Method for improving source-grid-load-storage integrated regulation capability

Abstract

The invention provides a method for improving source network load storage integrated adjustment capability, and belongs to the technical field of power system automation and energy management. Multi-side data acquisition terminals are deployed to acquire power grid operation parameters in real time and perform adaptive extended Kalman filtering noise elimination; a source-load-storage dynamic coupling matrix is constructed based on historical data, adjustment weight coefficients of all links are calculated, and an adaptive prediction fusion model is used for energy storage state estimation and new energy output prediction. An energy storage power compensation instruction, an adjustable load starting instruction and a source load storage scheduling instruction are generated in a grading manner according to the magnitude of power grid frequency deviation by adopting a hierarchical cooperative regulation and control strategy, and an adjustment weight coefficient and a control instruction are iteratively optimized according to an adjustment effect, so that the technical problem of insufficient source network load storage multi-link cooperative adjustment capability is solved.

Description

Technical Field

[0001] This invention belongs to the field of power system automation and energy management technology, and specifically relates to a method for improving the integrated regulation capability of power generation, grid, load and storage. Background Technology

[0002] In the field of power system frequency and voltage regulation control, traditional methods mainly rely on the automatic generation control and primary frequency regulation functions of thermal power units to achieve grid frequency and voltage stability. With the large-scale grid connection of new energy sources, integrated regulation of generation, grid, load, and storage has become an important means to improve grid flexibility. Existing technologies employ single-stage control strategies such as rapid response of energy storage, adjustable load demand-side management, and new energy power prediction, achieving grid balance by independently optimizing the regulation resources on each side. However, traditional single-stage control methods lack a multi-stage coordination mechanism involving generation, load, and storage. The regulation weights of each stage are fixed and cannot be dynamically adjusted according to the degree of grid disturbance, resulting in slow response speed during large frequency fluctuations and high regulation costs during small fluctuations. At the same time, existing prediction models are mostly single-task architectures, unable to simultaneously meet the accuracy requirements of energy storage state estimation and new energy output prediction. In other words, existing technologies suffer from insufficient multi-stage coordinated regulation capabilities involving generation, grid, load, and storage. Summary of the Invention

[0003] In view of this, the present invention provides a method for improving the integrated regulation capability of source, grid, load and storage, which can solve the technical problem of insufficient coordinated regulation capability of multiple links of source, grid, load and storage in the prior art.

[0004] This invention is implemented as follows: It provides a method for improving the integrated regulation capability of energy source, grid, load, and storage. This involves real-time data acquisition by data acquisition terminals deployed on the source side, load side, storage side, and grid side. These data include grid frequency deviation, grid voltage deviation, real-time renewable energy output, real-time load power, energy storage battery voltage, energy storage battery current, and energy storage battery temperature. After noise cancellation processing using an adaptive extended Kalman filter algorithm, the data is uploaded to a cloud platform. Based on historical and real-time data, a dynamic coupling matrix of source, load, and storage is constructed to calculate the energy storage regulation weight coefficient, load regulation weight coefficient, and renewable energy regulation weight coefficient. Finally, the energy storage battery voltage, current, and temperature are input into an adaptive predictive fusion model for energy storage. The state estimation calculation outputs the energy storage state of charge, energy storage health status, and predicted energy storage temperature distribution. After collecting and fusing multi-source meteorological data, the data is input into an adaptive prediction fusion model to calculate the new energy output forecast. The model outputs the new energy output forecast mean, upper quantile, and lower quantile, and calculates the prediction uncertainty quantification value. A hierarchical collaborative control strategy is constructed to generate energy storage power compensation commands, adjustable load start commands, or source-load-storage scheduling commands based on the absolute value of the grid frequency deviation. These commands are then sent to the corresponding actuators through the cloud platform, and the adjusted grid frequency deviation and adjusted grid voltage deviation are collected in real time. When the adjusted grid frequency deviation or adjusted grid voltage deviation does not reach the set threshold, the weighting coefficients are recalculated and new control commands are generated iteratively.

[0005] The adaptive extended Kalman filter algorithm includes a state prediction step, a covariance prediction step, a Kalman gain calculation step, a state update step, and a covariance update step. The adaptive mechanism dynamically adjusts the process noise covariance and the observation noise covariance based on the actual observation residuals.

[0006] The construction of the source-load-storage dynamic coupling matrix involves performing multiple regression analysis on historical grid frequency deviation, historical grid voltage deviation, historical renewable energy output, historical load power, and historical energy storage charging and discharging power. This analysis calculates the partial derivatives of energy storage charging and discharging power, adjustable load capacity, and predicted wind and solar power output with respect to grid frequency deviation and grid voltage deviation. After normalization, these derivatives are used to form the energy storage regulation weight coefficient, load regulation weight coefficient, and renewable energy regulation weight coefficient.

[0007] The adaptive prediction fusion model has a multi-branch temporal coding fusion architecture, which includes an input layer, a feature extraction layer, a temporal dependency modeling layer, an attention weighting layer, and a multi-task output layer.

[0008] The input layer receives multi-source heterogeneous time-series data and performs normalization preprocessing, mapping the input data of different physical quantities to a unified numerical space.

[0009] The feature extraction layer uses a one-dimensional convolutional neural network to extract local temporal features, and extracts feature representations of different temporal granularities through multi-scale convolution in parallel.

[0010] The temporal dependency modeling layer constructs a long short-term memory pathway based on the gating mechanism of recurrent neural units, including input gate units, forget gate units, output gate units, and memory units.

[0011] The output value of the input gate unit is equal to the weighted linear combination of the input data and the hidden state of the previous time step by the Sigmoid function, and the output value of the forget gate unit is equal to another weighted linear combination of the input data and the hidden state of the previous time step by the Sigmoid function.

[0012] The update value of the memory unit is the product of the output value of the forget gate unit and the value of the memory unit at the previous time step, plus the product of the output value of the input gate unit and the candidate memory value. The memory unit stores and transmits long-term dependency information in the interval from negative 1 to positive 1 through the tanh activation function.

[0013] The attention weighting layer calculates the importance weights of temporal features through a self-attention mechanism. The importance weights are obtained by normalizing the dot product of the query vector and the key vector using Softmax, and then summed by weight to form the context vector.

[0014] Among them, the multi-task output layer outputs the energy storage state of charge, energy storage health status, and predicted values ​​of energy storage temperature distribution for the energy storage state estimation task, and outputs the predicted mean value of new energy output, the upper quantile value of new energy output, and the lower quantile value of new energy output for the new energy output prediction task.

[0015] The fusion objects of multi-source meteorological data fusion processing include temperature, humidity, wind speed, and irradiance grid data from numerical weather prediction data, cloud cover, cloud height, and cloud type remote sensing data from satellite cloud image data, and real-time meteorological element data from ground observation station meteorological data.

[0016] Among them, the multi-source meteorological data fusion processing uses spatial interpolation to unify data of different resolutions into the target grid, Kalman filtering to fuse data from different sources, and ground observation station meteorological data to correct numerical forecast biases.

[0017] The method for calculating the quantified value of prediction uncertainty is the difference between the upper quantile and the lower quantile of the new energy power output prediction.

[0018] Among them, the hierarchical coordinated control strategy controls the grid frequency deviation in layers. When the absolute value of the grid frequency deviation is greater than 0.2Hz, the energy storage power compensation command is generated first according to the energy storage regulation weight coefficient to achieve millisecond-level power compensation.

[0019] Specifically, when the absolute value of the grid frequency deviation is in the range of 0.1Hz to 0.2Hz, an adjustable load start command is generated based on the load adjustment weight coefficient to achieve a second-level response. When the absolute value of the grid frequency deviation is less than 0.1Hz, the dual decomposition alternating direction multiplier method optimization algorithm is used to calculate the source-load-storage scheduling plan for the future period.

[0020] This invention quantifies the contribution of each link to the grid regulation demand by constructing a dynamic coupling matrix of source, load, and storage. It employs an adaptive predictive fusion model to simultaneously estimate the state of storage and predict the output of new energy sources. Furthermore, it designs a hierarchical collaborative regulation strategy that allocates different response resources according to the magnitude of grid frequency deviation. The dynamic coupling matrix calculates the weighting coefficients for energy storage regulation, load regulation, and new energy regulation based on historical data, enabling the regulation capacity of each link to adaptively change with grid operating conditions. The hierarchical collaborative regulation strategy sets up a three-level response mechanism for different disturbance levels: millisecond-level energy storage power compensation, second-level adjustable load initiation, and minute-level optimized scheduling of source, load, and storage, achieving a balance between response speed and regulation cost. In summary, this invention solves the technical problem of insufficient collaborative regulation capacity of multiple links (source, grid, load, and storage) mentioned in the background art. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a graph showing the predicted output of new energy sources in the embodiment.

[0023] Figure 3 This is a comparison chart of the power grid frequency response under different regulation strategies in the embodiments.

