Renewable energy integrated optimization system and method thereof

Abstract

The invention discloses a renewable energy source integrated optimization system and a method thereof. The method comprises the following steps: acquiring a data set acquired in a current period; inputting the data set into a pre-trained Gaussian model to obtain a prediction mean value and a prediction variance for predicting the renewable energy power in a preset future time window; carrying out difference calculation on the predicted mean value and a grid-connected power reference value to obtain a mismatch value; performing fluctuation component decomposition on the mismatch value, and estimating reference distribution power applied to the hybrid energy storage system; distributing renewable energy power under the condition that the difference value between the reference distribution power and the optimal distribution power does not exceed the limit; and projecting the optimal distribution power into the feasible region, judging whether the optimal distribution power crosses the boundary or not, if so, determining an adjustment amount for adjustment according to the deviation condition and the change condition of the power grid state data in the current period and the previous period, and outputting a control instruction containing the optimal distribution power in the feasible region. According to the invention, the stability and utilization rate of renewable energy grid connection can be improved.

Description

Technical Field

[0001] This invention relates to the field of renewable energy technology, specifically to a renewable energy integrated optimization system and method. Background Technology

[0002] With the large-scale integration of renewable energy sources such as photovoltaics and wind power, the power grid faces significant intermittency and volatility issues. Renewable energy output is severely affected by factors such as weather and climate, often resulting in rapid rises and falls in power output over short timescales. This not only causes disturbances in grid voltage and frequency but also leads to energy waste problems such as wind and solar power curtailment.

[0003] To mitigate these issues, existing technologies employ time series or machine learning models to predict renewable energy power. However, these models often only provide predictions, and for scenarios like wind and solar power, which are highly susceptible to weather disturbances, the prediction error distribution tends to be skewed and exhibits heavy-tailed characteristics. The lack of quantification of prediction uncertainty makes it difficult for downstream scheduling optimization to identify the uncertainty risks associated with the prediction results. Furthermore, renewable energy power sequences vary drastically with weather, seasons, and diurnal cycles, exhibiting non-stationarity. If rare weather conditions occur (such as extreme weather), the lack of corresponding samples can increase the bias in model predictions, easily leading to control mismatch. Summary of the Invention

[0004] To address the problems of uncertainty and poor adaptability to non-stationarity in existing technologies, this invention provides a renewable energy integrated optimization system and method.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, this application discloses a renewable energy integration optimization method for power allocation in a hybrid energy storage system having a first energy storage unit and a second energy storage unit. The method includes the following steps:

[0007] Acquire the preprocessed dataset collected in the current cycle. The dataset includes renewable energy power, grid status data, and power and charge status of hybrid energy storage systems.

[0008] The dataset is input into a pre-trained Gaussian model to obtain the predicted mean and variance of the predicted renewable energy power within a preset future time window;

[0009] The mismatch value is obtained by calculating the difference between the predicted mean and the grid-connected power reference value, which is obtained from the grid-connected power data of the previous cycle.

[0010] The mismatch value is decomposed into fluctuation components. The decomposition results are estimated based on the prediction variance and the allocation of the previous cycle to obtain the reference allocation power applicable to the hybrid energy storage system.

[0011] The optimal allocated power is obtained by allocating renewable energy power; wherein the optimal allocated power satisfies the condition that the difference between the reference allocated power and the optimal allocated power does not exceed the limit.

[0012] A feasible region is constructed based on the power and charge state of the hybrid energy storage system. The optimal power allocation is projected into the feasible region to determine whether it exceeds the limit. If it does, the adjustment amount is determined based on the deviation and change of the grid state data between the current cycle and the previous cycle. The optimal power allocation is adjusted until it is within the feasible region.

[0013] The output contains control commands for the optimal power allocation within the feasible region.

[0014] Secondly, this application discloses a renewable energy integrated optimization system, which is applied to the aforementioned renewable energy integrated optimization method, including a data acquisition module, a data prediction module, a mismatch value calculation module, a reference allocation power estimation module, a power allocation module, a power adjustment module, and a result output module.

[0015] The data acquisition module is used to acquire the dataset collected in the current cycle and preprocessed. The dataset includes renewable energy power, grid status data, and power and charge status of the hybrid energy storage system.

[0016] The data prediction module is used to input the dataset into a pre-trained Gaussian model to obtain the predicted mean and variance of the predicted renewable energy power within a preset future time window;

[0017] The mismatch value calculation module is used to calculate the mismatch value by comparing the predicted mean with the grid-connected power reference value, which is obtained from the grid-connected power data of the previous cycle.

[0018] The reference allocation power estimation module is used to decompose the mismatch value into fluctuation components, and estimate the decomposition results based on the prediction variance and the allocation situation of the previous cycle to obtain the reference allocation power applied to the first energy storage unit and the second energy storage unit.

[0019] The power allocation module is used to allocate renewable energy power to obtain the optimal allocation power; wherein, the optimal allocation power satisfies the condition that the difference between the reference allocation power and the optimal allocation power does not exceed the limit.

[0020] The power adjustment module is used to construct a feasible region based on the power and charge state of the hybrid energy storage system, project the optimal power allocation into the feasible region, determine whether it exceeds the limit, and if so, determine the adjustment amount based on the deviation and change of the grid state data between the current cycle and the previous cycle, and adjust the optimal power allocation until it is within the feasible region.

[0021] The result output module is used to output control commands containing the optimal power allocation within the feasible region.

[0022] Thirdly, this application discloses a computer terminal, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the aforementioned renewable energy integration optimization method.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] This application can simultaneously utilize the predicted mean and variance to quantify uncertainty, and combine fluctuation decomposition and the differentiated characteristics of hybrid energy storage to achieve dynamic power allocation; by constructing a feasible domain based on energy storage power and charge state and introducing a state deviation correction mechanism between adjacent cycles, the feasibility and robustness of the allocation results are guaranteed, thereby improving the stability and utilization rate of renewable energy grid connection. Attached Figure Description

[0025] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0026] Figure 1 This is a flowchart of the renewable energy integration optimization method described in this invention;

[0027] Figure 2 For based on Figure 1 The logical flowchart for estimating the decomposition results based on the predicted variance and the allocation of the previous period.

[0028] Figure 3 Simulation diagram of mismatch value from the time domain perspective;

[0029] Figure 4 The simulation plot shows the contour lines of the objective function and the trajectory of the solution.

[0030] Figure 5 For based on Figure 1 The flowchart for parameter optimization;

[0031] Figure 6 For based on Figure 1 Application scenarios;

[0032] Figure 7 This is a block diagram of the renewable energy integrated optimization system described in Example 2;

[0033] Figure 8 This is a block diagram of the computer terminal described in Example 3. Detailed Implementation

[0034] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0035] Application Overview

[0036] In existing technologies, renewable energy power prediction models typically output only a single prediction value, lacking quantitative analysis of the prediction error distribution. When encountering extreme weather or rare operating conditions, prediction deviations increase, leading to a mismatch between energy storage system dispatch commands and actual power demand. Traditional methods employ fixed allocation strategies to handle prediction errors, failing to dynamically adapt to changes in grid conditions. In scenarios with rapid power fluctuations, this can easily cause overcharging and over-discharging of energy storage units, resulting in accelerated equipment aging.

