A generator with an energy storage mechanism and methods of using the same

Abstract

The application relates to the technical field of power system operation and control, and discloses a generator with an energy storage mechanism and a use method thereof. The method constructs a multi-dimensional time sequence input matrix according to preset time window interception of historical power generation data, historical load data and energy storage mechanism historical state of charge data, extracts the space-time correlation features between different variables in the multi-dimensional time sequence input matrix by using a graph convolution network, inputs the space-time correlation features into a long short-term memory network, outputs a power generation power prediction curve and a load prediction curve of a next scheduling period, calculates a net power fluctuation interval according to a difference between the two, and dynamically generates a charge-discharge trigger threshold boundary and a maximum charge-discharge power constraint boundary through a nonlinear mapping function in combination with current health state data of the energy storage mechanism. The application eliminates the regulation dead zone and over-regulation phenomenon caused by the mismatch between the fixed threshold and the working condition, overcomes the problem of insufficient regulation depth, and reduces the invalid switching times of the charge-discharge state.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, and discloses a generator with an energy storage mechanism and its usage method. Background Technology

[0002] When generators with energy storage mechanisms perform energy dispatch, conventional technical solutions employ static rules based on pre-set fixed power thresholds or fixed state of charge (SOC) ranges. During operation, the control program continuously compares real-time collected power generation and load data with these preset fixed thresholds. When the difference between power generation and load exceeds the preset fixed power threshold, or when the SOC of the energy storage mechanism reaches the upper or lower limit of a preset fixed range, the control program directly issues fixed-value charging and discharging commands to the energy storage mechanism according to a static logic table stored in the storage medium. These static rules rely on manual experience to set parameters, remain constant throughout the system's lifespan, and lack the ability to adjust the regulation logic based on operating conditions.

[0003] The fixed power threshold and fixed state of charge range in the above technical solutions cannot reflect the coupled fluctuation characteristics of power generation and load over time. When the fluctuations in power generation and load are in a stable period, the system does not operate because the fixed threshold is set too wide. However, when small fluctuations accumulate and exceed the fixed threshold, charging and discharging commands are triggered, leading to frequent and meaningless charging and discharging switches in the energy storage mechanism. When there are sharp fluctuations in power generation and load, the fixed threshold is set too narrowly, causing the energy storage mechanism to reach the fixed protection limit and exit regulation after absorbing or releasing a small amount of electrical energy, resulting in insufficient regulation depth. At the same time, frequent small charging and discharging switches will exacerbate the irreversible loss of chemical materials inside the energy storage mechanism, leading to accelerated capacity decay. Summary of the Invention

[0004] The purpose of this invention is to provide a generator with an energy storage mechanism and a method for using it, which can effectively solve the problems in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a generator with an energy storage mechanism, comprising a data interception component, a feature extraction component, a boundary generation component, and a scheduling solution component; The data extraction component extracts historical power generation data, historical load data, and historical state of charge data of energy storage institutions according to a preset time window, and constructs a multi-dimensional time-series input matrix. The feature extraction component uses a graph convolutional network to extract the spatiotemporal correlation features between different variables in the multidimensional time-series input matrix, inputs the spatiotemporal correlation features into a long short-term memory network, and outputs the power generation prediction curve and load prediction curve for the next scheduling cycle. The boundary generation component calculates the net power fluctuation range based on the difference between the power generation prediction curve and the load prediction curve, and dynamically generates the charging and discharging trigger threshold boundary and the maximum charging and discharging power constraint boundary for the current scheduling cycle through a nonlinear mapping function, in conjunction with the current health status data of the energy storage mechanism. The scheduling solution component, within the boundary of the charge / discharge trigger threshold and the boundary of the maximum charge / discharge power constraint, uses the particle swarm optimization algorithm to generate the charge / discharge power scheduling timetable of the energy storage mechanism, with the objective function of minimizing the variance of the net power fluctuation range and minimizing the number of charge / discharge state switching.

[0006] Preferably, when constructing the multidimensional time-series input matrix, the data extraction component aligns the acquired historical power generation data, historical load data, and historical state of charge data of the energy storage mechanism according to timestamps, performs linear interpolation processing for missing values ​​and extreme value removal processing based on box plot statistics, extracts the processed data into equal-length segments according to the preset time window, concatenates the data vectors of multiple time steps in the feature dimension, adds a timestamp encoding vector containing periodic features and time position features to the end of the concatenated matrix, performs maximum and minimum value normalization processing on the concatenated and encoded matrix, and generates the multidimensional time-series input matrix.

[0007] Preferably, before extracting the spatiotemporal correlation features using the graph convolutional network, the feature extraction component constructs an adjacency matrix of a graph structure based on the Pearson correlation coefficients between different variables in the multidimensional temporal input matrix. Each column of feature variables in the multidimensional temporal input matrix is ​​used as a node in the graph structure, and the Pearson correlation coefficient is used as the edge weight connecting two nodes in the graph structure. The graph convolutional network performs graph convolution operations on the adjacency matrix and the feature matrix containing node features, aggregates the neighbor node features of each node, and outputs the spatiotemporal correlation features.

[0008] Preferably, when the feature extraction component inputs the spatiotemporal correlation features into the Long Short-Term Memory (LSTM) network, it expands the spatiotemporal correlation features according to time steps and inputs them into each hidden layer of the LTM network. An attention mechanism layer is connected to the output of the LTM network. The attention mechanism layer calculates the attention weights of the hidden state vectors output by the LTM network at each time step. The hidden state vectors and the corresponding attention weights are weighted and summed. The weighted and summed vectors are input into the fully connected layer, and the power generation prediction curve and the load prediction curve for the next scheduling cycle are output.

[0009] Preferably, when the boundary generation component acquires the current health status data of the energy storage mechanism, it collects the port voltage and port current of the energy storage mechanism based on the second-order Thevenin equivalent circuit model of the energy storage mechanism, and performs offline identification of the ohmic internal resistance parameter and polarization internal resistance parameter in the second-order Thevenin equivalent circuit model using the least squares method, and calculates the health status based on the ratio of the ohmic internal resistance parameter to the rated internal resistance parameter. When generating the charge / discharge trigger threshold boundary and the maximum charge / discharge power constraint boundary through the nonlinear mapping function, the health status and the net power fluctuation range are input into the nonlinear mapping function composed of piecewise linear functions, and the charge / discharge trigger threshold boundary and the maximum charge / discharge power constraint boundary are output.

[0010] Preferably, when the scheduling solution component uses the particle swarm optimization algorithm to generate the charging and discharging power scheduling time series table, it divides the next scheduling cycle into multiple discrete time steps, initializes a particle swarm composed of multiple particles, and the position vector of each particle represents the charging and discharging power sequence of the energy storage mechanism in the multiple discrete time steps. In each iteration of updating the particle velocity and position, an adaptive inertial weight is introduced. When the number of iterations increases, the value of the adaptive inertial weight is reduced. After each update of the particle position, it is detected whether the charging and discharging power value in the position vector exceeds the maximum charging and discharging power constraint boundary. If it exceeds the boundary, the value exceeding the boundary is replaced with the limit value corresponding to the maximum charging and discharging power constraint boundary. After the iteration terminates, the globally optimal position vector is output as the charging and discharging power scheduling time series table.