[0024] Figure 4 This is a comparison chart of the accuracy of energy storage state estimation in the embodiments. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0026] like Figure 1 As shown, the invention provides a method for improving the integrated regulation capability of power generation, grid, load, and storage, comprising: S01. Collect and filter source-load-storage grid data. Specifically, this involves collecting grid frequency deviation, grid voltage deviation, real-time output of new energy sources, real-time power of load, energy storage battery voltage, energy storage battery current, and energy storage battery temperature in real time through source-side data acquisition terminals, load-side data acquisition terminals, storage-side data acquisition terminals, and grid-side data acquisition terminals deployed on the source side, load-side data acquisition terminals, storage-side data acquisition terminals, and grid-side data acquisition terminals. Then, an adaptive extended Kalman filter algorithm is used to perform noise cancellation processing on the grid frequency deviation, grid voltage deviation, real-time output of new energy sources, real-time power of load, energy storage battery voltage, energy storage battery current, and energy storage battery temperature before uploading the data to the cloud platform. S02. Construct a source-load-storage coupling matrix and calculate the adjustment weight coefficients. Specifically, based on historical grid frequency deviation, historical grid voltage deviation, historical renewable energy output, historical load power, historical energy storage charging and discharging power, and the grid frequency deviation, grid voltage deviation, renewable energy real-time output, and load real-time power after noise cancellation processing, construct a source-load-storage dynamic coupling matrix. Use the grid frequency deviation and grid voltage deviation as the target variables of the coupling matrix, and use the predicted output of wind and solar power, adjustable load capacity, and energy storage charging and discharging power as the control variables of the coupling matrix to calculate the energy storage adjustment weight coefficient, load adjustment weight coefficient, and renewable energy adjustment weight coefficient. S03. Estimate the energy storage status and limit the charging and discharging power when the health status is low. Specifically, the energy storage battery voltage, the energy storage battery current, and the energy storage battery temperature are input into the adaptive prediction fusion model to estimate the energy storage status. The model outputs the energy storage state of charge, the energy storage health status, and the predicted values ​​of the energy storage temperature distribution. When the energy storage health status is lower than 85%, an energy storage charging and discharging power limiting command is generated. S04. Integrating meteorological data to predict new energy output and quantifying uncertainty, specifically: collecting numerical weather forecast data, satellite cloud image data, and ground observation station meteorological data, and performing multi-source meteorological data fusion processing to generate fused meteorological field data, inputting the fused meteorological field data into the adaptive prediction fusion model to calculate new energy output prediction, and outputting the predicted mean value, upper quantile, and lower quantile of new energy output for the next 1 to 24 hours; calculating the quantified value of prediction uncertainty based on the difference between the upper quantile and the lower quantile of new energy output prediction, and generating a standby capacity configuration instruction accordingly; S05. Layered control of frequency deviation and calculation of time-period scheduling plan: Specifically, a layered coordinated control strategy is constructed. When the absolute value of the grid frequency deviation is greater than 0.2Hz, the energy storage power compensation command is preferentially generated according to the energy storage adjustment weight coefficient to achieve millisecond-level power compensation. When the absolute value of the grid frequency deviation is in the range of 0.1Hz to 0.2Hz, the adjustable load start command is generated according to the load adjustment weight coefficient to achieve second-level response. When the absolute value of the grid frequency deviation is less than 0.1Hz, the dual decomposition alternating direction multiplier method optimization algorithm is used to calculate the future time-period source-load-storage scheduling plan and generate source-side scheduling command, load-side scheduling command, and storage-side scheduling command. S06. The instructions are issued to the executing agencies. If the deviation exceeds the limit, the instructions are recalculated iteratively. Specifically, the energy storage power compensation instructions, energy storage charging and discharging power limit instructions, standby capacity configuration instructions, adjustable load start instructions, source-side dispatch instructions, load-side dispatch instructions, and storage-side dispatch instructions are issued to the corresponding executing agencies through the cloud platform. The adjusted grid frequency deviation and adjusted grid voltage deviation are collected in real time. When the absolute value of the adjusted grid frequency deviation does not drop to within 0.05Hz or the adjusted grid voltage deviation exceeds 5% of the rated voltage value, the energy storage adjustment weight coefficient, the load adjustment weight coefficient, and the new energy adjustment weight coefficient are recalculated, and new energy storage power compensation instructions, adjustable load start instructions, source-side dispatch instructions, load-side dispatch instructions, and storage-side dispatch instructions are generated iteratively.

[0027] The adaptive prediction fusion model is structured as a multi-branch temporal coding fusion architecture, comprising an input layer, a feature extraction layer, a temporal dependency modeling layer, an attention weighting layer, and a multi-task output layer. The input layer receives heterogeneous temporal data from multiple sources and performs normalization preprocessing, mapping input data of different physical quantities to a unified numerical space. The feature extraction layer uses a one-dimensional convolutional neural network to extract local temporal features, with kernel sizes ranging from 3 to 7 and a stride of 1, extracting feature representations at different temporal granularities through multi-scale convolution in parallel. The temporal dependency modeling layer constructs a long short-term memory pathway based on the gating mechanism of recurrent neural units. This layer includes an input gate unit, a forget gate unit, an output gate unit, and a memory unit. The input gate unit determines the retention ratio of the current input information through a sigmoid activation function; the forget gate unit controls the degree of forgetting of historical memories; the output gate unit adjusts the amount of information output at the current moment; and the memory unit stores and transmits long-term dependency information in the range of -1 to +1 using a tanh activation function. The output value of the input gate unit is equal to the weighted linear combination of the input data and the previous hidden state applied by the Sigmoid function. The output value of the forget gate unit is equal to another weighted linear combination of the input data and the previous hidden state applied by the Sigmoid function. The update value of the memory unit is the product of the output value of the forget gate unit and the previous memory unit value, plus the product of the output value of the input gate unit and the candidate memory value. The output value of the output gate unit is equal to the third weighted linear combination of the input data and the previous hidden state applied by the Sigmoid function. The current hidden state is the product of the output value of the output gate unit and the update value of the memory unit after tanh activation. The attention weighting layer calculates the importance weights of temporal features through a self-attention mechanism, assigning higher weights to key moments and key features. The importance weights are obtained by taking the dot product of the query vector and the key vector and then normalizing it using Softmax, and finally summing them to form the context vector. The multi-task output layer sets up independent fully connected layers for different prediction tasks. For the energy storage state estimation task, it outputs the predicted values ​​of the energy storage state of charge, the energy storage health status, and the energy storage temperature distribution. For the renewable energy output prediction task, it outputs the predicted mean, the predicted upper quantile, and the predicted lower quantile of renewable energy output. The time-series dependency modeling layer effectively solves the gradient vanishing problem in long sequence processing of traditional recurrent neural networks through the gating mechanism of the input gate unit, the forget gate unit, and the output gate unit, enabling the adaptive prediction fusion model to maintain stable information transmission in sequences of hundreds of time steps.For energy storage battery state estimation, the memory unit can capture the long-term cumulative impact of charge-discharge cycles on the energy storage health status, the forget gate unit automatically filters short-term noise interference, and the output gate unit dynamically adjusts the sensitivity of state estimation according to the current operating conditions, resulting in a state estimation error reduction of more than 40% compared to the traditional extended Kalman filter algorithm under changing operating conditions. For renewable energy output prediction, the time-dependent modeling layer can simultaneously model the short-term fluctuation patterns and long-term seasonal trends of meteorological elements, and the attention-weighted layer automatically identifies the meteorological feature combinations that have the most significant impact on renewable energy output. Under extreme weather conditions, the prediction error is reduced by more than 30% compared to a single long short-term memory neural network model, and the prediction interval coverage is increased to more than 90%, effectively suppressing the propagation and amplification effect of prediction error in the rolling optimization of model prediction control. The multi-task output layer enables different prediction tasks to share the underlying feature representation, improving the generalization ability and parameter efficiency of the adaptive prediction fusion model. It can still maintain stable prediction performance even with limited training samples, and achieves synergistic enhancement of feature extraction and task prediction through end-to-end optimization.

[0028] The steps for establishing the training dataset for the adaptive prediction fusion model specifically include collecting at least one year of historical source-load-storage operation data. This historical data includes historical energy storage battery voltage, current, temperature, state of charge (SCC) tag data, energy storage health status (HS) tag data, historical renewable energy output data, and corresponding historical meteorological data. The data sampling frequency is from 1 second to 15 minutes. Outlier detection is performed on the historical SCC data to remove abnormal data points caused by sensor malfunctions or communication interruptions. For the energy storage state estimation task, charge and discharge condition segments are extracted as energy storage training samples. Each training sample has a length of 100 to 500 time steps. The SCC tag data and HHS health status (HS) tag data are processed using high-precision... The test equipment was used to measure the results under laboratory conditions. For the task of predicting renewable energy output, a renewable energy training sample was constructed, with the input being a historical 72-hour meteorological sequence and a historical 24-hour output sequence, and the output being the renewable energy output for the next 1 to 24 hours. The historical energy storage battery voltage, historical energy storage battery current, historical energy storage battery temperature, historical renewable energy output data, and historical meteorological data were standardized using the mean-standard-deviation normalization method, mapping them to a zero-mean-unit-variance distribution. The energy storage training sample and the renewable energy training sample were divided into a training set, a validation set, and a test set in an 8:1:1 ratio.

