtained, and the actual configured thermal power reserve capacity is calculated by combining it with the optimized power output prediction value.