[0037] To address the aforementioned issues, research has revealed that the distribution characteristics of prediction errors directly impact the robustness of scheduling decisions, necessitating the establishment of a dynamic mapping relationship between prediction uncertainty and energy storage response capabilities. Analysis of historical operational data reveals a strong coupling relationship between grid state parameters and the state of charge (SOC) of the energy storage system, leading to the proposal of a closed-loop control framework that jointly optimizes prediction results and real-time conditions. For non-stationary power sequences, a probabilistic model is employed to quantify prediction risks, combined with fluctuation component decomposition to achieve differentiated responses from energy storage units. Following the introduction of the basic concept of this invention, embodiments will be described in detail below with reference to the accompanying drawings.

[0038] Example 1

[0039] like Figure 1 The diagram illustrates an integrated optimization method for renewable energy, including the following steps:

[0040] S100. Obtain the dataset collected in the current cycle and preprocessed. The dataset includes renewable energy power, grid status data, and power and charge status of the hybrid energy storage system.

[0041] 1) Data collection:

[0042] Based on the built-in monitoring modules of photovoltaic inverters and wind turbine converters, the actual output active and reactive power are collected in the current cycle (e.g., 5s or 1min). At the same time, environmental data (e.g., light intensity, wind speed, wind direction, and temperature) are recorded synchronously based on information collected from weather stations as auxiliary features for prediction input.

[0043] The system collects grid state parameters such as grid voltage, frequency, and frequency change rate by installing voltage transformers, current transformers, and phasor measurement units at the grid connection point; it also collects operating constraints, electricity price signals, or power reference commands issued by the grid dispatch center.

[0044] The power, voltage, current, and state of charge of the hybrid energy storage system are obtained through the battery management system and supercapacitor controller, including the maximum charge and discharge power limits of each unit. 1. Safe SoC boundary. It should be noted that the first and second energy storage units included in the hybrid energy storage system are the Battery Energy Storage Unit (BESS) and the Supercapacitor Energy Storage Unit (SCES), respectively.

[0045] 2) Data synchronization:

[0046] The collected multi-source data are aligned according to a unified timestamp. If there is missing data, linear interpolation, moving average, or nearest neighbor is used to fill in the missing data to ensure that all input variables correspond one-to-one with the current period.

[0047] 3) Data preprocessing:

[0048] High-frequency noise is filtered using wavelet filtering or low-pass filtering to remove spikes and outliers. Anomaly detection is performed on power grid state variables and power data to eliminate invalid points. Various data types (e.g., power, frequency, voltage, and state of charge (SoC)) are normalized according to their physical dimensions, scaling data with different dimensions to a unified interval. The normalized data is then standardized to meet zero-mean and unit-variance distributions, improving the prediction accuracy and stability of the Gaussian model. It should be noted that for numerical continuous variables in the collected data, both normalization and standardization are required. However, for state variables, such as switch states, standardization is not performed; only encoding (0 / 1) is needed. For quantities already within a unified interval, only standardization is performed, omitting normalization.

[0049] 4) Dataset Construction:

[0050] The aligned and preprocessed data is organized into feature vectors. Build dataset , as the input for the current cycle.

[0051] S200. Input the dataset into the pre-trained Gaussian model to obtain the predicted mean and variance of the predicted renewable energy power within the preset future time window.

[0052] The specific training process of the Gaussian model (GPR) is as follows:

[0053] Obtain the training sample set from historical periods: ;

[0054] The input feature vector includes renewable energy power, grid status, energy storage power and SoC, environmental variables, etc. For the corresponding target output (actually observed renewable energy power), the data in the training set has undergone corresponding preprocessing operations such as time alignment, outlier removal, normalization, and standardization; N is the number of samples, and i is the sample code.

[0055] Assume the objective function Follows a Gaussian process:

[0056]

[0057] Represents a Gaussian process. Represents the mean function, Represents the kernel function.

[0058] The covariance matrix of the training samples is calculated based on the kernel function: That is, the covariance between the i-th input sample and the j-th input sample.

[0059] The expression after adding the noise variance term is as follows: ; For noise variance, It is an identity matrix.

[0060] The hyperparameters that need to be optimized in this model include:

[0061] Length scale (Determines the range of influence of input features on output);

[0062] Signal variance (Amplitude of function change);

[0063] noise variance .

[0064] Optimize hyperparameters by maximizing the marginal log-likelihood (MLL):

[0065]

[0066] The target output vector for the training samples; for Transpose of; The input set for the training samples contains all input feature vectors; For hyperparameter set; Covariance matrix The determinant; Covariance matrix The inverse matrix; N is the number of samples.

[0067] Optimize using gradient descent or quasi-Newton methods to obtain the optimal parameters:

[0068]

[0069] This refers to the optimized length scale, signal variance, and noise variance.

[0070] Once the Gaussian model is trained, the feature vectors are... Input, based on Constructing the input feature vector of the point to be predicted using historical data .

[0071] Input feature vector of the point to be predicted The predicted distribution is:

[0072]

[0073] This indicates a normal distribution.

[0074] The predicted mean and predicted variance are:

[0075]

[0076]

[0077] In the input Based on this, the future time window length H is set (e.g., 15 minutes or 1 hour, etc.), and the model outputs the predicted mean sequence of each prediction point within this window. With the corresponding predicted variance sequence The forecast mean reflects the most likely trend in renewable energy power output; the forecast variance reflects the confidence level and uncertainty of the forecast.

[0078] Package the predicted mean and variance into a single prediction result:

[0079]

[0080] S300. The mismatch value is obtained by calculating the difference between the predicted mean and the grid-connected power reference value, which is obtained from the grid-connected power data of the previous cycle stored in the database.

[0081] Before calculating the mismatch value, the load power at the grid connection point is obtained and it is determined whether it exceeds the preset third threshold. If so, the load power and the grid connection power reference value are used together with the predicted average value to calculate the mismatch value.

[0082] Obtain the load power at the grid connection point It is collected in real time by the data acquisition equipment installed at the grid connection point.

[0083] The grid-connected power data of the previous cycle is preferably stored in the controller storage unit on the grid connection point side or in the historical database of the station control server. This storage unit is used to call the grid-connected power of the previous cycle as a reference value to input into the Gaussian prediction model in the current cycle. At the same time, the data can also be uploaded to the cloud database for long-term Bayesian optimization and model retraining.

[0084] Time alignment and noise reduction are performed on the grid-connected power data from the previous cycle to obtain the grid-connected power reference value for the previous cycle. Calculate the reference value for grid-connected power:

[0085]

[0086] The exponential weighting coefficient (between 0 and 1) determines the weight of the current measurement and historical reference. The closer it is to 1, the greater the weight of the current data; the closer it is to 0, the greater the weight of historical smoothing. This represents the measured active power at the grid connection point during the current period t.

[0087] (Third threshold) indicates that the system load exceeds the neglected index, so a joint calculation is performed:

[0088] Mismatch value: ;

[0089] Combined reference power: ;

[0090] Adaptive weighting factor ; This is the parameter for the width of the buffer band. ; The rated power of the system (or the rated capacity of the grid connection point); This is a truncation function that restricts the input z to the interval [0,1].