[0029] The specific steps for training the adaptive prediction fusion model include: initializing model parameters using the Xavier initialization method to ensure that the variance of parameters at each layer remains stable during forward and backward propagation; setting the batch size to 32 to 128, the initial learning rate to 0.001, and using the Adam optimizer for gradient updates; defining the multi-task joint loss function as a weighted sum of the mean square error loss of energy storage state estimation, the quantile loss of new energy output prediction, and the coverage loss of the prediction interval, with the loss weight coefficients set according to the importance of the tasks; performing forward propagation to calculate the prediction output on the training set, calculating the value of the multi-task joint loss function, and calculating the result using the backpropagation algorithm. Calculate the gradients of each parameter and update the model parameters based on these gradients. Evaluate the model performance on the validation set every 10 training epochs. Trigger learning rate decay when the validation set loss does not decrease for 5 consecutive epochs, with a decay coefficient of 0.5. Stop training when the validation set loss does not decrease for 20 consecutive epochs, and select the model parameters with the best performance on the validation set as the final model parameters. Evaluate the mean absolute error, root mean square error, and prediction interval coverage of the adaptive prediction fusion model on the test set, ensuring that the energy storage state estimation error is less than 3%, the normalized root mean square error of the new energy output prediction is less than 15%, and the prediction interval coverage is greater than 85%.

[0030] The adaptive extended Kalman filter algorithm is a recursive estimation algorithm for a state-space model, comprising a state prediction step, a covariance prediction step, a Kalman gain calculation step, a state update step, and a covariance update step. The state prediction step predicts the current state based on the system state equation using the previous optimal estimate. The covariance prediction step synchronously updates the state estimation error covariance matrix. The Kalman gain calculation step calculates the optimal weighting coefficients by minimizing the estimation error variance criterion. The state update step fuses the predicted value and the actual observation value to obtain the current optimal estimate. The covariance update step corrects the estimation error covariance. The adaptive mechanism dynamically adjusts the process noise covariance and the observation noise covariance based on the actual observation residuals.

[0031] The source-load-storage dynamic coupling matrix is ​​a mathematical expression matrix that quantifies the contribution of each link in the source-grid-load-storage system to the grid regulation demand. The matrix elements of the source-load-storage dynamic coupling matrix are the sensitivity coefficients of each control variable to the target variable. The source-load-storage dynamic coupling matrix is ​​constructed by performing multiple regression analysis on the historical grid frequency deviation, the historical grid voltage deviation, the historical renewable energy output, the historical load power, and the historical energy storage charging and discharging power. The partial derivatives of the energy storage charging and discharging power, the adjustable load capacity, and the predicted wind and solar power output with respect to the grid frequency deviation and the grid voltage deviation are calculated, and then normalized to form the energy storage regulation weight coefficient, the load regulation weight coefficient, and the renewable energy regulation weight coefficient.

[0032] The dual decomposition alternating direction multiplier method optimization algorithm is a distributed optimization framework used to solve large-scale source-load-storage coordinated scheduling problems. This algorithm decomposes the global optimization problem into source-side, load-side, and storage-side subproblems, each with corresponding local optimization objectives and constraints. Dual variables are introduced to represent coupled constraints such as grid power balance, and an augmented Lagrangian function is constructed to incorporate constraint violation penalties into the objective function. The iterative process of the dual decomposition alternating direction multiplier method includes parallel solving of local optimal control strategies for each subproblem, gradient updates of the dual variables based on global constraint residuals, and determining whether the original residuals and dual residuals satisfy convergence criteria. The dual decomposition alternating direction multiplier method coordinates the behavior of each subsystem through dual updates, enabling it to converge to global consistency.

[0033] The multi-source meteorological data fusion processing is a data processing method that integrates meteorological information from different sources and with different spatiotemporal resolutions. The fusion objects of the multi-source meteorological data fusion processing include grid data of temperature, humidity, wind speed, and irradiance from the numerical weather prediction data; remote sensing data of cloud cover, cloud height, and cloud type from the satellite cloud image data; and real-time meteorological element data from the ground observation station meteorological data. The multi-source meteorological data fusion processing uses spatial interpolation to unify data of different resolutions to a target grid, fuses data from different sources using the Kalman filter method, and uses the ground observation station meteorological data to correct numerical forecast biases, generating the fused meteorological field data.

[0034] The quantified value of prediction uncertainty is a statistical characteristic quantification of the deviation between the predicted value and the actual value, reflecting the reliability of the prediction result. The quantified value of prediction uncertainty is calculated as the difference between the upper quantile and the lower quantile of the predicted new energy power output.

[0035] The hierarchical coordinated control strategy is a multi-level adjustment mechanism designed based on the degree of grid disturbance and response time requirements. The millisecond-level power compensation utilizes the rapid power response characteristics of the energy storage system to perform primary frequency regulation for large frequency deviations; the compensation power is calculated by multiplying the grid frequency deviation by the frequency droop coefficient. The second-level response initiates adjustable loads for demand-side response for medium frequency deviations, prioritizing loads according to load priority and interruption cost. The future-period source-load-storage scheduling plan addresses small frequency deviations or steady-state optimization needs, employing a model predictive control algorithm for rolling optimization calculations. The optimization objective is to minimize operating costs, with constraints including grid power balance constraints, equipment capacity limitations, ramp rate limitations, and energy storage state-of-charge constraints.

[0036] The energy storage state of charge is a normalized parameter characterizing the remaining capacity of the energy storage battery, defined as the percentage of the current remaining battery capacity to the rated capacity.

[0037] The energy storage health status is a normalized parameter characterizing the degree of capacity degradation of energy storage batteries, defined as the percentage of current available capacity to rated capacity. The energy storage health status gradually decreases with factors such as cycle count, operating temperature, and charge / discharge rate, reflecting the degree of battery aging.

[0038] The present invention also provides a method for forming an integrated source-grid-load-storage regulation system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions are used to execute the above method when the computer is run.

[0039] The specific implementation methods of the above steps are described in detail below.

[0040] The specific implementation of step S01 involves using a distributed data acquisition architecture to achieve real-time acquisition and preprocessing of multi-source heterogeneous data. First, data acquisition terminals deployed on the source side acquire real-time power output data from new energy sources via analog acquisition modules at a sampling period of 1 to 5 seconds. Data acquisition terminals deployed on the load side acquire real-time load power data via smart meter communication interfaces at a sampling period of 5 to 15 seconds. Data acquisition terminals deployed on the energy storage side synchronously acquire energy storage battery voltage, current, and temperature data via the battery management system communication protocol at a sampling period of 1 second. Data acquisition terminals deployed on the grid side acquire grid frequency deviation and voltage deviation data with millisecond-level accuracy via phasor measurement devices. The acquired raw data contains significant noise components due to various factors such as sensor measurement noise, environmental electromagnetic interference, and communication channel noise. Directly using this data for subsequent calculations can lead to control command jitter and system instability; therefore, noise cancellation processing is necessary. The Adaptive Extended Kalman Filter (ALED) algorithm achieves optimal estimation of the real signal by establishing a system state-space model. The algorithm first predicts the current state based on the system state equation, synchronously updating the state estimation error covariance matrix. Then, it calculates the Kalman gain coefficient based on the residuals between the actual observations and the predicted values. This gain coefficient reflects the reliability of the observation information. Finally, the predicted and observed values ​​are weighted and fused according to the Kalman gain coefficient to obtain the optimal estimate for the current moment. The adaptive mechanism dynamically adjusts the process noise covariance and the observation noise covariance based on the statistical characteristics of the observation residuals. When the observation residuals continuously increase, it indicates increased observation noise or inaccurate system model. In this case, the algorithm automatically increases the process noise covariance to improve the flexibility of system prediction. When the observation residuals continuously decrease, the algorithm increases the observation noise covariance to reduce dependence on potentially distorted observations. This adaptive adjustment mechanism enables the filtering algorithm to adapt to time-varying noise environments and changing operating conditions. After noise cancellation, the data is uploaded to the time-series database of the cloud platform via a 4G or 5G wireless communication module. The cloud platform performs timestamp alignment and data quality verification on the received data, and removes abnormal data packets with a communication delay of more than 100 milliseconds or a data missing rate of more than 5%.