[0011] Preferably, when the data extraction component performs extreme value removal processing based on box plot statistics, it introduces a density-based noise-applied spatial clustering algorithm to map the historical power generation data, historical load data, and historical state of charge data of the energy storage mechanism, which are aligned according to timestamps, into a set of data points in three-dimensional space. It calculates the distance from each data point in the set to other data points in a specified neighborhood, counts the number of data points whose distance is less than a preset neighborhood radius, marks the data points whose number is less than a preset minimum sample size as outliers, deletes the data marked as outliers from the set of data points, and restores the remaining data points after deleting the outliers to time series data.

[0012] Preferably, when calculating the Pearson correlation coefficient, the feature extraction component simultaneously calculates the time-delay cross-correlation function between different variables in the multidimensional time-series input matrix, traverses the preset time lag range, calculates the cross-correlation number at each time lag step, selects the time lag step corresponding to the largest cross-correlation number as the time-delay feature between variables, inputs the Pearson correlation coefficient and the time-delay feature into the preset fusion weight calculation formula, calculates the fusion correlation degree value, uses the fusion correlation degree value as the edge weight connecting two nodes in the graph structure, and generates the adjacency matrix containing time-series lag information.

[0013] Preferably, when the boundary generation component performs offline identification of the parameters in the second-order Thevenin equivalent circuit model using the least squares method, it employs a recursive least squares method with a forgetting factor. Each time a new energy storage mechanism port voltage and current are acquired, the gain vector at the current moment is calculated. Based on the gain vector, the parameter estimates and covariance matrix in the second-order Thevenin equivalent circuit model at the current moment are updated. During the calculation of the gain vector, the forgetting factor is introduced to attenuate the covariance matrix at historical moments. The updated parameter estimates are used as the ohmic resistance parameter and the polarization resistance parameter at the current moment.

[0014] The present invention also provides a method for using a generator with an energy storage mechanism, the method comprising: Historical power generation data, historical load data, and historical state of charge data of energy storage institutions are extracted according to a preset time window to construct a multi-dimensional time-series input matrix. The spatiotemporal correlation features between different variables in the multidimensional time-series input matrix are extracted using a graph convolutional network. These spatiotemporal correlation features are then input into a long short-term memory network to output the power generation prediction curve and load prediction curve for the next scheduling cycle. The net power fluctuation range is calculated based on the difference between the power generation prediction curve and the load prediction curve. Combined with the current health status data of the energy storage mechanism, the charging and discharging trigger threshold boundary and the maximum charging and discharging power constraint boundary of the current scheduling cycle are dynamically generated through a nonlinear mapping function. Within the boundary of the charge / discharge trigger threshold and the boundary of the maximum charge / discharge power constraint, the particle swarm optimization algorithm is used to solve and generate the charge / discharge power scheduling timetable of the energy storage mechanism, with the objective function being to minimize the variance of the net power fluctuation range and the number of charge / discharge state switching.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a multi-dimensional time-series input matrix by extracting historical data according to a preset time window. It then uses a graph convolutional network to extract spatiotemporal correlation features between different variables and inputs these features into a long short-term memory network to output a prediction curve. Based on the difference between the prediction curves, it calculates the net power fluctuation range. Combining this with the health status data of the energy storage mechanism, it dynamically generates charging and discharging trigger threshold boundaries and maximum charging and discharging power constraint boundaries through a nonlinear mapping function. Within these dynamically generated boundaries, the objective function is to minimize the variance of the net power fluctuation range and the number of charging and discharging state switching operations. A particle swarm optimization algorithm is then used to solve for and generate a charging and discharging power scheduling time series table. This scheme, through the extraction of spatiotemporal correlation features from multi-dimensional time-series data and dynamic boundary tuning, enables the charging and discharging trigger conditions and power upper limits of the energy storage mechanism to adaptively adjust according to the actual fluctuation characteristics of power generation and load. This eliminates the adjustment dead zone and over-adjustment phenomena caused by the mismatch between fixed thresholds and actual operating conditions, and overcomes the problem of insufficient adjustment depth.

[0016] 2. When generating the multidimensional time-series input matrix, linear interpolation and extreme value removal based on density noise spatial clustering algorithm are performed on the data, and a timestamp encoding vector is added to the end of the matrix for normalization, which improves the feature representation capability of the input data. When extracting spatiotemporal correlation features, the Pearson correlation coefficient and time delay cross-correlation function between variables are calculated, and the fused correlation degree value is used as the edge weight of the graph structure to construct an adjacency matrix containing time-series lag information. An attention mechanism layer is introduced at the output of the long short-term memory network to calculate the attention weight of the hidden state vector at each time step and perform weighted summation, which improves the fitting accuracy of the prediction curve to the time-series fluctuations. When obtaining the health status of the energy storage mechanism, the recursive least squares method with forgetting factor is used to identify the internal resistance parameters in the second-order Thevenin equivalent circuit model online. Adaptive inertial weight and out-of-bounds replacement mechanism are introduced in the particle swarm algorithm solution to ensure the execution reliability of the charging and discharging power scheduling time table and reduce the number of invalid switching of the charging and discharging state of the energy storage mechanism. Attached Figure Description

[0017] Figure 1 This is a flowchart of an adaptive scheduling method for a power generation system with an energy storage mechanism. Figure 2 Flowchart for constructing a multidimensional time series input matrix; Figure 3 Flowchart for spatiotemporal correlation feature extraction and prediction curve output; Figure 4 Dynamically generate flowcharts for constraint boundaries; Figure 5 Flowchart for generating scheduling timetables for particle swarm optimization (PSO) solution; Figure 6 This is a flowchart of the data interaction process for the adaptive dispatching system of the power generation system. Detailed Implementation

[0018] Please refer to Figure 1 This embodiment provides a generator with an energy storage mechanism, including a data extraction component, a feature extraction component, a boundary generation component, and a scheduling solution component. The data extraction component extracts historical power generation data, historical load data, and historical state of charge data of the energy storage mechanism according to a preset time window, constructing a multi-dimensional time-series input matrix. The preset time window has a length of T consecutive time steps, with each time step corresponding to a time interval of Δt. The value of Δt is consistent with the generator's sampling period, and the value of T corresponds to the coverage duration of the historical operating data. The data extraction component extracts historical power generation data, historical load data, and historical state of charge data of the energy storage mechanism from the generator's real-time database and historical archive database, according to the start and end timestamps corresponding to the preset time window. The historical power generation data is a time series of the total active power output value of each power generation unit in the generator; the historical load data is a time series of the total active power value of the electrical load connected to the generator; and the historical state of charge data of the energy storage mechanism is a time series of the cumulative state of charge percentage of individual energy storage units or battery clusters.