[0091] This method suppresses noise and over-response while ensuring real-time performance, and automatically increases conservatism when the load is abnormal.

[0092] Among them, regarding the third threshold You can choose to set it dynamically, specifically:

[0093]

[0094] in, The proportionality coefficient (between 0 and 1); The rated power of the system (or the rated capacity of the grid connection point); For dynamic adjustment coefficients, The variance is predicted from a Gaussian model.

[0095] The mismatch value is: .

[0096] When calculating the mismatch value, To achieve a uniform time step and smoothing, First-order filtering can be used. If there is a lack of load power... If it is less than the third threshold, then it is considered to be less than the third threshold. If the grid-connected power reference value is missing Replace with the most recent valid value and save. , This will be used for subsequent learning and review of the BO outer ring.

[0097] Final output mismatch value sequence .

[0098] S400. Decompose the mismatch value into fluctuation components, estimate the decomposition results based on the prediction variance and the allocation of the previous cycle, and obtain the reference allocation power applied to the hybrid energy storage system.

[0099] like Figure 2 As shown, specifically:

[0100] The mismatch value fluctuation is decomposed into a first component and a second component with different frequencies. The first component and the second component correspond to the first energy storage unit and the second energy storage unit, respectively.

[0101] The predicted variance is smoothed and used as the uncertainty weight, and the allocation index of the first and second energy storage units in the previous period is obtained.

[0102] The reference allocated power applied to the hybrid energy storage system is obtained by multiplying the uncertainty weight and the corresponding allocation index as coefficients with the first component and the second component, respectively.

[0103] In this embodiment, the first component For the slowly varying component and the second component Take the rapidly changing component as an example.

[0104] like Figure 3 The image shown is a simulation diagram of the mismatch value from a time-domain perspective. Figure 3 As can be seen from the graph, the horizontal axis represents time, and the vertical axis represents the mismatch value and the frequency corresponding to the component. Blue represents the original signal of the mismatch value, orange represents the slow-changing component, the curve is smooth, close to the "large trend line", the fluctuation amplitude is large (tens of kW), but it is almost unchanged within a short time window; green represents the fast-changing component, the period is short (a few seconds to tens of seconds), the jitter is very frequent; the amplitude is small (a few kW to tens of kW), the curve is like a sawtooth / wave, constantly undulating.

[0105] Low-pass filter produces slowly varying components:

[0106]

[0107] This is the mismatch value. For slowly varying components, For the previous time step ( Slowly varying components of ); smoothing factor , The time cutoff constant can be selected from 15 to 60 seconds in this embodiment; The sampling period.

[0108] The fast change component is: .

[0109] Smoothing prediction variance (suppressing spikes): Smoothing factor ; This represents the prediction variance at the previous time step.

[0110] Normalization to weights The greater the uncertainty, the smaller the weight. That is, the uncertainty weight:

[0111]

[0112] For the predicted standard deviation; For reference scale, (Rated power).

[0113] Retrieve the allocation index from the previous period from the controller storage unit or the historical database of the station control server stored at the grid connection point. .

[0114] Calculate the reference power allocation:

[0115]

[0116]

[0117] The reference power allocation is trimmed and transferred. If a power exceeds the limit, it is trimmed first, and then the excess part is preferentially transferred to the same-direction component of another unit to ensure that the reference power allocation and the component sign are consistent. A smooth transition is selected for synthesis verification according to actual needs.

[0118] Final output reference power allocation sequence: .

[0119] The purpose of this measure is to dynamically adapt the response characteristics of energy storage units according to the fluctuation frequency characteristics, suppress the impact of prediction uncertainty on allocation through variance weighting, and maintain the temporal stability of the strategy by using historical allocation index, thereby reducing power allocation deviation and improving the dynamic adjustment capability of hybrid energy storage system.

[0120] S500. Allocate renewable energy power to obtain the optimal allocated power; wherein the optimal allocated power satisfies the condition that the difference between the reference allocated power and the optimal allocated power does not exceed the limit.

[0121] Specifically:

[0122] Based on the predicted mean and predicted variance, the first constraint is obtained by tightening the opportunity constraint on the power grid state data; the power grid state data includes the grid-connected power and charge status collected in the current cycle.

[0123] Construct a second constraint that the difference between the reference allocated power and the optimal allocated power does not exceed the limit, and a third constraint that the optimal allocated power does not exceed the limit, which together with the first constraint form a constraint set;

[0124] Construct an objective function that includes at least reducing grid jitter and rationally allocating the power of the hybrid energy storage system. Based on the optimizer, solve the constraint set and the objective function pair to obtain the optimal power allocation.

[0125] 1) First constraint:

[0126] The probabilistic constraint of "grid state not exceeding limits" is transformed into deterministic tightening. Take grid connection smoothing as an example:

[0127]

[0128] in, (.) represents probability. To predict the mean, This represents the grid-connected power at the previous time step; This is the maximum ramp rate allowed for grid-connected power. This represents the failure probability of the ramp constraint, i.e., the probability that the constraint is allowed to be violated, and is usually taken as a small value (such as 0.05). The actual power allocated to the first and second energy storage units.

[0129] If the prediction error is approximately Gaussian, take the standard normal quantile:

[0130]

[0131] in, It is the inverse cumulative distribution function (quantile function) of the standard normal distribution, used to transform probability constraints into deterministic tightening terms.

[0132] Grid-connected power upper / lower limit probability constraints:

[0133]

[0134] in, , These are the upper and lower limit boundary values ​​for grid-connected power.

[0135] The corresponding deterministic tightening is:

[0136]

[0137] in, Let represent the failure probability of the upper and lower limits of grid-connected power constraints, and let represent the tolerance probability of power exceeding the limits.

[0138] 2) Second constraint:

[0139] Limit the deviation of the actual allocation from the reference allocation, and adapt to uncertainty:

[0140]

[0141]

[0142] in, The actual power allocated to the first or second energy storage unit. The reference power allocation for the first or second energy storage unit. This is the maximum allowable deviation for the first or second energy storage unit. For uncertain weights, This is the maximum allowed fixed deviation value for the first or second energy storage unit. This is the index for the energy storage unit.

[0143] 3) Third constraint:

[0144]

[0145]

[0146]

[0147] in, For the first The rated maximum charge and discharge power of each energy storage unit. For the previous time step The actual power allocated to each energy storage unit; For the first The maximum ramp rate (i.e., power change rate limit) of each energy storage unit. For the next time step The state of charge of each energy storage unit For the first The state of charge of each energy storage unit; The sampling period is For quota capacity, For charging efficiency, For discharge efficiency, This is a positive part operator used to distinguish between charging power (positive) and discharging power (negative). To allow SoC range.

[0148] If we consider the health of the temperature, we can increase the temperature rise, approximately as follows:

[0149]

[0150] in, For the first The temperature of each energy storage unit This represents the highest permissible temperature threshold.

[0151] Construct the objective function:

[0152]

[0153] in, The length of the future time window; For time indexing; To smooth out the weights during grid connection, This represents the change in grid-connected power. For reference tracking weights; For the energy storage degradation cost function, , For degenerate weights; The actual power allocation for the first and second energy storage units. The reference power allocation for the first and second energy storage units.