[0041] The specific implementation of step S02 involves constructing a dynamic coupling matrix based on historical operating data and real-time monitoring data to quantify the regulation capabilities of each link in the power generation, grid, load, and energy storage system. This matrix reflects the degree of influence of each control variable on the grid frequency and voltage under the current grid operating state. First, historical grid frequency deviation, historical grid voltage deviation, historical renewable energy output, historical load power, and historical energy storage charging and discharging power data for the most recent 30 days are extracted from the cloud platform's time-series database. These data are then aggregated and statistically analyzed at a 15-minute time granularity to obtain the historical operating dataset. Then, the real-time grid frequency deviation, real-time grid voltage deviation, real-time renewable energy output, and real-time load power, after noise reduction processing in step S01, are concatenated with the historical operating dataset to form a complete time-series data sequence. A quantitative relationship model is established between grid frequency deviation and energy storage charging and discharging power, adjustable load capacity, and predicted wind and solar power output using multiple linear regression analysis. Simultaneously, a quantitative relationship model is established between grid voltage deviation and the aforementioned control variables. The partial derivatives of each control variable with respect to the target variable, i.e., the sensitivity coefficients, are obtained by solving the regression equations. Due to the significant differences in the physical dimensions and numerical ranges of different control variables, the sensitivity coefficients need to be normalized. The normalization method involves dividing each sensitivity coefficient by the maximum value of all sensitivity coefficients. The normalized sensitivity coefficients are the energy storage regulation weight coefficients, load regulation weight coefficients, and renewable energy regulation weight coefficients. These weight coefficients reflect the contribution of each regulation resource to grid stability under the current grid operating conditions. A larger weight coefficient indicates a more significant effect of that resource on regulating grid frequency or voltage. In subsequent hierarchical coordinated control, regulation resources with larger weight coefficients should be prioritized. The dynamic coupling matrix is ​​updated every 15 minutes to ensure that the weight coefficients reflect changes in the grid operating status in real time.

[0042] The specific implementation of step S03 involves using a deep learning model to accurately estimate and predict the operating state of the energy storage system, providing a state constraint basis for energy storage charging and discharging control. The energy storage battery voltage, current, and temperature data collected in step S01 are organized into a time-series input sequence with 100 time steps, each time step being 1 second long; that is, each input sequence contains the energy storage operation data for the most recent 100 seconds. The input layer of the adaptive prediction fusion model performs normalization preprocessing on the three types of input data, mapping the voltage data to the 0-1 range, normalizing the current data according to the rated current, and normalizing the temperature data according to the operating temperature range. The feature extraction layer uses three parallel one-dimensional convolutional neural network branches to extract local temporal features. The kernel size of the first branch is 3, the kernel size of the second branch is 5, and the kernel size of the third branch is 7, with a stride of 1 for all three branches. Multi-scale convolution is used to extract feature representations at different time granularities in parallel. The temporal dependency modeling layer constructs a long short-term memory pathway based on the gating mechanism of a long short-term memory neural network. This layer includes an input gate, a forget gate, an output gate, and memory units. The input gate uses a sigmoid activation function to determine the retention ratio of the current input information, the forget gate controls the degree of forgetting of historical memories, the output gate adjusts the amount of information output at the current moment, and the memory units store and transmit long-term dependency information in the range of -1 to +1 using a hyperbolic tangent activation function. The attention weighting layer calculates the importance weights of temporal features through a self-attention mechanism, assigning higher weights to key moments and key features. The importance weights are obtained by normalizing the dot product of the query vector and the key vector using Softmax. The multi-task output layer outputs the energy storage state of charge, energy storage health status, and predicted values ​​of energy storage temperature distribution for the energy storage state of estimation task. The energy storage state of charge is the percentage of the current remaining battery capacity to the rated capacity, and the energy storage health status is the percentage of the current available capacity to the rated capacity. When the energy storage health status output by the model is below 85%, it indicates that the energy storage battery capacity has degraded to the threshold that requires limiting the charging and discharging power. At this time, the cloud platform generates an energy storage charging and discharging power limiting instruction, which limits the maximum charging and discharging power to 80% of the rated power to prevent further acceleration of battery aging.

[0043] The specific implementation of step S04 involves achieving high-precision, advanced prediction and uncertainty quantification of new energy power output through multi-source meteorological data fusion and deep learning prediction models, providing a decision-making basis for the coordinated scheduling of power generation, grid, load, and storage. First, grid data of temperature, humidity, wind speed, and irradiance for the next 72 hours are collected from the numerical weather prediction system via a meteorological data interface. The grid spatial resolution is 3 km to 9 km, and the temporal resolution is 1 hour. Simultaneously, remote sensing data of cloud cover, cloud height, and cloud type provided by meteorological satellites are collected, with a spatial resolution of 1 km to 4 km. Real-time meteorological element data from ground observation stations near the new energy power plants are also collected, with a sampling period of 10 minutes. Since meteorological data from different sources have different spatiotemporal resolutions and measurement accuracies, fusion processing is required to generate unified fused meteorological field data. The spatial interpolation method uses the Kriging interpolation algorithm to unify the grid data of different resolutions to the target grid, with the target grid resolution set to 1 km. Then, the Kalman filter method is used to perform temporal fusion of the numerical weather prediction data and satellite remote sensing data, utilizing high-precision real-time observation data from ground observation stations to correct systematic biases in the numerical weather prediction. An adaptive prediction fusion model, integrating meteorological field data, is used to calculate renewable energy output forecasts. The model inputs are a historical 72-hour meteorological sequence and a historical 24-hour power output sequence. The model outputs the predicted mean, upper quantile, and lower quantile of renewable energy output for the next 1 to 24 hours. The predicted mean reflects the expected value of renewable energy output. The upper and lower quantiles correspond to the upper and lower bounds of the forecast at a 90% confidence level, respectively. The model is trained using a quantile regression loss function to output the prediction interval. The difference between the predicted upper and lower quantiles of renewable energy output represents the quantified uncertainty of the forecast. A larger value indicates a lower confidence level in the forecast, requiring the power grid to allocate more reserve capacity to compensate for forecast errors. The cloud platform generates a standby capacity configuration instruction based on the quantified value of the predicted uncertainty. When the quantified value of the predicted uncertainty exceeds 20% of the installed capacity of new energy, the standby capacity is configured as 50% of the quantified value of the predicted uncertainty. When the quantified value of the predicted uncertainty is between 10% and 20%, the standby capacity is configured as 30% of the quantified value of the predicted uncertainty. When the quantified value of the predicted uncertainty is less than 10%, the standby capacity is configured as 15% of the quantified value of the predicted uncertainty.

[0044] The specific implementation of step S05 involves designing a hierarchical collaborative control strategy based on the severity of grid disturbances and the response time characteristics of different regulatory resources. This achieves multi-timescale collaborative regulation, ranging from millisecond-level power compensation to minute-level optimized scheduling. The cloud platform first determines whether the absolute value of the grid frequency deviation is greater than 0.2Hz. This threshold corresponds to the initiation threshold for primary frequency regulation. When the absolute value of the grid frequency deviation is greater than 0.2Hz, it indicates that the grid has suffered a significant active power disturbance. In this case, it is necessary to quickly inject or absorb active power to prevent the frequency from deviating further from its rated value. Because the energy storage system uses power electronic converter control, it has a millisecond-level power response speed. Therefore, in scenarios with large frequency deviations, the energy storage system is prioritized for power compensation. The cloud platform calculates the energy storage power compensation command based on the energy storage regulation weight coefficient obtained in step S02 and the grid frequency deviation. The compensation power is equal to the grid frequency deviation multiplied by the frequency droop coefficient and then multiplied by the energy storage regulation weight coefficient. The frequency droop coefficient reflects the power regulation demand corresponding to a unit frequency deviation, typically ranging from 20 MW / Hertz to 50 MW / Hertz. When the absolute value of the grid frequency deviation is between 0.1Hz and 0.2Hz, the grid frequency deviation is moderate. At this time, adjustable loads are activated for demand-side response, with response times ranging from seconds to minutes. The cloud platform generates adjustable load activation commands based on the load adjustment weight coefficients calculated in step S02. Adjustable loads are then prioritized according to load priority and interruption cost, with priority given to air conditioning and water heater loads with low interruption cost and fast response speed. Industrial loads with higher interruption cost are called when a larger adjustment capacity is needed. When the absolute value of the grid frequency deviation is less than 0.1Hz, the grid operation is close to steady state. At this time, model predictive control is used for source-load-storage coordinated optimization scheduling. The dual decomposition alternating direction multiplier method optimization algorithm is used to solve the large-scale source-load-storage coordinated scheduling problem. This algorithm decomposes the global optimization problem into source-side subproblems, load-side subproblems, and storage-side subproblems. Each subproblem is solved in parallel to find the local optimal control strategy under local constraints. Then, the dual variables are used to coordinate the subsystems to converge to the global optimal solution. The optimization objective is to minimize the operating cost over the next 24 hours. Constraints include grid power balance constraints, equipment capacity limitations, ramp rate limitations, and energy storage state of charge limitations. Dual variables represent coupled constraints such as grid power balance. An augmented Lagrangian function is constructed to incorporate constraint violation penalties into the objective function. During iteration, the dual variables are updated with gradients based on the global constraint residuals. The algorithm stops iterating when both the original residual and the dual residual are less than a set convergence threshold. The cloud platform generates source-side scheduling instructions, load-side scheduling instructions, and energy storage-side scheduling instructions based on the optimization results. These instructions include target values ​​for renewable energy output, adjustable load start-up and shutdown plans, and energy storage charging and discharging power plans for each time period.