[0019] Please refer to Figure 2 The data extraction component aligns the three sets of time series data according to timestamps, ensuring that the same row of data corresponds to the same sampling time. Linear interpolation is performed on missing values ​​in the aligned data, and extreme value removal based on box plot statistics is performed on the interpolated data. The processed data is then truncated to the same length according to a preset time window, resulting in three time step data vectors of length T. The data extraction component concatenates the three time step data vectors along the feature dimensions to obtain an initial matrix of dimension T×3. A timestamp encoding vector containing periodic and time position features is added to the end of the initial matrix. The dimension of the timestamp encoding vector is consistent with the time step dimension of the initial matrix. The periodic features include the sine and cosine encoding values ​​corresponding to hourly, daily, and weekly features. The time position features are the relative position encoding values ​​of the corresponding time step within the preset time window. After concatenation, a feature matrix of dimension T×F is obtained, where F is the total number of feature dimensions. The data extraction component performs maximum-minimum normalization on the concatenated feature matrix, mapping the values ​​of all features within the matrix to the [0,1] interval, generating a multi-dimensional time-series input matrix. The normalization calculation formula is as follows:

[0020] in, Let be the value of the i-th time step and the j-th feature dimension in the normalized multidimensional time series input matrix. This represents the original value of the i-th time step and the j-th feature dimension in the feature matrix before normalization. The minimum value of the j-th feature dimension within a preset time window. It represents the maximum value of the j-th feature dimension within a preset time window.

[0021] Table 1. Definition of Feature Dimensions of Multidimensional Temporal Input Matrix ; This table clarifies the physical meaning and numerical attributes of all feature dimensions contained in the multidimensional time-series input matrix, ensuring that all data input to subsequent feature extraction components have a unified time base and numerical range, and avoiding feature extraction deviations caused by differences in feature dimensions.

[0022] The feature extraction component receives the multi-dimensional time-series input matrix output by the data extraction component. It then uses a graph convolutional network to extract the spatiotemporal correlation features between different variables in the multi-dimensional time-series input matrix. These features are input into a long short-term memory network, outputting the power generation prediction curve and load prediction curve for the next scheduling cycle. The feature extraction component constructs a graph-structured adjacency matrix based on the Pearson correlation coefficients between different variables in the multi-dimensional time-series input matrix. Each column of feature variables in the multi-dimensional time-series input matrix serves as a node in the graph structure, and the Pearson correlation coefficient is used as the edge weight connecting two nodes. The formula for calculating the Pearson correlation coefficient is:

[0023] in, Let Pearson's correlation coefficient be the coefficient between the m-th feature variable and the n-th feature variable. Let m be the value of the feature variable at the i-th time step. Let m be the average value of the m-th feature variable within a preset time window. Let n be the value of the nth feature variable at the i-th time step. is the average value of the nth feature variable within a preset time window, and T is the total number of time steps contained in the preset time window.

[0024] The graph structure constructed by the feature extraction component contains F nodes, corresponding to the F feature dimensions of the multidimensional temporal input matrix. The adjacency matrix A has dimensions F×F, where the element in the m-th row and n-th column is... The diagonal elements of the adjacency matrix are all 1s, representing the correlation between nodes. The feature extraction component inputs the adjacency matrix A and the feature matrix X containing node features into the graph convolutional network, performs graph convolution operations, aggregates the features of each node's neighboring nodes, and outputs spatiotemporal correlation features. The layer propagation formula of the graph convolutional network is:

[0025] in, Let be the output feature matrix of the (l+1)th layer of the graph convolutional network. For non-linear activation functions, the ReLU activation function is used. The adjacency matrix for adding self-joins, where I is the identity matrix. for The degree matrix, where the element in the m-th row and m-th column is... The sum of all elements in the m-th row. Let be the input feature matrix of the l-th layer of the graph convolutional network, and let be the input of the 0th layer. X is a multidimensional time-series input matrix. Let be the trainable weight matrix of the l-th layer of the graph convolutional network.

[0026] The graph convolutional network is set up as a 2-layer stacked structure. The first layer outputs a feature dimension of 32, and the second layer outputs a feature dimension of 16. After two layers of graph convolution operations, the output is a node aggregation feature matrix with a dimension of F×16. After transposing this matrix, a spatiotemporal correlation feature matrix with a dimension of T×(F×16) is obtained, which completes the fusion and extraction of spatial correlation features and temporal dimension features between different variables.

[0027] The feature extraction component inputs spatiotemporal correlation features into a Long Short-Term Memory (LSTM) network, outputting the power generation prediction curve and load prediction curve for the next scheduling cycle. The length of the next scheduling cycle is an integer multiple of the single time step interval of the preset time window, divided into... Each time step. The feature extraction component expands the spatiotemporal correlation feature matrix according to the time steps and inputs it sequentially into each hidden layer of the Long Short-Term Memory (LSTM) network. The LTM network is configured with a two-layer stacked structure, with 64 neurons in each hidden layer. The LTM network uses a gating mechanism to achieve feature transfer and forgetting in the temporal dimension, completing the deep extraction of temporal features. The cell state update formula of the LTM network is:

[0028] in, The output of the forget gate at time t. The input gate output is at time t. Let t represent the candidate cell state at time t. Let t represent the cell state at time t. The output of the gate at time t is the output of the gate. Let be the hidden state vector at time t. For the trainable weight matrix corresponding to the gating structure, This is the trainable bias vector corresponding to the gating structure. It is the sigmoid activation function. For Hadamard product operations, Let be the hidden state vector at time t-1. The spatiotemporal correlation feature vector is input to the Long Short-Term Memory network at time t.

[0029] The output of the Long Short-Term Memory (LSTM) network is connected to an attention mechanism layer. This layer receives the hidden state vectors output by the LTM network at each time step, calculates the attention weight corresponding to each hidden state vector, and performs a weighted summation of the hidden state vectors and their corresponding attention weights to obtain the feature vector fused with temporal attention. The formula for calculating the attention weights is:

[0030] in, Let be the attention score value of the hidden state vector at time t. Here are the trainable weight matrix and bias vector for the attention mechanism layer. Let be the normalized attention weights corresponding to the hidden state vector at time t. The feature vector of the fused temporal attention is obtained after weighted summation, and T is the total number of time steps contained in the preset time window.

[0031] The feature extraction component inputs the weighted summed feature vector into a fully connected layer. The fully connected layer is configured with two layers: the first layer has 32 neurons, and the second layer has 2×100 neurons. ,in The total number of time steps in the next scheduling cycle is given. After reshaping, the output vector of the fully connected layer has a dimension of [value missing]. The output matrix is ​​2×2, where the first column corresponds to the power generation prediction value for each time step of the next scheduling cycle, and the second column corresponds to the load prediction value for each time step of the next scheduling cycle. The two columns of prediction values ​​are arranged in the order of time steps to generate the power generation prediction curve and the load prediction curve respectively.