[0154] The optimizer can be QP / SOCP, assembling a constraint set. With the objectives of minimizing grid-connected power jitter, reference allocation tracking, and suppressing energy storage degradation, a quadratic objective function is established. A quadratic programming / second-order cone optimizer is then used for rolling solution to obtain the optimal solution within the constraint set. .

[0155] Only step 1 is issued: The remaining storage.

[0156] like Figure 4 The figure shows the contour lines of the objective function and the trajectory of the solution. The horizontal axis P1 represents the power of the battery energy storage unit (kW), and the vertical axis P2 represents the power of the supercapacitor energy storage unit (kW). The contour lines represent the objective function. Red indicates the iterative trajectory of the optimizer, blue squares represent the starting point, and gold represents the ending point (a total of 20 iterations). The trajectory first rapidly descends to a trough region, then undergoes fine-tuning within the trough, and finally converges within the constraint boundaries.

[0157] S600. Construct a feasible region based on the power and charge state of the hybrid energy storage system, project the optimal power allocation into the feasible region, determine whether it exceeds the boundary, and if so, determine the adjustment amount based on the deviation and change of the grid state data between the current cycle and the previous cycle, and adjust the optimal power allocation until it is within the feasible region; the calculation process for adjusting the optimal power allocation is as follows:

[0158] Obtain the grid status data of the current cycle and the previous cycle, extract the grid connection frequency and calculate the frequency difference, and calculate the frequency change value by comparing it with the sampling period;

[0159] The deviation is calculated by the difference between the current grid connection frequency and the rated reference frequency.

[0160] The gain adjustment function assigns coefficients to the deviation and frequency change values, and then the two are summed to obtain the adjustment amount.

[0161] Existing fixed-weight allocation methods are ill-suited to the dynamic demands of different operating conditions. For instance, when extreme weather causes drastic fluctuations in grid frequency, fixed weights may lead to overcompensation or undercompensation. This application addresses this issue by generating composite adjustment quantities based on the instantaneous deviation and dynamic trend of the grid state, thus avoiding control mismatch caused by a single indicator under extreme conditions. A dynamic weight allocation mechanism balances the impact of current deviations and future trends, suppressing deviation expansion and improving the response speed of power allocation. A gain adjustment function is utilized to adapt to different grid operating scenarios, enhancing the robustness and adaptability of adjustment quantity calculation.

[0162] Constructing the feasible domain:

[0163] With the first Taking one energy storage unit as an example (i∈{1,2}):

[0164]

[0165]

[0166] This represents the charging / discharging power of the energy storage unit; positive indicates discharging, and negative indicates charging. For the first The maximum allowable power of each energy storage unit For the first The power of each energy storage unit at time t; For the first The maximum ramp rate of each energy storage unit; The state of charge at the next time t+1 is predicted under a given power P; Let represent the temperature state of the i-th energy storage unit under power P; For the first The maximum allowable temperature for each energy storage unit; For the first The state of charge of an energy storage unit at time t; Let be the charging and discharging efficiency of the i-th energy storage unit; To take the positive part of P, i.e., the charging power part; To obtain The positive part, i.e., the discharge power part.

[0167] The overall feasible region is a Cartesian product: .

[0168] Euclidean projection of the optimal power allocation:

[0169]

[0170] Let be the projected power of the i-th energy storage unit at time t+1 (the value after projection onto the feasible region). The optimal power allocation for the i-th energy storage unit at time t+1 (before projection). This is the projection operator.

[0171] like ,but Otherwise, it is projected to the boundary (power saturation / ramp limit / SoC boundary).

[0172] Calculate the frequency difference: ;

[0173] Indicates the grid connection frequency for the current cycle. Indicates the rated reference frequency.

[0174] Frequency change value: .

[0175] Proportional gain: ;

[0176] Derivative gain: ;

[0177] Frequency change rate risk indicator History | quantile estimates of | To predict uncertainty indicators, the standard deviation of the prediction variance can be used; in this embodiment, ∈[0.05,0.3]、 ∈[0.02,0.2].

[0178] First Energy Storage Unit (BESS) =1), the second energy storage unit (SCES, denoted as 1), =2), then:

[0179] BESS proportional allocation factor The distribution coefficient of the derivative term of BESS ;

[0180] SCES proportional term allocation factor The derivative term distribution coefficient of SCES .

[0181] Among them, allocation factor ∈[0.5,0.8] (can be the distribution index of the previous period) (Or obtained through BO outer ring learning).

[0182] The adjustment amount is: .

[0183] Superimposed on the projected baseline value and saturated: ; This is a joint saturation operator for boundaries such as power, ramp-up, SoC, and temperature. Then, an out-of-bounds check is performed. If no out-of-bounds error occurs, the process ends; otherwise, it backtracks and scales the adjustment amount, repeating the "projection → superposition → saturation → check" process until it is feasible or the maximum number of iterations (e.g., 3–5 times) is reached. If it is still not feasible, then the result is directly taken. (Pure projection solution).

[0184] S700 outputs control commands containing the optimal power allocation within the feasible region.

[0185] After obtaining the optimal power allocation within the feasible region, the system performs amplitude limiting and smoothing according to a unified unit and direction. Based on the communication protocol, it encapsulates control commands containing device identifiers, timestamps, serial numbers, validity periods, power setpoints, and boundary mirrors, and sends them to the corresponding hybrid energy storage system through the station control gateway. Furthermore, it reads back the execution results and performs deviation verification with the commands. If the deviation exceeds the limit or the confirmation fails, it executes degradation and safety measures to ensure the executability, real-time performance, and security of the control commands.

[0186] Compared with existing static power allocation methods that use deterministic prediction results, which are prone to conflicts between control commands and physical constraints when prediction errors are large, this scheme dynamically adjusts the allocation weights by adjusting the prediction variance, introduces opportunity constraints into the optimization objective to handle uncertainty risks, and adopts a feasible region projection mechanism to ensure the physical feasibility of control commands, so as to maintain the stable operation of the power grid even under extreme conditions.

[0187] Through the above technical solutions, this application effectively mitigates the dispatch risks caused by forecast uncertainty, adapts to the time-varying characteristics of the power grid through a dynamic adjustment mechanism, and avoids control command mismatch under extreme weather conditions. It utilizes fluctuation component decomposition to achieve coordinated response of energy storage units, reducing the impact of rapid power fluctuations on the power grid. Feasible domain projection and online adjustment mechanisms ensure the safe operation of the energy storage system, extend equipment lifespan, and improve the reliability and economy of renewable energy grid connection.

[0188] The main scheme of this embodiment has been described in detail above. The following section details the parameter optimization operation, which, after outputting a control command containing the optimal power allocation within the feasible region, also includes determining the parameter optimization operation based on the time interval between the current cycle and the previous parameter optimization cycle. Figure 5 As shown, the specific steps are as follows:

[0189] S801. Retrieve the period of the last parameter optimization from the database, and calculate the time interval by calculating the time difference between the current period and the previous period;

[0190] S802. Determine whether the time interval is greater than the preset update interval. If so, obtain the average of all predicted values ​​and the actual power from the last parameter optimization to the current cycle. Calculate the error by comparing the two values ​​and determine whether the error is within the preset tolerance range.