[0045] The specific implementation of step S06 involves sending various control commands generated by the cloud platform to the corresponding actuators and monitoring and correcting the adjustment effects in real time to form a closed-loop control architecture. The cloud platform sends energy storage power compensation commands to the energy management system of the energy storage system via IoT communication protocols. The energy management system adjusts the active power output value of the energy storage converter according to the commands, with a sending delay of no more than 50 milliseconds. Simultaneously, energy storage charging and discharging power limiting commands are sent to the energy management system. When the energy storage health status is below 85%, the maximum charging and discharging power is limited to extend battery life. Reserve capacity configuration commands are sent to the grid dispatch center, which reserves corresponding spinning reserve capacity and non-spinning reserve capacity based on the predicted uncertainty quantification value. Adjustable load start commands are sent to the load aggregator's energy management platform. The platform sends control signals to the user-side adjustable loads according to the commands to achieve start-stop control of flexible loads such as air conditioners, water heaters, and charging piles. Source-side dispatch commands are sent to the station control system of the new energy power plant. The control system adjusts the wind turbine yaw angle or photovoltaic inverter power factor according to the dispatch commands to achieve coordinated control of active and reactive power. The load-side dispatching command and the storage-side dispatching command are respectively sent to the corresponding energy management system. After receiving the control command, each actuator executes the corresponding control action, and the cloud platform continuously monitors the grid operation status after adjustment. The cloud platform collects the grid frequency deviation and grid voltage deviation after adjustment in real time to determine whether the adjustment effect has achieved the expected target. When the absolute value of the grid frequency deviation after adjustment does not drop to within 0.05Hz or the grid voltage deviation after adjustment exceeds 5% of the rated voltage value, it indicates that the current adjustment strategy has failed to effectively suppress the grid frequency and voltage deviation. At this time, it is necessary to recalculate the adjustment weight coefficient and iteratively generate new control commands. The cloud platform returns to step S02, reconstructs the source-load-storage dynamic coupling matrix based on the latest grid operation data, calculates the updated energy storage adjustment weight coefficient, load adjustment weight coefficient, and new energy adjustment weight coefficient, and then iteratively generates new energy storage power compensation commands, adjustable load start commands, source-side dispatching commands, load-side dispatching commands, and storage-side dispatching commands according to the hierarchical collaborative control strategy in step S05, forming a feedback correction mechanism to ensure that the grid frequency and voltage deviation converge to the allowable range.

[0046] It should be noted that the first key technical idea of ​​this invention is to adopt an adaptive predictive fusion model to realize a unified modeling framework for energy storage state estimation and new energy output prediction. Through a multi-branch time-series coding fusion architecture and a long short-term memory neural network gating mechanism, the gradient vanishing problem in long-sequence processing of traditional methods is effectively solved. This allows the model to simultaneously capture the long-term cumulative impact of charge-discharge cycles on the health status of energy storage and the short-term fluctuation patterns of meteorological factors. Compared with traditional extended Kalman filter algorithms and single long short-term memory neural network models, this significantly reduces prediction errors and improves prediction interval coverage, providing more accurate state and prediction information for source-grid-load-storage coordinated scheduling. The second key technical idea is to construct a source-load-storage dynamic coupling matrix to quantify the contribution of each regulating resource to the grid frequency and voltage. The sensitivity coefficients of each control variable to the target variable are calculated using multivariate regression analysis and normalized into weight coefficients. These weight coefficients can reflect changes in grid operating conditions in real time, providing a scientific basis for resource scheduling for hierarchical coordinated control strategies. Compared with traditional fixed scheduling priority methods, this approach can respond more flexibly to dynamic changes in grid operating conditions. The third key technical approach is to design a hierarchical collaborative control strategy to achieve multi-timescale coordinated regulation based on the degree of grid disturbance and response time requirements. For large frequency deviations, millisecond-level response energy storage systems are prioritized for primary frequency regulation; for medium frequency deviations, adjustable loads with second-level response are activated for demand-side response; and for small frequency deviations, model predictive control methods are used for source-load-storage coordinated optimal scheduling. This hierarchical control mechanism fully leverages the response speed advantages of different regulation resources, significantly improving the regulation speed and accuracy of grid frequency and voltage compared to traditional single-regulation methods. The synergistic effect of these three key technical approaches lies in the adaptive predictive fusion model providing accurate state estimation and prediction information for the dynamic coupling matrix, the dynamic coupling matrix providing real-time updated weight coefficients for the hierarchical collaborative control strategy, and the hierarchical collaborative control strategy generating differentiated control commands based on the weight coefficients and grid operating status. These three elements form a complete closed-loop control link from state perception to weight calculation to hierarchical regulation. Compared to traditional open-loop scheduling methods and single-timescale control methods, this approach can more effectively address grid frequency and voltage deviations caused by fluctuations in new energy output and load changes, significantly improving the regulation capability and operational stability of the integrated source-grid-load-storage system.

[0047] It should be noted that this invention also solves the following technical problem: the accuracy of traditional energy storage state estimation methods decreases when operating conditions change. This invention, through the time-dependent modeling layer of the adaptive predictive fusion model, constructs a long short-term memory path using a gating mechanism of input gate units, forget gate units, and output gate units. This effectively captures the long-term cumulative impact of charge-discharge cycles on the health state of energy storage, automatically filters short-term noise interference, and dynamically adjusts the state estimation sensitivity according to the current operating conditions, thus solving the problem of decreased estimation accuracy in traditional extended Kalman filter algorithms when operating conditions change. It also solves the problem that the regulation strategy cannot adaptively adjust according to the actual effect. In step S06, this invention collects the grid frequency deviation and voltage deviation in real time after regulation. When the regulation effect does not reach the set threshold, it recalculates the energy storage regulation weight coefficient, load regulation weight coefficient, and new energy regulation weight coefficient, and iteratively generates new control commands, forming a closed-loop feedback control mechanism. This adaptive adjustment mechanism enables the regulation strategy to dynamically optimize the regulation weights of each link according to the actual regulation effect, avoiding the problem of poor regulation effect of fixed weight strategies under complex operating conditions, and improving the system's robustness to grid disturbances.

[0048] Specifically, the principle of this invention is as follows: This invention can solve the technical problem of insufficient coordinated regulation capability of multiple links in the source-grid-load-storage system because, through the construction of a dynamic coupling matrix of source, load, and storage, the grid frequency deviation and voltage deviation are used as target variables, and the adjustable resources on each side are used as control variables. The sensitivity coefficients of each control variable to the target variable are calculated, quantifying the regulation contribution of each link under different operating conditions. The hierarchical coordinated regulation strategy sets different response thresholds according to the magnitude of the frequency deviation. For large deviations, the energy storage system is preferentially called to achieve millisecond-level response; for medium deviations, the adjustable load is activated to achieve second-level response; and for small deviations, the dual decomposition alternating direction multiplier method optimization algorithm is used for global optimization, achieving a balance between speed and economy. The adaptive prediction fusion model shares the underlying feature expression through the multi-task output layer, enabling coordinated optimization of energy storage state estimation and new energy output prediction, improving prediction accuracy and providing accurate decision-making basis for regulation strategies, thereby realizing coordinated regulation of multiple links and multiple time scales.