[0032] The boundary generation component receives the power generation prediction curve and load prediction curve output by the feature extraction component. Based on the difference between the power generation prediction curve and the load prediction curve, it calculates the net power fluctuation range. Combining this with the current health status data of the energy storage facility, it dynamically generates the charging / discharging trigger threshold boundary and the maximum charging / discharging power constraint boundary for the current scheduling cycle through a nonlinear mapping function. The boundary generation component calculates the numerical difference between the two curves at the same time step to obtain the net power time series. Based on the maximum and minimum values ​​of the net power time series, it calculates the net power fluctuation range using the following formula:

[0033] in, This represents the net power value at time step t of the next scheduling cycle. Let be the value of the power generation prediction curve at time step t. Let be the value of the load forecast curve at time step t. This refers to the net power fluctuation range. This represents the minimum value of the net power time series. The maximum value of the net power time series is t, which ranges from 1 to 1. , This represents the total number of time steps included in the next scheduling cycle.

[0034] The boundary generation component synchronously acquires the current health status data of the energy storage mechanism. Based on the second-order Thevenin equivalent circuit model of the energy storage mechanism, it collects the port voltage and port current of the energy storage mechanism. The ohmic internal resistance parameter and polarization internal resistance parameter in the second-order Thevenin equivalent circuit model are identified offline using the least squares method. The health status is calculated based on the ratio of the ohmic internal resistance parameter to the rated internal resistance parameter. The voltage equation of the second-order Thevenin equivalent circuit model is:

[0035] in, Let be the port voltage of the energy storage mechanism at time t. Let be the open-circuit voltage of the energy storage mechanism at time t. The ohmic internal resistance parameter of the energy storage mechanism, Let be the port current of the energy storage mechanism at time t. The current is positive during discharge and negative during charging. These are the two polarization internal resistance parameters of the energy storage mechanism. Let t be the current values ​​of the two polarization branches at time t.

[0036] The boundary generation component uses the least squares method to apply the ohmic internal resistance parameter in the above formula. Polarization internal resistance parameters Offline identification is performed to obtain the current ohmic internal resistance parameter identification value. The health status is calculated based on the ratio of the ohmic internal resistance parameter identification value to the rated internal resistance parameter of the energy storage mechanism at the time of manufacture. The calculation formula is as follows:

[0037] in, The health status of the energy storage mechanism, with a value ranging from 0 to 1. This represents the actual value of the ohmic internal resistance parameter identified at the current moment. This refers to the rated internal resistance parameter in ohms when the energy storage mechanism leaves the factory.

[0038] The boundary generation component inputs the calculated health status and net power fluctuation range into a nonlinear mapping function composed of piecewise linear functions, and outputs the charge / discharge trigger threshold boundary and the maximum charge / discharge power constraint boundary for the current scheduling period. The piecewise linear formula of the nonlinear mapping function is as follows:

[0039] in, The charging trigger threshold boundary, This is the discharge trigger threshold boundary. The piecewise linear scaling factor is related to health status. As the maximum charging power constraint boundary, As the maximum discharge power constraint boundary, For piecewise linear power coefficients related to health status, The rated charging power of the energy storage mechanism, This refers to the rated discharge power of the energy storage mechanism.

[0040] Both the piecewise linear proportional coefficient and the piecewise linear power coefficient are set to three piecewise functions. When the health level is in the range of [0.8, 1], the proportional coefficient and power coefficient take the rated values. When the health level is in the range of [0.6, 0.8), the proportional coefficient and power coefficient decrease linearly with the health level. When the health level is in the range of [0, 0.6), the proportional coefficient and power coefficient take the minimum values ​​to ensure that the charging and discharging actions of the energy storage mechanism are constrained in the low health level state and to reduce irreversible losses.

[0041] Within the boundaries of the charge / discharge trigger threshold and the maximum charge / discharge power constraint, the scheduling solution component uses the particle swarm optimization algorithm to generate the charge / discharge power scheduling timetable for the energy storage mechanism, with the objective function of minimizing the variance of the net power fluctuation range and the minimum number of charge / discharge state transitions. The scheduling solution component then divides the next scheduling cycle into... There are [number] discrete time steps, each corresponding to a charge / discharge power decision variable. Charging power is negative, discharging power is positive, and standby power is 0. The scheduling and solving component initializes a particle swarm of M particles, where each particle's position vector has a dimension of [dimension]. Each element in the position vector represents the charging / discharging power of the energy storage mechanism at the corresponding discrete time step. The velocity vector of each particle has the same dimension as the position vector, and the values ​​of the velocity vector elements are randomly generated within a preset range during initialization. The mathematical expression of the objective function is:

[0042] in, The objective function value, This is the variance calculation function. Let t represent the charging and discharging power of the energy storage mechanism at time step t, where discharging is positive and charging is negative. This refers to the number of times the energy storage mechanism switches between charging and discharging states during the next scheduling cycle. These are weighting coefficients, all of which are positive numbers. These are the upper and lower limit constraints for the state of charge of the energy storage mechanism. This represents the state of charge of the energy storage mechanism at the end of the t-th time step.

[0043] The number of charge / discharge state switching times is calculated as follows: when the sign of the charge / discharge power changes between two adjacent time steps, or when the power in one time step is 0 and the power in the other time step is not 0, it is considered a state switching. The number of switching times is accumulated throughout the entire scheduling cycle. The recursive formula for the state of charge is: ,in For charging and discharging efficiency, during charging Take charging efficiency, during discharge Take the reciprocal of the discharge efficiency. This refers to the rated capacity of the energy storage mechanism.

[0044] The scheduling and solving component uses a particle swarm optimization (PSO) algorithm to iteratively solve the objective function. In each iteration, the particle velocity and position are updated, and an adaptive inertia weight is introduced. The value of the adaptive inertia weight is decreased as the number of iterations increases. The velocity and position update formulas for the PSO algorithm are as follows:

[0045] in, Let be the velocity vector of the i-th particle in the (k+1)-th iteration. The adaptive inertia weights at the k-th iteration. Let be the position vector of the i-th particle at the k-th iteration. As a learning factor, It is a random number with a value in the interval [0,1]. Let be the individual optimal position vector of the i-th particle in the k-th iteration. This is the globally optimal position vector at the k-th iteration.

[0046] The formula for calculating adaptive inertia weight is:

[0047] in, This represents the maximum value of the inertia weight. This represents the minimum value of the inertia weight. The maximum number of iterations, This represents the current iteration number.