[0191] S803. Otherwise, adjust the weights of the performance evaluation targets based on error decomposition, and optimize the pre-stored proxy model with the adjusted performance evaluation targets; the performance evaluation targets shall include at least the grid-connected power smoothness with different weight ratios, the degradation degree of the hybrid energy storage system, and the default degree based on the first constraint;

[0192] S804. Obtain the dataset between the current cycle and the previous parameter optimization cycle and input it into the optimized surrogate model for training. Use the parameters learned by the surrogate model to update the Gaussian model and the objective function.

[0193] If the time interval is greater than the preset update interval (e.g., 6 hours or 24 hours, set according to actual needs), and the error between the predicted mean and the actual power of the previous optimization to the current period exceeds the tolerance range, then an optimization operation is performed. That is, the power smoothness error, energy storage degradation error, and constraint default error are calculated separately, and the weights in the objective function are adjusted according to the error. If the update interval is reached and the error does not exceed the tolerance range, then no optimization is performed to avoid overfitting.

[0194] The initial construction of the surrogate model is based on the performance evaluation objective, constructing a surrogate model (GPR or conditional GP). Historical data is collected within a calibration time window: input features include predicted power, energy storage status, grid status, etc., and output features include evaluation indicators such as power smoothness, degradation degree, and default degree. The data is divided into training and validation sets.

[0195] The surrogate model is trained using the training set, and the stopping condition is:

[0196] 1. The MLL increment of the surrogate model is less than the threshold. ( ),or

[0197] 2. Validation set error is below the threshold. (1%–10%), or

[0198] 3. Reach the maximum number of iterations (50-200 times).

[0199] Once the proxy model is obtained, subsequent updates can be based on this proxy model.

[0200] Unify the required optimization variables (such as objective function weights, allocation factors, etc.) into a design variable vector xs;

[0201] Weather / load levels, time periods, forecast variance statistics, and recent operating condition labels are used to construct conditional BOs, forming contextual features c.

[0202] Within the data collection window, i.e., within the time interval window, the following statistics are performed:

[0203] Grid connection jitter: ; This represents the difference in grid-connected power at adjacent times;

[0204] Degradation costs: (Quadratic or piecewise linear approximation); These are the degradation costs of the first and second energy storage units, respectively.

[0205] Default rate: (Opportunity constraint softening amount or number of times the boundary is exceeded);

[0206] Combined into the overall goal (performance evaluation goal):

[0207]

[0208] These are the adjusted weighting coefficients.

[0209] Constraints (for secure business operation): default rate not exceeding maximum tolerance, maximum temperature not exceeding limits, SoC hold-up range, etc.

[0210] Based on the parameters currently in use, i.e., the design variable vector xs0 from the last BO optimization, generate N0 sets of Latin hypercube / multi-startup perturbations {x j}, Execute steps S500-700 to obtain {y j}, forming a dataset .

[0211] Update the surrogate model, i.e., use a conditional Gaussian process: The kernel function can be RBF / composite kernel; the hyperparameters are trained using marginal log-likelihood (MLL).

[0212] For Gaussian processes, It is a mean function. This is the kernel function.

[0213] If BO needs to be constrained, then each constraint index Also train a GPR agent.

[0214] Then, candidate point determination is performed: The desired improvement function is set. ; For expectation operators; The best known observation (i.e., the best objective function value in history); The predicted mean (given by the GPR surrogate model); It is a positive part operator.

[0215] If constraints are applied, then the acquisition function of BO is constrained. ; This represents the probability of falling into the feasible region.

[0216] Multiple-start L-BFGS, particle swarm optimization, or mesh + local optimization methods can be used to solve the problem. And apply a trust domain: ;

[0217] To optimize the acquisition function The candidate solutions obtained; express With the current best Maximum difference across all dimensions; The initial trust domain is defined as the radius of the trust domain. If the performance of the new sampling point is improved during subsequent runs, the trust domain is expanded. If the improvement is insufficient, the trust domain is shrunk. If there is no improvement after M consecutive optimizations, or the maximum number of iterations or the time budget is reached, the trust domain adjustment is stopped.

[0218] If the EI value is below the threshold or reaches the maximum number of iterations, it indicates that the BO outer loop has converged. When applied to a real-world system, parameters are updated, and performance feedback and data write-back are performed, facilitating the updating of the proxy model.

[0219] Specifically, this technical solution addresses the insufficient adaptability of prediction models under non-stationary operating conditions through a periodic parameter update mechanism. It dynamically adjusts prediction model parameters based on system operating status, effectively addressing the non-stationary characteristics of renewable energy power. Under rare conditions such as extreme weather, error decomposition and weight adjustment mechanisms enhance the model's adaptability to abnormal data, reducing control mismatch caused by prediction bias. Incremental training and multiple termination conditions are employed to maintain timely model updates while avoiding excessive consumption of computational resources. Ultimately, this achieves long-term stability of the power allocation strategy for hybrid energy storage systems, improving the robustness of the power grid under complex operating conditions.

[0220] After adjusting the weights of the performance evaluation target based on error decomposition or updating the surrogate model, the adjusted performance evaluation target or the updated surrogate model is validated in different scenarios based on the digital twin model. If the validation fails, the weight correction of the performance evaluation target or the retraining of the surrogate model is triggered.

[0221] The digital twin model is obtained by coupling a physical model built on historical datasets and a perturbation generator trained on historical datasets.

[0222] This approach, through the synergy of a physical model and a perturbation generator, simultaneously covers the known data distribution and the boundary of the theoretical feasible domain during the verification phase. This solves the verification blind spot problem of single data-driven methods under rare operating conditions and avoids the risk of model overfitting due to historical data bias.

[0223] First, let's discuss the preparation for digital twins:

[0224] A cascaded model (Twin-P) of the power plant layer and energy storage layer was established based on historical operating data and equipment nameplate parameters.

[0225] Grid side: Equivalent grid + droop control / connection point model (PCC, line impedance, frequency / voltage dynamics);

[0226] Energy storage side: BESS / SCES power-SoC-temperature coupling model (including charging efficiency, equivalent series resistance, charge transfer resistance, thermal capacity parameters, etc.);

[0227] Control side: The filtering, constraint, model prediction control objectives and execution logic used in S300–S700 are mirrored implementations.

[0228] Historical data is used for parameter identification and calibration (least square / Bayesian estimation) to ensure that the mean absolute error (MAE) and root mean square error (RMSE) of the simulation output meet the standards of historical measurements, thereby obtaining the physical model.

[0229] Then, the perturbation generator (Twin-D) is trained, that is:

[0230] Historical weather, load, electricity price, and random fault sequences are used as the training set to train a time-series disturbance generator. The disturbance generator can employ ARIMA (Autoregressive Integral Moving Average), VAR (Multivariate Vector Autoregressive), VAE (Variational Autoencoder), LSTM (Long Short-Term Memory Neural Network), or Diffusion (Diffusion Model) / GAN (Generative Adversarial Network) models (selected based on on-site computing power and accuracy). The goal is to generate scenarios with controllable intensity consistent with historical statistics: wind / solar fluctuations, load spikes, communication delays / packet loss, component derating, etc. The statistical consistency of the time-series disturbance generator output is constrained: mean, variance, autocorrelation, spectral density, and extreme value quantiles deviating from historical values ​​must be ≤ a preset threshold (e.g., 10%).