[0049] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0050] The specific implementation of step S01 involves real-time data acquisition of grid frequency deviation, grid voltage deviation, real-time output of new energy sources, real-time load power, energy storage battery voltage, energy storage battery current, and energy storage battery temperature using data acquisition terminals deployed on the source side, load side, energy storage side, and grid side. The acquired data is then processed for noise reduction using an adaptive extended Kalman filter algorithm before being uploaded to the cloud platform. The state prediction steps of the adaptive extended Kalman filter algorithm are expressed as follows: ; In the formula, For the first The predicted state value at time 10:00. For the first The optimal estimate at time [time]. The system state transition function, For the first The control input at any given time. The formula for the covariance prediction step is expressed as follows: ; In the formula, For the first The covariance matrix of the prediction error at any given time. For the first Time-estimation error covariance matrix The state transition Jacobian matrix is ​​obtained by considering the state transition function. For state variables Taking the partial derivative, we get for The transpose of the matrix, Let be the process noise covariance matrix, with units of the square of each state variable. The Kalman gain calculation steps are expressed as follows: ; In the formula, For Kalman gain, The Jacobian matrix of the observation matrix is ​​obtained by applying the observation function. For state variables Taking the partial derivative, we get for The transpose of the matrix, The observation noise covariance matrix is ​​expressed in units of the square of each observed variable. The state update steps are formulated as follows: ; In the formula, For the first The optimal estimate at time. For the first Time-based observations For the observation function, To observe the residuals, the units are the same as the observed values. The units are the same. The formula for the covariance update step is expressed as follows: ; In the formula, For the first The estimated error covariance matrix updated at time step 1. This is the identity matrix. The adaptive mechanism dynamically adjusts the noise covariance based on the observed residuals, and the adjustment formula is as follows: , ; In the formula, and For the first The adaptive adjustment coefficient at time step is dimensionless and has an empirical value of 0.9 to 1.1. and For the first The noise covariance matrix at time t.

[0051] The specific implementation of step S02 is to construct a source-load-storage dynamic coupling matrix based on historical grid frequency deviation, historical grid voltage deviation, historical renewable energy output, historical load power, historical energy storage charging and discharging power, and the grid frequency deviation, grid voltage deviation, renewable energy real-time output, and load real-time power after noise cancellation processing. Grid frequency deviation and grid voltage deviation are used as the target variables of the coupling matrix, and wind and solar predicted output, adjustable load capacity, and energy storage charging and discharging power are used as the control variables. The energy storage regulation weight coefficient, load regulation weight coefficient, and renewable energy regulation weight coefficient are calculated. The sensitivity coefficient of the source-load-storage dynamic coupling matrix is ​​calculated through multiple regression analysis, and the formula is expressed as follows: ; In the formula, For the first The control variable affects the first... The sensitivity coefficient of each target variable, measured in units of the target variable divided by the units of the control variables. For the first One target variable, For the first One control variable, The number of historical data points. For the first Time of the first Historical values ​​of the target variable For the first The mean of the target variables, For the first Time of the first Historical values ​​of each control variable For the first The mean of each control variable is used. After normalization, the adjustment weight coefficients are obtained, as expressed by the following formula: ; In the formula, For the first The adjustment weighting coefficients for each control variable are dimensionless. To control the total number of variables, a value of 3 is used. Energy storage regulation weighting coefficient. Load adjustment weighting coefficient New energy adjustment weight coefficient Corresponding to time value.

[0052] The specific implementation of step S03 involves inputting the energy storage battery voltage, current, and temperature into an adaptive prediction fusion model to perform energy storage state estimation calculations. The model outputs predicted values ​​for the energy storage state of charge, energy storage health status, and energy storage temperature distribution. When the energy storage health status is below 85%, a power limitation command for energy storage charging and discharging is generated. The formula for defining the energy storage state of charge is as follows: ; In the formula, The energy storage state of charge is dimensionless. This represents the current remaining battery power, in ampere-hours. This refers to the battery's rated capacity, expressed in ampere-hours (AH). The formula for defining the energy storage health status is as follows: ; In the formula, For energy storage in a healthy state, dimensionless. This represents the currently available capacity, in ampere-hours.

[0053] The specific implementation of step S04 involves collecting numerical weather prediction data, satellite cloud image data, and ground observation station meteorological data, and performing multi-source meteorological data fusion processing to generate fused meteorological field data. This fused meteorological field data is then input into an adaptive prediction fusion model to calculate new energy power output forecasts. The model outputs the predicted mean, upper quantile, and lower quantile of new energy power output for the next 1 to 24 hours. Based on the difference between the upper and lower quantiles of the new energy power output forecasts, a prediction uncertainty quantification value is calculated, and a reserve capacity configuration instruction is generated accordingly. The formula for calculating the prediction uncertainty quantification value is as follows: ; In the formula, To quantify the uncertainty in the forecast, the unit is kilowatt. Upper quantile of the predicted contribution of new energy sources, in kilowatts. The lower quantile of the predicted power output for new energy sources is given, in kilowatts. The rated capacity for new energy sources is expressed in kilowatts (kW). The formula for reserve capacity configuration is as follows: ; In the formula, This is the reserve capacity configuration value, in kilowatts. This is the reserve capacity factor, dimensionless, with an empirical value of 1.2 to 1.5.

[0054] The specific implementation of step S05 involves constructing a hierarchical coordinated control strategy. When the absolute value of the grid frequency deviation is greater than 0.2 Hz, energy storage power compensation commands are preferentially generated based on the energy storage regulation weight coefficient to achieve millisecond-level power compensation. When the absolute value of the grid frequency deviation is between 0.1 Hz and 0.2 Hz, adjustable load start-up commands are generated based on the load regulation weight coefficient to achieve second-level response. When the absolute value of the grid frequency deviation is less than 0.1 Hz, the dual decomposition alternating direction multiplier method optimization algorithm is used to calculate the future time period source-load-storage scheduling plan and generate source-side scheduling commands, load-side scheduling commands, and storage-side scheduling commands. The calculation formula for the energy storage power compensation command is as follows: ; In the formula, This is the energy storage power compensation value, in kilowatts. This is the frequency droop factor, measured in kilowatts per hertz, with an empirical value ranging from 1000 to 5000. This refers to the power grid frequency deviation, measured in Hertz. This is the rated frequency of the power grid, measured in Hertz, and has a value of 50 or 60. This is the energy storage regulation weighting coefficient, which is dimensionless. The formula for calculating the starting power of adjustable loads is as follows: ; In the formula, This refers to the adjustable load starting power, measured in kilowatts. This is the load regulation factor, measured in kilowatts per hertz, with an empirical value ranging from 500 to 3000. This is the load adjustment weighting coefficient, which is dimensionless.

[0055] The specific implementation of step S06 involves sending energy storage power compensation instructions, energy storage charging and discharging power limitation instructions, reserve capacity configuration instructions, adjustable load start-up instructions, source-side dispatch instructions, load-side dispatch instructions, and storage-side dispatch instructions to the corresponding actuators via a cloud platform. Simultaneously, it collects the adjusted grid frequency deviation and adjusted grid voltage deviation in real time. When the absolute value of the adjusted grid frequency deviation does not drop below 0.05 Hz or the adjusted grid voltage deviation exceeds 5% of the rated voltage, it recalculates the energy storage adjustment weight coefficient, load adjustment weight coefficient, and new energy adjustment weight coefficient, and iteratively generates new energy storage power compensation instructions, adjustable load start-up instructions, source-side dispatch instructions, load-side dispatch instructions, and storage-side dispatch instructions. While sending the control instructions to the actuators at each stage, the cloud platform simultaneously collects the adjusted grid status data in real time. If the adjustment effect does not meet the target, it iterates the coupling matrix and control strategy again until the grid stability requirements are met.

[0056] The formula for calculating the output value of the input gate unit in the temporal dependency modeling layer of the adaptive prediction fusion model is as follows: ; In the formula, For the first The output value of the input gate unit at any given time is dimensionless and ranges from 0 to 1. The Sigmoid activation function is defined as follows: ,in For input independent variables, It is a natural exponential function. The input gate weight matrix, For the first The state vector is hidden at all times. For the first Input data vector at any time, This represents a vector concatenation operation. Let be the input gate bias vector. The formula for calculating the output value of the forget gate unit is as follows: ; In the formula, For the first The output value of the time-forget gate unit is dimensionless and ranges from 0 to 1. Here is the forget gate weight matrix. Let be the forget gate bias vector. The formula for calculating candidate memory values ​​is as follows: ; In the formula, For the first Candidate memory value vector at each time step Let hyperbolic tangent activation function be defined as follows: The value ranges from -1 to +1. The candidate memory weight matrix, This represents the candidate memory bias vector. The formula for calculating the update value of a memory cell is as follows: ; In the formula, For the first The updated value vector of the memory unit at each moment. It is the element-wise multiplication operator. For the first The memory unit value vector is used at all times. The formula for calculating the output value of the output gate unit is as follows: ; In the formula, For the first The output value of the output gate unit at any given time is dimensionless and ranges from 0 to 1. This is the output gate weight matrix. This is the output gate bias vector. The formula for calculating the hidden state at the current time step is as follows: ; In the formula, For the first The state vector is hidden at all times.

[0057] The formula for calculating the dot product of the query vector and the key vector in the attention-weighted layer of the adaptive prediction fusion model is as follows: ; In the formula, For the first The query vector and the first Similarity score of each key vector, dimensionless. For the first A query vector, For the first A key vector, for The transpose of , The dimension of the key vector is dimensionless and typically ranges from 64 to 512. Importance weights are obtained through Softmax normalization, expressed by the following formula: ; In the formula, For the first The query vector pairs the first Attention weights for each key vector, dimensionless. The number of time series steps. This is the natural exponential function. The context vector is formed by a weighted sum, as expressed in the following formula: ; In the formula, For the first Context vectors, For the first A vector of values.