[0048] The value is 0.9. The value is 0.4. The average value is 2, and the maximum number of iterations is... The value is set to 200, and the particle swarm size M is set to 50. After each update of the particle position, the scheduling solution component checks whether the charging and discharging power values ​​in the position vector exceed the maximum charging and discharging power constraint boundary. If they do, the values ​​exceeding the boundary are replaced with the limit values ​​corresponding to the maximum charging and discharging power constraint boundary. At the same time, it checks whether the charging and discharging power meets the constraint of the charging and discharging trigger threshold boundary. If it does not meet the constraint, the power value of the corresponding time step is set to 0. After each iteration, the scheduling solution component updates the individual optimal position and the global optimal position of each particle. When the number of iterations reaches the maximum number of iterations, or when the objective function value corresponding to the global optimal position has not changed for a preset number of consecutive iterations, the iteration process is terminated, and the global optimal position vector is output as the charging and discharging power scheduling time sequence table of the energy storage mechanism. Each element in the charging and discharging power scheduling time sequence table corresponds to the charging and discharging power setting value for each discrete time step in the next scheduling cycle.

[0049] Please refer to Figure 6 This embodiment achieves adaptive scheduling of generators with energy storage mechanisms by constructing a multi-dimensional time-series input matrix, extracting spatiotemporal correlation features and generating prediction curves, tuning dynamic charging and discharging boundaries, and solving scheduling time sequences based on multi-objective optimization. The charging and discharging trigger conditions and power constraint boundaries can be dynamically adjusted according to the fluctuation characteristics of power generation and load and the health status of the energy storage mechanism, eliminating the problem of mismatch between static fixed thresholds and actual operating conditions, and reducing the invalid switching of the charging and discharging states of the energy storage mechanism.

[0050] In one optional embodiment, when constructing the multi-dimensional time-series input matrix, the data extraction component aligns the acquired historical power generation data, historical load data, and historical state of charge data of energy storage institutions according to timestamps, performs linear interpolation processing for missing values, and extreme value removal processing based on box plot statistics. During the extreme value removal processing based on box plot statistics, a density-based noise spatial clustering algorithm is introduced to perform secondary outlier removal. The data extraction component first aligns the acquired raw data according to timestamps. During the alignment process, the timestamp of the data source with the highest sampling frequency is used as the benchmark, and the values ​​from other data sources are filled to the corresponding timestamp positions using linear interpolation, ensuring that all data sequences have a completely consistent number of time steps and time benchmark. For missing values ​​in the aligned data sequences, the data extraction component fills them using linear interpolation. When the number of consecutive missing time steps is less than a preset threshold, the filling value is calculated based on the two nearest non-missing values ​​before and after the missing value, according to the time interval ratio. When the number of consecutive missing time steps is greater than or equal to the preset threshold, the historical average value of the corresponding time step is used for filling. The preset threshold is set to 5 consecutive time steps.

[0051] The data extraction component performs extreme value removal based on box plot statistics on the data sequences after missing value imputation. It calculates the upper quartile (Q3) and lower quartile (Q1) of each data sequence, and calculates the interquartile range (IQR) = Q3 - Q1. Data points with values ​​greater than Q3 + 1.5 × IQR or less than Q1 - 1.5 × IQR are marked as preliminary extreme values, and the median of the corresponding data sequence is used as a temporary replacement for these preliminary extreme values. After preliminary extreme value removal, the data extraction component introduces a density-based noise-applied spatial clustering algorithm to perform secondary outlier detection and removal. The data extraction component maps historical power generation data, historical load data, and historical state of charge data of energy storage institutions, aligned according to timestamps, into a set of data points in three-dimensional space. The three dimensions of each data point correspond to the power generation value, load power value, and state of charge value at the same time step, respectively. The total number of data points in the set is consistent with the total number of time steps. The formula for calculating the neighborhood density of a data point is:

[0052] in, For data point p The neighborhood contains the set of data points, where D is the set of data points in three-dimensional space. Let be the Euclidean distance between data point p and data point q. This is the preset neighborhood radius.

[0053] The data extraction component calculates the Euclidean distance from each data point in the data point set to other data points within a specified neighborhood, and counts distances less than the preset neighborhood radius. The number of data points is determined by the number of data points. Data points with a number less than the preset minimum sample size (MinPts) are marked as outliers. These outliers are then removed from the data point set. The remaining data points are then restored to time series data in timestamp order, completing the quadratic extremum removal process. Neighborhood radius. The value of is adaptively adjusted based on the variance of the data sequence, and the minimum number of samples, MinPts, is set to 4.

[0054] After extreme value removal, the data truncation component truncates the processed data to an equal length according to a preset time window. It then concatenates the data vectors from multiple time steps along their feature dimensions and adds a timestamp encoding vector containing periodic and time position features to the end of the concatenated matrix. The periodic features in the timestamp encoding vector include hourly, daily, and weekly periodic features. Each periodic feature is generated using sine and cosine encoding methods. The encoding period for the hourly periodic feature is 24 hours, and the corresponding encoding formula is: , Where h is the number of hours at the corresponding time step; the encoding period for the daily periodic feature is 7 days, and the corresponding encoding formula is: , Where d is the number of days in a week at the corresponding time step; the encoding period for the weekly periodic feature is 12 months, and the corresponding encoding formula is: , , where m is the month number of the corresponding time step. The time position feature is the relative position encoding value of the corresponding time step within a preset time window, and the encoding formula is: , where i is the sequence number of the time step within the preset window, and T is the total number of time steps in the preset window.

[0055] The data extraction component performs maximum and minimum value normalization on the concatenated and encoded matrix to generate a multi-dimensional time series input matrix. During the normalization process, the maximum and minimum values ​​are calculated for each feature dimension, and all values ​​in that dimension are mapped to the [0,1] interval. For feature dimensions where the maximum and minimum values ​​are equal, all values ​​in that dimension are uniformly set to 0.5 to avoid division by zero errors.

[0056] Table 2 Data Preprocessing Parameter Configuration Table ; This table clarifies the parameter configurations and physical meanings of each stage in the data preprocessing process, ensuring the reproducibility of the data preprocessing process. Through two-level extreme value removal and multi-dimensional timestamp encoding, it improves the quality and feature representation of the input data and reduces the interference of outliers on subsequent prediction and scheduling processes.

[0057] This embodiment constructs a high-quality multidimensional time-series input matrix by combining timestamp alignment, two-level missing value filling, two-level extreme value removal using a combination of box plots and density-based noise applied spatial clustering algorithm, and multi-dimensional timestamp encoding and normalization. This effectively eliminates noise and outliers in the original data, enhances the temporal feature representation capability of the input data, and provides a reliable data foundation for subsequent feature extraction and prediction processes.