[0231] Finally, perform dual-mode coupling:

[0232] Perturb the scene generated by Twin-D Inject the corresponding Twin-P interfaces: resource output, load, equivalent power grid parameters, communication delay, etc., to form a repeatable virtual test bench.

[0233] These represent meteorological disturbances, load disturbances, power grid disturbances, and communication disturbances, respectively.

[0234] After adjusting the weights of the performance evaluation target based on error decomposition or after training / updating the surrogate model, determine the objects to be verified: the adjusted target weights and the model constructed from them to predict the control target, and the updated surrogate model.

[0235] Sampling N from Twin-D sc Grouping scenes:

[0236] Typical scenario: 50th / 75th percentiles of solar radiation / wind speed / load distribution by season;

[0237] Stressful scenarios include: rapid shading, wind shear, sudden load increase, voltage step, and communication delays of 200–500 ms.

[0238] Rare scenarios: extreme low / high temperatures, component derating, single PCS failure bypass.

[0239] Each scenario group has an evaluation window (e.g., 1–6 hours) and a control step size (consistent with the actual situation).

[0240] In each scenario, Twin-P is driven by "current online control parameters + object to be verified".

[0241] Record KPIs step by step:

[0242] 1. Grid connection jitter: ; This represents the difference in grid-connected power at adjacent times;

[0243] 2. Degradation costs: (Quadratic or piecewise linear approximation); These are the degradation costs of the first and second energy storage units, respectively.

[0244] 3. Degree of breach of contract: (Opportunity constraint softening amount or number of times the boundary is exceeded);

[0245] 4. Safety parameters: Whether the SoC is within limits, the maximum temperature of the hybrid energy storage system, and the number of power ramp-up / saturation cycles of the energy storage converter PCS;

[0246] 5. Robustness: Sensitivity curves for disturbance strengths ∈ [0.5, 1.5].

[0247] The following conditions must be met simultaneously:

[0248] For reference smoothness (historical baseline or target value); The allowable increase in smoothness is (10%~20%).

[0249] For reference degradation (baseline lifetime consumption value); The allowable increase in degradation rate (5%~15%).

[0250] The default rate shall not exceed the limit (1%~5%) and the maximum deviation shall not exceed the limit. );

[0251] No hard temperature overruns in SoC / temperature, no sustained saturation in PCS;

[0252] Predictive calibration of the surrogate model: Coverage exceeds the minimum predictive calibration threshold (preferably 0.85).

[0253] If the scenario pass rate is greater than or equal to the minimum scenario pass rate (e.g., ≥80% normal + ≥60% stress), it is considered "passed"; otherwise, it is considered "failed".

[0254] If the judgment fails, weight adjustment and surrogate model retraining will be performed. Weight adjustment is as follows: if high-frequency jitter is too large: increase the grid connection smoothing weight; if degradation indicators exceed the standard, adjust the energy storage degradation weight; if the default rate is too high, increase the default penalty weight, and perform weight normalization and pruning to avoid excessive weight of certain indicators.

[0255] Agent model retraining:

[0256] Insufficient coverage → Increase noise prior or modify kernel function;

[0257] Systematic bias → Introduce new features (temperature, wind shear, hysteresis, etc.), or extend the training data window;

[0258] Overfitting → Add a noise variance term or use regularization.

[0259] If it still fails in round R (e.g., round 3), roll back to the last successful parameter set and freeze the current round of updates (record the reason and scenario).

[0260] The above section details the steps involved in parameter optimization. The following section provides a detailed explanation of how to adjust data compression measures based on network conditions. Specifically, this includes:

[0261] When acquiring the pre-processed dataset collected in the current period, the network state parameters are acquired simultaneously. After normalization, the corresponding compression measures are obtained by referring to a pre-set lookup table that represents the network state and compression measures. The parameter data of the current period are then reported according to the corresponding compression measures.

[0262] The system collects current-cycle grid operation parameters from devices such as SCADA / PMU / RTU, including grid-connected power, voltage, frequency, RoCoF, and sampling period. Preprocessing operations such as time synchronization and data cleaning are performed on the collected data. Then, normalization is performed to unify data of different dimensions into the [0,1] interval, facilitating subsequent comparison and compression decision-making.

[0263] The lookup table is generated using offline training or empirical configuration. For example, when bandwidth is below a threshold, it is mapped to a reduction in sampling rate; when packet loss rate is above a threshold, it is mapped to redundant coding measures. Compression measures refer to data processing strategies adopted for different network states. Specifically, they can be implemented using one or more combinations of data downsampling, lossy compression, and priority marking. For example, data compression algorithms can be enabled to reduce transmission load when the network is congested.

[0264] Based on the normalized indicators, the current network status level is determined by referring to the table below:

[0265] Table 1: Network Status - Compression Measures Comparison Table

[0266] Network status level Voltage / frequency deviation RoCoF Data bandwidth status Compression measures normal Deviation ≤2% RoCoF≤0.1Hz / s Sufficient bandwidth Full report (uncompressed) Sub-health Deviation 2%–5% RoCoF 0.1–0.3 Hz / s Bandwidth is generally Only report key parameters (such as power, average frequency, and 10Hz downsampling). urgent Deviation ≥5% RoCoF ≥ 0.3 Hz / s Bandwidth limited Extreme compression (reports only power average and frequency extremes, downsampling at 1Hz)

[0267] If it is in a normal state:

[0268] Transmit complete data (high frequency, full data, high bandwidth utilization).

[0269] If the patient is in a sub-healthy state:

[0270] Compression measure 1: Key parameters are retained, and some samples are reduced in frequency.

[0271] If it is an emergency:

[0272] Initiate compression measure 2: extreme compression, uploading only a small number of key parameters (average power, extreme frequency), and trigger an alarm flag.

[0273] The compressed data packets are reported to the edge server or cloud center according to the preset communication protocol (such as MQTT, IEC61850, DL / T860), and a compression level identifier is attached to the packet header so that the central system can restore the context when decoding.

[0274] This application automatically reduces the transmission accuracy of non-critical data when network conditions deteriorate, prioritizing the timely delivery of control commands and avoiding power allocation lag caused by communication delays. When the network stabilizes, it automatically removes compression restrictions to ensure data integrity, solving the problem of delayed response from manual configuration strategies in traditional methods. For example, in the event of momentary network jitter, the system can switch to a simplified data format within milliseconds, improving the success rate of transmitting critical control parameters to an operable level.

[0275] To facilitate understanding of the above embodiments, a specific application scenario of the above embodiments will be used as an example for illustration below:

[0276] Please combine Figure 6 Let the installed capacity of a certain site's new energy power station be:

[0277] Photovoltaics (PV): 12MWp;

[0278] Wind power (WT): 8MW;

[0279] Hybrid energy storage system (BESS+SCES): 6MW / 12MWh (BESS, battery energy storage unit), 3MW / 0.5MWh (SCES, supercapacitor unit), grid connection rated capacity = 20MW (grid connection power ≡ grid power).