[0058] The specific implementation methods for establishing the training dataset of the adaptive prediction fusion model are the same as described above, and will not be repeated in detail here. The specific implementation methods for training the adaptive prediction fusion model are the same as described above, and will not be repeated in detail here. The specific implementation methods for the interpretation of the adaptive extended Kalman filter algorithm are the same as described above, and will not be repeated in detail here. The specific implementation methods for the interpretation of the source-load-storage dynamic coupling matrix are the same as described above, and will not be repeated in detail here. The specific implementation methods for the interpretation of the dual decomposition alternating direction multiplier method optimization algorithm are the same as described above, and will not be repeated in detail here. The specific implementation methods for the interpretation of the multi-source meteorological data fusion processing are the same as described above, and will not be repeated in detail here. The specific implementation methods for the interpretation of the quantified value of prediction uncertainty are the same as described above, and will not be repeated in detail here. The specific implementation methods for the interpretation of the hierarchical collaborative control strategy are the same as described above, and will not be repeated in detail here. The specific implementation methods for the interpretation of the energy storage state of charge are the same as described above, and will not be repeated in detail here. The specific implementation methods for the interpretation of the energy storage health status are the same as described above, and will not be repeated in detail here.

[0059] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2: To verify the effectiveness of this invention, technicians built a numerical simulation analysis environment and conducted simulation tests by constructing an integrated source-grid-load-storage regulation system for a regional power grid. This regional power grid includes a wind power installed capacity of 200MW, a photovoltaic installed capacity of 150MW, an energy storage system with a rated power of 80MW and a rated capacity of 160MWh, and an adjustable load capacity of 120MW. The grid's rated frequency is 50Hz, and the rated voltage is 110kV. The simulation duration was set to 24 hours, the data sampling frequency was 1 second, and the historical data spanned 12 months.

[0060] Technicians first collected real-time grid operation data using data acquisition terminals deployed on each side. Source-side terminals collected real-time output data from wind farms and photovoltaic power plants; load-side terminals collected real-time power data from industrial, commercial, and residential loads; storage-side terminals collected battery voltage, current, and temperature data from energy storage systems; and grid-side terminals collected grid frequency and voltage data. Initially, the grid frequency was 49.85Hz, the grid voltage was 108.9kV, the real-time wind power output was 135MW, the real-time photovoltaic output was 78MW, the real-time load power was 245MW, the energy storage battery pack voltage was 1520V, the energy storage battery pack current was 42A, and the average energy storage battery temperature was 28℃. The raw data collected contained sensor noise and communication interference. Technicians used an adaptive extended Kalman filter algorithm to remove noise from this data. The filtered grid frequency deviation was -0.15Hz, and the grid voltage deviation was -1.1kV. The noise-removed data was then uploaded to the cloud platform.

[0061] Technicians constructed a dynamic coupling matrix of source, load, and storage based on historical grid frequency deviation, historical grid voltage deviation, historical renewable energy output, historical load power, and historical energy storage charging and discharging power. Historical data statistics show that over the past 12 months, the grid frequency deviation ranged from -0.35Hz to +0.28Hz, the grid voltage deviation ranged from -5.2kV to +4.8kV, wind power output ranged from 15MW to 198MW, photovoltaic output ranged from 0MW to 148MW, load power ranged from 180MW to 320MW, and energy storage charging and discharging power ranged from -80MW to +80MW. Through multiple regression analysis of this historical data, the sensitivity coefficient of energy storage charging and discharging power to grid frequency deviation was calculated to be 0.0125. The sensitivity coefficient of the adjustable load capacity to grid frequency deviation is 0.0082. The sensitivity coefficient of new energy output to grid frequency deviation is 0.0058. After normalization, the weighting coefficients for energy storage regulation are 0.47, load regulation is 0.31, and new energy regulation is 0.22.

[0062] Technicians input battery voltage, current, and temperature into an adaptive prediction fusion model to estimate the energy storage state. This model employs a multi-branch temporal coding fusion architecture. The input layer receives historical data from 300 time steps, including voltage, current, and temperature sequences from the energy storage battery pack. The feature extraction layer uses a one-dimensional convolutional neural network to extract local temporal features, with kernel sizes of 3, 5, and 7 and a stride of 1, resulting in a 48-dimensional feature vector. The temporal dependency modeling layer constructs a memory path based on a long short-term memory (LSM) gating mechanism, with a hidden state dimension of 128. Input, forget, and output gates calculate gating signals using a Sigmoid activation function, while the memory unit stores long-term dependency information using a tanh activation function. The attention weighting layer calculates the importance weights of the temporal features, assigning higher weights to critical moments, and sums the weighted values ​​to form a context vector. The multi-task output layer outputs a state of charge of 68%, a health status of 91%, and a predicted temperature distribution showing an internal temperature range of 27°C to 31°C with good temperature uniformity. Since the energy storage health status is 91%, which is higher than the 85% threshold, the system did not generate an energy storage charging and discharging power limit command.

[0063] Technicians collected numerical weather prediction data, satellite cloud imagery data, and ground-based meteorological data for multi-source meteorological data fusion processing. The numerical weather prediction data provides a 24-hour temperature forecast range of 18°C ​​to 26°C, a humidity forecast range of 45% to 72%, a wind speed forecast range of 3.2 m / s to 8.5 m / s, and an irradiance forecast range of 0. Up to 850 The grid resolution is 10km. Satellite cloud imagery provides cloud cover distribution and cloud height data, showing that cloud cover will increase from 35% to 68% within the next 8 hours. Ground-based meteorological data provides real-time meteorological elements to correct numerical forecast biases. Data from different resolutions are unified to a 1km target grid using spatial interpolation methods, and Kalman filtering is used to fuse data from different sources, generating a fused meteorological field. Technicians input the fused meteorological field data into an adaptive prediction fusion model for new energy output prediction calculations, such as... Figure 2As shown, the output includes the predicted average, upper quantile, and lower quantile of renewable energy output for the next 1 to 24 hours. The prediction results show that the predicted average wind power output for the next 4 hours is 142MW, with an upper quantile of 165MW and a lower quantile of 118MW; the predicted average photovoltaic output is 52MW, with an upper quantile of 68MW and a lower quantile of 35MW. The prediction uncertainty is quantified based on the difference between the upper and lower quantiles of the renewable energy output predictions. The predicted uncertainty for wind power is 47MW, for photovoltaic power it is 33MW, and the total predicted uncertainty is 80MW. Based on the predicted uncertainty, a reserve capacity configuration instruction is generated, requiring the energy storage system to reserve 80MWh of reserve capacity and the adjustable load to reserve 48MW of response capacity.

[0064] Technicians constructed a hierarchical collaborative control strategy to adjust the power grid frequency deviation. Since the current absolute value of the power grid frequency deviation is 0.15Hz, falling within the 0.1Hz to 0.2Hz range, the system generates adjustable load initiation commands based on load adjustment weighting coefficients, achieving a second-level response. Adjustable loads are prioritized: air compressor loads in industrial loads have priority 1, circulating water pump loads have priority 2, and air conditioning loads in commercial loads have priority 3, lighting loads have priority 4. The system initiates a reduction of the air compressor load (priority 1) by 18MW, and a reduction of the circulating water pump load (priority 2) by 12MW, for a total reduction of 30MW. Three seconds after the adjustment is executed, the power grid frequency rises to 49.92Hz, and the frequency deviation drops to -0.08Hz. Since the absolute value of the frequency deviation has fallen below 0.1Hz, the system switches to a minute-level optimized scheduling mode, employing a dual decomposition alternating direction multiplier method optimization algorithm to calculate the source-load-storage scheduling plan for future time periods. The optimization algorithm decomposes the global optimization problem into source-side subproblems, load-side subproblems, and storage-side subproblems, each with its own local optimization objective and constraints. The source-side subproblem optimizes the renewable energy generation plan, the load-side subproblem optimizes the adjustable load response plan, and the storage-side subproblem optimizes the energy storage charging and discharging plan. The algorithm introduces dual variables to represent the grid power balance constraints and constructs an augmented Lagrangian function to incorporate constraint violation penalties into the objective function. After 15 iterations, the original residual is reduced to 0.005 MW, and the dual residual is reduced to 0.002, satisfying the convergence criterion. The optimized calculation outputs the scheduling plan for the next hour, as shown in Table 1.