[0058] In another optional embodiment, before extracting spatiotemporal correlation features using a graph convolutional network, the feature extraction component constructs an adjacency matrix of a graph structure based on the Pearson correlation coefficients between different variables in the multidimensional temporal input matrix. When calculating the Pearson correlation coefficients, the time-delay cross-correlation function between different variables in the multidimensional temporal input matrix is ​​calculated simultaneously. The preset time lag range is traversed, and the cross-correlation number under each time lag step is calculated. The time lag step corresponding to the cross-correlation number with the largest value is selected as the time delay feature between variables. The Pearson correlation coefficient and the time delay feature are input into a preset fusion weight calculation formula to calculate the fusion correlation value. The fusion correlation value is used as the edge weight connecting two nodes in the graph structure to generate an adjacency matrix containing temporal lag information.

[0059] The feature extraction component receives a multi-dimensional time-series input matrix with dimensions T×F, where T is the total number of time steps and F is the total number of feature dimensions. Each column of the feature variable in the matrix is ​​used as a node in a graph structure, generating a total of F nodes. For any two distinct nodes m and n, the feature extraction component calculates the corresponding Pearson correlation coefficient. Cross-correlation function with time delay, traversing the preset time lag range ,in The preset maximum lag step size is set to 10 time steps. The formula for calculating the time delay cross-correlation function is:

[0060] in, Let the feature variables corresponding to node m and the feature variables corresponding to node n have a time lag step of . Cross-correlation coefficients at time Let m be the value of the feature variable corresponding to the i-th time step node. Let m be the average value of the feature variable corresponding to node m over the entire time window. For the i+th The feature variable values ​​corresponding to each time step node n. Let be the average value of the feature variable corresponding to node n over the entire time window, and T be the total number of time steps.

[0061] The feature extraction component iterates through all preset time lag steps and selects the value with the largest absolute value of the cross-correlation coefficient as the maximum time delay cross-correlation coefficient. Record the time lag step corresponding to this value. As the optimal time delay feature among variables, the feature extraction component inputs the Pearson correlation coefficient and the maximum time delay cross-correlation coefficient into the preset fusion weight calculation formula to calculate the fusion correlation degree. The fusion weight calculation formula is as follows: ,in The fusion weighting coefficient has a value range of [0,1]. In this embodiment... The value is 0.6. Let be the fusion correlation score between node m and node n. The feature extraction component uses this fusion correlation score as the edge weight connecting the two nodes in the graph structure, constructing an adjacency matrix A of dimension F×F. The element in the m-th row and n-th column of the adjacency matrix is... The diagonal elements of the adjacency matrix are all 1, representing the degree of association of the nodes themselves, generating an adjacency matrix that includes time lag information.

[0062] The feature extraction component inputs the constructed adjacency matrix and multidimensional temporal input matrix into the graph convolutional network (GCNN), performs graph convolution operations, aggregates the features of each node's neighboring nodes, and outputs spatiotemporal correlation features. The GCNN adopts a three-layer stacked structure, with output feature dimensions of 64, 32, and 16 for each layer, respectively. Each layer uses the ReLU activation function, followed by a Dropout layer with a Dropout rate of 0.2 to avoid model overfitting. After three layers of graph convolution operations, the output is a node aggregation feature matrix with dimension F×16. After transposing and flattening this matrix, a spatiotemporal correlation feature matrix with dimension T×(F×16) is obtained, completing the fusion extraction of spatial correlation features and temporal dimension features.

[0063] The feature extraction component expands the spatiotemporal correlation features according to time steps and inputs them into the hidden layers of the Long Short-Term Memory (LSTM) network. The LTM network adopts a three-layer stacked structure, with 128, 64, and 32 neurons in each hidden layer, respectively. Dropout layers are connected between each layer, with a dropout rate of 0.2. The input sequence length of the LTM network is T, and the input feature dimension at each time step is F×16. The LTM network extracts temporal features through a gating mechanism and outputs a hidden state vector corresponding to each time step, with a hidden state vector dimension of 32.

[0064] The feature extraction component is connected to the attention mechanism layer at the output of the Long Short-Term Memory (LSTM) network. The attention mechanism layer calculates the attention weights of the hidden state vectors output by the LTM network at each time step. The hidden state vectors are then weighted and summed with their corresponding attention weights. This weighted sum is then input to the fully connected layer, outputting the power generation prediction curve and load prediction curve for the next scheduling cycle. The attention mechanism layer employs a multi-head attention structure with 4 attention heads, each with a dimension of 8. The output dimension of the multi-head attention is 32, consistent with the dimension of the hidden state vector. The output of the multi-head attention layer, after layer normalization, is input to two fully connected layers. The first fully connected layer has 64 neurons and uses the ReLU activation function, while the second fully connected layer has 2×100 neurons. ,in The second fully connected layer uses a linear activation function, where is the total number of time steps in the next scheduling cycle. After reshaping the output vector of the fully connected layer, a vector of dimension is obtained. A 2×2 prediction matrix is ​​used, with the first column corresponding to the predicted power generation value and the second column corresponding to the predicted load value. The predicted values ​​are arranged in order of time steps to generate continuous power generation prediction curves and load prediction curves.

[0065] Table 3. Structural parameters of convolutional networks and long short-term memory networks. ; This table clarifies the layer structure, output dimension, activation function, and additional structure configuration of each network module in the feature extraction component, ensuring the reproducibility of the network structure. By constructing an adjacency matrix that integrates temporal and mutual information, stacking multi-layer graph convolutional networks and long short-term memory networks, and introducing a multi-head attention mechanism, the accuracy of the prediction curve in fitting temporal fluctuations is improved, and the impact of prediction bias on scheduling results is reduced.

[0066] Please refer to Figure 3 This embodiment constructs an adjacency matrix that integrates Pearson correlation coefficient and time delay cross-correlation information, thereby realizing graph structure modeling that includes time-series lag information. It extracts spatiotemporal correlation features between multiple variables through a multi-layer graph convolutional network, and combines stacked long short-term memory network and multi-head attention mechanism to achieve high-precision prediction of power generation and load, providing an accurate prediction basis for subsequent dynamic boundary generation and scheduling optimization.

[0067] Please refer to Figure 4In another optional embodiment, the boundary generation component calculates the net power fluctuation range based on the difference between the power generation prediction curve and the load prediction curve. Combining this with the current health status data of the energy storage mechanism, it dynamically generates the charging / discharging trigger threshold boundary and the maximum charging / discharging power constraint boundary for the current scheduling cycle through a nonlinear mapping function. When acquiring the current health status data of the energy storage mechanism, the boundary generation component collects the port voltage and port current of the energy storage mechanism based on the second-order Thevenin equivalent circuit model. It then uses a recursive least squares method with a forgetting factor to identify the ohmic internal resistance parameter and polarization internal resistance parameter in the second-order Thevenin equivalent circuit model online, and calculates the health status based on the ratio of the ohmic internal resistance parameter to the rated internal resistance parameter.

[0068] The boundary generation component transforms the voltage equations of the second-order Thevenin equivalent circuit model into a linear regression form to facilitate parameter identification using the recursive least squares method. The discrete state equations of the second-order Thevenin equivalent circuit model are:

[0069] in, Let be the terminal voltages of the two polarization branches at time t. , The time constants of the two polarization branches, Polarized capacitor, Let be the sampling time interval. Transform the above equation into a linear regression form: ,in For the observation vector, For data vectors, The parameter vector to be identified.