[0280] Prediction model: Gaussian process regression (GPR) outputs the predicted mean and variance for the future H=60min, and backpropagates the uncertainty.

[0281] 1. Normal weather (load negligible), 12:00–13:00 sunny with light breeze:

[0282] GPR forecast: The forecast mean is between 9.2 and 10.1 MW, and the forecast variance is stable at 0.35 MW.

[0283] Grid connection reference: P ref=9.6MW;

[0284] The PCC load is 0.12MW, and the third threshold is 0.20MW, so it is ignored.

[0285] Mismatch value: The range is approximately −0.4–0.5MW.

[0286] The low-pass filter yields a slow-varying component of approximately +0.30MW (borne by BESS) and a fast-varying component of approximately +0.05MW (borne by SCES).

[0287] Calculate the reference power allocation:

[0288]

[0289]

[0290] The optimization objective is to minimize "grid jitter + reference tracking deviation + degradation cost" to obtain the actual allocation:

[0291] .

[0292] The solution has been verified to be feasible under power / ramp / SoC / temperature constraints.

[0293] After projecting onto the feasible region and superimposing a small frequency correction, control commands are issued; the readback deviation does not exceed the limit. Conclusion: The load is ignored, and the system mainly focuses on smoothing grid-connected power.

[0294] 15:00–16:00 Sudden rapid shading + wind shear + short-term PCC load surge.

[0295] GPR forecast: The forecast mean has decreased from 9.8MW to 6.5MW; the forecast variance has increased to 0.9MW.

[0296] Grid connection reference: P ref =9.4MW;

[0297] PCC load surged to 0.65MW, exceeding the third threshold of 0.25MW, indicating that it needs to be included.

[0298] Combined reference power: ≈8.9MW;

[0299] Mismatch value As low as −2.1MW. During extreme periods, with grid-connected power equal to grid power as the center, the load is considered a “conservative disturbance” to the reference, thus increasing the safety margin.

[0300] Slow-changing component ≈ −1.70MW (BESS preferentially absorbs / reduces load), fast-changing component ≈ −0.25MW. The uncertainty weight decreases as uncertainty increases, suppressing aggressive actions. Therefore:

[0301]

[0302]

[0303] The optimization objective is to minimize "grid jitter + reference tracking deviation + degradation cost" to obtain the actual allocation:

[0304] .

[0305] Grid connection and ramping are affected by Φ −1 (0.97) Tightened; SoC and temperature rise constraints are both satisfied.

[0306] After projecting onto the feasible region and superimposing a small frequency correction, control commands are issued; the readback deviation does not exceed the limit.

[0307] Control commands are issued through the station control gateway, and include a timestamp, device ID, and validity period to ensure executability and security.

[0308] Reviewing the operational data from the past three months, we found that:

[0309] Under similar weather conditions, increasing battery discharge 15 minutes earlier reduces power curve fluctuations.

[0310] Therefore, the system automatically updates the agent model and optimizes parameters, and will respond in advance when similar weather occurs again. Before deploying the new strategy, the system uses a digital twin model for virtual simulation:

[0311] Scenario 1: Normal sunshine + slight wind speed fluctuations;

[0312] Scenario 2: Sudden strong wind (wind speed increases by 12 m / s);

[0313] Scenario 3: Prolonged cloudy weather (solar power consumption drops by 70%);

[0314] The results show that:

[0315] In 98% of scenarios, the system can maintain a stable power supply with a default rate of ≤3%, and the SoC remains within a safe range (20%–90%). Verification passed; new strategy implemented.

[0316] This optimization method effectively smooths out renewable energy power fluctuations, maintaining grid-connected power within the range of 59.8–60.1 MW. The energy storage units have clearly defined roles: BESS handles slow-varying regulation, while SCES responds to fast-varying disturbances. Grid frequency deviation is controlled within the tolerable range, meeting grid connection standards. Simultaneously, energy storage degradation and health status are considered to avoid overuse of a single energy storage unit.

[0317] Example 2

[0318] like Figure 7 As shown in the figure, this embodiment introduces a renewable energy integrated optimization system, which is applied to the aforementioned renewable energy integrated optimization method. It includes a data acquisition module, a data prediction module, a mismatch value calculation module, a reference allocation power estimation module, a power allocation module, a power adjustment module, and a result output module.

[0319] The data acquisition module is used to acquire the dataset collected in the current cycle and preprocessed. The dataset includes renewable energy power, grid status data, and power and charge status of the hybrid energy storage system.

[0320] The data prediction module is used to input the dataset into a pre-trained Gaussian model to obtain the predicted mean and variance of the predicted renewable energy power within a preset future time window;

[0321] The mismatch value calculation module is used to calculate the mismatch value by comparing the predicted mean with the grid-connected power reference value, which is obtained from the grid-connected power data of the previous cycle.

[0322] The reference allocation power estimation module is used to decompose the mismatch value into fluctuation components, and estimate the decomposition results based on the prediction variance and the allocation situation of the previous cycle to obtain the reference allocation power applied to the first energy storage unit and the second energy storage unit.

[0323] The power allocation module is used to allocate renewable energy power to obtain the optimal allocation power under the condition that the difference between the reference allocation power and the actual allocation does not exceed the limit;

[0324] The power adjustment module is used to construct a feasible region based on the power and charge state of the hybrid energy storage system, project the optimal power allocation into the feasible region, determine whether it exceeds the limit, and if so, determine the adjustment amount based on the deviation and change of the grid state data between the current cycle and the previous cycle, and adjust the optimal power allocation until it is within the feasible region.

[0325] The result output module is used to output control commands containing the optimal power allocation within the feasible region.

[0326] In some specific implementations, Gaussian models can employ kernel function combinations to handle non-stationary data. For example, periodic kernel functions can be used to capture diurnal power variation trends, while radial basis function kernel functions can be used to fit short-term fluctuation characteristics. The feasible region construction can incorporate the energy storage unit's charge and discharge efficiency curves, for instance, limiting the state of charge to the 20%-90% range to avoid overcharging and over-discharging of the battery. Fluctuation component decomposition can utilize empirical mode decomposition methods, decomposing mismatch values ​​into high-frequency and low-frequency components, corresponding to the power allocation tasks of the supercapacitor and the battery, respectively.

[0327] This embodiment has the same beneficial effects as Embodiment 1.

[0328] Example 3

[0329] like Figure 8 As shown in the figure, this embodiment introduces a computer terminal, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the aforementioned renewable energy integration optimization method.

[0330] There can be one processor or multiple processors. Figure 8 Taking one example, in this embodiment, the processor and memory can be connected via a bus or other means, wherein, Figure 8 Taking the example of a bus connection, the corresponding input and output devices are also shown.

[0331] When applying integrated optimization methods for renewable energy, they can be implemented in software form, such as as a standalone program installed on a computer terminal, which could be a computer, smartphone, or similar device. Alternatively, they can be designed as an embedded program installed on a computer terminal, such as a microcontroller.