[0065] Table 1 Source-Load-Storage Dispatch Plan for the Next 1 Hour

[0066] Based on the dispatch plan, source-side dispatch instructions, load-side dispatch instructions, and storage-side dispatch instructions are generated and distributed to the corresponding execution agencies through the cloud platform. The source-side dispatch instructions require wind farms and photovoltaic power plants to track their output according to the forecast curve. The load-side dispatch instructions require adjustable loads to be reduced by 8MW in the first 15 minutes and by 5MW in the second 15 minutes. The storage-side dispatch instructions require energy storage systems to charge 12MW in the first 15 minutes, charge 8MW in the second 15 minutes, discharge 6MW in the third 15 minutes, and discharge 10MW in the fourth 15 minutes.

[0067] Technicians collected real-time data on the grid frequency and voltage deviations after adjustment to evaluate the effectiveness. Five minutes after adjustment, the grid frequency stabilized at 49.98Hz with a frequency deviation of -0.02Hz, and the grid voltage stabilized at 109.7kV with a voltage deviation of -0.3kV. Since the absolute value of the grid frequency deviation decreased to within 0.05Hz and the voltage deviation did not exceed 5% of the rated voltage, the system determined that the adjustment effect met the requirements and continued to execute the current dispatch plan. In subsequent operation, when new disturbances caused the grid frequency deviation to exceed the threshold again, the system automatically recalculated the energy storage adjustment weight coefficient, load adjustment weight coefficient, and new energy adjustment weight coefficient, and iteratively generated new control commands.

[0068] like Figure 3 As shown, technicians compared and analyzed the grid frequency response curves under different regulation strategies. Traditional single energy storage regulation strategies respond promptly to large frequency deviations but lack sustained regulation capability; single load regulation strategies have slower response speeds but larger regulation capacity; and single renewable energy regulation strategies are constrained by meteorological conditions and have poor regulation flexibility. The hierarchical collaborative regulation strategy of this invention comprehensively utilizes the rapid response characteristics of energy storage, the advantages of load regulation capacity, and the controllability of renewable energy output. It dynamically allocates regulation tasks to each link according to the magnitude of the frequency deviation, achieving an organic combination of response speed and regulation capacity. When the grid frequency deviates significantly, the energy storage system is prioritized for millisecond-level power compensation to quickly suppress frequency fluctuations. When the frequency deviation drops to a moderate level, adjustable loads are activated for second-level response, providing sustained regulation capability. When the frequency deviation drops to a small level, a global optimization algorithm is used for source-load-storage collaborative scheduling to achieve refined regulation.

[0069] like Figure 4As shown, the technical staff analyzed the performance of the adaptive predictive fusion model in energy storage state estimation. The traditional extended Kalman filter algorithm shows a significant increase in state estimation error when the charging and discharging conditions of the energy storage battery change, especially during high-current charging and discharging switching, where the estimation accuracy drops significantly. The adaptive predictive fusion model of this invention effectively captures the long-term cumulative impact of charging and discharging cycles on the energy storage health state through a gating mechanism in the time-dependent modeling layer. The input gate unit dynamically adjusts the proportion of new information received based on the current charging and discharging current, the forget gate unit automatically filters short-term noise interference to avoid the impact of instantaneous fluctuations on state estimation, and the output gate unit dynamically adjusts the sensitivity of state estimation based on the current operating conditions, ensuring stable estimation accuracy even under changing operating conditions. The attention-weighted layer assigns higher weights to charging and discharging data at key moments, enhancing the model's ability to capture state change trends. The multi-task output layer allows the energy storage state of charge estimation and energy storage health state estimation to share the underlying feature representation, improving the model's generalization ability and parameter efficiency.

[0070] This invention achieves adaptive calculation of adjustment weights for each component through the construction of a dynamic coupling matrix for source-load-storage systems. This enables the adjustment strategy to dynamically optimize resource allocation on each side according to the actual operating conditions of the power grid, solving the problem of poor adjustment performance of traditional fixed-weight strategies under complex conditions. The hierarchical collaborative control strategy sets different response thresholds and adjustment mechanisms based on the degree of power grid disturbance, achieving multi-timescale collaborative control. This ensures both rapid response capability under large disturbances and economic efficiency under small disturbances. The adaptive prediction fusion model simultaneously performs energy storage state estimation and new energy output prediction through a multi-task learning framework, improving prediction accuracy and providing accurate decision-making basis for the control strategy, avoiding the problem of insufficient accuracy in multi-objective prediction by traditional single-task models. The dual decomposition alternating direction multiplier method optimization algorithm decomposes the global optimization problem into parallel solutions to sub-problems on each side through a distributed optimization framework, improving the computational efficiency of large-scale source-load-storage collaborative scheduling. Global consistency coordination of the behavior of each subsystem is achieved through dual variable updates. The closed-loop feedback control mechanism iteratively optimizes the adjustment weight coefficients and control commands based on the actual adjustment effect, enabling the system to have adaptive adjustment capabilities and improving its robustness to power grid disturbances.

[0071] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0072] Table 2. Variable Explanation Table (Part 1)

[0073] Table 3. Variable Explanation Table (Part Two)

[0074] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for improving the integrated regulation capability of power generation, grid, load, and storage, characterized in that, Real-time data acquisition is achieved through data acquisition terminals deployed at the source, load, and storage sides, as well as the grid. These terminals collect data on grid frequency deviation, grid voltage deviation, real-time renewable energy output, real-time load power, energy storage battery voltage, energy storage battery current, and energy storage battery temperature. Noise is eliminated using an adaptive extended Kalman filter algorithm before the data is uploaded to the cloud platform. A dynamic coupling matrix between source, load, and storage is constructed based on historical and real-time data to calculate energy storage regulation weight coefficients, load regulation weight coefficients, and renewable energy regulation weight coefficients. Energy storage battery voltage, current, and temperature are input into an adaptive predictive fusion model for energy storage state estimation, outputting the energy storage state of charge and energy storage parameters. The system collects and integrates multi-source meteorological data to predict the health status and energy storage temperature distribution. This data is then input into an adaptive prediction fusion model to calculate the predicted output of new energy sources. The model outputs the predicted mean, upper quantile, and lower quantile of new energy output, and calculates the quantified value of the prediction uncertainty. A hierarchical collaborative control strategy is constructed to generate energy storage power compensation commands, adjustable load start commands, or source-load-storage scheduling commands based on the absolute value of the grid frequency deviation. These commands are then distributed to the corresponding actuators through a cloud platform. The system also collects the adjusted grid frequency deviation and adjusted grid voltage deviation in real time. When the adjusted grid frequency deviation or adjusted grid voltage deviation does not reach the set threshold, the weighting coefficients are recalculated and new control commands are generated iteratively.

2. The method according to claim 1, characterized in that, The adaptive extended Kalman filter algorithm includes a state prediction step, a covariance prediction step, a Kalman gain calculation step, a state update step, and a covariance update step. The adaptive mechanism dynamically adjusts the process noise covariance and the observation noise covariance based on the actual observation residuals.

3. The method according to claim 2, characterized in that, The dynamic coupling matrix of source, load and storage is constructed by performing multiple regression analysis on historical grid frequency deviation, historical grid voltage deviation, historical renewable energy output, historical load power and historical energy storage charging and discharging power. The partial derivatives of energy storage charging and discharging power, adjustable load capacity and wind and solar predicted output with respect to grid frequency deviation and grid voltage deviation are calculated and normalized to form energy storage regulation weight coefficient, load regulation weight coefficient and renewable energy regulation weight coefficient.

4. The method according to claim 3, characterized in that, The adaptive prediction fusion model has a multi-branch temporal coding fusion architecture, which includes an input layer, a feature extraction layer, a temporal dependency modeling layer, an attention weighting layer, and a multi-task output layer.

5. The method according to claim 4, characterized in that, The input layer receives heterogeneous time-series data from multiple sources and performs normalization preprocessing to map the input data of different physical quantities to a unified numerical space.

6. The method according to claim 5, characterized in that, The feature extraction layer uses a one-dimensional convolutional neural network to extract local temporal features, and extracts feature representations at different temporal granularities through multi-scale convolution in parallel.

7. The method according to claim 6, characterized in that, The temporal dependency modeling layer constructs a long short-term memory pathway based on the gating mechanism of recurrent neural units, including input gate units, forget gate units, output gate units, and memory units.

8. The method according to claim 7, characterized in that, The output value of the input gate unit is equal to the weighted linear combination of the input data and the previous hidden state applied by the Sigmoid function, and the output value of the forget gate unit is equal to another weighted linear combination of the input data and the previous hidden state applied by the Sigmoid function.

9. The method according to claim 8, characterized in that, The update value of the memory unit is the product of the output value of the forget gate unit and the value of the memory unit at the previous time step, plus the product of the output value of the input gate unit and the candidate memory value. The memory unit stores and transmits long-term dependency information in the interval from negative 1 to positive 1 through the tanh activation function.

10. The method according to claim 9, characterized in that, The attention-weighted layer calculates the importance weights of temporal features through a self-attention mechanism. The importance weights are obtained by normalizing the dot product of the query vector and the key vector using Softmax, and then summed by weight to form the context vector.

Images