[0070] The boundary generation component employs a recursive least squares method with a forgetting factor to identify the parameter vectors online. Each time new energy storage mechanism port voltage and current are acquired, the gain vector at the current moment is calculated. Based on the gain vector, the parameter estimates and covariance matrix in the second-order Thevenin equivalent circuit model at the current moment are updated. A forgetting factor is introduced during the gain vector calculation to attenuate the covariance matrix from historical moments. The updated parameter estimates are then used as the ohmic and polarization resistance parameters at the current moment. The parameter update formula for the recursive least squares method with a forgetting factor is as follows:

[0071] in, Let be the gain vector at time t. The parameter estimate covariance matrix at time t is... The forgetting factor has a value range of (0,1). In this embodiment... The value is 0.995. Let be the vector of parameter estimates at time t. Let be the observation vector at time t. Let be the data vector at time t, and I be the identity matrix.

[0072] The boundary generation component extracts the ohmic resistance parameter from the updated parameter estimate vector. The identification value is used to calculate the health status (SOH) of the energy storage mechanism based on the ratio of the identified ohmic internal resistance parameter to the rated internal resistance parameter. The calculation formula for health status is consistent with the aforementioned embodiment. The boundary generation component inputs the calculated health status and net power fluctuation range into a nonlinear mapping function composed of piecewise linear functions, and outputs the charge / discharge trigger threshold boundary and the maximum charge / discharge power constraint boundary for the current scheduling cycle. The nonlinear mapping function adopts a 5-segment piecewise linear structure, setting different proportional coefficients and power coefficients according to different health status intervals to ensure accurate matching between the charge / discharge boundaries and the health status of the energy storage mechanism.

[0073] Table 4. Piecewise Interval Parameter Table of Nonlinear Mapping Function ; This table clarifies the piecewise parameter values ​​of the nonlinear mapping function in different health ranges. The piecewise linear function enables dynamic tuning of the charge and discharge trigger threshold boundary and the maximum charge and discharge power constraint boundary. When the energy storage mechanism is in high health, the charge and discharge constraints are relaxed to improve the system's regulation capability. When the health is low, the charge and discharge constraints are tightened to reduce irreversible losses of the energy storage mechanism and extend its service life.

[0074] Please refer to Figure 5 After the boundary generation component outputs the charge / discharge trigger threshold boundary and the maximum charge / discharge power constraint boundary, the scheduling solution component, within these boundaries, uses the particle swarm optimization algorithm to generate the charge / discharge power scheduling timetable for the energy storage mechanism, with the objective function of minimizing the variance of the net power fluctuation range and minimizing the number of charge / discharge state transitions. The scheduling solution component then divides the next scheduling cycle into... For each discrete time step, a particle swarm of M particles is initialized. The position vector of each particle represents the charging and discharging power sequence of the energy storage mechanism over multiple discrete time steps. During initialization, the elements of the position vector of each particle are randomly generated within the maximum charging and discharging power constraint boundary, and the elements of the velocity vector are randomly generated within a preset velocity range. The preset velocity range is... ,in This refers to the rated power of the energy storage mechanism.

[0075] In each iteration of updating particle velocity and position, the scheduling solver introduces an adaptive inertia weight. The value of the adaptive inertia weight decreases as the number of iterations increases. The calculation formula for the adaptive inertia weight is consistent with the previous embodiment. Simultaneously, the learning factor is adaptively adjusted: c1 linearly decreases from 2.5 to 0.5 as the number of iterations increases, while c2 linearly increases from 0.5 to 2.5, achieving an optimization effect that emphasizes individual exploration in the early stages of iteration and global convergence in the later stages. After each particle position update, the scheduling solver checks whether the charging / discharging power value in the position vector exceeds the maximum charging / discharging power constraint boundary. If it does, the value exceeding the boundary is replaced with the limit value corresponding to the maximum charging / discharging power constraint boundary. Simultaneously, it checks whether the charging / discharging power meets the charging / discharging trigger threshold boundary constraint. When the charging / discharging power is in a charging state, if the power value is greater than the charging trigger threshold boundary, the power value for that time step is set to 0; when the charging / discharging power is in a discharging state, if the power value is less than the discharging trigger threshold boundary, the power value for that time step is set to 0, ensuring that the position vectors of all particles satisfy the constraint conditions.

[0076] After each iteration, the scheduling solution component calculates the objective function value corresponding to the position vector of each particle, updates the individual optimal position and the global optimal position of each particle, and terminates the iteration process when the maximum number of iterations is reached or the objective function value corresponding to the global optimal position has not changed for 50 consecutive iterations. The global optimal position vector is then output as the charging and discharging power scheduling timetable for the energy storage mechanism. After outputting the charging and discharging power scheduling timetable, the scheduling solution component synchronously verifies the trajectory of the energy storage mechanism's state of charge change corresponding to the timetable, ensuring that the state of charge remains within the preset upper and lower limits throughout the entire scheduling cycle. If the verification fails, the weight coefficients of the objective function are adjusted, and the solution process is re-executed until a charging and discharging power scheduling timetable that meets the constraints is generated.

[0077] This embodiment achieves online identification of the internal resistance parameters of the energy storage mechanism through a recursive least squares method with a forgetting factor, accurately obtains the real-time health status of the energy storage mechanism, and realizes dynamic tuning of the charging and discharging boundary by combining a piecewise linear nonlinear mapping function. The multi-objective optimization solution is completed by a particle swarm algorithm with adaptive inertia weight and learning factor. The generated charging and discharging power scheduling time table takes into account both the system power fluctuation smoothing effect and the service life of the energy storage mechanism, ensuring the reliability and economy of the scheduling scheme.

Claims

1. A generator with an energy storage mechanism, characterized in that: It includes data extraction components, feature extraction components, boundary generation components, and scheduling and solving components; The data extraction component extracts historical power generation data, historical load data, and historical state of charge data of energy storage institutions according to a preset time window, and constructs a multi-dimensional time-series input matrix. The feature extraction component uses a graph convolutional network to extract the spatiotemporal correlation features between different variables in the multidimensional time-series input matrix, inputs the spatiotemporal correlation features into a long short-term memory network, and outputs the power generation prediction curve and load prediction curve for the next scheduling cycle. The boundary generation component calculates the net power fluctuation range based on the difference between the power generation prediction curve and the load prediction curve, and dynamically generates the charging and discharging trigger threshold boundary and the maximum charging and discharging power constraint boundary for the current scheduling cycle through a nonlinear mapping function, in conjunction with the current health status data of the energy storage mechanism. The scheduling solution component, within the boundary of the charge / discharge trigger threshold and the boundary of the maximum charge / discharge power constraint, uses the particle swarm optimization algorithm to generate the charge / discharge power scheduling timetable of the energy storage mechanism, with the objective function of minimizing the variance of the net power fluctuation range and minimizing the number of charge / discharge state switching.