[0332] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A renewable energy integration optimization method, used for power allocation in a hybrid energy storage system having a first energy storage unit and a second energy storage unit, characterized in that, include: Acquire the dataset collected in the current cycle and preprocessed, which includes renewable energy power, grid status data, and power and charge status of the hybrid energy storage system; The dataset is input into a pre-trained Gaussian model to obtain the predicted mean and variance of the predicted renewable energy power within a preset future time window; The mismatch value is obtained by calculating the difference between the predicted mean and the grid-connected power reference value, which is obtained from the grid-connected power data of the previous cycle. The mismatch value is decomposed into fluctuation components, and the decomposition results are estimated based on the prediction variance and the allocation of the previous cycle to obtain the reference allocation power applied to the hybrid energy storage system. The optimal allocated power is obtained by allocating renewable energy power; wherein the optimal allocated power satisfies the condition that the difference between the reference allocated power and the optimal allocated power does not exceed the limit. A feasible region is constructed based on the power and charge state of the hybrid energy storage system. The optimal power allocation is projected into the feasible region to determine whether it exceeds the limit. If it does, the adjustment amount is determined based on the deviation and change of the grid state data between the current cycle and the previous cycle. The optimal power allocation is adjusted until it is within the feasible region. The output contains control commands for the optimal power allocation within the feasible region.

2. The renewable energy integrated optimization method according to claim 1, characterized in that, The specific steps for performing fluctuation component decomposition on the mismatch value, estimating the decomposition results based on the prediction variance and the allocation situation of the previous cycle, and obtaining the reference allocation power applicable to the hybrid energy storage system are as follows: The mismatch value fluctuation is decomposed into a first component and a second component with different frequencies, and the first component and the second component correspond to the first energy storage unit and the second energy storage unit, respectively. The predicted variance is smoothed and used as the uncertainty weight, and the allocation index of the first and second energy storage units in the previous period is obtained. The reference allocated power applied to the hybrid energy storage system is obtained by multiplying the uncertainty weight and the corresponding allocation index as coefficients with the first component and the second component, respectively.

3. The renewable energy integrated optimization method according to claim 1, characterized in that, The specific steps for obtaining the optimal allocated power by allocating renewable energy power, wherein the optimal allocated power satisfies the condition that the difference between the reference allocated power and the optimal allocated power does not exceed a limit, are as follows: Based on the predicted mean and predicted variance, the first constraint is obtained by tightening the opportunity constraint on the power grid state data; the power grid state data includes the grid-connected power and charge status collected in the current cycle. Construct a second constraint that the difference between the reference allocated power and the optimal allocated power does not exceed the limit, and a third constraint that the optimal allocated power does not exceed the limit, which together with the first constraint form a constraint set; Construct an objective function that includes at least reducing grid jitter and rationally allocating the power of the hybrid energy storage system. Based on the optimizer, solve the constraint set and the objective function pair to obtain the optimal power allocation.

4. The renewable energy integrated optimization method according to claim 3, characterized in that, The specific steps for determining the adjustment amount based on the deviation and changes in power grid status data between the current cycle and the previous cycle are as follows: Obtain the grid status data of the current cycle and the previous cycle, extract the grid connection frequency and calculate the frequency difference, and calculate the frequency change value by comparing it with the sampling period; The deviation is calculated by the difference between the current grid connection frequency and the rated reference frequency. The gain adjustment function assigns coefficients to the deviation and frequency change values, and then the two are summed to obtain the adjustment amount.

5. The renewable energy integrated optimization method according to claim 3, characterized in that, After outputting the control command containing the optimal power allocation within the feasible region, the process also includes determining the parameter optimization operation based on the time interval between the current cycle and the previous parameter optimization cycle. The specific steps are as follows: Retrieve the period of the last parameter optimization from the database, and calculate the time interval by comparing the current period with the time difference between them. If the time interval is greater than the preset update interval, then the average of all predicted values ​​from the last parameter optimization to the current cycle and the actual power are obtained. The difference between the two is used to calculate the error and determine whether the error is within the preset tolerance range. Otherwise, the weights of the performance evaluation objectives are adjusted based on error decomposition, and the pre-stored proxy model is optimized based on the adjusted performance evaluation objectives; the performance evaluation objectives include at least the grid-connected power smoothness with different weight ratios, the degradation degree of the hybrid energy storage system, and the default degree based on the first constraint; The dataset obtained between the current period and the previous parameter optimization period is input into the optimized surrogate model for training, and the parameters learned by the surrogate model are used to update the Gaussian model and the objective function.

6. The renewable energy integrated optimization method according to claim 5, characterized in that, After adjusting the weights of the performance evaluation target based on error decomposition or updating the surrogate model, the adjusted performance evaluation target or the updated surrogate model is verified in different scenarios based on the digital twin model. If the verification fails, the weight correction of the performance evaluation target or the retraining of the surrogate model is triggered. The digital twin model is obtained by coupling a physical model built on historical datasets and a perturbation generator trained on historical datasets.

7. The renewable energy integrated optimization method according to claim 1, characterized in that, Before calculating the mismatch value by subtracting the predicted mean from the grid-connected power reference value, the load power at the grid connection point is obtained and it is determined whether it exceeds the preset third threshold. If so, the load power and the grid-connected power reference value are used together to calculate the difference with the predicted mean.

8. The renewable energy integrated optimization method according to claim 1, characterized in that, When acquiring the pre-processed dataset collected in the current period, the network state parameters are acquired simultaneously. After normalization, the corresponding compression measures are obtained by referring to a pre-set lookup table that represents the network state and compression measures. The parameter data of the current period are then reported according to the corresponding compression measures.

9. A renewable energy integrated optimization system, applied to the renewable energy integrated optimization method as described in any one of claims 1-8, characterized in that, It includes: The data acquisition module is used to acquire the dataset collected in the current cycle and preprocessed. The dataset includes renewable energy power, grid status data, and power and charge status of the hybrid energy storage system. The data prediction module is used to input the dataset into a pre-trained Gaussian model to obtain the predicted mean and predicted variance of the predicted renewable energy power within a preset future time window. The mismatch value calculation module is used to calculate the mismatch value by comparing the predicted mean with the grid-connected power reference value, wherein the grid-connected power reference value is obtained from the stored grid-connected power data of the previous cycle; The reference allocation power estimation module is used to decompose the mismatch value into fluctuation components, estimate the decomposition result based on the prediction variance and the allocation situation of the previous cycle, and obtain the reference allocation power applied to the first energy storage unit and the second energy storage unit. A power allocation module is used to allocate renewable energy power to obtain the optimal allocation power; wherein the optimal allocation power satisfies the condition that the difference between the reference allocation power and the optimal allocation power does not exceed a limit. The power adjustment module is used to construct a feasible region based on the power and charge state of the hybrid energy storage system, project the optimal power allocation into the feasible region, determine whether it exceeds the limit, and if so, determine the adjustment amount based on the deviation and change of the grid state data between the current cycle and the previous cycle, and adjust the optimal power allocation until it is within the feasible region. The result output module is used to output control commands containing the optimal power allocation within the feasible region.

10. A computer terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the renewable energy integration optimization method as described in any one of claims 1-8.

Images