2. The generator with an energy storage mechanism according to claim 1, characterized in that, When constructing the multidimensional time-series input matrix, the data extraction component aligns the acquired historical power generation data, historical load data, and historical state of charge data of the energy storage mechanism according to timestamps, performs linear interpolation processing for missing values ​​and extreme value removal processing based on box plot statistics, extracts the processed data into equal-length segments according to the preset time window, concatenates the data vectors of multiple time steps in the feature dimension, adds a timestamp encoded vector containing periodic and time position features to the end of the concatenated matrix, and performs maximum and minimum value normalization processing on the concatenated and encoded matrix to generate the multidimensional time-series input matrix.

3. The generator with an energy storage mechanism according to claim 1, characterized in that, Before extracting the spatiotemporal correlation features using the graph convolutional network, the feature extraction component constructs a graph-structured adjacency matrix based on the Pearson correlation coefficients between different variables in the multidimensional temporal input matrix. Each column of feature variables in the multidimensional temporal input matrix is ​​used as a node in the graph structure, and the Pearson correlation coefficient is used as the edge weight connecting two nodes in the graph structure. The graph convolutional network performs graph convolution operations on the adjacency matrix and the feature matrix containing node features, aggregating the neighbor node features of each node and outputting the spatiotemporal correlation features.

4. The generator with an energy storage mechanism according to claim 1, characterized in that, When the feature extraction component inputs the spatiotemporal correlation features into the Long Short-Term Memory (LSTM) network, it expands the spatiotemporal correlation features according to time steps and inputs them into each hidden layer of the LTM network. An attention mechanism layer is connected to the output of the LTM network. The attention mechanism layer calculates the attention weights of the hidden state vectors output by the LTM network at each time step. The hidden state vectors and the corresponding attention weights are then weighted and summed. The weighted and summed vector is input into the fully connected layer, and the power generation prediction curve and the load prediction curve for the next scheduling cycle are output.

5. The generator with an energy storage mechanism according to claim 1, characterized in that, When the boundary generation component obtains the current health status data of the energy storage mechanism, it collects the port voltage and port current of the energy storage mechanism based on the second-order Thevenin equivalent circuit model of the energy storage mechanism, and performs offline identification of the ohmic internal resistance parameter and polarization internal resistance parameter in the second-order Thevenin equivalent circuit model using the least squares method, and calculates the health status based on the ratio of the ohmic internal resistance parameter to the rated internal resistance parameter. When generating the charge / discharge trigger threshold boundary and the maximum charge / discharge power constraint boundary through the nonlinear mapping function, the health status and the net power fluctuation range are input into the nonlinear mapping function composed of piecewise linear functions, and the charge / discharge trigger threshold boundary and the maximum charge / discharge power constraint boundary are output.

6. The generator with an energy storage mechanism according to claim 1, characterized in that, When the scheduling solution component uses the particle swarm optimization algorithm to generate the charging and discharging power scheduling time sequence table, it divides the next scheduling cycle into multiple discrete time steps, initializes a particle swarm composed of multiple particles, and the position vector of each particle represents the charging and discharging power sequence of the energy storage mechanism in the multiple discrete time steps. In each iteration of updating the particle velocity and position, an adaptive inertial weight is introduced. When the number of iterations increases, the value of the adaptive inertial weight is reduced. After each update of the particle position, it is detected whether the charging and discharging power value in the position vector exceeds the maximum charging and discharging power constraint boundary. If it exceeds the boundary, the value exceeding the boundary is replaced with the limit value corresponding to the maximum charging and discharging power constraint boundary. After the iteration terminates, the globally optimal position vector is output as the charging and discharging power scheduling time sequence table.

7. The generator with an energy storage mechanism according to claim 2, characterized in that, When performing extreme value removal processing based on box plot statistics, the data extraction component introduces a density-based noise-applied spatial clustering algorithm. This algorithm maps the historical power generation data, historical load data, and historical state of charge data of the energy storage mechanism, aligned according to timestamps, into a set of data points in three-dimensional space. It calculates the distance from each data point in the set to other data points in a specified neighborhood, counts the number of data points whose distance is less than a preset neighborhood radius, marks the data points whose number is less than a preset minimum sample size as outliers, deletes the data points marked as outliers from the set, and restores the remaining data points after deleting the outliers to time series data.

8. The generator with an energy storage mechanism according to claim 3, characterized in that, When calculating the Pearson correlation coefficient, the feature extraction component simultaneously calculates the time-delay cross-correlation function between different variables in the multidimensional time-series input matrix, traverses the preset time lag range, calculates the cross-correlation number at each time lag step, selects the time lag step corresponding to the largest cross-correlation number as the time-delay feature between variables, inputs the Pearson correlation coefficient and the time-delay feature into the preset fusion weight calculation formula, calculates the fusion correlation degree value, and uses the fusion correlation degree value as the edge weight connecting two nodes in the graph structure to generate the adjacency matrix containing time-series lag information.

9. The generator with an energy storage mechanism according to claim 5, characterized in that, When the boundary generation component performs offline identification of the parameters in the second-order Thevenin equivalent circuit model using the least squares method, it employs a recursive least squares method with a forgetting factor. Each time a new energy storage mechanism port voltage and current are acquired, the component calculates the gain vector at the current moment. Based on the gain vector, it updates the parameter estimates and covariance matrix in the second-order Thevenin equivalent circuit model at the current moment. During the calculation of the gain vector, the forgetting factor is introduced to attenuate the covariance matrix at historical moments. The updated parameter estimates are then used as the ohmic resistance parameter and polarization resistance parameter at the current moment.

10. A method of using a generator with an energy storage mechanism, characterized in that, The method includes: Historical power generation data, historical load data, and historical state of charge data of energy storage institutions are extracted according to a preset time window to construct a multi-dimensional time-series input matrix. The spatiotemporal correlation features between different variables in the multidimensional time-series input matrix are extracted using a graph convolutional network. These spatiotemporal correlation features are then input into a long short-term memory network to output the power generation prediction curve and load prediction curve for the next scheduling cycle. The net power fluctuation range is calculated based on the difference between the power generation prediction curve and the load prediction curve. Combined with the current health status data of the energy storage mechanism, the charging and discharging trigger threshold boundary and the maximum charging and discharging power constraint boundary of the current scheduling cycle are dynamically generated through a nonlinear mapping function. Within the boundary of the charge / discharge trigger threshold and the boundary of the maximum charge / discharge power constraint, the particle swarm optimization algorithm is used to solve and generate the charge / discharge power scheduling timetable of the energy storage mechanism, with the objective function being to minimize the variance of the net power fluctuation range and the number of charge / discharge state switching.

Images